Martin Larralde
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Martin Larralde has a total of 302 co-authors. Here are the top co-authors by number of shared publications:
- Georg Zeller
Leiden University Medical Center
11 shared publications - Laura M. Carroll
Department of Clinical Microbiology, Umeå University, Umeå, Sweden
5 shared publications - Christoph Steinbeck
Institute for Inorganic and Analytical Chemistry, Friedrich-Schiller-University, Lessingstr. 8, 07743 Jena, Germany
4 shared publications - Josefin Blom
Department of Clinical Microbiology, Umeå University , Umeå, Västerbotten 90187
4 shared publications - Kenneth Haug
European Molecular Biology Laboratory, European Bioinformatics Institute (EMBL-EBI), Wellcome Genome Campus, Hinxton, Cambridge CB10 1SD, United Kingdom
4 shared publications - Quinten Raymond Ducarmon
Centre for Microbiome Analyses and Therapeutics, Leiden University Center for Infectious Diseases (LUCID), Leiden University Medical Center , 2333ZA, Leiden
4 shared publications
Latest Publications
- Open Access
- 17 November 2026
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 22 (11) , e1014925
Ontology pre-training improves machine learning-based predictions for metabolites
Deep learning has recently led to multiple breakthroughs in longstanding biological problems. Recent advances have shown that integration of expert knowledge can improve deep learning model predictive performance in complex domains such as biology. Bio-ontologies offer a rich source of curated biological knowledge that can be harnessed to this end. Here, we describe an intuitive and generalisable approach to embed the knowledge contained in a classification hierarchy derived from a bio-ontology into a deep learning model as an intermediate training step between general-purpose pre-training and task-specific fine-tuning: a process that we call’ontology pre-training’. We show that this approach leads to an improvement in predictive performance and a reduction in training time for a broad range of prediction tasks related to metabolite functions, activities and properties in living systems, using a range of datasets derived from MoleculeNet. We see the biggest improvement for regression tasks, e.g., prediction of lipophilicity and aqueous solubility of molecules, and a robust improvement for most classification tasks. Our approach can be adapted for a wide range of knowledge sources, models and prediction tasks.
- 17 November 2026
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 22 (11) , e1014925
Open AccessOntology pre-training improves machine learning-based predictions for metabolites
Deep learning has recently led to multiple breakthroughs in longstanding biological problems. Recent advances have shown that integration of expert knowledge can improve deep learning model predictive performance in complex domains such as biology. Bio-ontologies offer a rich source of curated biological knowledge that can be harnessed to this end. Here, we describe an intuitive and generalisable approach to embed the knowledge contained in a classification hierarchy derived from a bio-ontology into a deep learning model as an intermediate training step between general-purpose pre-training and task-specific fine-tuning: a process that we call’ontology pre-training’. We show that this approach leads to an improvement in predictive performance and a reduction in training time for a broad range of prediction tasks related to metabolite functions, activities and properties in living systems, using a range of datasets derived from MoleculeNet. We see the biggest improvement for regression tasks, e.g., prediction of lipophilicity and aqueous solubility of molecules, and a robust improvement for most classification tasks. Our approach can be adapted for a wide range of knowledge sources, models and prediction tasks.
- Open Access
- 13 October 2026
- Published by Oxford University Press (OUP) in Nucleic Acids Research
- Vol. 54 (19)
Fast, flexible gene cluster family delineation with IGUA
Prokaryotic genomes harbor a variety of functional elements encoded as contiguous multi-gene clusters. To identify homologous gene clusters, computational tools that account for multi-gene architectures are needed to group gene clusters into Gene Cluster Families (GCFs). However, existing GCF delineation methods do not scale well to large datasets and are often limited to certain subclasses of gene clusters. Here, we present IGUA (Iterative Gene clUster Analysis; https://github.com/zellerlab/IGUA), a scalable, flexible GCF delineation method for genomic segments with multi-gene architectures. On a biosynthetic gene cluster (BGC) clustering task, IGUA is 29× and 10× faster than the state-of-the-art (BiG-SCAPE v2 and BiG-SLiCE v2, respectively; 28 CPUs), without sacrificing accuracy. To highlight its scalability, we use IGUA to cluster >2.8 million BGCs from ≈1 million prokaryotic genomes in <18 h (n = 2 829 071 BGCs to 56 960 GCFs). To showcase its utility beyond BGC clustering, we use IGUA to cluster five additional prokaryotic gene cluster types (secretion systems, prophages, genomic islands, phage-plasmids, and carbohydrate active enzyme [CAZyme] gene clusters). Overall, IGUA represents a versatile GCF delineation tool with unmatched computational efficiency and flexibility, enabling (meta)genomic mining applications at unprecedented scales.
