Category: NATURALEZA

Nota do editor A natureza de Springer permanece neutra em relação às reivindicações jurisdicionais em mapas publicados e afiliações institucionais.

Este é um resumo de: Laman Trip, DS et al. Um atlas específico de tecido das associações de proteínas-proteínas permite a priorização de genes de doenças candidatas. Nat. Biotechnol. https://doi.org/10.1038/s41587-025-02659-z (2025).

Nota do editor A natureza de Springer permanece neutra em relação às reivindicações jurisdicionais em mapas publicados e afiliações institucionais.

Este é um resumo de: George, BM et al. Tratamento de modelos agudos de leucemia mielóide, visando uma proteína de ligação à superfície celular. Nat. Biotechnol. https://doi.org/10.1038/S41587-025-02648-2 (2025).

Selection and curation of IscB orthologs

The extended database of IscB sequences generated previously¹⁴ was collected, resulting in diverse genomic and metagenomic loci containing IscB proteins along with their closest 50% sequence identity cluster representative in the previously described IscB, IsrB and Cas9 phylogenetic tree. For each locus that was selected to be experimentally tested, the full IscB system was generated as follows. First, the putative IscB coding sequences and ωRNAs, as previously determined, were refined as follows. All protein sequences within the same 50% sequence identity cluster were aligned to the candidate IscB protein using multiple alignment using fast Fourier transform (MAFFT)⁴³. If the candidate IscB protein contained a large (≥50 aa) C-terminal insertion relative to the other proteins within the cluster, the locus was discarded. If the candidate IscB protein contained an N-terminal insertion (≥10 aa) but contained a downstream start codon site that would eliminate the insertion without removing any of the conserved N-terminal PLMP domain (previously named after the frequently observed PLMP amino acid motif in this domain), the downstream start site was selected in place of the computationally determined start site. For metagenomic sequences with multiple related protein sequences within the same 95% sequence identity cluster, all proteins within the cluster were aligned using MAFFT⁴³. The most accurate IscB protein sequence was determined to be the one that most closely matched the consensus sequence of this alignment. If the candidate IscB locus did not contain the most accurate IscB protein sequence, the candidate locus was switched to the locus that contains the most accurate IscB protein sequence. For determination of ωRNA boundaries, the upstream region of the IscB protein coding sequence was aligned using MAFFT⁴³ to loci from the same 50% sequence identity IscB protein cluster as well as phylogenetically related loci as determined by the previously determined phylogenetic tree¹⁴. The 3′ boundary of the ωRNA was selected to be ~2 bp upstream from the protein start site to match experimentally observed ωRNA boundaries. The 5′ end of the ωRNA was selected as the 5′ end determined by the CMAlign covariance model for the corresponding ωRNA type¹⁴ if the first two bases of the 5′ of the CMAlign model matched the first two bases of the ωRNA coding sequence in the given locus. However, in cases in which the model and the candidate ωRNA did not agree at the 5′ end, the ωRNA 5′ location was determined to be the 5′ most position where a sharp increase in conservation in the alignment was observed, signifying the beginning of the ωRNA. In cases in which the ωRNA 5′ could not be resolved, the candidate locus was discarded.

Multiple criteria were used to select the initial set from the large set of possible IscBs to be tested experimentally. The main criterion was phylogenetic diversity—we sampled systems from representative branches across the previously described phylogenetic tree¹⁴. The next criterion was human-related pathogens. For this subset, NCBI taxon IDs were matched when available to candidate IscB loci when available. IscB loci belonging to bacteria known to have human hosts were selected for this round. For another criterion, IscBs with REC-like insertions were prioritized. For this subset, IscBs were aligned using MAFFT⁴³ and the alignment columns between the bridge helix and RuvC-II domains were inferred as REC-like insertions. Candidate IscBs with insertions (>20 aa) in this region were selected for their potential REC-like domains. For the second set, we selected IscB systems based on similarity along the tree to other systems that we found had genome editing activity in human cells.

AlphaFold2 models of tested orthologs were generated as follows. Each tested protein sequence was searched against the full dataset of all IscBs and Cas9s described previously¹⁴ using MMSeqs2 keeping the top 501 protein hits (sorted by e-value) beyond the query protein⁴⁴. Alignments were generated using clustal omega⁴⁵ and used as input multiple sequence alignments for AlphaFold2 running under ColabFold package^27,46 without the use of templates and with up to 16 recycles using model 3, stopping if the pLDDT exceeds 95 and using Amber relaxation for the side chains.

A phylogenetic tree of the main tested type II-D and IscB orthologs, excluding CasIscBs, though including TbaIscB due to its relationship as the founding member of type II-D Cas9s, was constructed as follows. All tested IscBs were aligned with MAFFT-einsi⁴³, and then alignment columns with >50% gaps were removed. The processed alignment was then used to create a phylogenetic tree using IQ-Tree2 default parameters with 2,000 ultra fast bootstraps and using the optimal substitution model determined by ModelFinder^47,48. Protein length and ωRNA length for each system were determined based on the manually curated sequences for the protein and ωRNA, respectively. IVTT-determined TAM sequences and human genome editing activity (through the multi-guide panel) were determined experimentally as described in the sections below.

Cell-free transcription–translation TAM interference assay

IscB protein sequences were human codon optimized using the GenScript codon optimization tool. IscB genes and ωRNA scaffolds were custom synthesized by Twist Biosciences, and transcription–translation templates were generated by PCR from custom synthesis products. Cell-free transcription–translation reactions were carried out using a PURExpress In Vitro Protein Synthesis Kit (NEB) as per the manufacturer’s protocol with half-volume reactions, using 75 ng of template for the protein of interest, 125 ng of template for the corresponding ωRNA with a guide targeting the TAM library and 25 ng of TAM library plasmid. Reactions were incubated at 37 °C for 4 h, then quenched by placing at 4 °C or on ice and adding 10 µg RNase A (Qiagen) and 8 units Proteinase K (NEB) each followed by a 5-min incubation at 37 °C. DNA was extracted by PCR purification columns and adaptors were ligated using an NEBNext Ultra II DNA Library Prep Kit for Illumina (NEB) using the NEBNext Adaptor for Illumina (NEB) as per the manufacturer’s protocol. Following adaptor ligation, cleaved products were amplified specifically using one primer specific to the TAM library backbone and one primer specific to the NEBNext adaptor with a 12-cycle PCR using NEBNext High Fidelity 2X PCR Master Mix (NEB) with an annealing temperature of 63 °C, followed by a second 18-cycle round of PCR to further add the Illumina i5 adaptor. Amplified libraries were gel extracted, quantified by qPCR using a KAPA Library Quantification Kit for Illumina (Roche) on a StepOne Plus machine (Applied Biosystems, Thermo Fisher Scientific) and subject to single-end sequencing on an Illumina MiSeq with read 1 80 cycles, index 1 8 cycles and index 2 8 cycles. TAMs were extracted and the enrichment score for each TAM or PAM was calculated by filtering for all TAMs or PAMs present more than once and normalizing to the TAM or PAM frequency in the input library subject to the same IVTT and quenching reactions. A position weight matrix based on the enrichment score was generated, and both WebLogos and Krona plots were visualized based on this position weight matrix using a custom Python script.

Mammalian cell culture and transfection

Mammalian cell culture experiments were performed in the HEK293FT line (Thermo Fisher Scientific) and AML12 line (CRL-2254, ATCC). HEK293FT cells were grown in Dulbecco’s modified Eagle medium with high glucose, sodium pyruvate and GlutaMAX (Thermo Fisher Scientific). AML12 cells were grown in Dulbecco’s modified Eagle medium/Nutrient Mixture F-12 (Thermo Fisher Scientific), supplemented with 40 ng ml⁻¹ dexamethasone (Sigma-Aldrich) and 1× Insulin–Transferrin–Selenium (Thermo Fisher Scientific). All cells were additionally supplemented with 1× penicillin–streptomycin (Thermo Fisher Scientific), 10 mM HEPES (Thermo Fisher Scientific) and 10% fetal bovine serum (VWR Seradigm). All cells were maintained at confluency below 80%.

For DNA transfection, all transfections were performed with Lipofectamine 3000 (Thermo Fisher Scientific). Cells were plated 16–20 h before transfection to ensure 90% confluency at the time of transfection. For 96-well plates, cells were plated at 20,000 cells per well, and for 24-well plates, cells were plated at 100,000 cells per well. For each well on the plate, transfection plasmids were combined with 2 µl of P3000 solution per every 1 µg DNA and OptiMEM I Reduced Serum Medium (Thermo Fisher Scientific) to a total of 25 µl. Separately, 23 µl of OptiMEM was combined with 2 µl of Lipofectamine 3000. Plasmid and Lipofectamine solutions were then combined and pipetted onto cells.

For RNA transfection, all transfections were performed with Lipofectamine MessengerMAX transfection reagent (Thermo Fisher Scientific). Cells were plated similarly to the DNA transfections described above. For each well of 96-well plates, a total amount of 200 ng RNA was combined with 0.6 µl Lipofectamine MessengerMAX reagent and OptiMEM to make 10 µl of transfection mixture, which was pipetted onto cells. For different mass ratios (1:4, 1:2, 1:1, 2:1, 4:1) of in vitro-transcribed ωRNAs to mRNAs (OMEGAoff construct), ωRNA and mRNA were combined at the indicated ratio in a total of 200 ng RNA. ωRNA templates were amplified using Q5 High-Fidelity DNA Polymerase (NEB) and purified with QIAquick spin columns (Qiagen), and RNA was transcribed using a HiScribe T7 Quick High Yield RNA Synthesis Kit (NEB) and purified using an RNA Clean & Concentrator-25 Kit (Zymo Research). For mRNA encoding the OMEGAoff protein, we first digested a plasmid encoding the protein using AanI (Thermo Fisher Scientific) to obtain a linear DNA fragment. In vitro transcription (IVT) reactions were assembled with T7 buffer (NEB), 100 mM ATP (NEB), 100 mM GTP (NEB), 100 mM CTP (NEB), 100 mM pseudo-UTP (Trilink), CleanCap AG (Trilink) and T7 RNA Polymerase (NEB) and incubated at 37 °C for 5 h. The reaction was further treated with TURBO DNAse enzyme (Thermo Fisher Scientific) followed by LiCl (Thermo Fisher Scientific)-based purification before transfection as described.

Mammalian genome editing

ωRNA scaffold backbones were cloned into a pUC19-based human U6 expression backbone by Gibson Assembly. Human codon-optimized IscB genes were cloned into an immediate early promoter enhancer of cytomegalovirus (CMV) expression backbone by Gibson assembly using 2X Gibson Assembly Master Mix (NEB) to generate pCMV-SV40 NLS-IscB protein-nucleoplasmin NLS-3xHA constructs. For initial testing, 12-guide libraries were cloned in a pool mixing primers to add each of the 12 guides in a given pool at equimolar ratios, and ωRNA scaffold backbones were subject to whole plasmid amplification with guide primers annealing to the U6 promoter and a second primer annealing to the start of the ωRNA scaffold using Phusion Flash High-Fidelity 2X Master Mix (Thermo Fisher Scientific). PCR products were gel extracted and eluted in 30 μl, then blunt-end ligated to circularize by addition of 5 units T4 PNK (NEB), 200 units T4 DNA Ligase (NEB) and final 1X T4 DNA Ligase Buffer (NEB) followed by incubation for 1.5 h at room temperature before transformation in Stbl3 chemically competent Escherichia coli (NEB). For individual guide constructs, oligos with appropriate overhangs were synthesized by Genewiz, annealed and phosphorylated using T4 PNK (NEB) and cloned into ωRNA backbones by restriction–ligation cloning. Human codon-optimized IscB genes were cloned into a CMV expression backbone by Gibson assembly using 2X Gibson Assembly Master Mix (NEB) to generate pCMV-SV40 NLS-IscB protein-nucleoplasmin NLS-3xHA constructs.

Before individual guides were tested, each tested IscB protein was screened for activity in HEK293FT cells using a pool of 12 guides cloned as described. For this 12-guide pooled initial screening of IscB proteins, 800 ng of protein expression construct and 800–1,200 ng of the corresponding guide pool with corresponding ωRNA scaffold were transfected in one well of a 24-well plate as described. After 60–72 h, genomic DNA was collected by washing the cells once in 1× Dulbecco’s phosphate buffered saline (DPBS) (Sigma-Aldrich) and dry trypsinizing cells using TrypLE (Thermo Fisher Scientific). Trypsinized cells were collected in 1 ml 1× DPBS and pelleted by centrifugation at 300 × g at 4 °C for 5 min. The supernatant was removed, and cells were resuspended in 50 μl QuickExtract DNA Extraction Solution (Lucigen) and cycled at 65 °C for 15 min, 68 °C for 15 min and then 95 °C for 10 min to lyse cells. Then, 2.5 µl of lysed cells was used as input into each PCR. Amplification of each region targeted by a guide in a given guide pool was performed individually.

For all experiments in which individual guide sequences were used, unless otherwise indicated below, 100 ng guide expression plasmid and 100 ng protein expression plasmid were transfected in each of 3 or 4 wells as indicated as biological replicates in a 96-well plate for each guide condition as described. For the experiments in Fig. 2g, 50 ng guide expression plasmid and 100 ng protein expression plasmid were transfected. After 60–72 h, genomic DNA was collected directly without any enrichment of editing events by washing the cells once in 1× DPBS (Sigma-Aldrich) and adding 50 μl QuickExtract DNA Extraction Solution (Lucigen). Cells were scraped from the plates to suspend in QuickExtract and cycled at 65 °C for 15 min, 68 °C for 15 min and then 95 °C for 10 min to lyse cells. Subsequently, 2.5 µl of lysed cells was used as input into each PCR.

For library amplification, target genomic regions were amplified with a 12-cycle PCR using NEBNext High Fidelity 2X PCR Master Mix (NEB) with an annealing temperature of 63 °C for 15 s, followed by a second 18-cycle round of PCR to add Illumina adapters and barcodes. The libraries were gel extracted and subject to single-end sequencing on an Illumina MiSeq with read 1 300 cycles, index 1 8 cycles and index 2 8 cycles. Indel frequency was analyzed using CRISPResso2 (ref. ⁴⁹), with a quantification window center of −9 and a window size of 6 based on a previous analysis of IscB cleavage patterns¹⁴. To eliminate noise from PCR and sequencing error, only indels with at least two reads or more than one base inserted or deleted were counted toward reported indel frequencies. For 12-guide pooled screens, read alignments were further inspected manually for presence of ‘true’ indels to select candidates for validation. For individual guide–ωRNA experiments, to assess statistical significance, two-tailed t-tests were performed using nontargeting guide–ωRNA conditions as a negative control. All indel data are available in Supplementary Table 2. Base editing frequency was analyzed using a previously reported Python script⁵⁰, and all data are available in Supplementary Table 7. All the primer sequences used for genome PCR were listed in Supplementary Table 8.

Cell-free transcription–translation cleavage assays

To test the cleavage activities of OrufIscB with various REC insertions and NovaIscB variants on target dsDNA with different TAM sequences, cell-free transcription–translation reactions were performed using a PURExpress In Vitro Protein Synthesis Kit (NEB) with half-volume reactions, using 75 ng of template for the NovaIscB variant of interest, 125 ng of template for the ωRNA and 50 ng of target dsDNA. IVTT templates for IscB variants and ωRNAs were prepared as described in Cell-Free Transcription–Translation TAM Interference Assay. Target DNA was prepared by amplifying a pUC19 plasmid containing the target sequence and adjacent TAM with Cy3 and Cy5 primers (IDT) using Q5 Hot Start Hi-Fidelity 2X Master Mix (NEB) as per the manufacturer’s protocol with 3% DMSO added. The reactions were incubated at 37 °C for 4 h, then quenched using 10 µg RNase A (Qiagen) and 1 μl Proteinase K (Qiagen) by a 5 min incubation at 37 °C. DNA was extracted by PCR purification (Qiagen), run on 4% E-gels (Thermo Fisher Scientific) and imaged on a BioRad Chemidoc in the Cy5 channel to visualize cleavage products.

Cell-free transcription–translation mismatch tolerance assay

The target library was designed by selecting 100 random sites from coding sequences in the human genome (hg38 assembly) adjacent to ATAAA TAMs. Target sites were selected to avoid homopolymeric sequences of four or more Ts or Gs to avoid guides with potential termination signals for later use with PolIII promoters and those with potential to form G quadruplexes. Targets were also selected to have an edit distance of at least 3 away from all other targets in the TAM-proximal 7 bp and an edit distance of at least 10 in total. All possible single mismatches for each target were then generated, and progressive mismatch targets were also generated by selecting 8 sequences each with TAM-distal mismatches ranging from 2 to 7 bp at the 5′ end of the target, with an edit distance of at least 8 from any of the nontemplate original targets in the library. A total of 316 random sequences with an edit distance of at least 10 from all other library members were added as negative controls. In addition, 10 N randomized barcodes were added to each individual target for distinguishing target identity after cleavage. The library was synthesized by GenScript and cloned into a pUC19 vector by Gibson Assembly.

An associated guide library was synthesized by IDT and cloned into a backbone containing the OrufIscB ωRNA scaffold by Gibson Assembly. In vitro transcription templates for the pooled guide library were then generated by PCR and were transcribed using a HiScribe T7 Quick High Yield RNA synthesis kit (NEB). IVTT templates for IscB proteins with REC domain insertions were prepared as described in Cell-Free Transcription–Translation TAM Interference Assay. Cell-free transcription–translation reactions were carried out using a PURExpress In Vitro Protein Synthesis Kit (NEB) as per the manufacturer’s protocol with half-volume reactions, using 75 ng of template for the protein of interest, 1.5 μM final concentration of the in vitro-transcribed ωRNA guide library and 25 ng of target library plasmid. Cas9 and a guide targeting a nonlibrary control were also included as an internal activity control for downstream library preparation and sequencing. Reactions were carried out and libraries were prepared as in Cell-Free Transcription–Translation TAM Interference Assay. Barcodes were extracted and used for quantifying reads using a custom Python script. The relative activity score was calculated by dividing the cleaved read count for each perfectly matched target by the read count of the Cas9 target cleaved in each reaction, then each of those quotients was divided by the normalized read count of the same cleaved target in the WT condition in the same sequencing run. The median for each REC insertion was plotted, and any variants with an activity score of 0.5 or greater was assessed for effective guide length as measured by increasing mismatch tolerance at the TAM-distal end of the guide. For mismatch targets, the read count of all mismatch targets was normalized to the read count of the associated perfectly matched target to generate relative scores for cleavage of ‘off-target’ substrates.

