Conserved protein sequence-structure signatures identify antibiotic resistance genes from the human microbiome Authors: Lachlan Bartrop, Emy Beauchemin-Lauzon, Frédéric Grenier, Sébastien Rodrigue, Louis-Patrick Haraoui Published: 2026 ABSTRACT AbstractBackgroundBacteria exhibiting antimicrobial resistance (AMR) are a significant public health challenge worldwide. The human microbiome is a known repository of key determinants of AMR, antibiotic resistance genes (ARGs). Identifying previously undetected ARGs in the human microbiome that confer resistance to clinical concentrations of antibiotics is a crucial component of addressing AMR, yet can be hindered by their low homology to existing ARGs. Here, we address this by focussing on functionally important protein regions.ResultsARG-PASS (ARG-PAirwise Sequence vs Structure) is a protein function prediction method which leverages a one-class support vector machine trained on pairwise primary and tertiary distributions of structurally conserved regions of proteins encoded by ARGs. ARG-PASS was applied to six reference strains of the Human Microbiome Project. Nine candidates were selected for experimental verification, and all were functionally confirmed when expressed in E.coli, belonging to ARG classes: APH, dfr, class B and C β-lactamases, and penicillin-binding proteins. We also used ARG-PASS directly on protein structures within the AlphaFold database and identified a phnP gene (metallo-β-lactamase fold) with low similarity to characterised β-lactamases and media-dependent ampicillin activity. In total, 80% of the tested genes confer resistance at CLSI resistant breakpoints and the remainder represent ‘pre-resistance’ genes, with activity but not at a clinically relevant minimum inhibitory concentration. We suggest pre-resistance genes may preferentially evolve into clinically relevant resistance determinants.ConclusionARG-PASS provides a precise computational approach that integrates sequence and structure to identify divergent ARGs that evade homology-based detection, supporting improved resistance surveillance and antibiotic stewardship. Video AbstractSupplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s40168-026-02487-6. Background The enthusiasm following the discovery of antibiotics was overshadowed by the rise of bacteria exhibiting antimicrobial resistance (AMR), defined as the ability of microbes to survive treatment with antibiotics that once effectively controlled them. AMR has since become one of the most pressing global public health challenges of our time [60]. Antibiotic resistance genes (ARGs), key drivers of AMR, predate the use of antibiotics [20]. For the most part, they are present in environmental bacteria but can be selected in response to anthropogenic antibiotic exposure, and subsequently pose a risk to human health once stably integrated into the human microbiome. This can occur through one of three main routes: (i) if they evolved in generalist or opportunist species which can also inhabit human microbiome niches; (ii) via horizontal gene transfer from environmental bacteria to human-associated bacteria by mobile genetic elements (MGEs); or (iii) by de novo evolution within the human microbiome [19, 40]. Previously unidentified (novel) ARGs can evade detection by computational algorithms designed for their identification, due to their low homology to existing ARGs. As such, identifying novel ARGs from human microbiomes that confer clinical levels of resistance to antibiotics is a crucial component of addressing AMR for accurate surveillance, diagnostics, and antibiotic stewardship. Existing computational tools for ARG discovery can be broadly categorised along two primary axes: (i) dependence on amino acid (aa) sequence similarity and (ii) strength of experimental (and clinical) validation. Traditional methods for identifying novel ARGs rely on a nearest neighbour approach using aa sequence alignment, typically with high sequence identity (seqID) cutoffs (> 60%) against databases of known ARGs [5]. Profile hidden Markov models (HMMs) extend this capability through probabilistic modelling of protein classes [24, 26, 99]. Experimental (wet-lab) validation of predictions [9], facilitates their deposition into well-curated ARG databases, such as the Comprehensive Antibiotic Resistance Database (CARD) [5]. However, the amount of microbial proteins with homology to ARGs increases with decreasing seqID and ARGs can often share lower seqID. For example, at the time of discovery blaNDM-1 and sul4 shared the highest seqID of ~ 33% to blaVIM and sul2, respectively [74, 100]. Although effective for detecting close homologues, these approaches are not as suited for the discovery of more divergent sequences supported by experimental and clinical validation [48]. When trained on the expanding volume of biological sequence data, ML-based classifiers have emerged as a potential alternative and/or complementary computational approach. Strong predictive performance has been reported with precisions exceeding 85% [1, 6, 31, 57, 96]. However, ML-based classifiers will also favour the detection of less divergent ARGs if they are trained and/or validated on databases which ultimately suffer the same inherent biases, such as ARGs with little sequence variation, only from clinical settings, or without experimental support [6, 31, 96]. Whilst performance is high on internal validation, their generalisability to new sequences from other sources remains less clear. Efforts to address this limitation expand training beyond sequences [25, 73]. Pairwise protein sequence vs (experimental) protein structure distributions revealed structures are more conserved than their sequences [34], potentially representing a different platform for ARG discovery. Pairwise comparative modelling (PCM) integrates protein structures encoded by ARGs and predicted novel ARGs from the intestinal resistome with an average seqID of 29.8% and was supported by experimental validation (Ruppé et al. [75]). Nonetheless, without comparison of minimum inhibitory concentrations (MICs) to clinical breakpoints, clinical relevance of predictions is difficult to infer. Since then, rapid developments in computational protein structure prediction tools [35, 97] has enabled progress in terms of accuracy (now in line with experimental solutions) and the number of structures available (> 200 million predicted by AlphaFold2) [93]. Protein structural similarity can be quantified using the template modelling (TM)-score, which ranges from 0 (no structural similarity) to 1 (a perfect structural match) [103]. Pairwise sequence vs (computational) protein structure distributions across the microbial protein landscape provide further support for protein structure conservation relative to protein sequence [38]. A nearest-neighbour pairwise analysis (using the maximum shared seqID per antibiotic resistant protein (ARP) structure and its corresponding TM-score) reveals ARP structures across well-known ARG classes exhibit coupling between sequence and structure, with an average median seqID of 83.4% and TM-score of 0.965 (Fig. 1a and b). This indicates that ARGs maintain both high sequence and structural conservation when compared to their closest homologue. However, ARGs can have high TM-scores despite lower seqID, with all classes containing representatives with a maximum seqID below 50% (Fig. 1a). For example, although sharing the maximum seqID of ~ 33%, ARP structures encoded by the aforementioned sul4 and blaNDM-1 share TM-scores of over 0.9 to their respective closest homologues, encoded by sul2 and blaVIM. Global pairwise analyses of the same dataset reveal lower seqIDs, with an average median of 31.7% and TM-score of 0.79, elucidating broader sequence and structure divergence among all representatives within each class (Fig. 1c and d). Importantly, existing computational tools potentially miss novel ARGs which have lower seqID to known ARGs. A computational approach which integrates the entire evolutionary context of the ARG class by a pairwise analysis, as opposed to the nearest-neighbour, may help to identify them. Sequence-structure similarities among antibiotic resistant protein (ARP) structures encoded by well-known ARG classes. a Boxplots of pairwise maximum aa sequence identities (seqIDs) and b their corresponding TM-scores. c Boxplots of pairwise (all-versus-all) seqIDs and d their corresponding TM-scores. ARG classes: aminoglycoside acetyl-transferases (AAC, number of pairwise comparisons (n) = 6228), aminoglycoside nucleotidyl-transferases (ANT, n = 2169), aminoglycoside phospho-transferases (APH, n = 1190), class A β-lactamases (class A BL, n = 87,320), class B1/B2 BL (n = 41,006), class B3 BL (n = 3782), Class C BL (n = 27,060), Class D BL (n = 48,180), dihydrofolate reductases (dfr, n = 2070), mobile dihydropteroate synthases (sul, n = 12), tetracycline resistance ribosomal protection proteins (TRPP, n = 342). Pairwise seqIDs and TM-scores were calculated using Foldseek under default settings, with TM-scores normalised by the subject length Nonetheless, ARP structures typically evolve from existing proteins (proto-resistance genes [55]) performing primary functions and are often involved in the target pathway of an antibiotic [7]. Proto-resistance genes exhibit high TM-scores to ARP structures however lack ARG function. Small changes in critical aa residues can make the difference between a true ARG, from a proto-resistance gene which does not confer clinical resistance but could evolve there (e.g., the evolution of mobile sul genes from chromosomal folP [94]). Therefore, simply prioritising proteins with high TM-scores to find novel ARP structures would likely