Over 46,000 putative beta-lactamases mined from environmental metagenomes, and two low-identity candidates confer multidrug resistance in E. coli
Synopsis
The functional diversity and mutational potential of beta-lactamases in environmental microbial communities were only partly understood. Combining protein language models with ensemble learning identified more than 46,000 putative beta-lactamases across all four Ambler classes from environmental metagenomes; two class C candidates sharing about 30% sequence identity with known enzymes conferred resistance to multiple beta-lactam antibiotics when expressed in E. coli, and mutations found by combining an inverse-folding model with protein language models altered resistance profiles, including increased carbapenem resistance.
Interpretation
Beta-lactamases in environmental metagenomes are no longer sought only by similarity to known sequences; sequence features encoded by a protein language model, paired with an ensemble of classifiers, identify them across all four Ambler classes. Earlier homology searches and profile-based methods become limited when candidates diverge substantially from characterized proteins, so the functional diversity of environmental enzymes stayed only partly characterized. On 1,778 experimentally characterized beta-lactamase sequences under five-fold cross-validation, the five base classifiers each performed well but varied across folds, while the ensemble model correctly classified every validation sample in all folds.
Candidates mined from the MGnify protein database span terrestrial and marine habitats worldwide, and their class composition differs from that of known reference sequences. Characterized beta-lactamases are dominated by class A, whereas the environmental candidates showed greater diversity and richness in classes B and C. Candidate counts were uneven by region, with 14,099 from North America, 11,248 from Asia and 10,665 from Europe, and class B accounted for 34.0% to 44.5% in every region.
Two class C candidates, Prot22 and Prot94, sharing about 30% identity with known enzymes, did confer resistance when expressed in E. coli, showing that the predicted divergent sequences carry real enzymatic activity. The blank control strain carrying the empty pET26b vector showed no appreciable resistance to the tested antibiotics or inhibitor combinations, only low-level tolerance. Prot22 had MICs above 256 μg/mL for penicillin, oxacillin, ampicillin and amoxicillin/clavulanic acid, while Prot94 reached its highest MIC of 512 μg/mL for penicillin but was more susceptible to the inhibitor combinations.
Combining structure-constrained ProteinMPNN with evolution-informed ESM1b can screen for single amino-acid substitutions that change resistance profiles. Wild-type Prot22 itself lacked a carbapenem resistance phenotype, and the mutant Mut2 (Y106F) acquired one. Mut2 of Prot22 gave MICs of 2 μg/mL for meropenem and 16 μg/mL for imipenem, and Mut5 (S73T) of Prot94 gave MICs of 512 μg/mL for both cefaclor and imipenem.
Perspective
The demonstrated scope is mining from the MGnify protein database release 2024_04 and validating two class C candidates, Prot22 and Prot94, plus several single-point mutants by MIC measurement in E. coli; the authors propose future work with broader experimental validation, experimental evolution and environmentally resolved metagenomic data.
The authors note that the ensemble framework's generalizability to highly divergent proteins still requires validation; experimental characterization covers only two class C candidates with resistance assessed mainly by MIC; mutation predictions were evaluated with a limited number of single substitutions; and ecological comparisons rely on available metagenomic sequences rather than direct measurements of selection pressure. They also state that higher entropy at predicted high-fitness sites cannot be read directly as evidence of positive selection, and that the share of predicted mutations matching the dominant natural residue was about 25% in ocean-derived and 30% in wastewater-derived sequences, a difference not statistically significant in pooled or class-specific analyses.