- 13 October 2026
- Published by Oxford University Press (OUP) in Nucleic Acids Research
- Vol. 54 (19)
Open AccessFast, flexible gene cluster family delineation with IGUA
Prokaryotic genomes harbor a variety of functional elements encoded as contiguous multi-gene clusters. To identify homologous gene clusters, computational tools that account for multi-gene architectures are needed to group gene clusters into Gene Cluster Families (GCFs). However, existing GCF delineation methods do not scale well to large datasets and are often limited to certain subclasses of gene clusters. Here, we present IGUA (Iterative Gene clUster Analysis; https://github.com/zellerlab/IGUA), a scalable, flexible GCF delineation method for genomic segments with multi-gene architectures. On a biosynthetic gene cluster (BGC) clustering task, IGUA is 29× and 10× faster than the state-of-the-art (BiG-SCAPE v2 and BiG-SLiCE v2, respectively; 28 CPUs), without sacrificing accuracy. To highlight its scalability, we use IGUA to cluster >2.8 million BGCs from ≈1 million prokaryotic genomes in <18 h (n = 2 829 071 BGCs to 56 960 GCFs). To showcase its utility beyond BGC clustering, we use IGUA to cluster five additional prokaryotic gene cluster types (secretion systems, prophages, genomic islands, phage-plasmids, and carbohydrate active enzyme [CAZyme] gene clusters). Overall, IGUA represents a versatile GCF delineation tool with unmatched computational efficiency and flexibility, enabling (meta)genomic mining applications at unprecedented scales.
- Open Access
- 6 August 2026
- Published by Wiley in Protein Science
- Vol. 35 (9)
Investigating enzyme function by geometric matching of catalytic motifs
The rapidly growing universe of predicted protein structures offers opportunities for data driven exploration but requires computationally scalable and interpretable tools. We developed a method to detect catalytic features in protein structures, providing insights into enzyme function and mechanism. A library of 6780 3D coordinate sets describing enzyme catalytic sites, referred to as templates, has been collected from manually curated examples of 762 enzyme catalytic mechanisms described in the Mechanism and Catalytic Site Atlas. For template searching we optimized the geometric‐matching algorithm Jess. We implemented RMSD and residue orientation filters to differentiate catalytically informative matches from spurious ones. We validated this approach on a non‐redundant set of high quality experimental ( n = 3751, <40% amino acid identity) enzyme structures with well annotated catalytic sites as well as predicted structures of the human proteome. We show matching catalytic templates solely on structure is more sensitive than sequence‐ and 3D‐structure‐based approaches in identifying homology between distantly related enzymes. Since geometric matching does not depend on conserved sequence motifs or even common evolutionary history, we are able to identify examples of structural active site similarity in highly divergent and possibly convergent enzymes. Such examples make interesting case studies into the evolution of enzyme function. Though not intended for characterizing substrate‐specific binding pockets, the speed and knowledge‐driven interpretability of our method make it well suited for expanding enzyme active‐site annotation across large predicted proteomes. We provide the method and template library as a Python module, Enzyme Motif Miner, at https://github.com/rayhackett/enzymm and as a webserver at https://www.ebi.ac.uk/thornton-srv/m-csa/enzymm .