Purification of IscB proteins

To purify OrufIscB, OrufIscB-REC and NovaIscB proteins, human codon-optimized IscB proteins were cloned into a pET45b(+) backbone with an N-terminal His14-Twin-strep-bdSUMO tag. These plasmids were transformed into BL21(DE3) competent cells (Thermo Fisher Scientific). Cells were grown at 37 °C in terrific broth (TB) medium supplemented with 100 μg ml⁻¹ ampicillin. Once the culture reached an optical density of approximately 0.6, the culture was shifted to 18 °C and supplemented with 0.2 mM isopropyl β-d-1thiogalactopyranoside (IPTG) for overnight induction at 18 °C. The pellet was collected by centrifugation and resuspended in the lysis buffer (50 mM Tris (pH 8), 1 M NaCl, 5 mM MgCl₂, 5% glycerol, 40 mM imidazole and 5 mM β-mercaptoethanol) with PMSF protease inhibitors. The pellets were lysed by passing twice through an LM20 Microfluidizer (Microfluidics) at 28,000 psi. The soluble fraction was collected after centrifugation at 15,060 × g for 30 min, then bound to Ni-NTA Agarose (Qiagen). The Ni beads were washed first with 12 column volumes (CV) of lysis buffer, then 5 CV of high-salt buffer (50 mM Tris (pH 8), 2 M NaCl, 5% glycerol, 5 mM MgCl₂, 40 mM imidazole and 5 mM β-mercaptoethanol) and 5 CV of low-salt buffer (50 mM Tris (pH 8), 500 mM NaCl, 5% glycerol, 5 mM MgCl₂, 40 mM imidazole and 5 mM β-mercaptoethanol) in turn. The proteins were eluted with elution buffer (50 mM Tris (pH 8), 500 mM NaCl, 5% glycerol, 5 mM MgCl₂, 300 mM imidazole and 5 mM β-mercaptoethanol), cleaved using bdSENP1 protease to remove the N-terminal tag and then dialyzed overnight in dialysis buffer (50 mM Tris (pH 8), 500 mM NaCl, 5% glycerol, 5 mM MgCl₂, and 5 mM β-mercaptoethanol). The proteins were concentrated, aliquoted and stored at −80 °C.

In vitro cleavage assay with purified proteins

For the in vitro cleavage assays, the labeled double-stranded DNA substrates were generated by PCR amplification of pUC19 plasmids containing the target and TAM sequences using Cy5 and DyLight800-conjugated DNA oligonucleotides (IDT) as primers as described in Cell-Free Transcription–Translation Cleavage Assays. All ωRNAs used in this assay were in vitro transcribed using the same protocol described in the RNA transfection protocol described in Mammalian Genome Editing. Each reaction of the cleavage assay contained 10 nM DNA substrate, 1.2 μM protein and 1.1 μM ωRNA in a final reaction buffer of 20 mM HEPES (pH 7.5), 50 mM NaCl and 5 mM MgCl₂. Reactions were incubated at 42 °C for 1 h, followed by RNase A treatment (Qiagen) and proteinase K treatment (NEB). DNA was then purified with QIAquick spin columns (Qiagen), resolved by gel electrophoresis on E-gels (Thermo Fisher Scientific) and imaged on a BioRad Chemidoc imager.

TTISS

TTISS assays were performed as described¹² with minor modifications as follows. Briefly, donor oligos (5′ – /5phos/G*T*TGTGAGCAAGGGCGAGGAGGATAACGCCTCTCTCCCAGCGACT*A*T – 3′ and 5′- /5phos/A*T*AGTCGCTGGGAGAGAGGCGTTATCCTCCTCGCCCTTGCTCACA*A*C – 3′, where * represents phosphothioate backbone modification) were annealed in nuclease-free duplex buffer (30 mM HEPES (pH 7.5), 100 mM potassium acetate) at a final concentration of 10 µM by incubating for 5 s at 95 °C and ramping down at 0.1 °C s⁻¹ to 4 °C. HEK293FT cells were transfected in 12-well plates using GeneJuice (MilliporeSigma) as per the manufacturer’s instructions with 1 µg protein expression plasmid, 2 µg combined omegaRNA expression plasmids and 1.5 µg annealed donor oligos. Subsequently, 72 h after transfection, each well was washed with 1 ml Dulbecco’s PBS (MilliporeSigma) and dry trypsinized with 200 µl TrypLE (Thermo Fisher Scientific). Cells were resuspended in 1 ml PBS and centrifuged at 300 × g for 5 min at 4 °C to pellet. The pellet was resuspended in 200 µl PBS and used as input to the Qiagen DNEasy Blood & Tissue kit (Qiagen) to extract genomic DNA. Then, 2 µg of purified genomic DNA from each sample was mixed with 20 µl of purified Tn5 enzyme loaded with a Tn5 adaptor (5′-CTGTCTCTTATACACATCTCCGAGCCCACGAGAC-3′) and 1× TAPS buffer (10 mM TAPS, 1 mM MgCl₂) and incubated at 55 °C for 10 min. Tagmented samples were purified using a QiaQuick DNA purification kit (Qiagen) and amplified twice using KOD Hot Start 2× PCR Master Mix (MilliporeSigma) with primers 5′-GTCGCTGGGAGAGAGGCGTTATC-3′ and 5′-GTCTCGTGGGCTCGGAGATGTGTATAAGAGACAG-3′ with 12 cycles and an annealing temperature of 60 °C in the first round of PCR, then with primers 5′-AATGATACGGCGACCACCGAGATCTACACTATAGCCTACACTCTTTCCCTACACGACGCTCTTCCGATCTTTATCCTCCTCGCCCTTGCTCAC-3′ and 5′-CAAGCAGAAGACGGCATACGAGATNNNNNNNNGTCTCGTGGGCTCGGAGATGTGT-3′ (NNNNNNNN refers to the barcode sequence) with 18 cycles and an annealing temperature of 65 °C in the second round of the PCR. Libraries were sequenced on an Illumina NextSeq. Within each experiment, the resulting FASTQ files were randomly downsampled so each sample had the same number of reads, for direct comparison. Across all experiments, at least 10 million and up to 50 million reads per sample were used for analysis. Reads were mapped using BrowserGenome.org (ref. ⁵¹). Off-targets for each guide were counted using a custom Python script, allowing for up to 7 mismatches in the guide sequence for 20-nt guides or 5 mismatches for 16-nt and 14-nt guides, and 1 mismatch in the TAM regardless of length. Libraries for associated on-target indel quantification were generated using the purified genomic DNA as described in Mammalian Genome Editing. For each variant, the specificity values (the percentages of total TTISS reads corresponding to detected off-targets) and activity values (the average indel fold changes of each variant versus WT OrufIscB across all guides included in each experiment) were plotted.

Mammalian base editing assays

Constructs expressing base editors and associated guide RNAs were transfected, genomic DNA was collected and libraries were prepared as described for individual protein–guide combinations in Mammalian Cell Culture and Transfection and Mammalian Genome Editing above. Editing was quantified by counting the number of reads at which the expected edited position in the amplicon was called as a G (on the top strand) or C (on the bottom strand) and dividing by the total number of reads in the sample using a previously described custom Python script⁵⁰. Unless otherwise noted, all reported data are the average of four biological replicates.

Assessment of OMEGAoff by qPCR

Constructs expressing OMEGAoff or CRISPRoff proteins and associated guide RNAs were transfected as described for individual protein–guide combinations in Mammalian Cell Culture and Transfection and Mammalian Genome Editing. RNA was extracted after 5 days unless otherwise specified using an RNeasy 96 Plus kit (Qiagen) and reverse transcribed using a RevertAid First Strand cDNA Synthesis Kit (Thermo Fisher Scientific) using random hexamer primers as previously described⁵². RNA expression was measured by qPCR using commercially available TaqMan probes (Thermo Fisher Scientific) on a LightCycler 480 II (Roche) with GAPDH as an endogenous internal control in 5-μl multiplexed reactions. Each of the four biological replicates is the average of four technical qPCR replicates, and relative expression was calculated using the double delta Ct (ddCt) method⁵³ with a negative control condition (average of all nontargeting replicates) consisting of the corresponding OMEGAoff or CRISPRoffv2.1 expression plasmid co-transfected with an AAVS1 targeting guide and GAPDH as the endogenous control. Statistical significance was assessed using a two-tailed t-test. All qPCR quantification data are available in Supplementary Table 7.

For gene activation experiments, we transfected 133 ng dOrufIscB-KRK–VPR of dCas9–VPR plasmids with 66-ng guide plasmids into 96-well plates of HEK293FT cells. RNA was extracted 5 days after transfection, then reverse transcribed and measured by qCPR as discussed above.

Western blot of PCSK9 protein after OMEGAoff repression

To perform a western blot to detect PCSK9 protein levels in AML12 cells after AAV transduction, cells were plated 16–20 h before infection to ensure 90% confluency at the time of transduction. For 12-well plates to be infected, cells were plated at 200,000 cells per well. Different AAV amounts (5 μl, 10 μl, 20 μl and 40 μl) were added. Then, the cells were transferred into 6-well plates after 2 days and collected 7 days after infection as follows. For each well, cells were washed with cold PBS and incubated for 30 min with 300 μl cold RIPA lysis buffer (Thermo Fisher Scientific) containing 3 μl Halt Protease Inhibitor Cocktail (Thermo Fisher Scientific) and 3 μl 0.5 M EDTA (Thermo Fisher Scientific). Samples were then prepared using LDS sample buffer (Thermo Fisher Scientific) and run on a 4–12% Bolt Bis–Tris gel (Thermo Fisher Scientific) in MOPS buffer for 30 min at 200 V. Membrane transfer was performed using iBlot2 Transfer Stacks (Thermo Fisher Scientific) on an iBlot2 machine for 7 min at 20 V. The membrane was blocked for 1 h at room temperature in 5% nonfat dry milk in TBS with 0.05% Tween-20 (TBS-T) buffer, then incubated with 1:1,000 Anti-PCSK9 antibody (Abcam, ab185194) and 1:10,000 Monoclonal Anti-β-Actin antibody (Sigma-Aldrich, A2228) in 2% nonfat dry milk in TBS-T buffer overnight. The membrane was then washed 5× in TBS-T buffer for 5 min each at room temperature, then incubated for an additional 2 h with secondary antibodies (Cell Signaling Technology, 7074S and 7076P2) at a 1:20,000 dilution in 2% nonfat dry milk in TBS-T buffer. The membrane was then washed an additional 3× as above and imaged on a BioRad Chemidoc imager.

RNA-seq

HEK293FT cells were transfected with OMEGAoff or CRISPRoff with nontargeting guides, CLTA-targeting guides and CALD1-targeting guides in 12-well plates. A total amount of 1.8 μg of plasmids was transfected per well including 1.2 μg OMEGAoff or CRISPRoff plasmid and 0.6 μg guide RNA plasmid. Total RNA was extracted using a Direct-zol RNA MiniPrep (Zymo) kit 14 days after transfection. The TruSeq Stranded mRNA Library Preparation Kit (Illumina) was used to prepare RNA-seq library samples starting with 1,000 ng RNA for each sample as per the manufacturer’s protocol. The library was quantified using a Qubit dsDNA HS Assay Kit (Thermo Fisher Scientific) and KAPA Library Quantification Kit (Roche). The final library was sequenced with read 1 50 cycles, index 1 6 cycles and index 2 6 cycles on an Illumina NextSeq. Spliced Transcripts Alignment to a Reference (STAR)^54,55 was used to align the sequencing reads to the human genome (GRCh37), and Salmon⁵⁴ was used to calculate the normalized transcripts per million (TPM) of each transcript. The transcripts whose TPM values were 0 in any one replicate and those showing more than twofold of difference between the two replicates were filtered out. The differentially expressed genes were identified using DESeq2 (ref. ⁵⁶) by the comparison of CLTA-targeting or CALD1-targeting samples to nontargeting samples. The differential expressed transcripts (−log₁₀(P value) > 5.5, log₂(fold change) < −1 or >1) were labeled. The significant P value was determined by Bonferroni correction.

AAV production

For AAV viral production, HEK293FT cells were used and maintained as described in Mammalian Cell Culture and Transfection. Five 15-cm plates were used per virus prep. For each prep, 60 μg adenoviral helper plasmid, 50 μg pAAV8 serotype packaging (AAV2/8) plasmid and 30 μg transgene plasmid carrying both ωRNA and OMEGAoff protein construct were added to 5 ml OptiMEM with 500 μl 1 mg ml⁻¹ PEI max solution (Polysciences). After mixing by vortexing, the mixture was incubated at room temperature for 5–10 min. Then, 1 ml of the transfection mixture was added to each plate dropwise immediately after mixing by pipetting to ensure distribution across the plates. Subsequently, 4 days after transfection, media were collected by PEG precipitation to isolate the particles as follows. Briefly, a solution containing 40% PEG 8000 (Promega) and 2.5 M NaCl was added into the medium at a ratio of 1:4. The mixture was incubated on ice for at least 2 h on a rocker, then centrifuged at 3,000 × g for 30 min. The large white pellet was then suspended in PBS and treated with 50 μl 100 mM MgCl₂ and 50 μl 10 mg ml⁻¹ DNAse for 60 min incubation at 37 °C. The solution was then subjected to an iodixanol gradient, and the virus-containing fractions were identified by qPCR and combined. Zeba Spin Desalting columns (Thermo Fisher Scientific) were used to purify and concentrate the viral particles. The titers of AAVs were determined by qPCR using ITR (inverted terminal repeat) primers (F: 5′-AACATGCTACGCAGAGAGGGAGTGG-3′, R: 5′-CATGAGACAAGGAACCCCTAGTGATGGAG-3′) with Roche Lightcycler 480.

In vivo experiments

All mice used for in vivo experiments were maintained at the vivarium facility of the Broad Institute with a standard diet, light cycles, temperature and humidity conditions. All the experiments were conducted on 5–6-week-old male C57/BL6 mice (The Jackson Laboratory) following IACUC (Institutional Animal Care and Use Committee)-approved protocols. The animals were made to fast for 12 h, and blood was collected through saphenous vein bleeds to measure the serum PCSK9 and total cholesterol levels. Each time, no more than 100 μl blood was collected. Twice, blood collection was performed before injection to evaluate pre-injection PCSK9 and cholesterol levels. After injection, blood collection was conducted once a week starting from 2 weeks postinjection. The AAV vectors, including Rosa26 targeting and two Pcsk9-targeting ωRNAs, were intravitreally injected into mice at a dosage of 2 × 10¹¹ total viral particles per animal. The injected volume was adjusted to 100 μl with sterile PBS. A PBS injection was included as a negative control. Five mice were injected for each condition. The animals were randomly chosen for each condition. To measure serum PCSK9 and cholesterol, blood samples were centrifuged at 2,000 × g for 15 min. The serum was then separated and stored at −20 °C for subsequent analysis. Serum PCSK9 levels were measured by ELISA using the Mouse Proprotein Convertase 9/PCSK9 Quantikine ELISA Kit (R&D Systems) using a 200-fold dilution as per the manufacturer’s instructions. Total cholesterol levels were measured using an Amplex Red Cholesterol Assay Kit (Thermo Fisher Scientific) following the manufacturer’s instructions. For liver function tests, the serum total bilirubin and ALT levels at the 24-week (6-month) time point were measured using a Bilirubin Assay Kit (Sigma-Aldrich) and an ALT Activity Assay Kit (Sigma-Aldrich) following the manufacturer’s instructions. All the assays were performed by BioTek Synergy Neo2 multi-mode reader (Thermo Fisher Scientific).

Cryo-EM sample preparation and data collection

The purified OrufIscB-REC–swap 49–ωRNA RNP complexes were loaded onto a Superose 6 Increase 10/300 column (Cytiva) equilibrated with a buffer containing 20 mM HEPES (pH 7.5), 150 mM NaCl, 2 mM MgCl₂ and 4.5 mM TCEP. The fractions of RNP were pooled and concentrated to 5 mg ml⁻¹ using Amicon Ultra-15 Centrifugal Filter Unit (50 kDa nominal molecular weight limit, Millipore UFC905024). To reconstitute the ternary complex, the RNP was mixed with double-strand target DNA, which was formed by the annealing of two DNA oligos encoding the target sequence, and incubated at 37 °C for 30 min. Then, 3 μl of ternary complex was applied onto glow-discharged CryoMatrix R1.2/1.3 300-mesh gold holey grids with amorphous alloy film (Zhenjiang Lehua Technology). The grids were blotted for 3 s under 100% humidity at 4 °C and then vitrified by plunging into liquid ethane using a Vitrobot Mark IV (Thermo Fisher Scientific).

The prepared grids were transferred to the EF-Krios (Thermo Fisher Scientific) operating at 300 kV with a GatanK3 imaging system and the data collected at 105,000× nominal magnification. The calibrated pixel size of 0.4125 Å was used for processing. Zero-loss images were taken using an energy filter slit width of 20 eV. Videos were collected using Leginon 3.6 (ref. ⁵⁷). Data were collected at a dose rate of 27.12 e⁻ Å⁻² s⁻¹ with a total exposure of 1.80 s, resulting in an accumulated dose of 48.82 e⁻ Å⁻². Intermediate frames were recorded every 0.05 s for a total of 40 frames per micrograph. A total of 7,580 images were collected at a nominal defocus range of 0.7–2.4 μm. Ice thickness was determined as described in refs. ^58,59.