result in the generation of a high false positive rate (FPR) [36]. Conversely, ARGs can also have reduced TM-scores to ARP structures, as depicted primarily for AAC’s, ANT’s, APH’s, and class B3 β-lactamases (Fig. 1d). This could be explained by ARP structures performing functions by similar localised structural motifs encoded by important aa’s [38]. Proteins which contain these motifs without structural conservation in other regions would have reduced TM-scores to ARP structures, so only prioritising proteins with high TM-scores would also result in a high false negative rate (FNR). We hypothesised that the sequence-structure landscape of conserved structural regions of ARP structures is distinct compared to proteins lacking a resistance function, and that this distinction could be used to infer the resistance function of a potentially novel ARP structure. As such, we developed a precise and simple method for identifying novel ARGs belonging to the following classes: aminoglycoside acetyltransferases (AAC), aminoglycoside nucleotidyltransferases (ANT), aminoglycoside phosphotransferases (APH), class A, B, C, and D β-lactamases, dihydrofolate reductases (Dfr), class 1, 2, and 3 β-lactam-resistant penicillin-binding proteins (PBPs), mobile sulfonamide-resistant dihydropteroate synthetases (Sul), and tetracycline resistance ribosomal protection proteins (TRPP). ARG-PASS (ARG-PAirwise Sequence vs Structure) leverages a one-class support vector machine (SVM) trained on pairwise primary and tertiary distributions of structurally conserved regions of proteins encoded by ARGs. We applied ARG-PASS to six reference strains of the Human Microbiome Project (HMP), and predicted 16 novel ARGs. Of these, 9 were selected for experimental verification in Escherichia coli and all had detectable activity: one APH(6’) from Pseudomonas aeruginosa; two class B3 β-lactamases from Burkholderiales sp. and Acinetobacter radioresistens; a cephalosporinase from Acinetobacter junii; four dfr genes; and a β-lactam-resistant penicillin-binding protein (PBP) from A. radioresistens. Seven reached MICs at Clinical and Laboratory Standards Institute (CLSI) resistant breakpoints [16], and we posit the remainder represent ‘pre-resistance’ genes, with activity below CLSI resistant breakpoints but at risk of becoming clinically relevant in the future. While our primary dataset was HMP reference strains, ARG-PASS also predicted a highly divergent phnP gene (metallo-β-lactamase superfamily protein) directly from the AlphaFold database (AFDB), which conferred a clinically relevant MIC against ampicillin when expressed in E. coli on solid media only. The average seqID of functionally confirmed genes to known ARGs is 46%, highlighting the ability of ARG-PASS to uncover genes with low homology to existing ARGs and potentially a large reservoir of previously uncharacterised ARGs with clinical relevance. Results Development of the ARG-PASS method Experimental validation remains the main bottleneck in the ARG discovery pipeline. High‑precision computational methods are therefore required to efficiently identify potential novel ARGs, including those with low sequence identity to known resistance genes. We hypothesised that the sequence-structure landscape of conserved structural regions of ARP structures is distinct compared to proteins lacking a resistance function, and that this distinction could be used to infer the resistance function of a potentially novel ARP structure. The proposed framework comprises three steps. For each ARP structure within a functionally coherent subgroup of an ARG class, extract a consensus structural fold conserved across the subgroup. For each ARP structure within a functionally coherent subgroup of an ARG class, extract a consensus structural fold conserved across the subgroup. For each conserved fold, quantify the pairwise structural similarity using the TM-score (to assess overall fold similarity) and the seqID at aligned positions (to capture conservation of functionally important residues). For each conserved fold, quantify the pairwise structural similarity using the TM-score (to assess overall fold similarity) and the seqID at aligned positions (to capture conservation of functionally important residues). A query protein structure is compared against the conserved folds of the ARPs within each cluster using the same metrics. If the query protein structure has any TM-score and seqID values comparable to those observed among the cluster, we infer that it also shares the conserved structural features and therefore likely encodes the same resistance function. A query protein structure is compared against the conserved folds of the ARPs within each cluster using the same metrics. If the query protein structure has any TM-score and seqID values comparable to those observed among the cluster, we infer that it also shares the conserved structural features and therefore likely encodes the same resistance function. We integrated these steps into ARG-PASS (ARG-PAirwise Sequence vs Structure) and validated predictions using standardised expression systems combined with antimicrobial susceptibility testing across CLSI breakpoints. In practice, ARG-PASS generates a two-dimensional distribution of pairwise seqIDs vs TM-scores of structurally conserved regions extracted from previously clustered known ARPs (high-lDDT ARP structures) and trains a one-class SVM to learn the decision boundary (DB) surrounding this distribution. A query ARP (qARP) structure is aligned to each high-lDDT ARP structure, and its derived seqID and TM-score are calculated. If any points lie within the DB, the qARP is predicted functional; if all points lie outside the DB, it is flagged as an outlier and predicted non-functional (Fig. 2). Overview of ARG-PASS. Step 1. Antibiotic resistant protein (ARP) structures for an ARG class are clustered at 20%, 30%, and 50% sequence identity (seqID) using Foldseek. Step 2. For each cluster, a multiple structure alignment (MSTA) is performed using FoldMason. Aligned residues with a per-column local Difference Distance Test (lDDT) score ≥ Q50, Q75, or Q90 percentiles of the per-column lDDT distribution are extracted to create high-lDDT ARP structure.pdb files for each ARP structure. Step 3. The pairwise (all-vs-all) seqIDs vs TM-scores are calculated for high-lDDT ARP structures within clusters using Foldseek TM-align. Step 4. Query ARP (qARP) structures are assigned to ARP cluster levels from Step 1 based on their maximum seqID to an ARP structure (determined by Foldseek). SeqIDs and TM-scores of the qARP structure to high-lDDT ARP structures in each cluster of the level are calculated using Foldseek TM-align. Step 5 (a). A one-class SVM is trained on the x, y distributions from Step 3, where the decision boundary (DB, in red) surrounds all training points. The seqIDs vs TM-scores of the aligned region of a single qARP structure against each high-lDDT ARP structure from Step 4 are plotted (data points in orange). If any data points lie on or within the DB, the qARP is classed as functional (b, example dfr qARP structure classed as functional), otherwise the qARP is classed as non-functional. c, d Maximum TM-score and seqID between qARP structures from HMP reference strains and ARP structures vs high-lDDT ARP structures after ARG-PASS (Mann-Whitney test between yes/no predictions of each condition: p < 0.001). e Numbers of qARPs from HMP reference strains analysed in this study. For further details see Methods. Figure created with BioRender.com ARG-PASS models were developed for clinically important ARG classes with characterised and conserved structural folds. Efflux pumps and regulatory mechanisms were excluded. One-class SVM models were trained on seqID vs TM-score (x,y) distributions of high-lDDT ARP structures from the following ARG classes: AAC, ANT, APH, class A, B, C, and D β-lactamases, Dfr, class 1, 2, and 3 β-lactam-resistant PBPs, mobile Sul, and TRPP. We validated ARG-PASS on an existing protein structure-based method used to identify ARGs from the human intestinal microbiota, called Pairwise Comparative Modelling (PCM) (Ruppé et al. [75]). ARG-PASS was evaluated against PCM sequences with functional synthesis results available and seqIDs below 95% to known ARGs (n = 59, Additional file 1: Table S5). ARG-PASS achieved a precision of 0.85 and specificity of 0.75, comparable to the precision of PCM ‘good’ predictions (0.853), and higher than PCM ‘fair’ predictions (0.44, Additional file 1: Tables S5 and S6). Furthermore, high precision of ARG-PASS was retained at low seqID, with 14 out of 17 (0.82) true positive predictions correct at seqIDs below 31% to known ARPs. The highest precision of ARG-PASS was achieved for AACs, and class B3 and D β-lactamases (1.00), but no functional predictions were made for ANT, class C β-lactamases, or TRPP, and contributed to a lower overall sensitivity of 0.51 (Additional file 1: Table S6). This potentially reflects the training of ARG-PASS on ARGs exclusively from the CARD, which predominantly confer clinical levels of resistance (see Discussion). The sensitivity of ARG-PASS was further evaluated on experimentally confirmed ARGs from the ResFinderFG v2.0 database, identified through functional metagenomics [30]. ARG-PASS functionally predicted 164 out of 204 sequences (sensitivity of 0.80), across a range of seqIDs (see Methods, Additional file 1: Table S7). Sensitivity for ANT (2/4), class C β-lactamases (16/21), and TRPP (14/15) was partially recovered relative to the PCM dataset, though these predictions have higher average seqIDs (42%, 70.9%, and 81%, respectively, Additional file 1: Table S7). Overall, the sensitivity was lower for hits with