- 6 August 2026
- Published by Wiley in Protein Science
- Vol. 35 (9)
Open AccessInvestigating enzyme function by geometric matching of catalytic motifs
The rapidly growing universe of predicted protein structures offers opportunities for data driven exploration but requires computationally scalable and interpretable tools. We developed a method to detect catalytic features in protein structures, providing insights into enzyme function and mechanism. A library of 6780 3D coordinate sets describing enzyme catalytic sites, referred to as templates, has been collected from manually curated examples of 762 enzyme catalytic mechanisms described in the Mechanism and Catalytic Site Atlas. For template searching we optimized the geometric‐matching algorithm Jess. We implemented RMSD and residue orientation filters to differentiate catalytically informative matches from spurious ones. We validated this approach on a non‐redundant set of high quality experimental ( n = 3751, <40% amino acid identity) enzyme structures with well annotated catalytic sites as well as predicted structures of the human proteome. We show matching catalytic templates solely on structure is more sensitive than sequence‐ and 3D‐structure‐based approaches in identifying homology between distantly related enzymes. Since geometric matching does not depend on conserved sequence motifs or even common evolutionary history, we are able to identify examples of structural active site similarity in highly divergent and possibly convergent enzymes. Such examples make interesting case studies into the evolution of enzyme function. Though not intended for characterizing substrate‐specific binding pockets, the speed and knowledge‐driven interpretability of our method make it well suited for expanding enzyme active‐site annotation across large predicted proteomes. We provide the method and template library as a Python module, Enzyme Motif Miner, at https://github.com/rayhackett/enzymm and as a webserver at https://www.ebi.ac.uk/thornton-srv/m-csa/enzymm .
- Open Access
Identification of novel cellular intermediates unveils unique enzymes for flagellar glycan biosynthesis in Clostridioides difficile
Glycosylation of bacterial surface proteins, such as flagellin (FliC), is important for their function and is often involved in virulence of pathogens. Glycans can be further modified by so-called post-glycosylation modifications (PGMs) often resulting in exclusive molecular structures. In Clostridioides difficile a unique glycan structure (Type A) decorates FliC (which forms the flagellar filament) that consists of an O -linked N -acetyl-β-D-glucosamine (GlcNAc) modified with an N -methyl-L-threonine via a phosphodiester linkage. This PGM is synthesized by a set of four enzymes encoded in one operon ( ftaABCD ), but the exact biosynthesis pathway and biosynthetic intermediates remain unknown. In this study, we chemically synthesized two hitherto undescribed biosynthetic intermediates that we predicted based on bioinformatic analyses, CDP-threonine and CDP- N -methylthreonine. We showed that they are involved in the Type A PGM biosynthesis, as evidenced by mass spectrometric analyses of extracts of a set of C. difficile mutant strains. Furthermore, we characterized FtaC to be a SAM-dependent CDP-threonine N -methyltransferase, that installs the methyl group on CDP-threonine prior to transfer of the PGM to GlcNAc-FliC, and we revealed FtaD as the CDP- N -methylthreonine:GlcNAc N -methylthreoninephosphotransferase. Finally, using recombinantly expressed FtaC and FtaD in combination with synthetic CDP-threonine, we reconstituted the biosynthesis pathway of the Type A PGM in vitro . Overall, our results open avenues to explore these unique biosynthesis enzymes in molecular detail to provide new points of entry for the development of biosynthesis inhibitors and tools to study the role of this PGM in virulence and flagellar assembly.