Cryo-EM data processing and model building

Image processing was performed on CryoSPARC v4.2.0 (ref. ⁶⁰) and RELION 4.0 (ref. ⁶¹). Image stacks were subjected to beam-induced motion correction using MotionCor2.0 (ref. ⁶²). Contrast transfer function parameters for each nondose-weighted micrograph were determined by CTFFIND4 (ref. ⁶³). On-the-fly particle picking was done by Warp⁶⁴. Automated particle picking yielded 2,701,471 particles, which were extracted on a binned dataset with a pixel size of 1.65 Å and were subjected to reference-free two-dimensional classification and 8 rounds of heterogeneous refinement, producing 258,164 particles with well-defined structural features of a ternary complex. These particles were re-extracted with a pixel size of 0.825 Å and subjected to nonuniform refinement⁶⁵, which generated a map with an indicated global resolution of 2.58 Å at a Fourier shell correlation of 0.143. The particles were subjected to three-dimensional classification, a subset with 68,817 particles showing features of the inserted REC domain, an extended guide–DNA heteroduplex and an HNH domain. These particles were then subjected to nonuniform refinement, generating a map with an indicated global resolution of 2.71 Å at a Fourier shell correlation of 0.143.

Protein models predicted by AlphaFold2 (refs. ^27,46) and an ωRNA model of OrufIscB–ωRNA–target DNA complex (Protein Data Bank: 7XHT) were used as initial models. The models were docked into the cryo-EM density maps using ChimeraX 1.7 (ref. ⁶⁶), followed by iterative manual adjustment and rebuilding in ISOLDE⁶⁷ and Coot 0.8.9 (ref. ⁶⁸), against the cryo-EM density. Real space refinements were performed using PHENIX 1.18 (ref. ⁶⁹). The model statistics were validated using MolProbity 4.5 (ref. ⁷⁰). The refinement statistics are provided in Supplementary Table 6. Structural figures were prepared in ChimeraX 1.7.

Selection and in silico testing of REC swaps

AlphaFold2 was used to create protein structural models of representative IscBs and type II-D Cas9s. Structures were then aligned along the bridge helix and RuvC-II region using PyMol’s super align algorithm to identify regions of homology that may exist near REC insertions. Additional type II-D Cas9 and IscB sequences were retrieved using an HMM search (HMMER) against all protein coding sequences (clustered at 100% sequence identity using MMSeqs2) from the genomic or metagenomic database as described in a previous study²⁸. Using MAFFT, the retrieved IscBs and early type II-D Cas9s were aligned, and all regions with REC-like insertions between the bridge helix and RuvC-II domains were selected. The regions corresponding to the structurally homologous regions were then identified and checked for sequence conservation. The positions of conserved residues near the bridge helix and the RuvC-II were used as anchor positions for recombining sequences. Proteins without homologous sequences at the anchor positions were not considered for REC swaps. REC insertions from other orthologs were then swapped into the OrufIscB ortholog by switching the region between the conserved residues. For a select set of divergent Cas9s with crystal structures, structural homology alone was used to select potential REC swaps with anchor points in similar regions as described above. However, because most Cas9s outside of type II-D do not have the conserved charged motif found after the bridge helix in most IscBs, the RECs in these cases were truncated at the N-terminus in a structurally similar location to maintain the overall folding of the REC domain. AlphaFold2 models of OrufIscB along with swapped REC domains were performed as described above for the native orthologs.

The sequence conservation Shannon information mapping was performed as follows. All type II-D Cas9s with REC domains were aligned using MAFFT along with a panel of IscBs, including OrufIscB. This alignment was then trimmed to match the REC domain boundaries of Nba-1 REC, which was then set as a reference for the alignment. For each position in the reference sequence of Nba-1 REC, the distribution of nongap sequences was determined, and the Shannon information was calculated for this distribution (using log base 2). High Shannon information positions indicate high conservation in the structure and were used to determine conserved residues for REC loop swaps for regions 1, 2 and 3. For each region (1, 2 and 3), swaps were created by exchanging the OrufIscB sequence within the conserved flanking residues with sequences from another REC ortholog, based on the multiple sequence alignment.

For Fig. 2a, AlphaFold3 (ref. ⁷¹) was used to generate a full RNP prediction with target DNA, nontarget DNA, RNA and protein. However, due to the low quality of the RNA prediction, only the guide portion of the RNA was shown. AlphaFold3 (ref. ⁷¹) models were generated for the WT OrufIscB system, OrufIscB-REC system and NovaIscB system.

RNA model

The secondary structure model of the OrufIscB ωRNA was generated by the RNAstructure webserver⁷² on the OrufIscB ωRNA scaffold. The prediction contained a spurious set of two base pairs between 69 and 141 as well as 70 and 140, which were removed, as positions 69 and 70 are thought to separate the adaptor hairpin and the large pseudoknot stemloop and are thus unlikely to form additional stemloop contacts in the structure.

Calculating the distribution of potential guides of OrufIscB within 500 bp of transcription start sites

Human transcription factor start sites were downloaded from refTSS_v4 (ref. ⁷³) and mapped onto the human genome (version GRCh38.110). As refTSS_v4 may contain multiple transcription start sites for the same mRNA or gene, only transcription start sites up to 100 bp into the predicted mRNA start site or upstream from the predicted mRNA start site were selected. Of these remaining transcription start sites, the one selected to correspond to a specific gene was the one with the closest distance to the gene. These resulting transcription start sites formed the processed refTSS_v4 dataset. Coding sequences (CDSs) were identified within each genomic chromosome (ignoring the mitochondrial genome) based on the GenBank annotation, along with its corresponding predicted mRNA. For each gene’s mRNA, the standard gene name was used to cross-reference with the refTSS_v4 database to identify the transcription start site, using the predicted mRNA start site from the human genome GenBank file as the default transcription start site if there was no corresponding entry in the processed refTSS_v4. For each gene with a CDS, a ±500 bp window was formed around the transcription start site and searched for NTAAA or TTTAN (reverse complement of NTAAA) sequences to identify the number of potential guides and target sites within a 500-bp window around the transcription start site. The distribution of the number of potential guides (within 500 bp of the transcription start site) was then calculated across all CDS-containing genes.

The same approach was used as above for calculating the number of genes available for NovaIscB for knockout (KO). This was done by taking each gene’s genome-mapped coding sequence segments and counting the number of TAM sites that would enable the guide to have position 10 of the guide inside the coding sequence. Position 10 was used as this results in guides in which the cleavage site will be located inside the CDS. The distribution of the number of theoretical guides that could target a given gene was then calculated across all CDSs.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Study design and participants

The TransEuro Transplant trial (NCT01898390) was a randomized, open-label study that recruited patients from five sites across the United Kingdom and Sweden (University of Cambridge, Cambridge, UK, Imperial College London, London, UK, the National Hospital for Neurology and Neurosurgery, London, UK, University of Cardiff, Cardiff, UK, and Skåne University Hospital, Lund, Sweden). Eligible participants were recruited from the ongoing observational TransEuro observational study⁹ and rescreened using the original observational study inclusion criteria with modifications (Supplementary Table 5) to ensure continued eligibility for transplantation. Thirty-six patients were recruited into the transplant trial and randomly allocated to the transplant arm of the study or the control arm. Six withdrew before any intervention, and three did not complete screening. Eleven patients, of the remaining 27, went on to receive hfVM transplants. Sixteen patients served as a control arm. These patients underwent the same PET and clinical examinations as the transplant arm but did not receive immunosuppression and did not undergo sham surgery.

Ethical approval

Ethical permission was received for fetal tissue preparation and use at Cambridge (96/085), Cardiff (13/WA/0210ADD) and Lund (2013/432 and 2016/535). The transplant study was approved by the relevant ethical authorities in the United Kingdom and Sweden (REC reference number 10/H0304/77 in the United Kingdom and the Swedish Ethical Review Authority (Etikprövningsmyndigheten) in Lund (reference numbers 2011/290, 2014/877 and 2019-06529)).

Changes in study design

The original study design was to transplant 20 patients in an open-label fashion, drawn from a larger natural history cohort of 150 participants, and the data obtained would then be used to calculate sample size needed for a larger double-blind placebo-controled trial. All of this was to be done over a 5-year period. However, 5 years into the natural history study, the first patient was grafted, and, at this point, a decision was made to stop the transplant trial when either all 20 patients had been grafted or 3 years had elapsed from the time of the first transplant. This decision was based on the following reasons:

1. (1)

   it seemed unethical to prolong the study beyond this time as it was clearly showing that using this tissue source in a trial was not feasible in the United Kingdom and Sweden,

2. (2)

   interpretation of the study would become extremely difficult if some patients had already reached their primary endpoint whereas others had still not been grafted, and

3. (3)

   advances in human stem cell-derived dopamine cells meant that trials using this new source of more readily available cells were already entering the clinic²⁴.

At the end of 3 years in 2018, a total of only 11 patients had been grafted (8 in Cambridge, UK, and 3 in Lund, Sweden). The reasons for this have been previously presented⁹. Thus, the time of the final data collection for our predefined primary endpoint for the last patient was in March 2021. Collection of the follow-up clinical and PET imaging data was delayed (and in some cases not possible) because of restrictions resulting from the COVID-19 pandemic that began in March 2020, and this included the 36-month PET imaging in the majority of patients. To try and standardize the timings to better align with the PET imaging analysis, we defined a pretransplant baseline as the visit immediately before the first transplant. We then elected to use the following as the key time points for the primary and secondary outcomes:

• 18 months: first visit at least 510 days after the last transplant surgery or after the baseline visit (control; 540 days = 30 × 18 months with 30 days leeway)

• 36 months: first visit at least 1,020 days after surgery (transplant) or after the baseline visit (control; 1,080 days = 30 × 36 months with 60 days leeway)

Clinical assessments

Once randomized, participants continued trial visits as part of the observational TransEuro trial schedule every 6 months. During each visit, participants underwent a battery of clinical tests during an OFF state (with the OFF state being defined as the patient not having had any dopaminergic medications for 12 h before assessments or 24 h for long-acting dopamine agonists) and ON state (defined as at least 1 h after the patient had taken their regular morning dose medications). These assessments included UPDRS Part III, RUSH Dyskinesia Scale and AIMS. Patients also completed the UPDRS Parts I, II and IV, Addenbrookes’ Cognitive Examination-Revised and a series of other cognitive and PD-related assessments (Supplementary Table 6). Study data were collected and managed using REDCap electronic data capture tools hosted at the University of Cambridge.

Imaging assessments

MRI and PET scanning were performed at Invicro, Hammersmith Hospital, London, UK. Patients had a structural MRI and [¹¹C]PE2I, [¹¹C]DASB and [¹⁸F]FDOPA PET scans at baseline and repeat scans just before surgery and at 18 months after their first transplant. The planned 36-month scanning could not be performed in sufficient numbers of patients because of the COVID-19 pandemic that began in 2020.

Image processing and kinetic modeling were conducted using MIAKAT v4.3.13 (Molecular Imaging and Kinetic Analysis Toolbox)²⁵ implemented within MATLAB 2016b (Mathworks), SPM12 v7487 (Statistical Parametric Mapping, Wellcome Trust Centre for Neuroimaging) and FSL v6.0 (FMRIB Image Analysis Group)²⁶.

Structural MPRAGE images were segmented and rigid registered to the MNI template, and for each patient, all visits were entered into serial longitudinal registration to create a midpoint average and associated deformation fields. The midpoint was used to define the putamen and caudate in accordance with previously published anatomy-based guidelines²⁷. Cerebellar gray matter was isolated using DARTEL to estimate flow fields from the MNI template to native space, applying this to CIC Atlas v1.2 and masking with a gray matter segment. Dynamic PET images were motion corrected and co-registered with their corresponding MPRAGE, using the summed PET images as an intermediary and normalized mutual information as a cost function, in one interpolation step. Parcellations were then applied to the dynamic PET data to generate regional time–activity curves. For [¹⁸F]FDOPA, Patlak graphical analysis was used to quantify the uptake rate constant (K_i), whereas for [¹¹C]PE2I, Logan graphical analysis was used to quantify BP_ND. For both, t* was set to 30 min. For [¹¹C]DASB, BP_ND was estimated with Simplified Reference Tissue Model 2 (SRTM2). Cerebellar gray matter was used as a reference region for all tracers.

Tissue preparation

Transplanted tissue was prepared from hfVM dissected from three fetuses collected after either medical or surgical abortions under full ethical approval. Tissue dissections were standardized across centers by established landmarks and documentation of each cut using photographs. The landmarks used for dissection are shown in Supplementary Fig. 1.

The collected tissue was stored for a maximum of 4 days in Hib(E) at 4 °C. On the day of surgery, three hfVMs were pooled and washed several times in DMEM (cGMP compliant, Life Technologies, A12861 01)/tirilazad mesylate (custom made to GMP grade, Rechon Life Sciences). The hfVMs were enzymatically digested in a mixture of Tryple E CTS (cGMP compliant, Life Technologies A12859-01) and Pulmozyme (Dornase-α, Roche) at 37 °C for 20 min. After incubation, the tissue was washed three to four times in dissociation medium DMEM/tirilazad mesylate/dornase-α to remove any Tryple E residue. The hfVMs were then dissociated very gently to produce a crude cell suspension, which was spun down and resuspended. Aliquots of 5 × 20 µl were prepared after confirmation of viability and transported to theater in a temperature-monitored cool box. Quality criteria to proceed with transplantation of hfVM cells was set at >80% cell viability on the day of implantation. Although insufficient tissue was a common problem given that the cell preparation had to be derived from at least three hfVMs per side grafted, only one cell preparation had a viability below that required for surgery. The crown rump length of the fetus varied between 15 mm (gestational age (weeks ± days) 7 + 6) and 35 mm (gestational age 10 + 2), and the final cell suspension viability was between 83 and 93%.

It is worth noting that in TransEuro, unlike previous hfVM transplant trials, the tissue was dissociated not using trypsin but using Pulmozyme, as the former could not be sourced at a clinical grade. Given the low number of patients transplanted, it could not be experimentally determined whether Pulmozyme was superior or not to trypsin and had impacted the final number of dopamine cells in the grafted tissue. In addition, we used tissue collected from medical terminations of pregnancy, not surgical terminations, as was the case in earlier trials²⁸. This may also have had an impact on the final number of surviving dopaminergic cells within the graft, as might the time spent in hibernation media before the final tissue preparation and transplant surgery.

Surgery

Neurosurgery was performed at one of two sites. All three Swedish patients underwent surgery at Skåne University Hospital, Lund, Sweden. All eight patients recruited at the UK sites had surgery performed at Cambridge University Hospital, Cambridge, UK. Each patient underwent two unilateral transplants within an interval of 1–5 months (3.88 ± 2.49 months) with imaging guidance for trajectory and stereotactic planning. Five trajectories were made per putamen using a transfrontal approach. Eight deposits of 2.5 μl were injected per trajectory for a total of 20 μl per trajectory. MRIs were performed after surgery to show the sites of tissue deposition (Supplementary Fig. 2), although only the needle tracts can be seen, not the transplant itself, as MRI cannot provide validated evidence for the integration of the grafted cells into the brain. Due to regulatory differences between countries, different surgical devices were used to deliver the transplants between the UK site and the Swedish site. The device used in Lund was the original R–L device used in previous open-label trials²⁸, whereas in Cambridge, an in-house-manufactured version of this device was made (TRN3), which was subsequently modified (TRN4) part way through the trial. This modification was undertaken in response to feedback from the neurosurgeon using the device in Cambridge, and the needle in both devices had an internal diameter of 0.82 mm and an external diameter of 1.07 mm.

After surgery, patients were given prophylactic antibiotics and were started on a standard whole-organ immunosuppressant regimen of cyclosporin (titrated to serum levels of between 100 ng ml^–1 and 200 ng ml^–1), 2 mg per kg (body weight) per day azathioprine and 40 mg of prednisolone weaning to 5 mg over 12 weeks after a one-off dose of 1 g at the time of surgery. Immunotherapy was maintained for 12 months after the last transplant and was then stopped. During this time, patients also took all recommended prophylactic treatments for patients on this immunosuppressive regimen, namely co-trimoxazole three times a week, omeprazole, calcichew daily and alendronic acid once a week.

Postsurgical visits

Patients were followed up 12, 24 and 48 h after surgery for routine postsurgical observations and blood tests. Two days after surgery, a postoperative brain MRI scan was performed to verify graft placement and examined for any perioperative hemorrhage. In addition to their regular study clinical assessment visits, patients also had safety visits at 7, 14, 21, 28 and 42 days and then 2, 3, 4, 5, 6, 9 and 12 months after surgery as well as blood testing to monitor their immunosuppression.

Video rescoring

To maintain intersite and inter-rater reliability, all UPDRS Part III assessments were videotaped. A random selection (n = 25) of these corresponding to the key time points (pretransplant visit and 36-month post-transplant visit or equivalent for controls) were examined by an independent rater blinded to the patient’s transplant status and rescored. These videos did not reveal the patient’s surgical status as all patients were required to wear a hat at all assessments to hide the presence or absence of surgical scars so that their group allocation (transplant or control) could not be identified from the videos. These rescored UPDRS Part III scores were used in the evaluation instead of the original score; however, the overall concordance rate between the two UPDRS scores was high (Supplementary Fig. 3).

Clinical outcomes

The primary outcome measure was defined as change in the UPDRS Part III score in the defined OFF state at 36 months after surgery compared to baseline. As the treatment is a dopamine therapy, and most likely to affect motor outcome, a motor score was felt to be the most appropriate measure, and the 36-month time point was chosen to allow sufficient time for any post-transplant benefits to evolve.

Secondary outcomes included a range of motor, nonmotor, quality of life and cognitive measures as well as changes in dopaminergic medication. These are summarized in the main text.

Statistical analysis

Given the limitations of sample size, site to site variability and procedural and patient heterogeneity, it is debatable whether inferential analyses are relevant and interpretable. Therefore, no inferential statistics were used to assess the efficacy of the transplants on the primary or secondary outcomes.

Statistical analysis of the imaging data included all bilaterally transplanted patients who completed the multi-PET protocol at three time points (n = 8). We performed two-way mixed ANCOVAs with repeated measures to examine whether differences in mean putamenal [¹⁸F]FDOPA K_i, [¹¹C]PE2I BP_ND and [¹¹C]DASB BP_ND values depended on group (transplant or control) and visit. Visit included pretransplant and 18 months post-transplant time points for the transplant group (n = 8) or baseline and 18-month follow-up time points for the control group, for which data were available for 16 patients for [¹⁸F]FDOPA K_i and [¹¹C]PE2I BP_ND and for 14 patients for [¹¹C]DASB BP_ND. We also conducted a series of two-way repeated measures ANCOVAs to examine whether differences in mean striatal [¹⁸F]FDOPA K_i, [¹¹C]PE2I BP_ND and [¹¹C]DASB BP_ND values depended on visit (baseline, pretransplant and post-transplant) and striatal region (putamen and caudate). The caudate was included as an internal control region given that it is also known to exhibit substantial dopaminergic neurodegeneration over time in PD. For both analyses, mean-centered age and disease duration at the first included time point were entered as continuous covariates where possible as they have been shown to be related to dopaminergic and serotonergic loss. Post hoc Tukey-adjusted pairwise comparisons of the estimated marginal means were conducted to evaluate pairwise differences where appropriate.