a seqID > 50% to known ARGs and assigned to 50% clusters according to the ARG-PASS pipeline, with only 47 out of 61 correctly identified (0.77). This potentially reflects tightened decision boundaries of seqID vs TM-score distributions from clusters with a low number of cluster members and thus training data, and will be addressed in future work (see Discussion). Finally, it should be noted that 13 out of 21 (0.62) candidates which fell in the 0–35% seqID range were identified by ARG-PASS, however 4 missed APH candidates were predicted against clusters with reduced training data and thus lower confidence (see Discussion, Additional file 1: Table S7). Specificity of ARG-PASS was further assessed by cross-class evaluation of β-lactamase models. ARP structures from each serine β-lactamase class (A, C, and D) and PBP’s were evaluated against models of the other serine β-lactamase classes (e.g., class A and C β-lactamases and PBP were evaluated against class D β-lactamases and so on). Class A, C, and D β-lactamases and PBPs share the DD-peptidase/serine β-lactamase CATH superfamily and structural similarity, so they represent challenging negatives [41]. Similarly, class B MBL classes share the MBL αββα fold and have related zinc coordination chemistry [39]. Thus the class B1/B2 MBL structures were evaluated against the class B3 model and vice versa. The only cross-class β-lactamase ARP structure predicted functional was LRA-13 (AlphaFold (AF)/UniProt accession: B5L5V6), a class C/class D bifunctional fusion β-lactamase predicted functional by the class D β-lactamase model and thus is expected behaviour. As such, the cross-class testing demonstrates that ARG-PASS discriminates class- and subclass-specific structural features. Furthermore, with a specificity of 0.75 against the PCM dataset, this suggests ARG-PASS sufficiently rejects non-functional proteins with similar folds. Notably, five proteins classified as class B1 MBLs presented in the PCM validation dataset were also predicted functional by the B1/B2 ARG-PASS model (Additional file 1: Table S5), yet possess class B3-like architecture. These proteins were both (i) experimentally confirmed to be functional MBLs in the PCM paper and (ii) have below 30% seqID to both B1/B2 and B3 MBL structures, and probably occupy a challenging prediction space for the B1/B2 MBL model to correctly assign the subclass, unlike the B3 MBLs correctly rejected in cross-class testing (discussed further below). ARG-PASS identified novel ARGs from Human Microbiome Project reference strains Protein structures encoded by query antibiotic resistant proteins (qARPs) originating from six HMP reference strains [89] (Additional file 1: Table S1), were downloaded directly from the AFDB if available or computed by AlphaFold2 (AF2) or AlphaFold3 (AF3), representing qARP structures for functional prediction by ARG-PASS. In total, 16 qARPs were classed as functional by ARG-PASS, predicted to provide resistance to streptomycin, β-lactams, or trimethoprim (Fig. 2e, Additional file 1: Table S2). The average maximum TM-score and seqID of qARP structures predicted as functional shared with high-lDDT ARP structures increased from comparison to full-length ARP structures (TM-score of 0.94 to 0.97 and seqID of 45.7% to 54%). Conversely, the average maximum TM-score and seqID of qARP structures predicted as non-functional to high-lDDT ARP structures from ARP structures decreased (TM-score of 0.79 to 0.75 and seqID of 28.9 to 26.6), which perhaps underscores the ability of the algorithm to distinguish from qARPs which do not have functionally important primary and tertiary structural features (Fig. 2c and d). Nine qARPs predicted as functional by ARG-PASS and belonging to five ARG classes (APH, Dfr, class B3 and C β-lactamases, and β-lactam-resistant PBPs), 5 qARPs belonging to ARG classes assessed by ARG-PASS without any predicted as functional, apart from TRPP (Fig. 2f, class B1/B2 β-lactamases, AAC, and Sul), and positive controls (not already available in the Minimal Antibiotic Resistance Platform), were synthesised and cloned into appropriate plasmids (pGDP2 or pGDP3) and expressed in Escherichia coli BW25113 (ΔbamBΔtolC) for functional verification [18]. We used the efflux deficient, hypersusceptible E. coli strain for expression of qARPs, so any antibiotic resistance activity observed could not be attributed to efflux pumps. Activity was detected in all qARPs predicted as functional against their respective antibiotics (Additional file 1: Table S2), with minimum inhibitory concentrations (MICs) determined in liquid media, by broth microdilution (BM) and solid media, by a spot plating assay (SPA) (Table 1). Genomic context analysis suggests qARPs with detected activity represent stable chromosomally encoded components of gut microbiome genomes rather than within MGEs. No activity was detected in the tested qARPs which were predicted as non-functional by ARG-PASS (Additional file 1: Table S2, MIC data not shown). Table 1Functional confirmation results of qARPs and controls. All MICs were determined by Broth Microdilution (BM) unless markedMIC (µg/mL) for E. coli BW25113 (ΔbamBΔtolC) harbouringqARPControlAmpicillinCefotaximeMeropenemTrimethoprimSMXTMP/SMXStreptomycinKanamycinCLSI resistant breakpoint ≥ 32 ≥ 4 ≥ 4 ≥ 16 ≥ 512 ≥ 4/76 ≥ 32$ ≥ 64pGDP3-aph01Aph01128; > 64*pGDP2-bla01Bla01 > 256; > 32*32; 16* ≤ 0.25; ≤ 2*pGDP2-pbp01Pbp01256; > 32*16; 8* ≤ 0.25; ≤ 2*pGDP2-bla02Bla024; ≤ 16* ≤ 0.5; ≤ 2* ≤ 0.25; ≤ 2*pGDP2-bla03Bla034; ≤ 16* ≤ 0.5; ≤ 2* ≤ 0.25; ≤ 2*pGDP1-bla04Bla04 ≤ 2; 32* ≤ 0.5; ≤ 2* ≤ 0.25; ≤ 2*pGDP2-bla04Bla04 ≤ 2; ≤ 16* ≤ 0.5; ≤ 2* ≤ 0.25; ≤ 2*pGDP3-dfr01Dfr01 > 256; > 128* ≤ 2/38*pGDP3-dfr02Dfr02 > 256; > 128* ≤ 2/38*pGDP3-dfr03Dfr0332; ≤ 8* ≤ 2/38*pGDP3-dfr04Dfr0464; 32* ≤ 2/38*pGDP3-folAfolA64; 32* ≤ 2/38*pGDP3-AAC(6')-Ie-APH(2'')AAC(6')-Ie-APH(2'') > 256; > 64*pGDP1-ACT-28ACT-28 > 256; > 32*32; > 4*2; 4*pGDP2-ccrAccrA64; > 32*32; > 4*16; > 4*pGDP1-AIM-1AIM-1 > 256; > 32* > 64; > 4* > 32; > 4*pGDP3-dfrKdfrK > 256; > 128* > 4/76*pGDP3-sul3sul3 > 512* ≤ 2/38*pGDP2-APH(6)IaAPH(6)-Ia256; > 64*pGDP2/3/4-empty vector ≤ 2; ≤ 16* ≤ 0.5; ≤ 2* ≤ 0.25; ≤ 2* ≤ 4; ≤ 8* ≤ 256* ≤ 2/38*4; ≤ 16*2; ≤ 16*SMX sulfamethoxazole, TMP/SMX trimethoprim/sulfamethoxazole*Measured by the spot plating assay (SPA)$NARMS breakpoint; Bold text = MIC > MIC of negative control Functional confirmation results of qARPs and controls. All MICs were determined by Broth Microdilution (BM) unless marked SMX sulfamethoxazole, TMP/SMX trimethoprim/sulfamethoxazole *Measured by the spot plating assay (SPA) $NARMS breakpoint; Bold text = MIC > MIC of negative control qARPs conferred MICs at CLSI resistant breakpoints β-lactamases For ampicillin, the negative control strain exhibited an MIC of 2 µg/mL. According to CLSI guidelines, susceptibility is defined as MIC ≤ 8 µg/mL and resistance as MIC ≥ 32 µg/mL. Expression of bla01 increased the MIC to > 256 µg/mL measured by BM and to > 32 µg/mL measured by the SPA, representing up to a 128-fold increase relative to the negative control, and places the strain in the resistant category for ampicillin (Table 1). For cefotaxime, the negative control strain exhibited an MIC of ≤ 2 µg/mL, while the CLSI susceptibility and resistance breakpoints are ≤ 1 µg/mL and ≥ 4 µg/mL, respectively. Expression of bla01 increased the cefotaxime MIC to 32 µg/mL by BM and 16 µg/mL by the SPA, also placing the strain in the resistant category for cefotaxime (Table 1). Bla01 is a potential cephalosporinase from Acinetobacter junii SH205 with 64.2% seqID to ADC-8 (Additional file 1: Table S3). ARG-PASS predicted the maximum TM-score (0.99) and seqID (73.8%) of Bla01 is shared with RSC1-1Q50_lddt (Fig. 3a and b). RSC1-1 was identified from Indian river sediments [49]. Based on low sequence homology to existing cephalosporinases and elevated MICs against β-lactams, we tentatively propose a name for Bla01 of ADC-n (where n = next available number). Catalytic motifs known to be responsible for β-lactam hydrolysis are conserved in ADC-n (64SXSK, 150YXN, and 315KTG) [69] and were captured as conserved regions of high-lDDT ARP structures in clusters of known cephalosporinases by ARG-PASS (Fig. 3b, Additional file 3: Alignment 1). a, c, e, g, i One-class SVM models for query antibiotic resistant protein (qARP) structures predicted as functional by ARG-PASS, each depicted with at least one data point within the DB (in red) surrounding the training distribution. b, d, f, h, j qARP structures with activity following functional verification are overlaid with the high-lDDT antibiotic resistant protein (ARP) structure with the maximum TM-score. Protein sequences of qARPs and fasta files of Foldseek structural alignments of qARP structures and high-lDDT ARP structures with the maximum TM-score are available in Additional files 2 and 3. One-class SVM plots were made using scikit-learn [67] and AlphaFold structures visualised with UCSF Chimera [51]. Abbreviations: SVM, support vector machine; BL, β-lactamases; APH, aminoglycoside phosphotransferases; PBP, penicillin-binding proteins; dfr, dihydrofolate reductases Acinetobacter sp. are common sources of ampC cephalosporinases, in particular ADC (Acinetobacter-derived cephalosporinases) variants which are often identified in A. baumannii [32]. ADC-n represents the first cephalosporinase functionally confirmed from HMP reference strain A. junii SH205, which may be the naturally occurring cephalosporinase from this species. Strains of A. junii are not commonly associated with disease, but it has been reported as an emerging One-Health pathogen [4]. Furthermore, the earliest observations of blaNDM-1 and tet(X3) originate from A. junii [28, 29]. Phylogenetic