Open AccessIdentification of novel cellular intermediates unveils unique enzymes for flagellar glycan biosynthesis in Clostridioides difficile
Glycosylation of bacterial surface proteins, such as flagellin (FliC), is important for their function and is often involved in virulence of pathogens. Glycans can be further modified by so-called post-glycosylation modifications (PGMs) often resulting in exclusive molecular structures. In Clostridioides difficile a unique glycan structure (Type A) decorates FliC (which forms the flagellar filament) that consists of an O -linked N -acetyl-β-D-glucosamine (GlcNAc) modified with an N -methyl-L-threonine via a phosphodiester linkage. This PGM is synthesized by a set of four enzymes encoded in one operon ( ftaABCD ), but the exact biosynthesis pathway and biosynthetic intermediates remain unknown. In this study, we chemically synthesized two hitherto undescribed biosynthetic intermediates that we predicted based on bioinformatic analyses, CDP-threonine and CDP- N -methylthreonine. We showed that they are involved in the Type A PGM biosynthesis, as evidenced by mass spectrometric analyses of extracts of a set of C. difficile mutant strains. Furthermore, we characterized FtaC to be a SAM-dependent CDP-threonine N -methyltransferase, that installs the methyl group on CDP-threonine prior to transfer of the PGM to GlcNAc-FliC, and we revealed FtaD as the CDP- N -methylthreonine:GlcNAc N -methylthreoninephosphotransferase. Finally, using recombinantly expressed FtaC and FtaD in combination with synthetic CDP-threonine, we reconstituted the biosynthesis pathway of the Type A PGM in vitro . Overall, our results open avenues to explore these unique biosynthesis enzymes in molecular detail to provide new points of entry for the development of biosynthesis inhibitors and tools to study the role of this PGM in virulence and flagellar assembly.
- Open Access
Investigating Enzyme Function by Geometric Matching of Catalytic Motifs
The rapidly growing universe of predicted protein structures offers opportunities for data driven exploration but requires computationally scalable and interpretable tools (1–3) . We developed a method to detect catalytic features in protein structures, providing insights into enzyme function and mechanism. A library of 6780 3D coordinate sets describing enzyme catalytic sites, referred to as templates, has been collected from manually curated examples of 762 enzyme catalytic mechanisms described in the Mechanism and Catalytic Site Atlas (4–6) . For template searching we optimised the geometric-matching algorithm Jess (7) . We implemented RMSD and residue orientation filters to differentiate catalytically informative matches from spurious ones. We validated this approach on a non-redundant set of high quality experimental (n=3751, (8) . Such examples make interesting case studies into the evolution of enzyme function. Though not intended for characterizing substrate-specific binding pockets, the speed and knowledge-driven interpretability of our method make it well suited for expanding enzyme active-site annotation across large predicted proteomes. We provide the method and template library as a Python module, Enzyme Motif Miner, at https://github.com/rayhackett/enzymm .
Open AccessInvestigating Enzyme Function by Geometric Matching of Catalytic Motifs
The rapidly growing universe of predicted protein structures offers opportunities for data driven exploration but requires computationally scalable and interpretable tools (1–3) . We developed a method to detect catalytic features in protein structures, providing insights into enzyme function and mechanism. A library of 6780 3D coordinate sets describing enzyme catalytic sites, referred to as templates, has been collected from manually curated examples of 762 enzyme catalytic mechanisms described in the Mechanism and Catalytic Site Atlas (4–6) . For template searching we optimised the geometric-matching algorithm Jess (7) . We implemented RMSD and residue orientation filters to differentiate catalytically informative matches from spurious ones. We validated this approach on a non-redundant set of high quality experimental (n=3751, (8) . Such examples make interesting case studies into the evolution of enzyme function. Though not intended for characterizing substrate-specific binding pockets, the speed and knowledge-driven interpretability of our method make it well suited for expanding enzyme active-site annotation across large predicted proteomes. We provide the method and template library as a Python module, Enzyme Motif Miner, at https://github.com/rayhackett/enzymm .