Spearman’s rank-order correlations were conducted to evaluate the relationship between the changes in PET parameters (posttransplant–pretransplant) and primary and secondary outcome scale change scores. For this purpose, observational data collected closest in time to the PET acquisitions were included for analysis.

Statistical analyses and visualizations were computed in R version 4.2.2 using the following packages: afex 1.3.0, car 3.1.2, emmeans 1.8.6, moments 0.14.1, geoR 1.9.2, rcompanion 2.4.30, rstatix 0.7.2, Hmisc 5.1.0, FSA 0.9.5, ggplot2 3.4.1 and ggpubr 0.6.0.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Protein coabundance scores genome-wide protein associations

We started by collecting protein abundance data from proteomics studies of cohorts of participants with cancer. In total, we compiled a dataset of 50 studies across 14 human tissues, encompassing 5,726 samples of tumors and 2,085 samples of adjacent healthy tissue^{26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75} (Fig. 1a and Supplementary Table 1). We further included the mRNA expression data paired to the proteomics for 2,930 of the tumor and 722 of the healthy samples. Following previous studies, we used the fact that protein complex members are strongly transcriptionally and post-transcriptionally coregulated to compute probabilities of protein–protein associations from the abundance data^5,17,20 (Fig. 1b, Methods and Supplementary Fig. 1). In short, we preprocessed the abundance data to obtain a log-transformed and median-normalized abundance across participants. For each study, we then computed a coabundance estimate of a protein pair as the Pearson correlation when both proteins were quantified in at least 30 samples (Supplementary Fig. 2). Lastly, with pairs of subunits for curated stable protein complexes as ground-truth positives (CORUM⁷⁶), we used a logistic model for each study to convert the coabundance estimates to probabilities of protein–protein associations (Supplementary Figs. 3–5).

To test the ability of the association probabilities to recover known complex members, we computed receiver operating characteristic (ROC) curves for probabilities derived from protein coabundance, mRNA coexpression and protein cofractionation^6,7,77 (Fig. 1c). We found that protein coabundance (area under the curve (AUC) = 0.80 ± 0.01 (mean ± s.e.m.)) outperformed protein cofractionation (AUC = 0.69 ± 0.01) and mRNA coexpression (AUC = 0.70 ± 0.01) data for recovering known interactions (Fig. 1d and Methods). In addition, the combination of mRNA and protein abundance data did not significantly improve the recovery of known complex members (Fig. 1e; AUC = 0.82 ± 0.01, P = 0.15, according to a one-sided Welch’s t-test). Therefore, with roughly half of all cohorts having paired mRNA expression data available, we chose to only use protein coabundance for computing association probabilities. Additionally, we found similar AUCs when regressing gene expression out of the protein abundance before computing protein coabundance estimates (AUC = 0.78 ± 0.01, P = 0.18), suggesting that post-transcriptional processes but not regulation of gene expression drive most of the predictive power for protein associations.

Having established that the association probabilities derived from protein coabundance data recover known interactions of protein complex members, we sought to test whether replicate studies of the same tissue yielded association probabilities that were representative for each tissue. As a starting point, we used the gene expression data to establish that the association probabilities were not driven by cell-type composition⁷⁸ (Supplementary Fig. 6). Next, using the 1,115,405 association probabilities that were quantified for all studies, we found that the replicate cohorts from the same tissue generally clustered together (Fig. 1f; for example, blood, brain, liver and lung). Next, we selected the associations that were tissue specific, that is associations whose average probability exceeded the 95th percentile for a given tissue (0.68 ± 0.01 across tissues) and whose average probability remained below 0.5 across all other tissues. Through a hold-one-out methodology, we found that the tissue-specific associations were primarily recovered by cohorts of the same tissue of origin (AUC = 0.71 ± 0.01) compared to cohorts from different tissues (AUC = 0.56 ± 0.00, P < 0.05 for all tissues, according to a one-sided Welch’s t-test) (Fig. 1g, Methods and Supplementary Fig. 7). Together, these observations suggest that the tissue of origin is a major driver of differences between cohorts.

An atlas of protein associations in human tissues

With the replicate cohorts representing the tissue of origin, we aggregated the association probabilities from cohorts of the same tissue into single association scores for 11 human tissues (Fig. 2a and Methods). Aggregating the replicate cohorts was advantageous, as all but one of the individual cohorts were outperformed by the tissue-level scores for recovering known protein interactions (P = 1.3 × 10⁻⁹, according to a one-sided Welch’s t-test). Moreover, the tumor-derived scores outperformed the healthy-tissue-derived scores for all tissues (Fig. 2b; AUC = 0.87 ± 0.01 and 0.82 ± 0.01, respectively, P = 8.3 × 10⁻⁵, according to a one-sided Welch’s t-test). In addition to the biopsy, where the genetic heterogeneity of tumors increased variation between samples (Supplementary Fig. 8), we found several other factors affecting the recovery of known interactions, such as the available number of cohorts per tissue, the number of samples per cohort, the tissue of origin and the MS methodology (Supplementary Figs. 2 and 9). The healthy-tissue-derived and tumor-derived scores originated from separate dissections of the same tissues and participants and could, thus, serve as independent replicates. Analogous to the cohorts, we computed tissue-specific associations from the healthy-tissue-derived scores, which we then recovered with the tumor-derived scores (Fig. 2c). For all tissues, we found that the tumor-derived scores primarily recovered the tissue-specific associations of the same healthy tissue (AUC = 0.74 ± 0.02) compared to the other healthy tissues (AUC = 0.53 ± 0.01, P = 5.9 × 10⁻⁵, according to a one-sided Welch’s t-test). These analyses show that the coabundance-derived tissue-level association scores recover known protein interactions and are reproducible and representative of the tissue of origin (Supplementary Fig. 10).

We defined a protein association atlas with association scores for all quantified protein pairs by averaging the association probabilities over the cohorts of each tissue. The resulting association atlas scores the association likelihood for 116 million protein pairs across 11 human tissues (Fig. 2d). On average, each tissue contains association scores for 56 ± 6.2 million protein pairs, of which 10 ± 1.0 million are likely to be associated (score > 0.5, average accuracy = 0.81 over all tissues, recall = 0.73 and diagnostic odds ratio = 13.0) and 0.49 ± 0.08 million are ‘confident’ associations (score > 0.8, average accuracy = 0.99 across tissues, recall = 0.21 and diagnostic odds ratio = 31.9) (Supplementary Fig. 11). These protein associations tended to be likely and confident in only a few tissues, with 99,103 protein pairs having likely associations in all tissues (Fig. 2d and Supplementary Fig. 12).

Differences between tissues not driven by gene expression

One of the well-known drivers of differences in protein interactions between tissues is gene expression; proteins can interact only if their gene is expressed in a tissue. Indeed, the proteins that were quantified in a given tissue were generally enriched for genes with elevated expression for that same tissue but not the other tissues (Supplementary Fig. 13; P = 1.3 × 10⁻⁶, according to a one-sided Mann–Whitney U-test). However, only up to 7% of differences in (likely) associations between tissues can be explained by differences in gene expression and only through the lack of detection (Supplementary Fig. 14). These observations demonstrate that the likely associations for each tissue reflect but are not defined by differences in gene expression, further supporting our previous observation that protein coabundance is primarily driven by post-transcriptional processes.

Having established that association scores generally reproduce well and that differences between tissues are not driven by gene expression, we sought to measure the share of tissue-specific associations. To do so, we used a threshold association score to quantify the percentage of a tissue’s associations that were likely (score > 0.5) for the replicate (Fig. 2e, orange curve, comparing healthy-tissue-derived and tumor-derived replicates). As expected, we found that the percentage of likely associations increases with the threshold score, with 46.3% of likely associations and 90.2% of confident associations (score > 0.8) also being likely for the replicate tissue. Similarly, we found that these percentages decreased to 32.9% and 54.6%, respectively, when comparing associations between pairs of tissues from the association atlas (Fig. 2e, green curve). Lastly, depending on threshold scores, we found between 18.8% and 34.0% (interquartile range) of likely associations to be tissue specific between pairs of tissues, given the difference in probabilities between the curves for replicates and tissues. Therefore, with up to 7% of likely associations not quantified in other tissues because of gene expression (Supplementary Fig. 14), we estimated over 25.8% (18.8% + 7%) of likely associations to be tissue specific.

Tissues recover tissue-specific cellular components

We sought to characterize the likely associations that were shared between tissues. Compared to all likely associations (average Jaccard index = 0.19), we found that the similarity between pairs of tissues increased as we restricted the likely associations to interactions identified through high-throughput screens such as yeast two-hybrid (Jaccard index = 0.30; HuRI)³ or AP (Jaccard index = 0.41; BioPlex)⁴ experiments (Fig. 2f). Likewise, the similarity between pairs of tissues increased when restricting likely associations to known interactions reported for signaling (Jaccard index = 0.32; SIGNOR)⁷⁹, biological pathways (Jaccard index = 0.48; Reactome)⁸⁰, physical associations (Jaccard index = 0.56; STRING)⁸¹ or human protein complexes (Jaccard index = 0.74; CORUM)⁷⁶. Lastly, we found that the quantified differences and similarities between tissues were not sensitive to the choice of score cutoffs (Supplementary Fig. 15). Thus, known protein interactions are typically shared by the tissues in our association atlas, with signaling interactions being less commonly recovered between tissues than stable protein complexes. These observations reflect the divergence between tissues for different types of interactions and may also reflect differences in accuracy for recovering associations for stable protein complexes compared to spatiotemporal interactions that are dynamic.

Well-characterized protein complexes were generally preserved across tissues, becoming more variable as the complex-averaged association scores decreased (ρ = −0.77, P = 6.2 × 10⁻¹²⁵) (Supplementary Fig. 16). As seen in other proteomics datasets⁸², more variable complexes are typically involved signaling and regulation (for example, tumor necrosis factor and emerin), while more stable complexes are involved central cellular structures (for example, ribosomes and the respiratory chain) (Supplementary Fig. 16). While protein complexes varied little between tissues, we found that associations varied strongly for tissue-specific cellular components, for example, for the brain (synapse-related components), throat (structural components of muscle fiber), lung (motile cilia) and liver (peroxisomes) (Supplementary Fig. 17). This suggests that tissue-specific and cell-type-specific cellular components are an important driver of tissue-specific protein associations that are independent of simple expression differences.

Association atlas reveals cell-type-specific associations

To explore cell-type-specific associations in our association atlas, we took the AP2 adaptor complex as a well-known example. The AP2 complex has neuron-specific functions in addition to functions that are general to all cells⁸³. Indeed, the subunits of the AP2 complex were coabundant in all tissues (average association score between subunits = 0.80). We found 91 proteins that had association scores with all AP2 subunits in all tissues and were known to associate with AP2 (STRING score > 400). Among these, the 51 synaptic proteins (SynGO⁸⁴) had higher association scores with the AP2 complex in the brain (average score = 0.54) compared to the other tissues (average score = 0.48 ± 0.00, P = 6.7 × 10⁻⁶, according to a one-sided Mann–Whitney U-test). Conversely, the nonsynaptic interactors had lower association scores with the AP2 complex in the brain (average score = 0.33) compared to the other tissues (average score = 0.43 ± 0.00, P = 1.1 × 10⁻²¹, according to a one-sided Mann–Whitney U-test) (Fig. 3a). We explored further examples by focusing on cell-type-specific associations in the context of disease. We found that proteins of hemoglobin are related to anemia and have likely associations with anemia proteins but only in the blood (Fig. 3b and Methods). Likewise, we found that subunits of chylomicron, which transports dairy lipids from the intestines, contain and have likely associations with proteins related to Crohn’s disease but only in the colon^85,86. Lastly, we found that subunits of fibrinogen, synthesized in the liver, contain and have liver-only likely associations with proteins related to liver disease^87,88. For the other tissues, we could find many examples of tissue-specific and cell-type-specific associations for protein complexes, cellular components and disorders such as diabetes and asthma (Supplementary Fig. 18). These examples demonstrate that our association atlas can be used to study tissue-specific functions of protein complexes and context-dependent associations for disease genes.

Tissue-specific relations of traits and cellular components

We sought to generalize these examples of context-specific associations by systematically mapping the relationships amongst traits and multiprotein structures such as complexes or cellular components. As sets of proteins, we defined cellular components by Gene Ontology (GO) and defined human traits on the basis of the genome-wide association studies (GWAS) Open Targets (OTAR) locus-to-gene (L2G) score (≥0.5)^89,90,91. We then scored the relationship between sets of proteins with the median association score of all possible protein pairs between the sets (Fig. 3c, schematic, omitting the intersection between gene sets). In total, we scored the relationships between traits (107,306 pairs), between traits and cellular components (240,967 pairs) and between components (134,002 pairs) across all tissues (Fig. 3d and Supplementary Tables 2–4). The relationship scores that were high in all tissues were primarily of core cellular components such as the ribosome and spliceosome (72% of relationships with relative average score > 1.75), while the relationship scores that varied most across tissues often involved tissue-specific structures such as synaptic components (61% of relationships with a coefficient of variation (CV) > 0.4). These observations suggest that the relationship scores recapitulate the relatedness of protein sets in a tissue-specific manner, particularly for the brain.

Relationship scores for prioritizing disease genes

Additionally, we unbiasedly scored the tissue specificity of each protein set as the median association score between all pairs of its proteins (Methods and Supplementary Tables 5 and 6). Using these scores, we then selected the 15 traits most specific to the brain, 13 of which were indeed related to the brain (Supplementary Table 7). Clustering these traits using the trait–trait relationship scores from the brain revealed a hierarchical organization of traits with co-occurring conditions such as anorexia nervosa, obsessive–compulsive disorder (OCD) and Tourette syndrome closely clustering together^92,93 (Fig. 3e, left dendrogram). As an example, we further determined the 15 cellular components that had the strongest brain-specific relationships with OCD, all but one of which were related or specific to neurons (Fig. 3e, right dendrogram, and Methods). The majority of these cellular components had few genes in common with the genes confidently associated with OCD (Fig. 3e, orange heat map; Jaccard indices < 0.04). However, after removing the few genes confidently associated with OCD through GWAS, we found that almost all components were still enriched with or contained OCD-related genes (Fig. 3e, purple–green heat maps), that is, drug targets for OCD (odds ratio = 8.4 ± 1.8; ChEMBL clinical stage 2 or higher)⁹⁴, genes related to OCD from mouse deletion phenotypes (odds ratio = 4.8 ± 0.8; International Mouse Phenotyping Consortium (IMPC) score ≥ 0.5)⁹⁵ or genes less confidently linked to OCD through GWAS (odds ratio = 1.6 ± 0.2, OTAR L2G score < 0.5) (Methods). Moreover, these 15 components with the strongest OCD relationships in the brain were more strongly enriched with OCD-related genes than other cellular components that contained OCD-linked genes (P = 4.3 × 10⁻¹¹ (drug targets), 3.4 × 10⁻¹⁵ (genes related to OCD in mice) and 6.6 × 10⁻⁷ (genes less confidently linked to OCD through GWAS), according to one-sided Mann–Whitney U-tests).

Together, the results above demonstrate that the proposed relationship scores can prioritize cellular components that are enriched with trait-relevant genes. Analogously, we found that we could use the relationship scores for reconstructing the hierarchical organization of the cell, including maps of subcellular structures and modules of tissue-specific relations between cellular components (Supplementary Fig. 19). These observations demonstrate the potential for our association atlas to facilitate the systematic mapping of relations among traits, cellular compartments and likely other ontology terms.

Validated brain interactions for schizophrenia-related genes

The results above indicate that the tissue-specific associations could facilitate the prioritization of disease-linked genes by functional association. Indeed, direct interactors of disease-linked genes have been used to prioritize causal genes in genetically linked loci and shown to be enriched in successful drug candidates^96,97,98. To explore this in more detail, we constructed a network of brain interactions for schizophrenia (SCZ)-related genes. Specifically, we sought to prioritize highly ranked associations for the brain that involve SCZ-related genes and that have additional evidence from orthogonal methodologies.

We started by taking n = 369 genes associated with SCZ through GWAS studies (‘starting genes’, L2G scores ≥ 0.5) and computed the top 25 traits and cellular components that had the strongest tissue-specific relation to SCZ in each tissue (Fig. 4a and Methods). This gave us a collection of genes related to SCZ in a tissue-specific manner. For each tissue, we then filtered for protein pairs that had one SCZ starting gene and one SCZ-related gene and required these protein pairs to have association scores exceeding the 97th percentile of the tissue scores (association score = 0.70 ± 0.01, 0.73 in the brain), leading to tissue-specific networks of associations for SCZ-related genes (Supplementary Fig. 20 and Supplementary Table 8). After removing the SCZ starting genes from the brain network, the remaining genes were still enriched for genes associated with SCZ in mice (Benjamini–Hochberg (BH)-adjusted P value = 1.5 × 10⁻⁵, IMPC score ≥ 0.5), drug targets for SCZ (BH-adjusted P value = 9.8 × 10⁻⁵, ChEMBL clinical stage 2 or higher) and other variants associated with SCZ (BH-adjusted P value = 1.0 × 10⁻⁷, OTAR L2G scores < 0.5), according to one-sided Fisher exact tests. This enrichment was specific to the brain compared to any of the other tissues, suggesting that the proposed methodology presents a systematic approach for prioritizing disease genes of tissue-specific traits (Fig. 4b).

Indeed, for brain-related disorders, including autism, attention deficit hyperactivity disorder (ADHD), Tourette syndrome and others, we found that the proposed methodology of prioritizing protein associations for disease-linked genes enriched, specifically in the brain, for other genes associated with the respective disorder through mouse phenotypes or drug targets, with many also being enriched for GWAS variants that are less confidently linked to the disorder (Supplementary Fig. 21).