inference of ADC-n with ADCs and other closest related cephalosporinases from Acinetobacter sp. (including those without MIC data obtained from the Beta-Lactamase DataBase (BLDB) [56] and NCBI) confirms its phylogenetic proximity to ADC-8, functionally confirmed in 2007 from Acinetobacter baylyi however is closest to the potential AmpC cephalosporinase from A. beijerinckii (not present in CARD). Both ADC-n and ADC-8 are divergent from other functionally confirmed cephalosporinases [104] (Additional file 4: Figure S1). Streptomycin resistance The negative control strain exhibited an MIC of 4 µg/mL against streptomycin as measured by BM. There are currently no CLSI breakpoints established for streptomycin; however, the National Antibiotic Resistance Monitoring Network for Enteric Bacteria (NARMS) record an interpretative susceptible MIC of ≤ 8 µg/mL and resistant MIC of ≥ 32 µg/mL for years tested between 2014 and 2019 [14]. Expression of aph01, a potential APH(6’) from HMP reference strain P. aeruginosa 2_1_26, increased the streptomycin MIC to 128 µg/mL by BM and MIC > 64 µg/mL by the SPA, similar to the positive control, APH(6’)-Ia. Aph01 shares a maximum seqID of 52.6% with APH(6)-Ic (Additional file 1: Table S3) and ARG-PASS identified a TM-score of 0.96 and seqID of 66.7% of Aph01 to APH(6)-IcQ50_lddt (Fig. 3e and f). Four known streptomycin-resistant APH(6’)’s are listed in CARD and all clustered at 20% seqID in the ARG-PASS pipeline (Fig. 3e). Along with APH(6’)-Ic and APH(6’)-Id, Aph01 likely represents the third known from P. aeruginosa, and fifth known APH(6’) gene. As such, we tentatively designate its name as APH(6’)-Ie. All three APH(6’) enzymes identified from P. aeruginosa (including APH(6’)-Ie presented here) are phylogenetically closer than APH(6’)-Ia and APH(6’)-lb, from Streptomyces griseus and S. glaucescens, respectively (Additional file 4: Figure S2). Trimethoprim and trimethoprim-sulfamethoxazole resistance The CLSI defines susceptibility and resistance to trimethoprim (TMP) as MIC’s of ≤ 8 µg/mL and ≥ 16 µg/mL, respectively. When expressed in E. coli, dfr01, dfr02 and dfr04, all encoding potential TMP-resistant dihydrofolate reductases (DHFRs) from Burkholderiales sp. 1_1_47, P. aeruginosa 2_1_26, and K. pneumoniae WGLW3, respectively, conferred clinical resistance to TMP by BM (MIC > 256 µg/mL for Dfr01 and Dfr02, and MIC 64 µg/mL for Dfr04) and by the SPA (MIC > 32 µg/mL for Dfr01 and Dfr02, and MIC 32 µg/mL for Dfr04). These MICs were similar to the positive control, dfrK, and up to 64-fold higher than the negative control (MIC ≤ 4 µg/mL by BM) (Table 1). Dfr03 from E. coli 53FAA also conferred elevated MICs to TMP however only when measured by BM. Differential growth dynamics in liquid versus solid media potentially influenced the observed activity of Dfr03 and has been reported previously for the susceptibility of Campylobacter sp. to trimethoprim-sulfamethoxazole (TMP/SMX) [45]. ARG-PASS determined Dfr01 and Dfr02 have maximum TM-scores of ~ 0.99 to DfrA39Q50_lddt and DfrA12Q50_lddt, respectively, and Dfr03 and Dfr04 share their maximum TM-score of ~ 0.99 with seqID of 53–56% to DfrA3Q50_lddt (Fig. 3j, Additional file 1: Table S2). The resistant category to TMP/SMX is defined by the CLSI as an MIC of ≥ 4/76 µg/mL. The strain harbouring positive control dfrK also conferred clinical resistance to TMP/SMX. As a structural analogue of p-aminobenzoic acid, sulfamethoxazole (SMX) competitively inhibits dihydropteroate synthetase (DHPS), which is an earlier step than where TMP acts in the folate synthesis pathway [61]. Clinical resistance to TMP/SMX is commonly reported by the combined presence of both a dfr and a sul gene (sul1-sul4) and less likely by the presence of only either [71, 80]. Interestingly, a clinical MIC to TMP/SMX was not observed for the strain harbouring a known mobile sul (sul3), as opposed to dfrK, which may have clinical relevance (Table 1). We investigated whether elevated TMP MICs were caused by overproducing DHFR which can titrate out TMP due to its mechanism as a competitive inhibitor, particularly under the control of the bla promoter in the pGDP3 plasmid [64, 84]. We cloned and expressed the wild-type (WT) E. coli chromosomal EcDHFR (encoded by folA) into pGDP3. EcDHFR conferred a similar TMP MIC to dfr04 by both measurement protocols, and conferred a higher MIC than dfr03. EcDHFR actually shares 100% seqID with Dfr03, apart from an additional 37 residues constituting an α-helix at the N-terminus in Dfr03 (Fig. 3j). This may have played a role in restricting the expression of dfr03 and thus the lower MICs recorded compared to EcDHFR. Furthermore, Dfr04 shares a seqID of 92% to Dfr03. After sequence alignment, substitutions present in Dfr04 from Dfr03 are V41I, H45L, I61V, L62I, Q65K,T68S, T73Q, K76S, D79E, P89E, E129D, D132E, and H149N. Further research is needed to deduce if these substitutions affected the variable TMP MICs observed. As with the genomic context analysis completed for all qARPs, the absence of exclusively MGE-associated sequences surrounding dfr01-dfr04, prevented confident inference of mobility for these enzymes. Moreover, contigs containing dfr01-dfr04 from their respective hosts are predicted to originate from chromosomal DNA by PlasmidHunter [91]. This suggests Dfr01-Dfr04 are Dfr products of chromosomally encoded folA in their respective HMP reference strains. E. coli BW25113 (ΔbamBΔtolC) expressing dfr01 and dfr02 exhibited three-fold higher TMP MICs than dfr03, dfr04, and EcDHFR, and conferred comparable MICs to the positive control dfrK. This phenotype is likely explained by intrinsic enzyme properties such as reduced TMP affinity (higher Ki) and/or increased catalytic efficiency (kcat) relative to Dfr03, Dfr04, and EcDHFR. Steady-state kinetic assays would be required to confirm these mechanisms. When Dfr01-Dfr04 are structurally aligned using FoldMason with the remaining Dfrs in their one-class SVM cluster for which they were predicted functional (Additional file 3: Alignment 7), no known resistance-conferring mutations are present across both queries and subjects [47, 86]. This provides support for the common mechanism of TMP-resistance in known mobile dfr genes: expression of decontextualised folA. It is also unsurprising that ARG-PASS predicts Dfr01-Dfr04 as functional, given it was trained on decontextualised folA genes (i.e., Dfr proteins from the CARD). There is alarming potential for these enzymes to confer clinical resistance to TMP when expressed in a low-copy plasmid by a promoter of similar strength to the bla (as observed under these experimental conditions), and may provide evidence for the ubiquity of mobile Dfrs currently known. Phylogenetic analysis with the remaining Dfr proteins in the cluster for which Dfr01-Dfr04 were predicted functional, indicates they are distinct from known TMP-resistant Dfr’s (Additional file 4: Figure S3). Dfr02 from P. aeruginosa is sister to Dfr03 and Dfr04, and is closest related to the integron-encoded DfrA3 from E. coli [11]. Strains of P. aeruginosa are often intrinsically resistant to TMP by low outer membrane permeability, efflux pumps, and the integron-encoded dfrB5 [23, 65]. Being the most divergent identified in this study, Dfr01 is outgrouped from Dfr02-Dfr04 (Additional file 1: Table S3; Additional file 4: Figure S3). Burkholderiales sp. may represent a source of dfrB genes and to the best of our knowledge, Dfr01 represents the first report of folA from Burkholderiales sp. conferring clinical resistance to TMP when expressed in E. coli [37]. A PBP3 homologue from A. radioresistens was functionally confirmed to provide resistance to β-lactams The strain harbouring Pbp01, a predicted penicillin-binding protein (PBP) from A. radioresistens SH164, conferred MICs above CLSI resistant breakpoints against ampicillin by BM (MIC 256 µg/mL) and the SPA (MIC > 32 µg/mL), and cefotaxime by BM (MIC 16 µg/mL) and the SPA (MIC 8 µg/mL, Table 1). Based on results of the Diamond search against the custom CARD database, Pbp01 (ACPO01000036.1_161) shares a seqID of 40.7% to the WT PBP3 from Klebsiella pneumoniae (Additional file 1: Table S3). Computational prediction of membrane-associated proteins have lower relative confidence [3] and PBPs are longer than most ARPs (i.e., > 500 aa’s). Foldseek (used in the ARG-PASS pipeline) structurally aligned distant residues of Pbp01 to the high-lDDT ARP structure of PBP3 from K. pneumoniae (AlphaFold: W9BBU1) with a TM-score of 0.96 and seqID of 48.6% (Fig. 3h, Additional file 1: Table S2). To the best of our knowledge, Pbp01 represents the first functionally verified report of a PBP3 from A. radioresistens conferring β-lactam resistance when expressed in E. coli. Point mutations in PBPs reduce the affinity of β-lactam antibiotics for their targets [101]. ARP structures used to create one-class SVM models for PBPs, were encoded by the representative WT sequences of PBP1–PBP3 in the protein-variant model in CARD, without introducing resistance-conferring mutations. This approach was considered appropriate because structural alignment of PBPs containing large numbers of diverse point mutations, often with highly localised effects, would be unlikely to yield a reliable consensus of structurally conserved regions (high-lDDT). For example, CARD lists 30 resistance-conferring point mutations/insertions for Helicobacter pylori PBP1 mutants (ARO: 3007060). The experimental confirmation of a predicted resistant PBP3 homologue, Pbp01, using a TM-score vs seqID distribution of conserved structural regions of WT PBPs, suggests these regions possess a consensus structural backbone, which, in addition to point mutations, could be