Deep learning-guided discovery reveals the diversity and mutational potential of environmental beta-lactamases
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Deep learning-guided discovery reveals the diversity and mutational potential of environmental beta-lactamases".
Marcus: Here's the gist of the paper. Deep learning-guided discovery reveals the diversity and mutational potential of environmental β-lactamases by integrating protein language models with ensemble learning to characterize environmental β-lactamase diversity and investigate how sequence variation influences antibiotic resistance Characterization of Environmental Diversity The study aimed to characterize environmental β-lactamase diversity and investigate how sequence variation influences antibiotic resistance By integrating protein language model (PLMs) with ensemble learning, the researchers identified more than 46,000 putative β-lactamases across all four Ambler classes from environmental metagenomes, which were widely distributed across diverse habitats and possessed distinct habitat-associated class compositions The phylogenetic analysis of experimentally characterized sequences showed that the four Ambler classes formed well-separated clades, indicating pronounced sequence divergence among them Furthermore, the distribution of functional phenotypes displayed clade-level specificity, with carbapenemase phenotype concentrated in classes B and D and ESBL phenotype occurring predominantly within class A clades Identification of Potential Candidates The ensemble learning framework was developed by employing ESM2-t33-650M as the feature extraction backbone to generate a 1,280-dimensional embedding for every amino acid residue Five complementary classifiers—MLP, CNN, BiLSTM, SVM, and RF—were designed to discriminate among Ambler classes and negative samples The ensemble model achieved high confidence in classifying every validation sample across all folds, indicating a stable and reliable tool for mining potential β-lactamases from environmental metagenomic data This framework was applied to the MGnify Protein Database, where HMM profiles were used for initial screening before ensemble filtering refined the predictions Ecological Distribution and Habitat Association The predicted candidate sequences were examined for their global distribution and habitat specificity across terrestrial, marine, and wastewater environments The distribution was markedly uneven, with the majority of candidates originating from North America, Asia, and Europe Habitat-level analysis revealed that candidate sequences were primarily associated with three broad categories: engineered water-treatment systems, natural aquatic ecosystems, and host-associated microbiomes Specifically, class B maintained a high proportion across diverse environments, suggesting broad environmental adaptability for these enzymes Exploration of Mutational Potential Protein inverse-folding models and PLMs were combined to predict potential high-fitness (HF) mutations by computing the log-likelihood ratio (LLR) of each mutant residue relative to the wild-type using ESM1b The LLR scores showed a significant positive correlation with experimentally measured mutation fitness, with an optimal threshold of −6.59 selected for identifying potentially beneficial mutations The combination of ProteinMPNN and ESM1b identified 31 and 42 putative HF mutation sites for Prot22 and Prot94, respectively, which were then experimentally validated Experimental results showed that mutations like Mut2 (Y106F) in Prot22 enhanced broad-spectrum resistance, including carbapenem-related resistance absent in the wild type Evolutionary Plasticity Across Habitats Site entropy profiling indicated that HF sites showed significantly higher site entropy compared to background positions across all Ambler classes Crucially, wastewater-derived enzymes exhibited significantly higher site entropy than ocean-derived enzymes in all four classes This suggests that the complex and variable chemical and microbial exposures characteristic of wastewater environments may intensify localized adaptive selection, resulting in greater sequence plasticity at relevant sites However, the study found no statistically significant evidence that habitat type consistently biases the specific amino acid substitution favored at individual HF sites Limitations and Future Directions Limitations acknowledged include the generalizability of the ensemble-learning framework to highly divergent proteins and experimental characterization being limited to two class C candidates Future studies are needed to integrate broader experimental validation, experimental evolution, and environmentally resolved metagenomic data to clarify how predicted mutational potential translates into realized evolutionary adaptation The work establishes an integrated framework combining PLMs, structure-informed sequence exploration, and experimental characterization to investigate the functional diversity of environmental β-lactamases Pruning redundant reasoning tokens This study aimed to characterize environmental β-lactamase diversity and investigate how sequence variation influences antibiotic resistance By integrating protein language model (PLMs) with ensemble learning, the researchers identified more than 46,000 putative β-lactamases across all four Ambler classes from environmental metagenomes, which were widely distributed across diverse habitats and possessed distinct habitat-associated class compositions The phylogenetic analysis of experimentally characterized sequences showed that the four Ambler classes formed well-separated clades, indicating pronounced sequence divergence among them How the training signal is built Protein language models (PLMs) learn and encode features related to protein