Having established that the presented methodology prioritizes protein associations that enrich for disease-related genes in a tissue-specific manner, we sought to validate the prioritized protein associations through additional evidence. Specifically, to further validate the networks of predicted associations for SCZ-related genes, we assembled a curated dataset of brain interactions established experimentally through pulldowns using human brain cells (that is, AP–MS or coimmunoprecipitation–MS in human microdissected brain tissue or human induced pluripotent stem cell-derived neurons^{99,100,101,102,103}). This dataset contained 7,887 human brain interactions for 30 bait proteins and has been incorporated into the IntAct database⁹ (Supplementary Table 9). We filtered these brain interactions for the bait proteins that were associated with SCZ (OTAR L2G score ≥ 0.5) and further filtered the tissue-specific networks of SCZ-related genes for associations involving at least one bait protein. The remaining associations of SCZ-related genes were strongly enriched with the interactions from pulldowns with SCZ-related baits, especially for the brain (log BH-adjusted P value = 84.3, according to a one-sided Fisher exact test) compared to the other tissues (log BH-adjusted P value = 1.8 on average) (Fig. 4c). Thus, the brain associations of SCZ-related genes but not the other tissues were experimentally validated by pulldown studies and enriched with SCZ-related association partners.

Next, we filtered the prioritized brain associations for SCZ-related genes for interactions that were also found through the pulldown studies to obtain a network of 205 validated brain interactions for SCZ-related genes (average association scores = 0.86 in the brain), which we simplified to only show synaptic (SynGO) genes having prior evidence (Fig. 4d, Methods and Supplementary Table 10). The visualized network contained 56 proteins connected to three bait proteins through a collection of 66 validated brain interactions. These connected proteins included SCZ drug targets (3 proteins; clinical stage 2 or higher), proteins associated with SCZ in mice (12 proteins; IMPC score ≥ 0.5) or proteins linked with SCZ with weaker prior evidence (15 proteins; OTAR L2G scores < 0.5). Surprisingly, only four of the visualized brain interactions were confidently reported in any of the major protein interaction databases before our curation effort (CACNA1C with CACNB1, CACNB2 and CACNB4 (STRING scores > 750) and SRC (Signor)) and only the interactions of SHANK3 with AP2B1 and CLTB were quantified in more than three other tissues (association scores = 0.34 ± 0.08 and 0.29 ± 0.04 in the other tissues). These observations support our earlier analysis that the network of validated interactions of SCZ-related genes is specific to the brain.

The visualized network contained several groups of highly interconnected proteins (STRING scores > 750). These groups were enriched with genes of cellular components typical for neuronal functioning and SCZ, such as a group of proteins for the postsynaptic cytoskeleton¹⁰⁴ (BH-adjusted P value = 2.3 × 10⁻⁸) or for clathrin-coated vesicles (9.4 × 10⁻¹⁴), according to one-sided Fisher exact tests. For the clathrin vesicle coat, the network connected all subunits of the AP2 complex and clathrin proteins to HCN1. Interestingly, previous pulldown studies showed that HCN channels directly interact with TRIP8b (refs. ^105,106). TRIP8b regulates the trafficking of HCN channels¹⁰⁶ and mainly associates with the AP2 complex¹⁰⁷. Moreover, while AP2 and clathrin are not cell type specific, both HCN1 and TRIP8b were found to be enriched at parvalbumin (PV)-positive synapses¹⁰⁸, with HCN channels being specific to PV neurons and important for their high firing frequencies^109,110. Given the link of PV neurons with SCZ^{111,112,113,114,115}, these observations suggest that AP2 and clathrin may be involved in a PV neuron-specific disruption of HCN channel trafficking with SCZ.

To suggest putative interface models for the validated brain interactions for SCZ-related genes, we used AlphaFold2 to predict the structures for 205 protein interactions, including the entire visualized network (Fig. 4d and Supplementary Table 11). The predicted models had more confident interactions compared to models for known complex members from CORUM¹¹⁶ (average pDockQ scores = 0.20 and 0.13, respectively, P = 1.6 × 10⁻¹⁹, according to a one-sided Mann–Whitney U-test). In total, we identified 15 moderate-confidence interactions (interface predicted template modeling (ipTM) scores > 0.5). These included the brain-specific binding of all three 14-3-3 proteins (YWHAG, YWHAH and YWHAZ) with HCN1 (Fig. 4e; average association score = 0.82, average ipTM score = 0.65). The interfaces of these three models overlap and are located in the C-terminal disordered region of HCN1 (residues 775–802). This consensus interface includes a predicted 14-4-3-binding site (centered around S789; average ipTM score = 0.75 at the putative binding site, 14-3-3-Pred score = 0.457)¹¹⁷ that has been verified through pulldown experiments¹¹⁸. Indeed, the binding of 14-3-3 proteins with HCN1 was found to be dependent on the phosphorylation of S789, with the interaction between 14-3-3 and HCN1 likely inhibiting HCN1 degradation¹¹⁸.

Lastly, the network contained 15 genes within loci genetically associated with SCZ that had weaker evidence supporting them as the causal genes at each locus. Given their interaction with other SCZ-related genes, these could be prioritized as more likely causal because of their functional roles. Some of these genes (AP2B1, ATP2B2 and SYNGAP1) had the highest L2G score for their respective locus with single-nucleotide polymorphisms (SNPs) linked to SCZ but had scores below the 0.5 cutoff used (0.457, 0.264 and 0.251, respectively)^90,91. In addition to the AP2 complex, we found a member of the AP3 complex (AP3B2). AP3B2 had the second highest score for the locus with SNPs (rs783540), which had splicing and expression quantitative trait locus (QTL) associations with AP3B2 but was ranked higher for disruption of CPEB1 given that the variant lies within a CPEB1 intron. Similarly, MARK2 was ranked second for two SNPs (rs7121067 and rs11231640), both having splicing and promoter capture Hi-C associations with MARK2 but being ranked higher for disrupting RCOR2 because of proximity to its transcription start site. CDC42 and NTRK2 had the second highest L2G scores for their locus but the top associated genes (WNT4 and AGTPBP1, respectively) had lower and less specific expression in the brain¹².

Cofractionation-derived synapse-specific interactome

As a final application of our protein association atlas, we focused on the interactome of synapses. To do so, we prepared and purified synaptosomes from rat brains as an orthogonal approach for validating the brain associations. We fractionated the synaptosomes with size-exclusion chromatography (SEC) into 75 fractions that were then subjected to liquid chromatography (LC)–MS/MS. A total of 3,409 unique proteins were detected, including well-known protein complexes such as the CCT complex subunits, whose profiles correlate across the fractions (Fig. 5a; average correlation coefficient = 0.96).

We preprocessed the fractionation profiles of the synaptosome and computed the coabundance of proteins across fractions to score the cofractionation of 4,276,350 protein pairs in rat synapses (Methods). To increase the confidence of the interaction scores, we then combined the rat synaptosome with other cofractionation profiles from in vivo mouse brains⁶ and subcellular fractionation profiles from human glioblastoma cells⁷. We merged the cofractionation studies by orthologs that were quantified in all three datasets and computed interaction probabilities with a logistic model using the CORUM database as positives (Methods and Supplementary Table 12). The resulting synaptic interactome quantified 1,309,771 interaction probabilities for 1,619 proteins and improved the recovery of known interactions compared to the cofractionation studies individually (Fig. 5b; AUC = 0.80 and 0.72). Of the 1,619 proteins in the interactome, 24% are annotated as synaptic proteins in the SynGO database, 49% have been reported as synapse-enriched in mouse brains and 56% have previously been identified through crosslinking MS (XL-MS) of the mouse synaptosome^84,108,119. All XL-MS interactions for these proteins were quantified in our synaptic interactome, having association scores of 0.59 compared to 0.37 for other associations (P = 2.0 × 10⁻⁴⁴, according to a one-sided Mann–Whitney U-test). Moreover, the synaptic interactions between synapse-enriched proteins were more likely compared to other interactions (average association scores = 0.40 and 0.36, P < 1 × 10⁻³⁰⁰, according to a one-sided Mann–Whitney U-test). Together, these observations suggest that our synaptic interactome largely aligns with current state-of-the-art synaptic interaction resources and scores the likelihood of interactions for a wide range of synaptic proteins, with interactions being more likely for synaptic proteins compared to other proteins.

Validated synaptic interactions for brain disease genes

The interactions between synapse-enriched proteins formed more likely associations compared to associations of nonsynaptic proteins, especially for the brain compared to the other tissues of our association atlas (Fig. 5c). Given this brain-specific elevated likelihood of synaptic interactions, we constructed a network of interactions between synaptic proteins (in SynGO or synapse-enriched in mouse brains) that were both likely coabundant in the brain (109,913 associations) and likely cofractionated for the synaptosome (121,732 interactions). The resulting network consisted of a collection of 37,318 validated protein interactions between synaptic proteins (Supplementary Table 13). These synaptic interactions were primarily specific to the brain because few of the interactions were likely for the majority of other tissues in our association atlas (20%) and only a small fraction have been reported in any protein interaction database (5.8% in STRING, 1.3% in HuMAP, 3.6% in IntAct and 1.6% in BioPlex).

We were particularly interested in the validated interactions of synaptic proteins that are associated with brain disorders. As before, we filtered for GWAS traits that had associated genes through mouse phenotyping (IMPC) or known drug targets (ChEMBL) and whose trait-level association scores were elevated in the brain compared to the other tissues (z score > 1; Methods). We found that 10 of the resulting 13 traits were disorders clearly related to the brain and selected the 727 confident synaptic interactions between genes associated with these brain-specific traits (association scores = 0.7 and 0.81 on average in the synaptosome and brain, respectively). Additionally, to suggest putative interface models, we used AlphaFold2 to predict the structures for these 727 interactions. The predicted models had more confident interactions compared to models for known interactors in CORUM (average pDockQ scores = 0.28 and 0.13, P = 3.7 × 10⁻¹⁴⁷, according to a one-sided Mann–Whitney U-test) or HuMAP (pDockQ scores = 0.28 and 0.25, P = 2.4 × 10⁻¹⁷) and high-confidence models (ipTM scores > 0.7) were enriched for additional evidence from XL-MS experiments in mouse synaptosomes¹¹⁹ (Supplementary Fig. 22). In total, we identified 105 moderate-confidence interactions (ipTM scores > 0.5; Supplementary Table 14). Lastly, we visualized (simplified) the networks of validated synaptic interactions between trait-related genes of the brain disorders (Fig. 5d and Methods).

Prioritizing synaptic disease genes for brain disorders

Several genes in the networks had weaker prior evidence supporting them as the causal genes at loci genetically associated with the brain disorders. These genes could be prioritized as more likely to be causal for the disorders because of their validated synaptic interactions with other genes that were confidently associated with the same disorders. As before, we looked for genes with the highest (below cutoff) L2G scores for their respective loci with SNPs linked to the brain disorders. We found likely causal genes for ADHD (MDH1, score = 0.491; CADPS2, score = 0.340; PIK3C3, score = 0.323), SCZ (TOM1L2, score = 0.492; AP2B1, score = 0.457; PSD3, score = 0.429; MYO18A, score = 0.428; ATP2B2, score = 0.264; TMX2, score = 0.232), Alzheimer disease (CLPTM1, score = 0.378; MADD, score = 0.315), autism (ATP2B2, score = 0.264; ATP2A2, score = 0.243), unipolar depression (MADD, score = 0.244) and bipolar disorder (ATP2A2, score = 0.206), with the last variant (MYO18A, score = 0.428) additionally being likely causal for Tourette syndrome, OCD, ADHD and unipolar depression^90,91. Of these genes, all but ATP2A2, MDH1, MYO18A and PIK3C3 had the highest or second highest expression in brain tissue¹².

Additionally, we found genes with synaptic interactions that did not have the highest L2G score for the respective SNP linked to the brain disorders but still had additional evidence. For example, we found that PAFAH1B1 associated with RHOA in our synaptic interactome, with both having weak prior evidence for being causal to depression. PAFAH1B1 has a role in neural mobility and is required for activation of Rho guanosine triphosphatases such as RhoA¹²⁰. PAFAH1B1 was the second highest scoring gene for the locus with an SNP (rs12938775) linked to unipolar depression. This variant has expression QTL associations with PAFAH1B1 and lies within a PAFAH1B1 intron, despite being more distant to the transcription start site of PAFAH1B1 compared to the highest scoring gene (CLUH). However, PAFAH1B1 encodes a synaptic protein, whereas CLUH does not, with the expression of PAFAH1B1 being higher and more specific to the brain compared to CLUH¹². Overall, this example demonstrates how orthogonal approaches targeting subcellular structures and individual tissues can provide tissue-specific protein interaction networks and aid in the prioritization of genes likely to be causal for tissue-related disorders.

Nota do editor A natureza de Springer permanece neutra em relação às reivindicações jurisdicionais em mapas publicados e afiliações institucionais.

Este é um resumo de: Griffiths, Ca et al. O spray de precursor de trealose 6-fosfato permeável à membrana aumenta o rendimento do trigo em ensaios de campo. Nat. Biotechnol. https://doi.org/10.1038/S41587-025-02611-1 (2025).

Optimized DMNB-T6P synthesis

General chemical materials and methods

All reagents were purchased from commercial sources and were used without further purification unless noted. Molecular sieve (4 Å, powder) used in reactions was activated at 350 °C for more than 12 h. Dry solvents for reactions were purchased from Sigma-Aldrich; the following abbreviations are used: PE, petroleum ether (boiling point (bp) 40–60 °C); EtOAc, ethyl acetate; THF, tetrahydrofuran. Thin layer chromatography (TLC) was carried out using Merck aluminium-backed sheets coated with Kieselgel 60-F₂₅₄ silica gel. Visualization of the reaction components was achieved using UV fluorescence (254 nm) and/or by charring with an acidified p-anisaldehyde solution in ethanol. Organic solvents were evaporated under reduced pressure, and the products were purified by flash column chromatography on silica gel (230–400 mesh). Proton nuclear magnetic resonance (¹H NMR) spectra were recorded on Bruker AVG400, AVH400 or AVB400 (400 MHz) spectrometers, and the chemical shifts are referenced to residual CHCl₃ (7.26 ppm, CDCl₃), CHD₂OD (3.30 ppm, CD₃OD) and C₆HD₅ (7.16 ppm, C₆D₆). Carbon nuclear magnetic resonance (¹³C NMR) spectra were recorded on Bruker AVG400 (100 MHz) spectrometers and are proton decoupled, and the chemical shifts are referenced to CDCl₃ (77.0 ppm) or CD₃OD (49.0 ppm). Assignments of NMR spectra were based on two-dimensional experiments (1H-1H COSY, DEPT-135, HSQC and HMBC) if required. Chemical shift for ³¹P NMR is reported with reference to phosphoric acid (0.00 ppm). Reported splitting patterns are abbreviated as follows: s, singlet; d, doublet; t, triplet; q, quartet; p, pentet; hept, heptet; m, multiplet; br, broad. Low-resolution mass spectra (LRMS) were recorded on a Micromass Platform 1 spectrometer using electrospray ionization (ESI) or on a Bruker Daltronic MicroTOF spectrometer. High-resolution mass spectra (HRMS) were recorded on a Bruker Daltronic MicroTOF spectrometer using ESI (m/z values are reported in Daltons). Optical rotations were measured on a PerkinElmer 241 polarimeter at 589 nm (Na D-line) with a path length of 1.0 dm at ambient temperature and are in units of degree ml g⁻¹ dm⁻¹. Infrared spectra were recorded on a Bruker Tensor 27 Fourier Transform spectrophotometer using attenuated total reflectance (ATR), and absorption maxima (ν max) are reported in wavenumbers (cm⁻¹). X-ray powder diffraction was recorded on a PANalytical Empyrean Series 2 powder diffractometer.

Preparation of K₂CO₃ solution for deprotection under anhydrous conditions

Anhydrous K₂CO₃ (solid, 50 mg, 0.1% (w/v) to methanol) was added into dry methanol (50 ml) under inert condition, and the mixture was stirred at room temperature for 30 min, followed by the addition of dry CH₂Cl₂ (10 ml, 20% (v/v) to methanol). The resulting solution was used directly for selective deprotection of 1 (2 g) at cold temperature without filtration. Concentration proves important to selectivity, and dosage must be increased proportionally for large-scale reactions.

2,3,4,6,2′,3′,4′,6′-Octakis-O-(trimethylsilyl)-d-trehalose (1)

Under inert atmosphere (argon), to a stirred solution of d-(+)-trehalose dihydrate (25.00 g, 66.08 mmol) in dry pyridine (250 ml) were added chlorotrimethylsilane (100.64 ml, 792.96 mmol) and hexamethyldisilazane (110.23 ml, 528.64 mmol) successively at cold temperature (ice water mixture, 2–4 °C), and the resulting mixture was allowed to warm to room temperature. After stirring for 6 h, the thick solution was concentrated under vacuum, and the crude residue was suspended in CH₂Cl₂ (300 ml) and washed with saturated NaCl (aq.) solution (300 ml). The organic layer was separated; the aqueous layer was extracted with CH₂Cl₂ (150 ml × 3); and the combined organic layers were dried over Na₂SO₄, filtered and concentrated and dried in vacuum to give the desired compound 1 (60.75 g, quant.) as an amorphous white solid: R_f = 0.68 (PE–EtOAc, 20:1); melting point (mp) 81–82 °C; [α]_D²⁵ + 96.9 (c 1.0, CH₂Cl₂); literature (lit.) mp 80–82 °C; [α]_D²⁵ + 94 (c 1.5, CHCl₃); ¹H NMR (400 MHz, CDCl₃): δ 4.91 (d, J_1,2 = 3.1 Hz, J_1′,2′ = 3.1 Hz, 2H, H-1, H-1′), 3.88 (t, J = 9.0 Hz, 2H), 3.78 (ddd, J = 9.4 Hz, J = 4.0 Hz, J = 2.4 Hz, 2H), 3.69 (dd, J = 11.3 Hz, J = 2.4 Hz, 2H), 3.65 (dd, J = 11.3 Hz, J = 4.0 Hz, 2H), 3.43 (t, J = 9.0 Hz, 2H), 3.38 (dd, J = 9.4 Hz, J = 3.1 Hz, 2H), 0.139 (s, 18H), 0.135 (s, 18H), 0.11 (s, 18H), 0.09 (s, 18H) ppm.