important for the function of β-lactam resistant PBPs. Five other PBPs were predicted as functional in this study, originating from A. junii, P. aeruginosa., K. pneumoniae subsp. pneumoniae, and Burkholderiales. sp., with TM-scores over 0.93 and seqIDs over 45% to known PBP3’s. It is difficult to assert if these likely represent true β-lactam-resistant PBP3’s given only one PBP3 was experimentally tested; however, further research is warranted into the β-lactam susceptibility of these strains. Reports linking PBPs to β-lactam resistance need further elucidation in Acinetobacter sp. [79, 101]. Further research is needed to deduce if the observed phenotype in Pbp01 is derived from point mutations and/or a reduced affinity for β-lactams [101]. OXA-905, only known from P. aeruginosa in CARD, was also identified in A. radioresistens in this study (Additional file 1: Table S3) suggesting this strain could already be resistant to penicillins and/or cephalosporins [22]. A Blastp search against the NCBI database reveals Pbp01 shares the maximum seqID of 85% to a PBP3 from the clinical pathogen A. baumannii (NCBI: MFY0112749.1). We could not ascertain whether this PBP3 from A. baumannii also provides resistance to β-lactams. Their significant homology shared poses a risk for pbp01 to horizontally transfer to A. baumannii, as occurred with OXA-23, where it is now a common carbapenemase [72]. Pre-resistance genes encode qARPs with activity but MICs below CLSI resistant breakpoints E. coli strains expressing bla02 and bla03, encoding potential class B3 metallo-β-lactamases (MBLs) from Burkholderiales sp. and A. radioresistens, respectively, conferred a two-fold increase in the ampicillin MIC determined by BM (MIC 4 µg/mL) against the negative control (MIC ≤ 2 µg/mL) (Table 1). ARG-PASS determined both Bla02 and Bla03 share a maximum structural similarity in conserved protein regions to class B3 MBLs, SPR-1Q75_lddt and MBL1bQ75_lddt, respectively (Fig. 3d, Additional file 1: Table S2). Whilst Bla02 and Bla03 share the conserved structural fold of class B3 MBLs as determined by ARG-PASS, they conferred sub-clinical levels of ampicillin resistance in the heterologous E. coli expression system used, and fall within the CLSI susceptible category (MIC ≤ 8 µg/mL). The Zn2+-binding active site encoded by residues H116, H118, D120, H121, H196, and H263 in class B3 MBLs [10] is conserved in Bla02 and Bla03 when aligned with AIM-1 (tested as a positive control in this study, Fig. 4a, Additional file 3: Alignment 13). Based on their structural and sequence homology to known ARGs and sub-clinical resistance activity, we term these ‘pre-resistance’ genes. Where other studies typically assign functionality to genes with any detected activity compared to a negative control after heterologous expression in E.coli, here we segregate them for a number of reasons. First, although heterologous expression in E.coli under laboratory conditions does not reflect the expression within a native host context, E.coli is one of the most common clinical human pathogens, and accurately quantifying MICs for pre-resistance classification is relevant if they were to mobilise to a pathogenic E.coli strain. Furthermore, we speculate pre-resistance genes constitute a distinct component of the resistome and have higher potential to evolve into ARPs (i.e., with clinical relevance). They are potentially further along the evolutionary path to clinical resistance than proto-resistance genes, for example in this case due to catalytic inefficiency compared to clinically relevant MBLs [54, 55]. Finally, pre-resistance genes like Bla02 and Bla03, could occupy a critical area within the fitness landscape of ARGs, where they are maintained in bacterial populations with relatively low fitness cost (e.g., compared to canonical MBLs [44], under commonly encountered, subclinical exposure to antibiotics (Fig. 4b). Conceptual Fitness-MIC landscape of the resistome as determined by Zn active site architecture of class B3 MBL’s. a AIM-1, Bla03, PCM1, and PCM2 exhibit overall fold similarity; however, PCM1 and PCM2 each lack a Zn- binding histidine residue (full Chimera structural alignment is available in Additional file 3: Alignment 11). b Fitness-MIC landscape comprising components of the Resistome within a continuum, ranging from proto-resistance elements with no activity (e.g., PCM2), through pre-resistance genes that confer detectable MIC increases yet without clinical relevance (e.g., Bla03 and PCM1), to fully evolved resistance determinants (e.g., AIM-1) Nonetheless, it should be noted an efflux-pump deficient E. coli BW25113 (ΔbamBΔtolC) was used to functionally verify qARPs to ensure all elevated MICs are attributed directly to the query enzyme. Although Bla02 and Bla03 identified in this study confer pre-resistance under the laboratory conditions in this study, in other hosts, they may act in concert with other molecular determinants when selected by higher antibiotic dosing regimes and subsequently reach CLSI resistant breakpoints, for example alongside increased efflux of carbapenems [62, 88]. Furthermore, the yield of functional periplasmic enzymes may have also negatively impacted the observed phenotype [83] and can be influenced by a less compatible signal peptide with the expression system [44]. Taken together, this suggests higher MICs may be observed in the native host. Here, ‘pre-resistance’ can be defined operationally as genes conferring elevated but sub-clinical MICs under heterologous laboratory expression conditions within E.coli. Further experimental validation in native host contexts is warranted to determine whether this reflects the native biological state. Structural alignment suggests a PCM-confirmed class B3 MBL is a candidate pre-resistance gene After structural alignment of a functional class B3 MBL candidate from the PCM validation dataset [75], designated PCM1 (query MC3.MG291.AS1.GP1.C1053.G2) and a non-functional class B3 MBL candidate designated PCM2 (query MC3.MG13.AS1.GP1.C10091.G28), with the known class B3 MBL, AIM-1, and Bla03, both tested in this study, reveals that PCM1 lacks histidine residue H263, which coordinates the Zn2 site [39]. This points to a structurally intermediate, class B3 MBL-like architecture. A tryptophan residue occupies this position (equivalent to H366 in the alignment Fig. 4a), and would likely compromise Zn2 coordination and reduce catalytic efficiency, analogous to the Lys263 substitution characterised in CSR-1 which reduced activity and sensitivity to clavulanic acid [68]. Therefore, PCM1 may represent a pre-resistance gene rather than a bona fide B3 MBL, and is a distinction not experimentally verifiable by nitrocefin testing alone (as performed in the PCM study). Moreover, PCM2 lacks histidine residue: H196 (equivalent to H234 in the alignment), which coordinates Zn1 for primary catalysis, and likely is more crucial for β-lactamase activity than H263 (absent from PCM1) [39]. Although PCM2 is structurally similar to AIM-1 (Fig. 4a), the absence of H196 and no detected activity in the PCM study suggests it represents a proto-resistance gene. Together with catalytic inefficiency and/or potentially incompatible signal peptides discussed above for Bla02 and Bla03, these examples provide a putative molecular basis for delineating the boundaries between proto-resistance, pre-resistance, and resistance (Fig. 1a and b). MIC testing of PCM1 against CLSI breakpoints would nonetheless be required to support this hypothesis. Structural alignment of the putative class B1 MBLs identified by PCM [75] with canonical class B1 MBLs, NDM-1, VIM-1, and IMP-1, reveals an absence of C221 (critical for Zn2 coordination) across all confirmed proteins, with all other zinc-binding residues conserved (Additional file 3: Alignment 12). After structural alignment to known class B3 MBLs and Bla02 and Bla03, these proteins also possess H121 at the equivalent Zn1 coordination site, which also suggests B3-like active site architecture (Additional file 3: Alignment 13). Notably, they were also predicted as functional class B1 MBLs by ARG-PASS. Although the use of the FoldMason lDDT as a threshold for conserved structural regions of proteins was sufficient to identify most functional MBLs (including potential pre-resistance genes) from both the PCM study and in this study with high precision, it may lack the resolution to discriminate functional MBL subclasses at higher levels of sequence divergence and is a limitation planned to be addressed in future work. Relaxing ARG-PASS thresholds predicted an evolutionarily divergent MBL-fold protein with ampicillin activity We next sought to further optimise functional predictions of qARPs by ARG-PASS by skipping the initial Diamond screening of a DNA database against the created database of ARG sequences from the CARD. This was achieved by analysing protein structures directly from the AFDB. We also sought to test the capability of ARG-PASS at predicting proteins with less than 20% seqID to known ARP structures and prioritised MBLs given their importance in conferring resistance to last-resort carbapenems. Finally, we prioritised annotations originating from E. coli to maximise chances of successful expression in our expression system E. coli BW25113 (ΔbamBΔtolC). The UniProtKB database (release 2025_02) was queried with the following filter: (taxonomy_id:562) AND (protein_name: Metallo-beta-lactamase). After discarding qARPs with > 20% seqID to known MBLs, redundant structures, and those not within 70–130% of the subject ARP structure length, 44 qARPs were analysed by ARG-PASS. Two qARPs were predicted as functional class B1/B2 MBLs; however, no qARPs were predicted as