function and evolution from large-scale sequence data, facilitating the identification of distantly related homologues Inspired by this, an ensemble learning framework was developed based on a PLM to identify potential novel βlactamases from environmental metagenomic data ESM2-t33-650M was employed as the feature extraction backbone to generate a 1,280-dimensional embedding for every amino acid residue Five complementary classifiers—MLP, CNN, BiLSTM, SVM, and RF—were designed to discriminate among Ambler classes and negative samples Where the gains come from The ensemble learning framework achieved high confidence in classifying every validation sample across all folds, indicating a stable and reliable tool for mining potential β-lactamases from environmental metagenomic data The combination of ProteinMPNN and ESM1b identified 31 and 42 putative HF mutation sites for Prot22 and Prot94, respectively, which were then experimentally validated Experimental results showed that mutations like Mut2 (Y106F) in Prot22 enhanced broad-spectrum resistance, including carbapenem-related resistance absent in the wild type The combination of sequence- and structure-based models can effectively identify HF mutations that alter β-lactamase activity and substrate specificity Pruning redundant reasoning tokens The phylogenetic diversification showed that the four Ambler classes formed well-separated clades, indicating pronounced sequence divergence among them The distribution of functional phenotypes displayed clade-level specificity, with carbapenemase phenotype concentrated in classes B and D and ESBL phenotype occurring predominantly within class A clades How the training signal is built The ensemble learning framework was next applied to mine potential β-lactamases from the MGnify Protein Database, where HMM profiles were first constructed for each Ambler class and used to scan the entire database with HMMER After ensemble filtering, each Ambler class retained a set of high-confidence candidates, with classes B and C showing greater diversity and richness Where the gains come from The predicted candidate sequences were examined for their global distribution and habitat specificity across terrestrial, marine, and wastewater environments Habitat-level analysis revealed that candidate sequences were primarily associated with three broad categories: engineered water-treatment systems, natural aquatic ecosystems, and host-associated microbiomes Class B maintained a high proportion across diverse environments, suggesting broad environmental adaptability for these enzymes Pruning redundant reasoning tokens The study aimed to systematically investigate the diversity, ecological distribution, functional properties, and evolutionary characteristics of β-lactamases in environmental microbial communities Specifically, the objectives included training a deep learning-based model integrating PLMs and ensemble learning to identify candidates Investigating their ecological distribution and habitat-associated diversity, and experimentally characterizing representative divergent β-lactamases was another goal Combining protein inverse folding and PLMs to predict potential high-fitness mutations, followed by experimental validation was the third objective Where the gains come from The study established an integrated framework combining protein language models, structure-informed sequence exploration, and experimental characterization to investigate the functional diversity of environmental β-lactamases This framework provides a foundation for exploring how uncharacterized environmental enzymes may contribute to the emergence and diversification of antibiotic resistance The identification of mutations that alter resistance does not by itself establish their evolutionary accessibility or selective advantage under natural conditions, but it provides experimentally testable hypotheses about the functional flexibility of environmental β-lactamases Pruning redundant reasoning tokens The growing availability of environmental metagenomic data has greatly expanded the known diversity of microbial proteins, yet translating this sequence diversity into functional and evolutionary understanding remains a substantial challenge Conventional sequence-based approaches such as homology searches and profile-based methods have been instrumental in characterizing antibiotic resistance genes PLMs offer a complementary strategy by representing proteins in a learned sequence space that captures information beyond direct pairwise similarity How the training signal is built PLMs learn and encode features related to protein function and evolution from large-scale sequence data, facilitating the identification of distantly related homologues This inspired the development of an ensemble learning framework based on a PLM to identify potential novel βlactamases from environmental metagenomic data ESM2-t33-650M was employed as the feature extraction backbone to generate a 1,280-dimensional embedding for every amino acid residue Where the gains come from The ensemble model achieved high confidence in classifying every validation sample across all folds, indicating a stable and reliable tool for mining potential β-lactamases from environmental metagenomic data The distribution of functional phenotypes displayed clade-level specificity, with carbapenemase phenotype concentrated in classes B and D and ESBL phenotype occurring predominantly within class A clades Pruning redundant reasoning tokens To determine whether the candidate β-lactamases identified by the ensemble learning framework exhibited functional activity, two candidates with low similarity to the reference database were first selected from each Ambler class The characterized class C β-lactamase AmpC served as a positive control (PC), and E.