2,3,4,2′,3′,4′,6′-Heptakis-O-(trimethylsilyl)-d-trehalose (2a) and 2,3,4,2′,3′,4′-hexakis-O-(trimethylsilyl)-d-trehalose (2b)

Potassium carbonate (165 mg, 1.19 mmol) was added into methanol (165 ml, high-performance liquid chromatography (HPLC) grade, 5.5 ml g⁻¹); the resulting suspension was stirred at room temperature for 30 min, and then CH₂Cl₂ (33 ml, HPLC grade, 1.1 ml g⁻¹) was added, followed by the addition of the substrate 1 (30 g, 32.62 mmol, ground, white powder) in one portion. After stirring for 2 h at room temperature, the clear solution was quenched by acetic acid (136.7 μl, 2.39 mmol) and pyridine (193 μl, 2.39 mmol) successively. After removal of the solvent under vacuum, the crude residue was then suspended in CH₂Cl₂ (200 ml) and washed with saturated NaCl solution (200 ml), and the organic layer was separated; the aqueous layer was extracted with CH₂Cl₂ (50 ml × 3); and the combined organic layers were dried over Na₂SO₄, filtered and concentrated and dried under high vacuum overnight, giving a mixture of 2a and 2b (25.7 g, quant., 2a:2b = 1:4) as a white foam, which was used directly in the phosphorylation.

2a when isolated is a colorless syrup: R_f = 0.21 (PE–EtOAc, 20:1); [α]_D²⁵ + 96.4 (c 1.0, CH₂Cl₂); lit. [α]_D²⁵ + 113 (c 2.5, PE); ¹H NMR (400 MHz, CDCl₃): δ 4.93 (d, J_1′,2′ = 3.1 Hz, 1H, H-1′), 4.88 (d, J_1,2 = 3.1 Hz, 1H, H-1), 3.91–3.82 (m, 3H), 3.79 (ddd, J = 9.4 Hz, J = 4.6 Hz, J = 2.0 Hz, 1H), 3.74–3.63 (m, 4H), 3.48–3.38 (m, 4H), 1.75 (br s, 1H), 0.16 (s, 9H), 0.140 (s, 9H), 0.138 (s, 18H), 0.12 (s, 9H), 0.11 (s, 9H), 0.10 (s, 9H) ppm.

2b when isolated is an amorphous white solid: R_f = 0.48 (PE–EtOAc, 3:1); mp 115–116 °C; [α]_D²⁵ + 99.8 (c 1.0, CH₂Cl₂); lit. mp 114–115 °C; [α]_D²² + 99.5 (c 2.7, CHCl₃); ¹H NMR (400 MHz, CDCl₃): δ 4.90 (d, J_1,2 = 3.1 Hz, J_1′,2′ = 3.1 Hz, 2H, H-1, H-1′), 3.89 (t, J = 9.0 Hz, 2H), 3.85 (dt, J = 9.5 Hz, J = 3.2 Hz, 2H), 3.74–3.66 (m, 4H), 3.48 (t, J = 9.1 Hz, 2H), 3.42 (dd, J = 9.3 Hz, J = 3.1 Hz, 2H), 1.73 (br s, 2H), 0.16 (s, 18H), 0.14 (s, 18H), 0.12 (s, 18H) ppm.

Synthesis of 4,5-dimethoxy-2-nitrobenzaldehyde (3)

Nitric acid (100 ml, 70%) was cooled by an ice water bath (2–4 °C) for 30 min; veratraldehyde (20 g, 120.35 mmol, ground) was added portion-wise with stirring; and the mixture was brought to 10 °C and stirred until a clear solution was obtained (around 1 h). Then, the mixture was poured into an ice water mixture (1,000 ml) while stirring vigorously. The resultant yellow solid was collected by filtration and washed with cold water to remove nitric acid completely, and the solid was recrystallized from boiling ethanol (300 ml), affording 3 (20 g, 79%) in the form of yellow needle crystals: R_f = 0.56 (PE–EtOAc, 3:1); mp 131–132 °C; ¹H NMR (400 MHz, CDCl₃): δ 10.45 (s, 1H, CHO), 7.62 (s, 1H, H-3), 7.42 (s, 1H, H-6), 4.04 (s, 3H, OCH₃), 4.03 (s, 3H, OCH₃) ppm.

4,5-Dimethoxy-2-nitrobenzyl alcohol (4)

Sodium borohydride (3.8 g, 100.6 mmol) was added to an ice-cooled solution of 4,5-dimethoxy-2-nitrobenzaldehyde (3) (17.7 g, 83.8 mmol) in anhydrous tetrahydrofuran (THF) (400 ml), and the mixture was stirred at 2–4 °C for 3 h. The reaction was quenched by addition of water (400 ml); the organic layer was separated; and the aqueous layer was extracted with CH₂Cl₂ (150 ml × 3). Then, the combined organic layers were dried over anhydrous Na₂SO₄ and filtered and concentrated to dryness to give alcohol 4 (17.8 g, quant.) as an amorphous yellow solid: R_f = 0.24 (PE–EtOAc, 3:1); mp 151–152 °C; ¹H NMR (400 MHz, CDCl₃): δ 7.71 (s, 1H, H-3), 7.18 (s, 1H, H-6), 4.96 (d, J = 6.5 Hz, 2H, ArCH₂OH), 4.01 (s, 3H, OCH₃), 3.96 (s, 3H, OCH₃), 2.60 (t, J = 6.5 Hz, 1H, OH) ppm.

Bis-(4,5-dimethoxy-2-nitrobenzyl)-N,N-diisopropylphosphoramidite (6)

Under inert atmosphere (argon), to a stirred solution of phosphorus trichloride (13.09 ml, 150 mmol) in dry THF (400 ml) were added diisopropylethylamine (52.25 ml, 300 mmol) and diisopropylamine (42.05 ml, 300 mmol) successively at cold temperature (ice water bath, 2–4 °C). After stirring for 4 h at the same temperature, the suspended solution was cooled to −15 °C. Then, triethylamine (46.00 ml, 330 mmol) and 4,5-dimethoxy-2-nitrobenzyl alcohol (4) (64.0 g, 300 mmol) were added successively. The resulting mixture was allowed to warm to room temperature and stirred for a further 20 h in the dark. Saturated NaHCO₃ (aq.) solution (200 ml) was added, and the resulting suspension was filtered, washed with water (50 ml × 2) and CH₃CN (50 ml × 2) and completely dried under vacuum to give the desired phosphoramidite 6 (73.5 g, 88%) as an amorphous yellow solid: R_f = 0.43 (PE–EtOAc, 3:1); mp 142–143 °C (melts and decomposes); ¹H NMR (400 MHz, CDCl₃): δ 7.63 (s, 2H, H-3, H-3′), 7.30 (s, 2H, H-6, H-6′), 5.150 (dd, J = 16.4 Hz, J = 6.9 Hz, 1H, ArCH₂O), 5.149 (dd, J = 16.4 Hz, J = 6.9 Hz, 1H, ArCH₂O), 5.061 (dd, J = 16.4 Hz, J = 6.9 Hz, 1H, ArCH₂O), 5.059 (dd, J = 16.4 Hz, J = 6.9 Hz, 1H, ArCH₂O), 3.87 (s, 6H, OCH₃ × 2), 3.86 (s, 6H, OCH₃ × 2), 3.73–3.64 (m, 2H, NCH(CH₃)₂ × 2), 1.19 (d, J = 6.8 Hz, 12H, NCH(CH₃)₂ × 2) ppm; ¹³C NMR (100 MHz, CDCl₃): δ 153.9 (C-5, C-5′), 147.6 (C-4, C-4′), 138.8 (C-2, C-2′), 131.74, 131.66 (C-3, C-3′), 109.4 (C-1, C-1′), 107.9 (C-6, C-6′), 62.6 (ArCH₂O), 62.4 (ArCH₂O), 56.39 (OCH₃ × 2), 56.35 (OCH₃ × 2), 43.5 (NCH(CH₃)₂), 43.4 (NCH(CH₃)₂), 24.8 (NCH(CH₃)₂); 24.7 (NCH(CH₃)₂) ppm; ³¹P NMR (162 MHz, CDCl₃): δ + 147.41 ppm.

6-O-Bis-(4,5-dimethoxy-2-nitrobenzyloxyphosphoryl)-d-trehalose (DMNB-T6P)

Potassium carbonate (165 mg, 1.19 mmol) was added into methanol (165 ml, HPLC grade, 5.5 ml g⁻¹), and the resulting suspension was stirred at room temperature for 30 min. Then, CH₂Cl₂ (33 ml, HPLC grade, 1.1 ml g⁻¹) was added, followed by the addition of the substrate 1 (30 g, 32.62 mmol) in one portion. After stirring for 2 h at room temperature, the clear solution was quenched by acetic acid (136.7 μl, 2.39 mmol) and pyridine (193 μl, 2.39 mmol) successively. After removal of the solvent under vacuum, the crude residue was then suspended in CH₂Cl₂ (200 ml) and washed with saturated NaCl solution (200 ml). The organic layer was separated; the aqueous layer was extracted with CH₂Cl₂ (50 ml × 3); and the combined organic layers were dried over Na₂SO₄, filtered and concentrated and dried under high vacuum overnight, giving a mixture of 2a and 2b as a white foam, which was used directly in the phosphorylation. Under inert environment (argon), a mixture of the residue from above and molecule sieve (32.6 g, 4 Å MS, powder, 100 mg ml⁻¹) in dry CH₂Cl₂ (326 ml, 10 ml mmol⁻¹) was stirred for 30 min at room temperature, and then 5-phenyl-1H-tetrazole (10.0 g, 68.50 mmol, 2.10 eq.) was added, followed by the addition of phosphoramidite 6 (19.0 g, 34.25 mmol, 1.05 eq.) in five portions over 2.5 h. After stirring for 30 min at room temperature, the solution was cooled to −78 °C, and meta-chloroperbenzoic acid (8.85 g, 35.88 mmol, 1.1 eq., ~70%) was added slowly, and the resulting mixture was allowed to warm to room temperature and was stirred for 30 min. Then, the reaction was quenched by dimethyl sulfide (479 μl, 6.52 mmol, 0.2 eq.) slowly. After stirring for 30 min, the mixture was filtered, and the filtrate was concentrated under vacuum and purified by flash column chromatography (PE–EtOAc, 1:1) to give a mixture of the trimethylsilyl (TMS)-protected intermediates as a white foam. The resulting foam was dissolved in CH₂Cl₂ (652 ml, HPLC grade, 20 ml mmol⁻¹), and trifluoroacetic acid (32.6 ml, 5%, v/v) was added. After stirring for 30 min at room temperature, the reaction solution was completely concentrated under vacuum, giving a yellow foam (around 17 g). Recrystallization: methanol (100 ml) was added, and the suspension was heated to 55 °C to facilitate a clear solution and then cooled to room temperature slowly. After repeating this ‘heating–cooling’ operation three times, yellow powder appeared. After that, it was left at 4 °C overnight, and the yellow solid was collected by filtration, giving the desired product DMNB-T6P (13.3 g, 50%) as an amorphous yellow powder that was then recrystallized to give a yellow solid: R_f = 0.23 (EtOAc–CH₃OH, 2:1; or EtOAc–CH₃OH, 3:1, plus 0.1% of formic acid (v/v)); mp 124–125 °C (melts and decomposes); [α]_D²⁵ + 63.2 (c 1.0, CH₃OH); ¹H NMR (400 MHz, CD₃OD): δ 7.63 (d, J_p = 0.8 Hz, 2H, ArH), 7.13 (s, 2H, ArH), 5.48–5.44 (m, 4H, ArCH₂O × 2), 5.04 (d, J_1,2 = 3.7 Hz, 1H, H-1), 5.00 (d, J_1′,2′ = 3.7 Hz, 1H, H-1′), 4.42–4.32 (m, 2H, H-6a, H-6b), 4.05–4.01 (m, 1H, H-5), 3.91 (s, 6H, OCH₃ × 2), 3.88 (s, 6H, OCH₃ × 2), 3.82–3.72 (m, 4H, H-5′, H-6′a, H-3, H-3′), 3.66 (dd, J_6′b,6′a = 12.0 Hz, J_6′b,5′ = 5.3 Hz, 1H, H-6′b), 3.44 (dd, J_2,3 = 8.2 Hz, J_2,1 = 3.7 Hz, 1H, H-2), 3.41 (dd, J_2′,3′ = 8.2 Hz, J_2′,1′ = 3.7 Hz, 1H, H-2′), 3.34 (dd, J_4,3 = 9.9 Hz, J_4,5 = 9.0 Hz, 1H, H-4), 3.30 (t, J_4′,3′ = 9.5 Hz, J_4′,5′ = 9.5 Hz, 1H, H-4′H) ppm; ¹³C NMR (100 MHz, CD₃OD) δ 155.2 (qCAr), 150.0 (qCAr), 140.8 (qCAr), 140.7 (qCAr), 127.73 (d, ³J_P,C = 6.6 Hz, qCAr), 127.66 (d, ³J_P,C = 6.6 Hz, qCAr), 111.6 (ArC), 111.5 (ArC), 109.3 (ArC), 95.34 (C-1), 95.27 (C-1′), 74.6 (C-3), 74.4 (C-3′), 73.9 (C-5′), 73.15 (C-2), 73.06 (C-2′), 72.0 (d, ³J_P,C5 = 6.4 Hz, C-5), 71.9 (C-4), 71.2 (C-4′), 68.8 (d, ²J_P,C6 = 5.7 Hz, C-6), 67.91 (d, ²J_P,C = 4.4 Hz, ArCH₂O), 67.88 (d, ²J_P,C = 4.4 Hz, ArCH₂O), 62.6 (C-6′), 57.0 (OCH₃), 56.8 (OCH₃) ppm; high-resolution mass spectrometry (HRMS) (ESI): m/z was calculated for C₃₀H₄₁O₂₂N₂NaP [M+Na]⁺ 835.1781. Found: 835.1772.

DMNB-T6P treatment

DMNB-T6P was dissolved in DMSO with Tween 20 as adjuvant (Supplementary Table 1) fresh for delivery to the crop using a backpack CO₂ sprayer with flat fan type nozzle at a flow of 200 L per hectare, covering the whole plot.

Field trial at CIMMYT, Mexico

Seeds were sown at the CIMMYTʼs Campo Experimental Norman E. Borlaug (CENEB) outside of Ciudad Obregon, Sonora, Mexico (27.372035, −109.924919). The soil type at the experimental station is a coarse sandy clay, mixed montmorillonitic typic caliciorthid, low in organic matter and slightly alkaline (pH 7.7)⁵⁹. Appropriate weed disease and pest control were implemented to avoid yield limitations. Plots were fertilized with 50 kg N per hectare (urea) and 50 kg P per hectare at soil preparation, 50 kg N per hectare with the first irrigation and another 150 kg N per hectare with the second irrigation. Four high-yielding, modern, semi-dwarf, spring wheat genotypes were grown: BACANORA T 88, KAUZ*2/MNV//KAUZ, KAMBARA2 and BORLAUG100 F2014. The plants were sown on 16 December 2021 in a randomized split plot design with DMNB-T6P treatments applied to main plots and cultivars randomized to subplots. Each plot consisted of two beds with two rows, 3.5 m in length. Dose was varied by concentration; four DMNB-T6P treatments were applied in the field: the control (0 T6P), 0.5 mM, 1 mM and 2 mM DMNB-T6P in the volume per m² equivalent to dose 2 adjusted for the sprayed area (Supplementary Table 1). Preparation of the T6P solution was as previously described²³ and as for the field trials in Argentina. The DMNB-T6P solution was applied once to the canopy of the wheat crop in the late afternoon at 10 DAA.

The field trial was harvested on 31 May 2022, after reaching full maturity. Yield components were evaluated following the CIMMYT Wheat Physiology Handbook⁶⁰. Fifty tillers were harvested at random per plot and then brought to the field station at CENEB for further processing. After harvesting the tillers, the spikes were removed from the stems and dried in an oven until reaching a dry constant weight. Seeds were then threshed and used to calculate thousand grain weight (TGW) and grain number (GN). Border plants were excluded from both the final and yield component harvests to minimize border effects between genotypes and treatments.

LEF was measured using a MultispeQ 2.0 (PhotosynQ) and the pre-programmed RIDES protocol. No significant difference was observed between treatments for ambient photosynthetic photon flux density (PPFD) at the time of measurement, indicating that differences in light intensity are not a contributing factor to differences seen between genotypes or treatments (Extended Data Fig. 8). Measurements were made in the field between 10:30 and 14:30 on the wheat flag leaf 3 d after the foliar application of the DMNB-T6P solution. In total, six plants (n = 6) were measured per genotype and treatment. Six plots were measured per genotype and treatment (n = 6). Within the plot, two plants were measured.

Field trials in Argentina

Over four seasons (2018, 2020, 2021 and 2022), field trials were performed under rainfed conditions at the National Institute of Agricultural Research (INTA) Oliveros Research Station, Santa Fe, Argentina (32° 3′ S, 60° 51′ W), in an argiudoll soil with more than 50 years of agricultural history⁶¹. High-yielding commercial Argentinian spring wheat bread-making varieties were chosen with 13–15% grain protein: Buck Saeta, DM Ceibo and MS INTA 415. Buck Saeta is Group 1, suitable for industrial baking. Ceibo is Group 2, suitable for traditional baking (more than 8 h of fermentation). MS INTA is Group 3, suitable for direct baking (less than 8 h of fermentation). No tillage conditions were used following soybean as the previous crop. Dose was varied by spray volume. DMNB-T6P was applied once at 1 mM in two or three separate doses (different volumes) (doses 1–3, at 220 ml, 438 ml or 656 ml per 7-m² plot; Supplementary Table 1). Application was at 10 DAA in 2018, 2020 and 2022 and at 16 DAA in 2021 (due to late delivery of DMNB-T6P), applied in the morning. Calendar timings are shown in Supplementary Table 2. Treatments were arranged in a randomized complete block design with 4–6 replications. Each experimental unit was seven rows spaced 20 cm and 7 m long. The central five rows of each plot were sprayed, giving a spray area of 7 m², of which 3 m² (three central rows 0.6 m × 5 m long) was harvested for grain yield. Phosphorus, sulphur and nitrogen fertilization was performed using super triple phosphate (20% P), calcium sulphate (18% S) and urea, applied at planting at a rate of 100 kg per hectare. N fertilization was estimated by summing pre-plant soil N test as nitrates at 0–60-cm depth (PPNT) plus N added as fertilizer to reach 140 kg per hectare as urea–ammonium nitrate (32% N). N rates were 130, 119, 77 and 101 for years 1, 2, 3 and 4, respectively. Soil organic matter was 2.3% in year 1, 2.5% in year 2, 2.6% in year 3 and 1.9% in year 4, and pH was 5.5, 6.1, 5.9 and 5.8 in the 4 years, respectively.