functional class B3 MBLs. Given class B3 MBLs are more diverse than class B1/B2 MBLs [39], this might reflect a limitation of homology-based clustering at a high level of divergence. We therefore conducted an additional analysis excluding prior clustering before obtaining high-lDDT ARP structures for class B3 MBLs (see Methods). Subsequently, 27 qARPs were predicted as functional class B3 MBLs by ARG-PASS, sharing an average maximum TM-score of 0.8 and seqID of 17.3% with high-lDDT class B3 MBLs. The two qARPs predicted by ARG-PASS as functional class B1/B2 MBLs as previously mentioned, also form part of these 27, and they share an average maximum TM-score of 0.875 and seqID of 24% with high-lDDT class B1/B2 MBLs (Additional file 1: Table S4). One qARP was selected for functional verification based on its annotation as a phosphonate metabolism protein (phnP, carbon-phosphorus lyase complex accessory protein, MBL superfamily) with no prior association with β-lactam resistance. The selected qARP (AF/UniProt accession: A0A377FG49), designated Bla04 here, was one of 12 phnP annotations in the AFDB predicted as a functional class B3 MBL by ARG-PASS, representing the most frequent annotation among the 27, and the basis for selecting a representative from this group for experimental verification (Additional file 1: Table S4). No activity compared to the negative control was initially observed when bla04 was expressed with the weaker lac promoter (within pGDP2), so bla04 was expressed by the stronger bla promoter (within pGDP1). In this context, Bla04 conferred an increased MIC against ampicillin when measured by the SPA (MIC 32 µg/mL) compared to the negative control (MIC ≤ 16 µg/mL), with growth observed at the CLSI intermediate breakpoint of 16 µg/mL. However, an increased ampicillin MIC was not observed compared to the negative control when measured by BM (Table 1). Bla04 is known from E. coli strains 412057 and NCTC9706 (non-HMP reference strains) in UniProt, and the phn operon in E. coli breaks down phosphonates when phosphate is scarce [98]. After structural alignment using Foldseek of Bla04 to known MBLs, Bla04 shares a maximum seqID of 17.4% with SFB-1, which makes it highly divergent from known MBLs. ARG-PASS revealed Bla04 shares a maximum TM-score of 0.8 with a seqID of 19.3% to PNGM-1Q75-lddt. (Fig. 5a, Additional file 1: Table S4). PNGM-1 was isolated from a deep-sea sediment and has a similar ampicillin MIC of 16 µg/mL when expressed in E. coli and determined using MH agar [66]. Only sites H116 and H121 were captured as high-lDDT residues in PNGM-1Q75-lddt. Noting this dataset was not previously clustered, this likely reflects structural variability in the active site across all class B3 MBLs, and flexible zinc coordination can correspond to different substrate profiles [39]. Sites H116 and H121 structurally aligned to the active site present in Bla04 by Foldseek in the ARG-PASS pipeline (encoded by residues H76, H80, Fig. 5a). a The Foldseek TM-align alignment (used in ARG-PASS) of Bla04 and the high-lDDT ARP structure, PNGM-1Q75_lddt (TM-score of 0.84 and seqID of 19.3%) and to PNGM-1Q50_lddt* (TM-score of 0.95 and seqID of 26%). Aligned active site on Bla04 and PNGM-1Q50_lddt* are displayed. *High-lDDT residues were extracted after pairwise alignment between Bla04 and PNGM-1. b Bla04 aligned to the full length ARP structure encoded by PNGM-1 (TM-score of 0.74 and seqID of 16.1%). Foldseek structural alignments are presented in Additional file 3: Alignment 14 and 15 When Bla04 and the full-length ARP structure of PNGM-1 are aligned using Foldseek, they share a TM-score of 0.74 and seqID of 16.1% (Fig. 5b). After pairwise alignment of Bla04 and PNGM-1 using Chimera, the Zn2+-binding active site in class B3 MBLs is conserved in Bla04 (Additional file 3: Alignment 16). To determine if it is potentially contained within a conserved fold relative to PNGM-1, after this pairwise alignment, high-lDDT residues of PNGM-1 were extracted to create a new structure file, designated PNGM-1Q50_lddt*. Here, the TM-score of Bla04 and PNGM-1Q50_lddt* increased to 0.95 with seqID 26% over the aligned region. The active site was captured as high-lDDT residues between these two structures, suggesting a conserved structural and functional core (Fig. 5a). Expression of phnP in E. coli has not been previously associated with β-lactam activity, and the metal-binding site actually favours Mn2+ (over Zn2+) [70]. Interestingly, de Souza et al. [21] have suggested promiscuous β-lactam resistance within ESKAPE species could be a result of β-lactamase homologues which share the MBL fold. No signal peptide was detected in Bla04 using SignalP 6.0 [87]; how a predicted cytoplasmic enzyme gives rise to the observed ampicillin phenotype therefore remains unresolved, though non-canonical cytoplasmic activity or partial mislocalisation by strong over-expression are possibilities. The absence of activity in liquid media aligns with such non-canonical behaviour, as a conventional periplasmic MBL would be expected to function in broth. This discordance may relate to differences between formats including drug diffusion. For example, low solid–liquid agreement for ampicillin susceptibility in E. coli has been reported previously (~ 20%) [63], as well as growth observed on agar and not in LB [78]. Moreover, the low copy number of the pGDP series may have also contributed to these effects [18, 95]. Finally, no signal peptide was detected using SignalP 6.0 in PNGM-1. It has been suggested the additional rRNase Z activity of PNGM-1 may point to the origin of class B3 MBLs [42], and phnP shares the tRNase Z endonuclease fold [70]. Discussion Recent studies have demonstrated the success of in silico prediction of novel ARGs with high precision on testing datasets [73, 96], as well after experimental verification of ARGs with less than 30% homology to known ARGs [75]. Nonetheless, novel ARGs that are computationally predicted by other methods and also clinically important (i.e., confer MICs at CLSI resistant breakpoints), more commonly have higher seqIDs to known ARGs [5, 8, 9, 46, 48], and false positive predictions increase at lower seqIDs. Furthermore, high-throughput functional verification methods remain limited and continue to represent the principal bottleneck in the ARG discovery pipeline. To address this, we leveraged recent advances in computational protein structure prediction to distinguish functional ARGs from proto-resistance genes lacking key amino acid residues and localised structural features associated with antibiotic resistance. This approach prioritised conserved structural regions of known ARGs and the residues that code for them. By adopting automated MIC testing where possible, experimental validation confirmed the method as a novel and precise strategy for ARG discovery. Sixteen potential ARGs identified from HMP reference strains were predicted functional by ARG-PASS. Nine of these were functionally verified and all had activity against streptomycin, trimethoprim, or β-lactams. Five proteins from HMP reference strains predicted non-functional by ARG-PASS were also experimentally confirmed to lack a resistance function, suggesting a low FNR for potential AAC(6’), sul, and class B1 β-lactamases. In addition, a phnP gene with only 17.3% seqID shared to known MBL structures was functionally predicted from the AFDB and had ampicillin activity. In total, eight out of 10 (80%) genes conferred an MIC at or above CLSI resistant breakpoints, with an average seqID of 46% to known ARGs. The high precision of ARG-PASS may stem from its use of the superposition-free metric (lDDT) employed in FoldMason to identify structurally conserved regions. Subsequent calculation of the pairwise TM-score vs seqID across these structurally conserved regions means the algorithm captures the congruence of the overall (conserved) fold for use in downstream machine learning prediction. Inferring function by structural and sequence comparison of distantly modelled residues by restricting alignments to structurally conserved regions (high-lDDT) reduces noise from disordered and/or structurally divergent regions, which can hamper full protein structure alignments [34, 38]. Clustering ARP structures at 20%, 30%, and 50% seqID before multiple structural alignments (MSTAs) is thus a critical component of the method as it limits pairwise seqID and TM-score distances between high-lDDT ARP structures. Subsequent assignment and comparison of a query protein's TM-score and seqID features to a cluster (depending on its existing homology to known ARP structures) contributes to algorithmic precision at a range of seqIDs. However, ARP structures which do not cluster due to distant homology (i.e., less than 20% homology) can limit direct comparison of qARP structure features to these ARPs. This limitation was partly addressed by omitting prior clustering in the ARG-PASS pipeline, enabling the identification and functional confirmation of a highly divergent protein sharing a maximum of only 17.3% similarity to known MBL structures. We anticipate this limitation will be minimised as models are updated with ARGs identified through complementary approaches (e.g., functional metagenomics, other computational tools). The only required input to ARG-PASS is a qARP structure. For qARPs of ~ 300 amino acids in length, protein structure predictions can be completed in under 5 min with a GPU and under 20 h on 8 CPUs. As such, AF structure prediction is computationally heavy, particularly when automation for large-scale annotation is required. Nonetheless, the AFDB can be queried for a vast amount of protein sequences already with protein structures, reducing