- Open Access
- 4 November 2025
- Published by Microbiology Society in Microbial Genomics
- Vol. 11 (11) , 001545
Genomic investigations in animal enteric disease-associated Clostridium perfringens
The anaerobic bacterium Clostridium perfringens is commonly found in the intestinal tract of humans and animals. However, there are marked differences in virulence between isolates and toxinotypes, which largely depend on various virulence factors produced by these strains. Studying C. perfringens genomes has been limited by fragmented assemblies from short-read sequencing and incomplete clinical metadata. Here, we present a high-quality collection of 236 isolates from animal hosts that underwent detailed pathomorphological examination. From 220 of them, genomes were generated by PacBio long-read sequencing, enabling comprehensive, structural level analysis of their virulome, plasmids, conjugative elements and biosynthetic gene clusters (BGCs). One hundred forty isolates were collected from animals with signs of C. perfringens -associated enteric disease, 32 from animals with no signs of C. perfringens -associated disease and 64 from healthy animals. An additional 383 publicly available C. perfringens complete or draft genomes were included for comparative analyses. We discovered 2 previously undescribed pore-forming toxin (PFT) homologues and 11 novel haemolysin and aerolysin-type PFT variants. Both findings expand the known spectrum of the C. perfringens virulome. Moreover, we defined two novel putative plasmid conjugative loci in a collection of 888 here assembled and circularized plasmids. They may facilitate HGT, supporting the dissemination of virulence and metabolic traits. We predicted 414 BGCs that were frequently toxinotype specific and encoded centrally for a bacteriocin that could help their carriers to outcompete other bacteria in shared environments. Furthermore, comparative analysis of C. perfringens plasmids revealed three distinct clusters based on their conjugation system. Altogether, these findings significantly expand the landscape of animal-associated C. perfringens through high-quality genome data and highlight novel virulence-associated features that provide a foundation for future studies of this important pathogen.
- 4 November 2025
- Published by Microbiology Society in Microbial Genomics
- Vol. 11 (11) , 001545
Open AccessGenomic investigations in animal enteric disease-associated Clostridium perfringens
The anaerobic bacterium Clostridium perfringens is commonly found in the intestinal tract of humans and animals. However, there are marked differences in virulence between isolates and toxinotypes, which largely depend on various virulence factors produced by these strains. Studying C. perfringens genomes has been limited by fragmented assemblies from short-read sequencing and incomplete clinical metadata. Here, we present a high-quality collection of 236 isolates from animal hosts that underwent detailed pathomorphological examination. From 220 of them, genomes were generated by PacBio long-read sequencing, enabling comprehensive, structural level analysis of their virulome, plasmids, conjugative elements and biosynthetic gene clusters (BGCs). One hundred forty isolates were collected from animals with signs of C. perfringens -associated enteric disease, 32 from animals with no signs of C. perfringens -associated disease and 64 from healthy animals. An additional 383 publicly available C. perfringens complete or draft genomes were included for comparative analyses. We discovered 2 previously undescribed pore-forming toxin (PFT) homologues and 11 novel haemolysin and aerolysin-type PFT variants. Both findings expand the known spectrum of the C. perfringens virulome. Moreover, we defined two novel putative plasmid conjugative loci in a collection of 888 here assembled and circularized plasmids. They may facilitate HGT, supporting the dissemination of virulence and metabolic traits. We predicted 414 BGCs that were frequently toxinotype specific and encoded centrally for a bacteriocin that could help their carriers to outcompete other bacteria in shared environments. Furthermore, comparative analysis of C. perfringens plasmids revealed three distinct clusters based on their conjugation system. Altogether, these findings significantly expand the landscape of animal-associated C. perfringens through high-quality genome data and highlight novel virulence-associated features that provide a foundation for future studies of this important pathogen.
- Open Access
- 28 October 2025
- Published by Springer Nature in Research Square
Machine learning inference of natural product chemistry across biosynthetic gene cluster types
With ever-increasing volumes of sequencing data for biosynthetic gene clusters (BGCs), computational methods to accurately predict which secondary metabolites result from these are critically lacking. Here, we present CHAMOIS, a machine learning-based tool for predicting chemical properties of secondary metabolites from protein domains annotated in the input BGCs. CHAMOIS infers 485 chemical properties from the ChemOnt ontology using logistic regression. It accurately predicts 111 such properties (AUPRC > 0.5) in cross-validation against known instances. Although CHAMOIS is not explicitly trained on biosynthetic knowledge, many of the inferred links between protein domains and metabolite properties are consistent with scientific literature, others suggest new biochemical functions of uncharacterized biosynthetic domains. Finally, CHAMOIS can pinpoint which BGC within a given genome produces a pre-specified metabolite (correct BGC in 69% of cases ranked among the top 5), which holds great potential for prioritising experimental BGC characterisation and discovery of novel biosynthetic enzymes.