Weather conditions during the wheat cycle

Cumulative rainfall from May (before crop planting and important for recharging the soil profile) to middle November (when physiological maturity was reached) was 544 mm, 119 mm, 290 mm and 130 mm in years 1, 2, 3 and 4, respectively (Extended Data Fig. 3). These values were 40% above, 70% below, 26% below and 67% below historical records. During the grain filling period (late October to early November), rainfall in years 1, 2, 3 and 4 averaged 125 mm, 58 mm, 74 mm and 39 mm, respectively (9% higher and 49%, 35% and 66% lower than historical records). Maximum and minimum temperature during the cycle ranged averaged from 22.0 °C to 23.7 °C and from 7.1 °C to 8.5 °C in the 4 years. During the grain filling period, maximum temperatures averaged 27.4 °C, and minimum temperatures averaged 12.2 °C. Maximum temperatures were 9% above historical values, and minimum temperatures were 6% below the historical records.

Protein determination

Protein was determined using a NIRS DS2500 analyzer (FOSS Analytical) and fitted to 14% moisture.

Data plotting and statistical analyses

Data are plotted as box plots (Figs. 2–4), which plot the data with medians but not the statistical tests. Statistical analysis of each Argentinian field trial was performed using a two-way factorial ANOVA accounting for the randomized complete block layout in R version 4.2.1. Additionally, a combined analysis over all 4 years was performed using a mixed model framework fitted using REML (Supplementary Table 3). The model consisted of variance components for both block and the blockplot residual separately for each year. Approximate (sequential) F statistics were calculated using Kenward–Roger degrees of freedom. Additionally, standard errors of the difference (SEDs) of the means are plotted as supplementary data (Extended Data Fig. 10). SEDs are shown for comparisons between pairs of overall T6P treatment means and for comparisons between pairs of means for combinations of genotype and T6P treatment based on 34, 24, 33 and 33 degrees of freedom for 2018, 2020, 2021 and 2022 experiments, respectively. Pairwise t-tests were conducted for Fig. 5ciii,d. Analysis of the Mexico field trial was conducted using multi-strata ANOVA to account for the split plot design. Models were fitted in Genstat 22nd edition. We avoid strict thresholding of P values and use of terms ‘significant’ and ‘non-significant’, as biological significance is best understood through examination of statistical tests and P values as a whole over the trialing period incorporating ANOVA analyses (Figs. 2–4) and combined analysis (Supplementary Table 3). We include P values lower than P < 0.1 and do not consider values higher than this. Although, of note, for acceptance as a new biostimulant in the European Union under regulations 2019/1009, such as DMNB-T6P, P < 0.15 values are required (European Document CEN/TS 17700-1:2022, ‘Plant Biostimulants – Claims – Part1: General Principles’ Annex A ‘P-value choice and impact on the results quality’).

Transcriptome analyses

Whole ears were sprayed 10 DAA with 1 mM DMNB-T6P on Cadenza wheat grown in a controlled environment as in ref. ²³. The middle-third of each ear was frozen in liquid nitrogen and stored at −80 C. Whole grain tissue was ground to a fine powder under liquid nitrogen, and total RNA was extracted using the TRIzol method for four independent biological replicates per condition at time 0, 4 h and 24 h after treatment with DMNB-T6P. After RNA integrity analysis and quantitation (Agilent, Bioanalyzer), poly(A)-enriched cDNA libraries were generated and sequenced on an Illumina NovaSeq 6000 sequencing platform generating 30–50 million 150-bp paired-end reads per sample. Low-quality reads and adaptor sequences were removed with Trimmomatic (trimmomatic-0.39.jar PE ILLUMINACLIP:TruSeq3-PE.fa:2:30:10:2:True TRAILING:30 MINLEN:40)⁶². The reads were aligned to the wheat reference genome (Triticum aestivum iwgsc_refseqv2.1 (ref. ⁶³)) using HISAT2/2.2.1-foss-2019b with default parameters⁶⁴ and converted to BAM format with SAMtools⁶⁵. Gene or transcript abundance was quantified using featureCounts⁶⁶ with the High Confidence iwgsc_refseqv2.1 annotation (counting only primary alignments of read pairs with a quality cutoff of 10). The RNA-seq data were deposited under BioProject in the National Center for Biotechnology Information (NCBI) Sequence Read Archive (SRA) (Supplementary Table 5)⁶⁷. Raw counts were normalized using the trimmed mean of M-values method by DESeq2 (ref. ⁶⁷). Differentially expressed genes (DEGs) were identified based on DESeq2 3.15 with adjusted P value (P_adj) < 0.05 and |log₂ fold change (FC)| > 1 as selection criteria. Further statistical analyses and visualizations were conducted in R, and plots and heatmaps were created using the ggplot2 3.4.0, ComplexHeatmap 2.14.0 and tidyheatmap 1.10.0 packages in R^68,69,70.

Microscopy

Whole grains were fixed in 4% paraformaldehyde with 2.5% glutaraldehyde, dehydrated in an ethanol series and embedded in LR White resin (TAAB Laboratories Equipment, Ltd.). Transverse sections of the medial region were imaged after staining with toluidine blue. All samples were imaged with a ×10 objective using an Axio Imager.Z2 (Zeiss). Sieve tube areas in the vascular bundle were manually traced and quantified with ImageJ⁷¹.

Gas exchange

Leaf gas exchange measurements of Cadenza wheat were made with a portable infrared open gas exchange system (LI-COR, LI-6400XT) under the following growing conditions: ambient CO₂ (400 µl l–1), leaf temperature 22 °C, PPFD 500 µmol m⁻² s⁻¹ and relative air humidity 65 ± 5% with an air flow rate of 200 µmol s⁻¹. The middle region of each flag leaf reached a steady state of CO₂ uptake in the leaf chamber before measurements were taken. Data are of four measurements taken at 10 DAA (before treatment), 11 DAA, 12 DAA, 15 DAA and 20 DAA from four separate plants treated with 1 mM DMNB-T6P applied to the spike at 10 DAA after growing under previously described conditions²³.

Treatment of sorghum with DMNB-T6P in controlled environment

Sweet sorghum seeds were grown in 30-cm pots containing Rothamsted compost²³ under 28 °C/22 °C, 12-h day/night cycles, 500 µmol m⁻² s⁻¹ quanta and 60% relative humidity. Regular watering was continued throughout the experiment except for drought stress treatments where watering was reduced to 60% of pot weight at anthesis and maintained at that level of drought until harvest. Eight milliliters per spike of 2 mM DMNB-T6P or control without DMNB-T6P with spray composition as for Argentinian experiments (Supplementary Table 2) was applied to spike only at 7 DAA and 14 DAA. Spikes were harvested at maturity; grain yield was measured; and significance was calculated by Studentʼs t-test. Each treatment contained six biological replicates.

Treatment of barley with DMNB-T6P in controlled environment

Spring barley seeds were grown in 21-cm pots containing Rothamsted compost²³ under 22 °C/18 °C, 16-h day/night cycles, 500 µmol m⁻² s⁻¹ quanta and 60% relative humidity. Regular watering was continued throughout the experiment except for drought stress treatments where watering was reduced at anthesis to 60% of pot weight and maintained at that level of drought until harvest. Twenty milliliters per plant of 2 mM DMNB-T6P or control without DMNB-T6P with spray composition as for Argentinian experiments (Supplementary Table 2) was applied to the upper canopy, including spikes, at 6 DAA and 11 DAA. Spikes were harvested at maturity; grain yield was measured; and significance was calculated by Studentʼs t-test. Each treatment contained six biological replicates.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Hungate, R. E., Smith, W. & Clarke, R. T. Suitability of butyl rubber stoppers for closing anaerobic roll culture tubes. J. Bacteriol. 91, 908–909 (1966).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Overmann, J., Abt, B. & Sikorski, J. Present and future of culturing bacteria. Annu. Rev. Microbiol. 71, 711–730 (2017).

Article 
CAS 
PubMed 

Google Scholar 

Lewis, W. H., Tahon, G., Geesink, P., Sousa, D. Z. & Ettema, T. J. G. Innovations to culturing the uncultured microbial majority. Nat. Rev. Microbiol. 19, 225–240 (2021).

Article 
CAS 
PubMed 

Google Scholar 

Barcenilla, A. et al. Phylogenetic relationships of butyrate-producing bacteria from the human gut. Appl. Environ. Microbiol. 66, 1654–1661 (2000).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Collins, M. D. et al. The phylogeny of the genus Clostridium: proposal of five new genera and eleven new species combinations. Int. J. Syst. Bacteriol. 44, 812–826 (1994).

Article 
CAS 
PubMed 

Google Scholar 

Gibson, G. R., Macfarlane, G. T. & Cummings, J. H. Occurrence of sulphate-reducing bacteria in human faeces and the relationship of dissimilatory sulphate reduction to methanogenesis in the large gut. J. Appl. Bacteriol. 65, 103–111 (1988).

Article 
CAS 
PubMed 

Google Scholar 

Holdeman, L. V., Good, I. J. & Moore, W. E. Human fecal flora: variation in bacterial composition within individuals and a possible effect of emotional stress. Appl. Environ. Microbiol. 31, 359–375 (1976).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Sakamoto, M. et al. Prevotella shahii sp. nov. and Prevotella salivae sp. nov., isolated from the human oral cavity. Int. J. Syst. Evol. Microbiol. 54, 877–883 (2004).

Article 
CAS 
PubMed 

Google Scholar 

Schwiertz, A. et al. Anaerostipes caccae gen. nov., sp. nov., a new saccharolytic, acetate-utilising, butyrate-producing bacterium from human faeces. Syst. Appl. Microbiol. 25, 46–51 (2002).

Article 
CAS 
PubMed 

Google Scholar 

Walker, A. W. & Hoyles, L. Human microbiome myths and misconceptions. Nat. Microbiol. 8, 1392–1396 (2023).

Article 
CAS 
PubMed 

Google Scholar 

Lagier, J. C. et al. Microbial culturomics: paradigm shift in the human gut microbiome study. Clin. Microbiol. Infect. 18, 1185–1193 (2012).

Article 
CAS 
PubMed 

Google Scholar 

Rahi, P., Prakash, O. & Shouche, Y. S. Matrix-assisted laser desorption/ionization time-of-flight mass-spectrometry (MALDI–TOF MS) based microbial identifications: challenges and scopes for microbial ecologists. Front. Microbiol. 7, 1359 (2016).

Article 
PubMed 
PubMed Central 

Google Scholar 

Afrizal, A. et al. Enhanced cultured diversity of the mouse gut microbiota enables custom-made synthetic communities. Cell Host Microbe 30, 1630–1645 (2022).

Article 
CAS 
PubMed 

Google Scholar 

Browne, H. P. et al. Culturing of ‘unculturable’ human microbiota reveals novel taxa and extensive sporulation. Nature 533, 543–546 (2016).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Forster, S. C. et al. A human gut bacterial genome and culture collection for improved metagenomic analyses. Nat. Biotechnol. 37, 186–192 (2019).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Furuichi, M. et al. Commensal consortia decolonize Enterobacteriaceae via ecological control. Nature 633, 878–886 (2024).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Groussin, M. et al. Elevated rates of horizontal gene transfer in the industrialized human microbiome. Cell 184, 2053–2067 (2021).

Article 
CAS 
PubMed 

Google Scholar 

Hitch, T. C. A. et al. Broad diversity of human gut bacteria accessible via a traceable strain deposition system. Preprint at bioRxiv https://doi.org/10.1101/2024.06.20.599854 (2024).

Liu, C. et al. Enlightening the taxonomy darkness of human gut microbiomes with a cultured biobank. Microbiome 9, 119 (2021).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Seshadri, R. et al. Cultivation and sequencing of rumen microbiome members from the Hungate1000 Collection. Nat. Biotechnol. 36, 359–367 (2018).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Sorbara, M. T. et al. Functional and genomic variation between human-derived isolates of Lachnospiraceae reveals inter- and intra-species diversity. Cell Host Microbe 28, 134–146 (2020).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Wylensek, D. et al. A collection of bacterial isolates from the pig intestine reveals functional and taxonomic diversity. Nat. Commun. 11, 6389 (2020).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Zenner, C. et al. Early-life immune system maturation in chickens using a synthetic community of cultured gut bacteria. mSystems 6, e01300-20 (2021).

Article 
PubMed 
PubMed Central 

Google Scholar 

Zhang, Z. J. et al. Comprehensive analyses of a large human gut Bacteroidales culture collection reveal species- and strain-level diversity and evolution. Cell Host Microbe 32, 1853–1867 (2024).

Article 
CAS 
PubMed 

Google Scholar 

Zou, Y. et al. 1,520 reference genomes from cultivated human gut bacteria enable functional microbiome analyses. Nat. Biotechnol. 37, 179–185 (2019).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Hitch, T. C. A. et al. Recent advances in culture-based gut microbiome research. Int. J. Med. Microbiol. 311, 151485 (2021).

Article 
PubMed 

Google Scholar 

Martiny, A. C. The ‘1% culturability paradigm’ needs to be carefully defined. ISME J. 14, 10–11 (2020).

Article 
PubMed 

Google Scholar 

Rodriguez Del Rio, A. et al. Functional and evolutionary significance of unknown genes from uncultivated taxa. Nature 626, 377–384 (2024).

Article 
CAS 
PubMed 

Google Scholar 

Thomas, A. M. & Segata, N. Multiple levels of the unknown in microbiome research. BMC Biol. 17, 48 (2019).

Article 
PubMed 
PubMed Central 

Google Scholar 

Blanco-Miguez, A. et al. Extension of the Segatella copri complex to 13 species with distinct large extrachromosomal elements and associations with host conditions. Cell Host Microbe 31, 1804–1819 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Yin, Q. et al. Ecological dynamics of Enterobacteriaceae in the human gut microbiome across global populations. Nat. Microbiol. 10, 541–553 (2025).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Afrizal, A. et al. Anaerobic single-cell dispensing facilitates the cultivation of human gut bacteria. Environ. Microbiol. 24, 3861–3881 (2022).

Article 
CAS 
PubMed 

Google Scholar 

Huang, Y. et al. High-throughput microbial culturomics using automation and machine learning. Nat. Biotechnol. 41, 1424–1433 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Muller, P. et al. High-throughput anaerobic screening for identifying compounds acting against gut bacteria in monocultures or communities. Nat. Protoc. 19, 668–699 (2024).

Article 
PubMed 

Google Scholar 

Sanders, J. G. et al. A low-cost genomics workflow enables isolate screening and strain-level analyses within microbiomes. Genome Biol. 23, 212 (2022).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Goodman, A. L. et al. Extensive personal human gut microbiota culture collections characterized and manipulated in gnotobiotic mice. Proc. Natl Acad. Sci. USA 108, 6252–6257 (2011).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Lee, K. S. et al. An automated Raman-based platform for the sorting of live cells by functional properties. Nat. Microbiol. 4, 1035–1048 (2019).

Article 
CAS 
PubMed 

Google Scholar 

Terekhov, S. S. et al. Ultrahigh-throughput functional profiling of microbiota communities. Proc. Natl Acad. Sci. USA 115, 9551–9556 (2018).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Watterson, W. J. et al. Droplet-based high-throughput cultivation for accurate screening of antibiotic resistant gut microbes. eLife 9, e56998 (2020).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Bellais, S. et al. Species-targeted sorting and cultivation of commensal bacteria from the gut microbiome using flow cytometry under anaerobic conditions. Microbiome 10, 24 (2022).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Pauvert, C. maldipickr: dereplicate and cherry-pick mass spectrometry spectra. https://doi.org/10.32614/cran.package.maldipickr (2023).

Helleckes, L. M. et al. From frozen cell bank to product assay: high-throughput strain characterisation for autonomous design–build–test–learn cycles. Microb. Cell Fact. 22, 130 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Aranda-Diaz, A. et al. Establishment and characterization of stable, diverse, fecal-derived in vitro microbial communities that model the intestinal microbiota. Cell Host Microbe 30, 260–272 (2022).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Cheng, A. G. et al. Design, construction, and in vivo augmentation of a complex gut microbiome. Cell 185, 3617–3636 (2022).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Goldman, S. L. et al. Culture-enriched community profiling improves resolution of the vertebrate gut microbiota. Mol. Ecol. Resour. 22, 122–136 (2022).

Article 
CAS 
PubMed 

Google Scholar 

Rettedal, E. A., Gumpert, H. & Sommer, M. O. Cultivation-based multiplex phenotyping of human gut microbiota allows targeted recovery of previously uncultured bacteria. Nat. Commun. 5, 4714 (2014).

Article 
CAS 
PubMed 

Google Scholar 

Tramontano, M. et al. Nutritional preferences of human gut bacteria reveal their metabolic idiosyncrasies. Nat. Microbiol. 3, 514–522 (2018).

Article 
CAS 
PubMed 

Google Scholar 

van den Berg, N. I. et al. Emergent survival and extinction of species within gut bacterial communities. Preprint at bioRxiv https://doi.org/10.1101/2024.04.29.591619 (2024).

Wornell, K. et al. High-throughput method for novel medium development for culture of anaerobic gut bacteria. Curr. Protoc. 2, e463 (2022).

Article 
CAS 
PubMed 

Google Scholar 

Koblitz, J., Reimer, L. C., Pukall, R. & Overmann, J. Predicting bacterial phenotypic traits through improved machine learning using high-quality, curated datasets. Preprint at bioRxiv https://doi.org/10.1101/2024.08.12.607695 (2024).