the need for computing a protein structure. In contrast, subsequent alignment of the qARP structure to high-lDDT ARP structures within each cluster by Foldseek and prediction by one-class SVMs are designed for computational efficiency and scalability [92] and are completed in the ARG-PASS pipeline on average in less than a second on a standard desktop or laptop computer. Improving the scalability of ARG-PASS with respect to protein structure prediction requirements are planned to be addressed in future work. Although limited to six HMP reference strains, this study functionally confirmed nine novel ARGs from the human gut, supporting the view that the human microbiome serves as a reservoir of resistance (and potentially pre-resistance) genes [76]. Although they appear to be chromosomally encoded, which aligns with findings from Ruppé et al. [75], horizontal transfer to pathogenic strains remains possible, particularly under intense or prolonged antibiotic exposure. Two genes did not confer MICs at CLSI resistant breakpoints after expression in E. coli; however, their stable integration within human microbiome-resident strains increases the risk of accumulating mutations and/or transferring to human pathogens, where clinically relevant MICs may be conferred. Finally, a novel streptomycin-resistant APH(6’) gene was identified from P. aeruginosa and folA genes from P. aeruginosa, K. pneumoniae subsp. pneumoniae, and E. coli, all provided TMP-resistance when expressed in E. coli. These species have known pathogenic strains, which increases their compatibility to both spread and function in these hosts. Limitations of the in silico validation of ARG-PASS There are a number of limitations regarding the in silico validation of ARG-PASS worth addressing. ARG-PASS precision, sensitivity, and specificity were assessed using the PCM dataset, which includes experimentally verified positive and negative functional synthesis results. However, PCM candidates were identified from metagenomic data of the human intestinal tract, so the heterologous expression of genes in E. coli may not reflect their possible true functionality in native host contexts. In addition, ARG-PASS identified 20 false negatives (FNs) from the PCM dataset, all with seqIDs below 30% to known ARP structures, aside from the two ANTs with seqIDs of 32.5% and 39% (Additional file 1: Table S5). Several factors may contribute to these FNs. Firstly, ARG-PASS was developed using ARG sequences strictly from the CARD, where resistance determinants are well-characterised and peer-reviewed [5]. As a result, most ARGs used in training one-class SVM models likely confer clinically relevant MICs, thus optimising ARG-PASS for their identification and potentially discriminating proteins which confer sub-clinical resistance. MIC determination for PCM proteins in the experimental validation pipeline used in this study would enable a more direct comparison. The two ANT FNs hold maximum seqIDs to ARP structures ANT(6)-Ia and ANT(6)-Ib, which comprise a cluster at 30% seqID with only four members. The misclassification of the PCM ANTs potentially reflects lower confidence in decision boundaries which arise from limited training data in small clusters. The same effect likely also contributed to the 40 FNs identified from the ResfinderFG 2.0 dataset. The lowest sensitivity was reported for the APH class (four out of nine identified, Additional file 1: Table S7). The five APH FNs (four of which fell within the 0–35% seqID range noted in the in silico analysis) all have highest seqIDs to APH(2″) genes, which form a cluster of 5 members at 30% seqID, thus with a sequence-structure distribution and decision boundary of reduced confidence. Furthermore, 14 FNs from the ResfinderFG v2.0 dataset have a seqID > 50% to known ARP structures and were also assigned to 50% clusters according to the ARG-PASS pipeline. However, most clusters at the 50% seqID level also have less than 10 members (Additional file 1: Table S9). Moreover, a larger number of clusters are omitted at more stringent seqID clustering (i.e., clusters with < 3 ARP structures, Additional file 1: Table S9), of which the closest homologue of a query may be contained and can thus limit best comparison of sequence-structure features and prediction. ARG-PASS was primarily designed for the prediction of novel ARGs at low seqIDs while prioritising precision, with the 50% seqID tier introduced to discriminate against high-seqID qARPs which lack resistance activity. Clusters with a low number of members and training data will be reduced as ARGs are identified through complementary approaches; however, this is a known limitation to be addressed in future versions of ARG-PASS. No dedicated negative test set was developed to assess the specificity and FPR of ARG-PASS which is a limitation of the current analysis. This is commonly achieved by predicting proteins with similar annotations to known ARGs which lack confirmed resistance function (e.g., MBL superfamily proteins). The assumed non-functionality of these proteins is difficult to assert without direct experimental validation; illustrated by the functional confirmation of a phnP gene (MBL superfamily protein) in this study and previously not associated with resistance. Instead, five proteins predicted non-functional by ARG-PASS from HMP reference strains were experimentally confirmed to lack resistance activity. Although a small dataset, these represent challenging cases for ARG-PASS given their ≥ 20% seqID to known ARGs, as per the Diamond screening (Additional file 1: Tables S2 and S3). Two of these were negative sul predictions (Additional file 1: Table S2). This was particularly important because the sul class was the only ARG class not represented in either the PCM or ResFinderFG v.2.0 datasets, which limited their in silico assessment. However, a specificity of 1.0 was established for the prediction of β-lactamases and PBPs across β-lactamase classes, 12 out of 16 non-functional proteins were correctly predicted from the PCM dataset (Additional file 1: Table S6), and all qARPs from HMP reference strains predicted as functional and selected for experimental validation were confirmed to have activity. Taken together, these findings support that ARG-PASS can reliably distinguish non-functional proteins from functional ARGs. Conclusions The computational prediction and functional confirmation of novel ARGs in this study provides a proof-of-concept for training a one-class SVM on pairwise primary vs tertiary protein structure similarities in conserved regions of ARP structures. High-precision algorithms are essential for identifying novel ARGs while minimising false positives revealed by antibiotic susceptibility testing. In this study, this was achieved through experimental verification of both functional and non-functional predictions, using a standardised expression system, and a range of susceptibility testing methods. Comparison of MICs to CLSI resistant breakpoints also delineated pre-resistance from resistance. Future large-scale investigations using ARG-PASS on human and environmental microbiome data may paint a more complete picture of numbers of potentially clinically relevant ARGs circulating in these environments, particularly those with low homology to existing ARGs, as well as genes not previously associated with resistance. Finally, future work will also expand the computational approach to other ARG classes, including beyond predicting ARG functionality, and we encourage researchers to experiment with other protein datasets. Methods Training set Subject antibiotic resistance protein (ARP) structures ARP structures of well-known ARG classes obtained from the CARD v4.0.1 [5] (AAC, APH, ANT, Dfr, Sul, class A, B, C, and D β-lactamases, PBP1—PBP3’s, and TRPP) were downloaded from the AFDB (highest available model quality), or if unavailable in the AFDB were predicted either by AlphaFold2 (AF2) [35], using full databases from 2023-05−30 and prior to the release of AlphaFold3 (AF3) in May 2024, or the AF3 web server since it became available [2]. AF2 predictions were performed using Digital Research Alliance of Canada clusters or the COSMIC2 gateway web server [15]. For PBP1-PBP3’s, which confer resistance by point mutations, the WT sequence was used as a representative sequence for protein structures because individual mutations are not expected to have major structural impact. Class A, C, and D β-lactamases have a large number of available structures (> 400 per class) which permitted filtering by model quality without losing a significant number of training examples, with AF2/AF3 structures only with a pLDDT ≥ 90 retained. Moreover, class A, C, and D β-lactamases comprise many highly similar variants, so redundancy was also reduced by clustering at 99% seqID over 90% length (of both query and subject) using Foldseek (v.10-941cd33) [92]. The retained high-confidence structures were verified to span the sequence diversity of the represented families. However, families for which no member had a structure meeting the final thresholds are as follows: BcIII, BKC, CBP, BCL, PME, ASU1, MUN, DES, Bla1, and BRO. Across all ARG classes, this resulted in 1203 ARP structures used for creation of the one-class SVM models. AlphaFold/UniProt accessions for all ARP structures are provided in Additional file 1: Table S8. Clustering of subject ARP structures ARP structures were first sorted into classes (i.e., AAC, APH, ANT, Dfr, class A, B1/B2, B3, C and D β-lactamases, Sul, PBP, and TRPP) and clustered by Foldseek under default settings at 20%, 30%, and 50% seqID over any length. Clustering of ARP structures at seqIDs of 30% and 50% were deemed