Ines: coli harbouring the empty pET26b vector served as a blank control (BC) Strains expressing AmpC, Prot22, or Prot94 each displayed β-lactam resistance phenotypes of varying magnitude These results demonstrate that Prot22 and Prot94 each confer resistance to multiple β-lactam antibiotics when expressed in E.
Marcus: coli, confirming that the two low-homology class C candidates identified by the ensemble learning framework possess functional β-lactamase activity Where the gains come from Protein inverse-folding models and PLMs were combined to predict potential high-fitness (HF) mutations by computing the log-likelihood ratio (LLR) of each mutant residue relative to the wild-type using ESM1b The combination of ProteinMPNN and ESM1b identified 31 and 42 putative HF mutation sites for Prot22 and Prot94, respectively, which were then experimentally validated Mutations located in flexible regions appear more likely to represent sites with elevated evolutionary potential The study establishes an integrated framework combining protein language models, structure-informed sequence exploration, and experimental characterization to investigate the functional diversity of environmental β-lactamases Pruning redundant reasoning tokens Across the four representative sequences, a total of 97 HF sites were identified Compared with all non-predicted background positions, HF sites showed significantly higher site entropy (median HMM = 3.20 vs.
Ines: 2.93; p = 1.2 × 10−4; n = 97 vs.
Paper summary: Ines: We're starting off by looking at this paper now called "Deep learning-guided discovery reveals the diversity and mutational potential of environmental beta-lactamases". The main thesis is that by integrating protein language models with ensemble learning, researchers can characterize the diversity of these enzymes and investigate how sequence variation influences antibiotic resistance.
Marcus: They claim they identified more than forty-six thousand putative beta-lactamases across all four Ambler classes from environmental metagenomes, and these candidates were widely distributed across diverse habitats, possessing distinct habitat-associated class compositions <ref:2611.14568#pg1,lactamases across all four Ambler classes from environmental metagenomes>.
Yuki: That wide distribution is key because the phylogenetic analysis showed that the four Ambler classes formed well-separated clades, which indicates pronounced sequence divergence among them.
Ines: Beyond just diversity in location, the study points out clade-level specificity in functional phenotypes; specifically, carbapenemase phenotype was found concentrated in classes B and D, while ESBL phenotype occurred predominantly within class A clades.
Marcus: The real significance is that they used this deep learning approach to find candidates that share approximately thirty percent sequence identity with known beta-lactamases, and when expressed in E coli, these candidates conferred resistance to multiple beta-lactam antibiotics <ref:2611.14568#pg1>.
Conclusion: Ines: So, to wrap up this discussion on "Deep learning-guided discovery reveals the diversity and mutational potential of environmental beta-lactamases," the authors are using advanced AI tools to systematically map out how diverse these enzymes are and where they might evolve into new resistance mechanisms.
Marcus: The work by Ge et al. is important because it provides a computational method for mining massive environmental metagenomic datasets, moving beyond older sequence-based approaches that can only find what’s directly similar to known sequences.
Yuki: For me, the implication is that we need to use these AI frameworks not just for discovery but to understand the functional flexibility of environmental enzymes and how they are poised to adapt under antibiotic selection pressures.
Ines: It establishes an integrated framework where structure-informed exploration and experimental validation work together, which gives us a foundation for exploring how uncharacterized environmental enzymes might contribute to the emergence of resistance.