Koblitz, J. et al. MediaDive: the expert-curated cultivation media database. Nucleic Acids Res. 51, D1531–D1538 (2023).

Article 
PubMed 

Google Scholar 

Blount, Z. D., Barrick, J. E., Davidson, C. J. & Lenski, R. E. Genomic analysis of a key innovation in an experimental Escherichia coli population. Nature 489, 513–518 (2012).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Riedel, T. et al. Genome resequencing of the virulent and multidrug-resistant reference strain Clostridium difficile 630. Genome Announc. 3, e00276-15 (2015).

Article 
PubMed 
PubMed Central 

Google Scholar 

Garcia-Bayona, L. & Comstock, L. E. Streamlined genetic manipulation of diverse Bacteroides and Parabacteroides isolates from the human gut microbiota. mBio 10, e01762-19 (2019).

Article 
PubMed 
PubMed Central 

Google Scholar 

Tripathi, S. et al. Randomly barcoded transposon mutant libraries for gut commensals I: strategies for efficient library construction. Cell Rep. 43, 113517 (2024).

Article 
CAS 
PubMed 

Google Scholar 

Davey, L. E. et al. A genetic system for Akkermansia muciniphila reveals a role for mucin foraging in gut colonization and host sterol biosynthesis gene expression. Nat. Microbiol. 8, 1450–1467 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Jin, W. B. et al. Genetic manipulation of gut microbes enables single-gene interrogation in a complex microbiome. Cell 185, 547–562 (2022).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Liu, H. et al. Magic pools: parallel assessment of transposon delivery vectors in bacteria. mSystems 3, e00143-17 (2018).

Article 
PubMed 
PubMed Central 

Google Scholar 

Mutalik, V. K. et al. Precise and reliable gene expression via standard transcription and translation initiation elements. Nat. Methods 10, 354–360 (2013).

Article 
CAS 
PubMed 

Google Scholar 

Whitaker, W. R., Shepherd, E. S. & Sonnenburg, J. L. Tunable expression tools enable single-cell strain distinction in the gut microbiome. Cell 169, 538–546 (2017).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Galardini, M. et al. Phenotype inference in an Escherichia coli strain panel. eLife 6, e31035 (2017).

Article 
PubMed 
PubMed Central 

Google Scholar 

Liuu, S. et al. Identification of a muropeptide precursor transporter from gut microbiota and its role in preventing intestinal inflammation. Proc. Natl Acad. Sci. USA 120, e2306863120 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Nasseri, S. A. et al. An alternative broad-specificity pathway for glycan breakdown in bacteria. Nature 631, 199–206 (2024).

Article 
CAS 
PubMed 

Google Scholar 

Zimmermann, M., Zimmermann-Kogadeeva, M., Wegmann, R. & Goodman, A. L. Mapping human microbiome drug metabolism by gut bacteria and their genes. Nature 570, 462–467 (2019).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Vercauteren, S. et al. The rise and future of CRISPR-based approaches for high-throughput genomics. FEMS Microbiol. Rev. 48, fuae020 (2024).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Voogdt, C. G. P. et al. Randomly barcoded transposon mutant libraries for gut commensals II: applying libraries for functional genetics. Cell Rep. 43, 113519 (2024).

Article 
CAS 
PubMed 

Google Scholar 

Nichols, R. J. et al. Phenotypic landscape of a bacterial cell. Cell 144, 143–156 (2011).

Article 
CAS 
PubMed 

Google Scholar 

Price, M. N. et al. Mutant phenotypes for thousands of bacterial genes of unknown function. Nature 557, 503–509 (2018).

Article 
CAS 
PubMed 

Google Scholar 

Gibson, G. R., Cummings, J. H. & Macfarlane, G. T. Use of a three-stage continuous culture system to study the effect of mucin on dissimilatory sulfate reduction and methanogenesis by mixed populations of human gut bacteria. Appl. Environ. Microbiol. 54, 2750–2755 (1988).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

McDonald, J. A. et al. Evaluation of microbial community reproducibility, stability and composition in a human distal gut chemostat model. J. Microbiol. Methods 95, 167–174 (2013).

Article 
CAS 
PubMed 

Google Scholar 

Possemiers, S., Verthe, K., Uyttendaele, S. & Verstraete, W. PCR–DGGE-based quantification of stability of the microbial community in a simulator of the human intestinal microbial ecosystem. FEMS Microbiol. Ecol. 49, 495–507 (2004).

Article 
CAS 
PubMed 

Google Scholar 

Silverman, J. D., Durand, H. K., Bloom, R. J., Mukherjee, S. & David, L. A. Dynamic linear models guide design and analysis of microbiota studies within artificial human guts. Microbiome 6, 202 (2018).

Article 
PubMed 
PubMed Central 

Google Scholar 

Probert, H. M. & Gibson, G. R. Development of a fermentation system to model sessile bacterial populations in the human colon. Biofilms 1, 13–19 (2004).

Article 

Google Scholar 

Jin, X. et al. Culturing of a complex gut microbial community in mucin-hydrogel carriers reveals strain- and gene-associated spatial organization. Nat. Commun. 14, 3510 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Liu, L. et al. Establishing a mucosal gut microbial community in vitro using an artificial simulator. PLoS ONE 13, e0197692 (2018).

Article 
PubMed 
PubMed Central 

Google Scholar 

Tran, T. H. et al. Adding mucins to an in vitro batch fermentation model of the large intestine induces changes in microbial population isolated from porcine feces depending on the substrate. FEMS Microbiol. Ecol. 92, fiv165 (2016).

Article 
PubMed 

Google Scholar 

Van den Abbeele, P. et al. Incorporating a mucosal environment in a dynamic gut model results in a more representative colonization by lactobacilli. Microb. Biotechnol. 5, 106–115 (2012).

Article 
PubMed 

Google Scholar 

Reese, A. T. et al. Antibiotic-induced changes in the microbiota disrupt redox dynamics in the gut. eLife 7, e35987 (2018).

Article 
PubMed 
PubMed Central 

Google Scholar 

Auchtung, J. M., Robinson, C. D. & Britton, R. A. Cultivation of stable, reproducible microbial communities from different fecal donors using minibioreactor arrays (MBRAs). Microbiome 3, 42 (2015).

Article 
PubMed 
PubMed Central 

Google Scholar 

Mahnic, A., Auchtung, J. M., Poklar Ulrih, N., Britton, R. A. & Rupnik, M. Microbiota in vitro modulated with polyphenols shows decreased colonization resistance against Clostridioides difficile but can neutralize cytotoxicity. Sci. Rep. 10, 8358 (2020).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Hobson, C. A. et al. MiniBioReactor Array (MBRA) in vitro gut model: a reliable system to study microbiota-dependent response to antibiotic treatment. JAC Antimicrob. Resist. 4, dlac077 (2022).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Naimi, S., Viennois, E., Gewirtz, A. T. & Chassaing, B. Direct impact of commonly used dietary emulsifiers on human gut microbiota. Microbiome 9, 66 (2021).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Yu, Y. et al. Emerging microfluidic technologies for microbiome research. Front. Microbiol. 13, 906979 (2022).

Article 
PubMed 
PubMed Central 

Google Scholar 

Ma, L. et al. Individually addressable arrays of replica microbial cultures enabled by splitting SlipChips. Integr. Biol. 6, 796–805 (2014).

Article 
CAS 

Google Scholar 

Ge, Z., Girguis, P. R. & Buie, C. R. Nanoporous microscale microbial incubators. Lab Chip 16, 480–488 (2016).

Article 
CAS 
PubMed 

Google Scholar 

Ma, L. et al. Gene-targeted microfluidic cultivation validated by isolation of a gut bacterium listed in Human Microbiome Project’s Most Wanted taxa. Proc. Natl Acad. Sci. USA 111, 9768–9773 (2014).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Tan, J. Y. et al. Co-cultivation of microbial sub-communities in microfluidic droplets facilitates high-resolution genomic dissection of microbial ‘dark matter’. Integr. Biol. 12, 263–274 (2020).

Article 

Google Scholar 

Kehe, J. et al. Massively parallel screening of synthetic microbial communities. Proc. Natl Acad. Sci. USA 116, 12804–12809 (2019).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Kehe, J. et al. Positive interactions are common among culturable bacteria. Sci. Adv. 7, eabi7159 (2021).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Westermann, A. J. & Vogel, J. Cross-species RNA-seq for deciphering host–microbe interactions. Nat. Rev. Genet. 22, 361–378 (2021).

Article 
CAS 
PubMed 

Google Scholar 

Wilmes, P., Heintz-Buschart, A. & Bond, P. L. A decade of metaproteomics: where we stand and what the future holds. Proteomics 15, 3409–3417 (2015).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Krautkramer, K. A., Fan, J. & Backhed, F. Gut microbial metabolites as multi-kingdom intermediates. Nat. Rev. Microbiol. 19, 77–94 (2021).

Article 
CAS 
PubMed 

Google Scholar 

Ricaurte, D. et al. High-throughput transcriptomics of 409 bacteria–drug pairs reveals drivers of gut microbiota perturbation. Nat. Microbiol. 9, 561–575 (2024).

Article 
PubMed 
PubMed Central 

Google Scholar 

Dinglasan, J. L. N., Otani, H., Doering, D. T., Udwary, D. & Mouncey, N. J. Microbial secondary metabolites: advancements to accelerate discovery towards application. Nat. Rev. Microbiol. https://doi.org/10.1038/s41579-024-01141-y (2025).

Scherlach, K. & Hertweck, C. Triggering cryptic natural product biosynthesis in microorganisms. Org. Biomol. Chem. 7, 1753–1760 (2009).

Article 
CAS 
PubMed 

Google Scholar 

Gabrielli, N. et al. Unravelling metabolic cross-feeding in a yeast-bacteria community using ¹³C-based proteomics. Mol. Syst. Biol. 19, e11501 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Hofer, K. & Jaschke, A. Epitranscriptomics: RNA modifications in bacteria and archaea. Microbiol. Spectr. 6, 10.1128/microbiolspec.rwr-0015-2017 (2018).

Birk, M. S., Charpentier, E. & Frese, C. K. Automated phosphopeptide enrichment for Gram-positive bacteria. J. Proteome Res. 20, 4886–4892 (2021).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Homberger, C., Barquist, L. & Vogel, J. Ushering in a new era of single-cell transcriptomics in bacteria. Microlife 3, uqac020 (2022).

Article 
PubMed 
PubMed Central 

Google Scholar 

Vegvari, A., Zhang, X. & Zubarev, R. A. Toward single bacterium proteomics. J. Am. Soc. Mass. Spectrom. 34, 2098–2106 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Duetz, W. A. et al. Methods for intense aeration, growth, storage, and replication of bacterial strains in microtiter plates. Appl. Environ. Microbiol. 66, 2641–2646 (2000).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Minich, J. J. et al. Quantifying and understanding well-to-well contamination in microbiome research. mSystems 4, e00186-19 (2019).

Article 
PubMed 
PubMed Central 

Google Scholar 

Enning, D. et al. Marine sulfate-reducing bacteria cause serious corrosion of iron under electroconductive biogenic mineral crust. Environ. Microbiol. 14, 1772–1787 (2012).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Zund, J. N. et al. A flexible high-throughput cultivation protocol to assess the response of individuals’ gut microbiota to diet-, drug-, and host-related factors. ISME Commun. 4, ycae035 (2024).

Article 
PubMed 
PubMed Central 

Google Scholar 

Bartscht, K., Cypionka, H. & Overmann, J. Evaluation of cell activity and of methods for the cultivation of bacteria from a natural lake community. FEMS Microbiol. Ecol. 28, 249–259 (1999).

Article 
CAS 

Google Scholar 

Diaz-Colunga, J., Catalan, P., San Roman, M., Arrabal, A. & Sanchez, A. Full factorial construction of synthetic microbial communities. Preprint at bioRxiv https://doi.org/10.1101/2024.05.03.592148 (2024).

Raj, K. et al. Automation assisted anaerobic phenotyping for metabolic engineering. Microb. Cell Fact. 20, 184 (2021).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Perez-Sepulveda, B. M. et al. An accessible, efficient and global approach for the large-scale sequencing of bacterial genomes. Genome Biol. 22, 349 (2021).

Article 
PubMed 
PubMed Central 

Google Scholar 

Zhang, J. et al. High-throughput cultivation and identification of bacteria from the plant root microbiota. Nat. Protoc. 16, 988–1012 (2021).

Article 
CAS 
PubMed 

Google Scholar 

Mortier, T., Wieme, A. D., Vandamme, P. & Waegeman, W. Bacterial species identification using MALDI–TOF mass spectrometry and machine learning techniques: a large-scale benchmarking study. Comput. Struct. Biotechnol. J. 19, 6157–6168 (2021).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Uvarova, Y. E. et al. Accurate noise-robust classification of Bacillus species from MALDI–TOF MS spectra using a denoising autoencoder. J. Integr. Bioinform. 20, 20230017 (2023).

Article 
PubMed 
PubMed Central 

Google Scholar 

Wang, H. Y. et al. Rapid classification of group B Streptococcus serotypes based on matrix-assisted laser desorption ionization–time of flight mass spectrometry and machine learning techniques. BMC Bioinformatics 20, 703 (2019).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

King, R. D., Costa, V. S., Mellingwood, C. & Soldatova, L. N. Automating sciences: philosophical and social dimensions. IEEE Technol. Soc. Mag. 37, 40–46 (2018).

Article 

Google Scholar 

Jing, J., Garbeva, P., Raaijmakers, J. M. & Medema, M. H. Strategies for tailoring functional microbial synthetic communities. ISME J. 18, wrae049 (2024).

Article 
PubMed 
PubMed Central 

Google Scholar 

Kumar, N., Hitch, T. C. A., Haller, D., Lagkouvardos, I. & Clavel, T. MiMiC: a bioinformatic approach for generation of synthetic communities from metagenomes. Microb. Biotechnol. 14, 1757–1770 (2021).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Raghu, A. K., Palanikumar, I. & Raman, K. Designing function-specific minimal microbiomes from large microbial communities. NPJ Syst. Biol. Appl. 10, 46 (2024).

Article 
PubMed 
PubMed Central 

Google Scholar 

Connors, B. M. et al. Control points for design of taxonomic composition in synthetic human gut communities. Cell Syst. 14, 1044–1058 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Predl, M., Miesskes, M., Rattei, T. & Zanghellini, J. PyCoMo: a Python package for community metabolic model creation and analysis. Bioinformatics 40, btae153 (2024).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

van der Lelie, D. et al. Rationally designed bacterial consortia to treat chronic immune-mediated colitis and restore intestinal homeostasis. Nat. Commun. 12, 3105 (2021).

Article 
PubMed 
PubMed Central 

Google Scholar 

Haby, B. et al. Integrated robotic mini bioreactor platform for automated, parallel microbial cultivation with online data handling and process control. SLAS Technol. 24, 569–582 (2019).

Article 
CAS 
PubMed 

Google Scholar 

Halle, L. et al. Robotic workflows for automated long-term adaptive laboratory evolution: improving ethanol utilization by Corynebacterium glutamicum. Microb. Cell Fact. 22, 175 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Heux, S., Poinot, J., Massou, S., Sokol, S. & Portais, J. C. A novel platform for automated high-throughput fluxome profiling of metabolic variants. Metab. Eng. 25, 8–19 (2014).

Article 
CAS 
PubMed 

Google Scholar 

Louie, T. et al. VE303, a defined bacterial consortium, for prevention of recurrent Clostridioides difficile infection: a randomized clinical trial. JAMA 329, 1356–1366 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Minkoff, N. Z. et al. Fecal microbiota transplantation for the treatment of recurrent Clostridioides difficile (Clostridium difficile). Cochrane Database Syst. Rev. 4, CD013871 (2023).

PubMed 

Google Scholar 

Baunwall, S. M. D. et al. Faecal microbiota transplantation for recurrent Clostridioides difficile infection: an updated systematic review and meta-analysis. EClinicalMedicine 29–30, 100642 (2020).

Article 
PubMed 
PubMed Central 

Google Scholar 

Eberl, C. et al. E. coli enhance colonization resistance against Salmonella Typhimurium by competing for galactitol, a context-dependent limiting carbon source. Cell Host Microbe 29, 1680–1692 (2021).

Article 
CAS 
PubMed 

Google Scholar 

Lawley, T. D. et al. Targeted restoration of the intestinal microbiota with a simple, defined bacteriotherapy resolves relapsing Clostridium difficile disease in mice. PLoS Pathog. 8, e1002995 (2012).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Osbelt, L. et al. Klebsiella oxytoca causes colonization resistance against multidrug-resistant K. pneumoniae in the gut via cooperative carbohydrate competition. Cell Host Microbe 29, 1663–1679 (2021).

Article 
CAS 
PubMed 

Google Scholar 

Spragge, F. et al. Microbiome diversity protects against pathogens by nutrient blocking. Science 382, eadj3502 (2023).

Article 
CAS 
PubMed 
PubMed Central 

Google Scholar 

Kurt, F. et al. Co-cultivation is a powerful approach to produce a robust functionally designed synthetic consortium as a live biotherapeutic product (LBP). Gut Microbes 15, 2177486 (2023).

Article 
PubMed 
PubMed Central 

Google Scholar 

DuPont, H. L., DuPont, A. W. & Tillotson, G. S. Microbiota restoration therapies for recurrent Clostridioides difficile infection reach an important new milestone. Ther. Adv. Gastroenterol. 17, 17562848241253089 (2024).

Article 
CAS 

Google Scholar 

El Hage Chehade, N. et al. Efficacy of fecal microbiota transplantation in the treatment of active ulcerative colitis: a systematic review and meta-analysis of double-blind randomized controlled trials. Inflamm. Bowel Dis. 29, 808–817 (2023).

Article 
PubMed 

Google Scholar 

Feuerstadt, P. et al. SER-109, an oral microbiome therapy for recurrent Clostridioides difficile infection. N. Engl. J. Med. 386, 220–229 (2022).

Article 
CAS 
PubMed 

Google Scholar 

Radlinski, L. C. & Baumler, A. J. Microbiome science needs more microbiologists. Nat. Microbiol. 10, 263–264 (2025).

Article 
CAS 
PubMed 

Google Scholar