- 28 October 2025
- Published by Springer Nature in Research Square
Open AccessMachine learning inference of natural product chemistry across biosynthetic gene cluster types
With ever-increasing volumes of sequencing data for biosynthetic gene clusters (BGCs), computational methods to accurately predict which secondary metabolites result from these are critically lacking. Here, we present CHAMOIS, a machine learning-based tool for predicting chemical properties of secondary metabolites from protein domains annotated in the input BGCs. CHAMOIS infers 485 chemical properties from the ChemOnt ontology using logistic regression. It accurately predicts 111 such properties (AUPRC > 0.5) in cross-validation against known instances. Although CHAMOIS is not explicitly trained on biosynthetic knowledge, many of the inferred links between protein domains and metabolite properties are consistent with scientific literature, others suggest new biochemical functions of uncharacterized biosynthetic domains. Finally, CHAMOIS can pinpoint which BGC within a given genome produces a pre-specified metabolite (correct BGC in 69% of cases ranked among the top 5), which holds great potential for prioritising experimental BGC characterisation and discovery of novel biosynthetic enzymes.
- Open Access
Large-scale insights into the biosynthetic potential of the Bacillus cereus group
Bacterial secondary metabolites (SMs) are a critical source of natural product-derived drugs. However, SM discovery efforts have focused overwhelmingly on Actinomycetes , potentially overlooking other key producers. Here, we explore the biosynthetic potential of the Bacillus cereus group, an underexplored complex of SM producers. Using a combined rule- and machine learning-based approach, we mine an unprecedented number of B. cereus group genomes ( n = 9,744) for SM-producing biosynthetic gene clusters (BGCs; n = 200,196). Notably, 158,678 B. cereus group BGCs (78.2%) did not cluster with previously described BGCs, suggesting new chemical scaffolds to be explored. B. pseudomycoides was particularly prolific in terms of its SM production potential (30.8 BGC families/genome, Kruskal-Wallis p < 0.0001), and we identify a previously uncharacterized, B. pseudomycoides -unique peptide. Overall, our study represents the largest survey of B. cereus group biosynthetic potential to date and posits the complex as an under-queried SM resource.
Open AccessLarge-scale insights into the biosynthetic potential of the Bacillus cereus group
Bacterial secondary metabolites (SMs) are a critical source of natural product-derived drugs. However, SM discovery efforts have focused overwhelmingly on Actinomycetes , potentially overlooking other key producers. Here, we explore the biosynthetic potential of the Bacillus cereus group, an underexplored complex of SM producers. Using a combined rule- and machine learning-based approach, we mine an unprecedented number of B. cereus group genomes ( n = 9,744) for SM-producing biosynthetic gene clusters (BGCs; n = 200,196). Notably, 158,678 B. cereus group BGCs (78.2%) did not cluster with previously described BGCs, suggesting new chemical scaffolds to be explored. B. pseudomycoides was particularly prolific in terms of its SM production potential (30.8 BGC families/genome, Kruskal-Wallis p < 0.0001), and we identify a previously uncharacterized, B. pseudomycoides -unique peptide. Overall, our study represents the largest survey of B. cereus group biosynthetic potential to date and posits the complex as an under-queried SM resource.
- Open Access
Ontology pre-training improves machine learning-based predictions for metabolites
Recent advances in the field of machine learning have shown that integration of expert knowledge improves performances, in particular for complex domains such as biology. Bio-ontologies offer a rich source of curated biological knowledge that can be harnessed to this end. Here, we describe an intuitive and generalisable approach to embed the knowledge contained in a classification hierarchy derived from a bio-ontology into a machine learning model as an intermediate training step between general-purpose pre-training and task-specific fine-tuning in a process that we call ‘ontology pre-training’. We show that this approach leads to an improvement in predictive performance and a reduction in training time for a broad range of prediction tasks relevant to understanding metabolite functions in living systems, using a range of datasets derived from MoleculeNet. We see the biggest improvement for regression tasks, e.g. prediction of lipophilicity and aqueous solubility of molecules, and a robust improvement for most classification tasks. Our approach can be adapted for a wide range of knowledge sources, models and prediction tasks.