Center for Biological Science and Technology, Advanced Institute of Natural Sciences, Beijing Normal University
q-bio.GN, cs.AI
Submitted: 2026-11-09
Updated: 2026-11-09
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 82/100
The gist: Deep learning-guided discovery reveals the diversity and mutational potential of environmental β-lactamases by integrating protein language models with ensemble learning to characterize…
Key concepts
- Protein Language Models (PLMs)
- PLMs are deep learning models trained on massive protein sequence data to learn the underlying language of proteins. They can encode complex information about a protein's structure and evolutionary history, helping researchers find distant homologs that traditional methods might miss.
- Ensemble Learning
- This technique combines the predictions from several different machine learning models (like MLP, CNN, SVM) to make a more robust and reliable decision. In this study, it was used to accurately classify and filter potential beta-lactamases from large environmental databases.
- High-Fitness (HF) Mutations
- These are specific amino acid changes predicted by models that are likely to improve an enzyme's function, such as increasing its ability to resist antibiotics. The study found mutations that significantly increased the likelihood of beneficial resistance, which were then tested experimentally.
- Ambler Classes
- These are a system used to classify beta-lactamases based on their structural features. The study found distinct groupings (classes A, B, C, and D) in environmental samples, showing that different groups have specific roles in antibiotic resistance.
Terminology
Summary
Deep learning-guided discovery reveals the diversity and mutational potential of environmental β-lactamases by integrating protein language models with ensemble learning to characterize environmental β-lactamase diversity and investigate how sequence variation influences antibiotic resistance
Characterization of Environmental Diversity
The study aimed to characterize environmental β-lactamase diversity and investigate how sequence variation influences antibiotic resistance By integrating protein language model (PLMs) with ensemble learning, the researchers identified more than 46,000 putative β-lactamases across all four Ambler classes from environmental metagenomes, which were widely distributed across diverse habitats and possessed distinct habitat-associated class compositions The phylogenetic analysis of experimentally characterized sequences showed that the four Ambler classes formed well-separated clades, indicating pronounced sequence divergence among them Furthermore, the distribution of functional phenotypes displayed clade-level specificity, with carbapenemase phenotype concentrated in classes B and D and ESBL phenotype occurring predominantly within class A clades
Identification of Potential Candidates
The ensemble learning framework was developed by employing ESM2-t33-650M as the feature extraction backbone to generate a 1,280-dimensional embedding for every amino acid residue Five complementary classifiers—MLP, CNN, BiLSTM, SVM, and RF—were designed to discriminate among Ambler classes and negative samples The ensemble model achieved high confidence in classifying every validation sample across all folds, indicating a stable and reliable tool for mining potential β-lactamases from environmental metagenomic data This framework was applied to the MGnify Protein Database, where HMM profiles were used for initial screening before ensemble filtering refined the predictions
Ecological Distribution and Habitat Association
The predicted candidate sequences were examined for their global distribution and habitat specificity across terrestrial, marine, and wastewater environments The distribution was markedly uneven, with the majority of candidates originating from North America, Asia, and Europe Habitat-level analysis revealed that candidate sequences were primarily associated with three broad categories: engineered water-treatment systems, natural aquatic ecosystems, and host-associated microbiomes Specifically, class B maintained a high proportion across diverse environments, suggesting broad environmental adaptability for these enzymes
Exploration of Mutational Potential
Protein inverse-folding models and PLMs were combined to predict potential high-fitness (HF) mutations by computing the log-likelihood ratio (LLR) of each mutant residue relative to the wild-type using ESM1b The LLR scores showed a significant positive correlation with experimentally measured mutation fitness, with an optimal threshold of −6.59 selected for identifying potentially beneficial mutations The combination of ProteinMPNN and ESM1b identified 31 and 42 putative HF mutation sites for Prot22 and Prot94, respectively, which were then experimentally validated Experimental results showed that mutations like Mut2 (Y106F) in Prot22 enhanced broad-spectrum resistance, including carbapenem-related resistance absent in the wild type
Evolutionary Plasticity Across Habitats