Open AccessOntology pre-training improves machine learning-based predictions for metabolites
Recent advances in the field of machine learning have shown that integration of expert knowledge improves performances, in particular for complex domains such as biology. Bio-ontologies offer a rich source of curated biological knowledge that can be harnessed to this end. Here, we describe an intuitive and generalisable approach to embed the knowledge contained in a classification hierarchy derived from a bio-ontology into a machine learning model as an intermediate training step between general-purpose pre-training and task-specific fine-tuning in a process that we call ‘ontology pre-training’. We show that this approach leads to an improvement in predictive performance and a reduction in training time for a broad range of prediction tasks relevant to understanding metabolite functions in living systems, using a range of datasets derived from MoleculeNet. We see the biggest improvement for regression tasks, e.g. prediction of lipophilicity and aqueous solubility of molecules, and a robust improvement for most classification tasks. Our approach can be adapted for a wide range of knowledge sources, models and prediction tasks.
- Open Access
Scalable and interpretable secretion system annotation with Sismis
Secretion systems play critical roles in bacterial growth, survival, and pathogenesis. Genes encoding secretion system components often co-occur together as gene clusters in bacterial (meta)genomes. However, existing tools for secretion system annotation are unable to utilize genomic context to make predictions and are unable to detect secretion systems of novel architecture. Here, we present Sismis ( s ecretion s yste m d is covery tool; https://github.com/lmc297/Sismis ), a scalable, interpretable, machine learning-based tool, which detects and classifies single-locus secretion systems in bacterial (meta)genomes with high accuracy (test set area-under-the-curve [AUC] values of 0.71 and 0.92 for precision-recall [PR] and receiver operating characteristic [ROC] curves, respectively). When applied to ≈700k prokaryotic (meta)genomes, Sismis identifies 747, 439 total secretion systems comprising 15, 612 major secretion system families, >80% of which contain no previously known/annotated secretion systems. To facilitate further exploration of these data, we present the Sismis Atlas ( https://sismis.microbe.dev/ ), an interactive secretion system database, which we use to identify a largely uncharacterized cluster of secretion systems with tight adherence (Tad) pili-like characteristics. Altogether, Sismis and its companion atlas enable accurate and interpretable secretion system annotation, exploration, and discovery at an unprecedented scale. GRAPHICAL
Open AccessScalable and interpretable secretion system annotation with Sismis
Secretion systems play critical roles in bacterial growth, survival, and pathogenesis. Genes encoding secretion system components often co-occur together as gene clusters in bacterial (meta)genomes. However, existing tools for secretion system annotation are unable to utilize genomic context to make predictions and are unable to detect secretion systems of novel architecture. Here, we present Sismis ( s ecretion s yste m d is covery tool; https://github.com/lmc297/Sismis ), a scalable, interpretable, machine learning-based tool, which detects and classifies single-locus secretion systems in bacterial (meta)genomes with high accuracy (test set area-under-the-curve [AUC] values of 0.71 and 0.92 for precision-recall [PR] and receiver operating characteristic [ROC] curves, respectively). When applied to ≈700k prokaryotic (meta)genomes, Sismis identifies 747, 439 total secretion systems comprising 15, 612 major secretion system families, >80% of which contain no previously known/annotated secretion systems. To facilitate further exploration of these data, we present the Sismis Atlas ( https://sismis.microbe.dev/ ), an interactive secretion system database, which we use to identify a largely uncharacterized cluster of secretion systems with tight adherence (Tad) pili-like characteristics. Altogether, Sismis and its companion atlas enable accurate and interpretable secretion system annotation, exploration, and discovery at an unprecedented scale. GRAPHICAL