Site entropy profiling indicated that HF sites showed significantly higher site entropy compared to background positions across all Ambler classes Crucially, wastewater-derived enzymes exhibited significantly higher site entropy than ocean-derived enzymes in all four classes This suggests that the complex and variable chemical and microbial exposures characteristic of wastewater environments may intensify localized adaptive selection, resulting in greater sequence plasticity at relevant sites However, the study found no statistically significant evidence that habitat type consistently biases the specific amino acid substitution favored at individual HF sites
Limitations and Future Directions
Limitations acknowledged include the generalizability of the ensemble-learning framework to highly divergent proteins and experimental characterization being limited to two class C candidates Future studies are needed to integrate broader experimental validation, experimental evolution, and environmentally resolved metagenomic data to clarify how predicted mutational potential translates into realized evolutionary adaptation The work establishes an integrated framework combining PLMs, structure-informed sequence exploration, and experimental characterization to investigate the functional diversity of environmental β-lactamases
Pruning redundant reasoning tokens
This study aimed to characterize environmental β-lactamase diversity and investigate how sequence variation influences antibiotic resistance By integrating protein language model (PLMs) with ensemble learning, the researchers identified more than 46,000 putative β-lactamases across all four Ambler classes from environmental metagenomes, which were widely distributed across diverse habitats and possessed distinct habitat-associated class compositions The phylogenetic analysis of experimentally characterized sequences showed that the four Ambler classes formed well-separated clades, indicating pronounced sequence divergence among them
How the training signal is built
Protein language models (PLMs) learn and encode features related to protein function and evolution from large-scale sequence data, facilitating the identification of distantly related homologues Inspired by this, an ensemble learning framework was developed based on a PLM to identify potential novel βlactamases from environmental metagenomic data ESM2-t33-650M was employed as the feature extraction backbone to generate a 1,280-dimensional embedding for every amino acid residue Five complementary classifiers—MLP, CNN, BiLSTM, SVM, and RF—were designed to discriminate among Ambler classes and negative samples
Where the gains come from
The ensemble learning framework achieved high confidence in classifying every validation sample across all folds, indicating a stable and reliable tool for mining potential β-lactamases from environmental metagenomic data The combination of ProteinMPNN and ESM1b identified 31 and 42 putative HF mutation sites for Prot22 and Prot94, respectively, which were then experimentally validated Experimental results showed that mutations like Mut2 (Y106F) in Prot22 enhanced broad-spectrum resistance, including carbapenem-related resistance absent in the wild type The combination of sequence- and structure-based models can effectively identify HF mutations that alter β-lactamase activity and substrate specificity
The phylogenetic diversification showed that the four Ambler classes formed well-separated clades, indicating pronounced sequence divergence among them The distribution of functional phenotypes displayed clade-level specificity, with carbapenemase phenotype concentrated in classes B and D and ESBL phenotype occurring predominantly within class A clades
The ensemble learning framework was next applied to mine potential β-lactamases from the MGnify Protein Database, where HMM profiles were first constructed for each Ambler class and used to scan the entire database with HMMER After ensemble filtering, each Ambler class retained a set of high-confidence candidates, with classes B and C showing greater diversity and richness
The predicted candidate sequences were examined for their global distribution and habitat specificity across terrestrial, marine, and wastewater environments Habitat-level analysis revealed that candidate sequences were primarily associated with three broad categories: engineered water-treatment systems, natural aquatic ecosystems, and host-associated microbiomes Class B maintained a high proportion across diverse environments, suggesting broad environmental adaptability for these enzymes
The study aimed to systematically investigate the diversity, ecological distribution, functional properties, and evolutionary characteristics of β-lactamases in environmental microbial communities Specifically, the objectives included training a deep learning-based model integrating PLMs and ensemble learning to identify candidates Investigating their ecological distribution and habitat-associated diversity, and experimentally characterizing representative divergent β-lactamases was another goal Combining protein inverse folding and PLMs to predict potential high-fitness mutations, followed by experimental validation was the third objective
The study established an integrated framework combining protein language models, structure-informed sequence exploration, and experimental characterization to investigate the functional diversity of environmental β-lactamases This framework provides a foundation for exploring how uncharacterized environmental enzymes may contribute to the emergence and diversification of antibiotic resistance The identification of mutations that alter resistance does not by itself establish their evolutionary accessibility or selective advantage under natural conditions, but it provides experimentally testable hypotheses about the functional flexibility of environmental β-lactamases
The growing availability of environmental metagenomic data has greatly expanded the known diversity of microbial proteins, yet translating this sequence diversity into functional and evolutionary understanding remains a substantial challenge Conventional sequence-based approaches such as homology searches and profile-based methods have been instrumental in characterizing antibiotic resistance genes PLMs offer a complementary strategy by representing proteins in a learned sequence space that captures information beyond direct pairwise similarity
PLMs learn and encode features related to protein function and evolution from large-scale sequence data, facilitating the identification of distantly related homologues This inspired the development of an ensemble learning framework based on a PLM to identify potential novel βlactamases from environmental metagenomic data ESM2-t33-650M was employed as the feature extraction backbone to generate a 1,280-dimensional embedding for every amino acid residue
The ensemble model achieved high confidence in classifying every validation sample across all folds, indicating a stable and reliable tool for mining potential β-lactamases from environmental metagenomic data The distribution of functional phenotypes displayed clade-level specificity, with carbapenemase phenotype concentrated in classes B and D and ESBL phenotype occurring predominantly within class A clades
To determine whether the candidate β-lactamases identified by the ensemble learning framework exhibited functional activity, two candidates with low similarity to the reference database were first selected from each Ambler class The characterized class C β-lactamase AmpC served as a positive control (PC), and E. coli harbouring the empty pET26b vector served as a blank control (BC) Strains expressing AmpC, Prot22, or Prot94 each displayed β-lactam resistance phenotypes of varying magnitude These results demonstrate that Prot22 and Prot94 each confer resistance to multiple β-lactam antibiotics when expressed in E. p. 8 coli, confirming that the two low-homology class C candidates identified by the ensemble learning framework possess functional β-lactamase activity
Protein inverse-folding models and PLMs were combined to predict potential high-fitness (HF) mutations by computing the log-likelihood ratio (LLR) of each mutant residue relative to the wild-type using ESM1b The combination of ProteinMPNN and ESM1b identified 31 and 42 putative HF mutation sites for Prot22 and Prot94, respectively, which were then experimentally validated Mutations located in flexible regions appear more likely to represent sites with elevated evolutionary potential The study establishes an integrated framework combining protein language models, structure-informed sequence exploration, and experimental characterization to investigate the functional diversity of environmental β-lactamases
Across the four representative sequences, a total of 97 HF sites were identified Compared with all non-predicted background positions, HF sites showed significantly higher site entropy (median HMM = 3.20 vs. 2.93; p = 1.2 × 10−4; n = 97 vs. 1,053) The same direction of effect was observed within each Ambler class, reaching significance in classes A (p = 4. p. 48 × 10−2), C (p = 3.8×10−3), and D (p = 3.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Protein language models can be integrated with ensemble learning to identify and classify putative beta-lactamase candidates from large-scale environmental metagenomic datasets through a deep learning-based framework (Fig 1B). This allows the system to systematically mine large databases and filter candidates based on learned features, reducing false positive rates.
-
The AI system can predict potential high-fitness mutations in beta-lactamases by combining structure constrained ProteinMPNN with evolution informed ESM1b scores to identify Putative High Fitness (HF) sites (Fig 5D and E). This enables the prediction of amino acid substitutions that substantially alter resistance phenotypes, such as enhanced broad spectrum resistance or substrate profile shifts.
-
The system can assess the evolutionary potential of predicted mutations by calculating site entropy and comparing it across contrasting environmental habitats like wastewater versus ocean environments (Fig 7C). This helps in understanding whether sequence variability is driven by localized adaptive selection or broader environmental factors.
Sources
Related papers
- SeekRBP: Leveraging Sequence-Structure Integration with Reinforcement Learning for Receptor-Binding Protein Identification
- Genome-Factory: A Library for Tuning, Deploying, and Interpreting Genomic Foundation Models
- AVA: Agentic Variant Adjudication for Automated, Auditable Read-Level Review in Cancer Diagnostic Sequencing
- C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation
- Leveraging genomic deep learning models for the prediction of non-coding variant effects
- Adaptive Allele-Haplotype Incidence Indexing for Pangenome Queries: Representation, Reconstruction, and Benchmarking