appropriate to group functional ARG homologues and are performed elsewhere to reduce functional redundancy [53, 85]. To achieve a more inclusive threshold, 20% seqID was introduced to account for the high sequence divergence across ARG classes and to ensure most ARGs were clustered for downstream analysis. However, any clusters with less than 3 ARP structures were omitted from subsequent analysis. The amount of clusters produced/omitted and numbers of ARP structures within cluster levels are presented in Additional file 1: Table S9. ARP structures used in training one-class SVM models are available for download from high-lDDT ARP structure databases at https://zenodo.org/records/21209970. Structurally conserved regions of ARP structures To identify structurally conserved regions within clusters of ARP structures, we employed FoldMason (v.2-7bd21ed) [27], a tool designed for structural alignment and consensus modelling of protein structures at scale. This release was current at the time of pipeline development; subsequent releases introduced changes to the alignment algorithm and lDDT scoring that reduced accuracy on the PCM dataset used for validation. FoldMason v.2-7bd21ed is therefore retained for reproducibility. FoldMason utilises a per-column local distance difference test (lDDT) score [50] to assess the consistency of atomic positions across aligned structures. The lDDT can estimate local structural conservation across multiple structural alignments (MSTAs) without relying on superposition, and enables identification of conserved structural regions even in protein structures with global conformational variability. There are limited established lDDT score thresholds that correspond to conserved structural columns within MSTAs, and they are often context-dependent. In this study, we analysed the distribution of lDDT scores (excluding values of −1, where only one protein structure is aligned in the column) and extracted residues ≥ the upper 50th percentile (Q50), the upper 75th percentile (Q75), or the upper 90th percentile (Q90), depending on the number of protein structures in the MSTA (Table 2). Thresholds were selected after manual inspection of extracted residues for ARG classes with well studied active sites, in particular class A to D β-lactamases, to verify that functionally important residues were captured while minimising inclusion of less conserved regions. These thresholds were then applied to the remaining ARG classes. Table 2lDDT threshold used to extract residues from ARP structures in MSTA to create high-lDDT ARP structuresNumber of structures in MSTAlDDT thresholdExample notation used for high-lDDT structure0–25Upper 50th percentile (median)AAC(6’)laQ50_lddt > 25–250Upper 75th percentiledfrA17Q75_lddt > 250Upper 90th percentileTEM-1Q90_lddt lDDT threshold used to extract residues from ARP structures in MSTA to create high-lDDT ARP structures Scaling the lDDT threshold with the number of structures in the MSTA (cluster size) represents a deliberate trade-off between capturing functionally important residues and retaining sufficient alignment length of the high-lDDT region, and ensures that pairwise seqID vs TM-score distributions reflect functionally relevant structural conservation rather than global structural similarity. This also extracted structurally and functionally conserved residues within the context of the particular MSTA, as opposed to applying a pre-defined lDDT threshold across all MSTAs, which fluctuate in numbers of ARP structures and structural variability. We performed further validation of these metrics using an existing protein structure-based method called PCM, with available positive and negative functional synthesis results of potential ARGs from various classes, as well as ARGs identified from functional metagenomics (see below). After residue extraction, high-lDDT ARP structures were created, represented as new.pdb structure files for each ARP structure. The Foldseek createdb command was used to create databases of high-lDDT ARP structures within each cluster for downstream creation of one-class SVMs. Training The training data represents pairwise (all-versus-all) primary and tertiary protein structure comparisons for high-lDDT ARP structures within 20%, 30%, and 50% seqID ARP structure clusters. In other words, an x,y distribution of seqIDs (x) vs TM-scores (y) of high-lDDT ARP structures. This was computed using Foldseek TM-align on Foldseek databases created in the previous section with parameters –alignment-type 1 and –exhaustive-search specified, and TM-scores normalised by the subject length. SeqID vs TM-score distributions were scaled using scikit-learn’s ‘StandardScaler’ [67], which standardises each feature independently by removing its mean and scaling to unit variance. The raw training data is available within Foldseek TM-align outputs from high-lDDT ARP structure databases available from https://zenodo.org/records/21209970. Parts of this workflow development were assisted using large language models (Anthropic), with outputs reviewed by authors. One-class SVM hyperparameter selection One-class SVMs are semi-supervised machine learning algorithms commonly used for outlier detection [81] and were deemed suitable for this classification task due to their effectiveness in scenarios where negative labels are unavailable or unreliable. The one-class SVM algorithm learns a decision function that defines a boundary in a transformed feature space around the training data. This boundary is called the decision boundary (DB). New data points are classified as inliers or outliers based on their relative position to the DB [77]. The radial basis function (RBF) kernel was selected as it produces smooth DBs suitable for continuous two-dimensional feature spaces, and is the standard kernel for one-class SVM applications. For two-dimensional x,y distributions, the shape and tightness of the DB are controlled by the hyperparameters ν(nu) and γ(gamma), and values for these parameters were chosen algorithmically rather than by manual tuning. The hyperparameter ν∈(0,1] sets an upper bound on the fraction of outliers and a lower bound on the number of support vectors. Here, the training data represents a pairwise distribution of seqIDs (x) vs TM-scores (y) for high-lDDT ARP structures within clusters. One-class SVMs were trained on each distribution using scikit-learn [67], and ν was selected as 1/N, where N is the number of training points and assumes that all training data are inliers. Assuming all training data are inliers was deemed appropriate to avoid underfitting the model given our hypothesis that functional ARPs have a different pairwise seqID vs TM-score distribution in conserved structural regions than non-functional proteins. Furthermore, ensuring the DB surrounds training points was important to reduce model bias from over-representation of variants with low variability within clusters (e.g., TEM β-lactamase variants). The kernel hyperparameter, γ, which defines the influence radius of a single data point in the RBF kernel, was selected automatically using scikit-learn’s ‘scale’ option, which sets γ=1/(number of features × data variance), and provides a standard balance between overfitting and underfitting. Query antibiotic-resistant protein (qARP) structures for functional prediction qARP structures from HMP reference strains Protein coding sequences (CDS) were predicted from nucleotide sequences of whole genome sequences (WGS) originating from six Gram-negative HMP reference strains [89] isolated from the human gut (Additional file 1: Table S1) by Prodigal v2.6.3 using default settings [33]. We selected these six strains to capture diversity across Gamma- and Betaproteobacteria while focusing on Gram-negative species, which increase the likelihood of successful expression in E. coli. The selection encompasses both fermentative and non-fermentative bacteria, and includes species with clinically relevant pathogenic strains, such as E. coli, K. pneumoniae, and P. aeruginosa, of which the latter two are represented among ESKAPE pathogens [52]. CDS were then subjected to relaxed homology alignment using Diamond v2.1.16 [13] against a database we created of ARG sequences obtained from CARD v4.0.1 [5]. ARG sequences in this database include those structures used to create one-class SVM models, however also ARG sequences from redundancy/quality filtered structures. The following arguments specified were used for this Diamond screen: -e 10 --sensitive --query-cover 70 --subject-cover 70. Results for this screening against the HMP reference strains are presented in Additional file 1: Table S3. Protein structures of query ARGs not considered known ARGs (with seqID < 90%) were either downloaded from the AFDB or predicted by AF2/AF3 in the same manner as subject ARP structures, representing qARP structures for functional prediction by ARG-PASS. AlphaFold/UniProt accessions or computed AF2/AF3 structures of qARPs predicted functional and experimentally verified are listed in Additional file 1: Table S2. Potential MBLs downloaded from the AFDB To identify potential metallo-β-lactamases (MBLs) with low homology to known MBLs and skip the initial Diamond screening against the custom CARD database, we obtained UniProt accessions from the UniProtKB database (release 2025_02) [90] with the following filter: (taxonomy_id:562) AND (protein_name:Metallo-beta-lactamase). qARP structures were then downloaded directly from the AFDB. qARP structures with > 20% seqID to known MBLs and not 70–130% of the subject ARP structure length were discarded. The remaining qARPs were deduplicated by clustering at 99% seqID over 90% length (of both query and subject) using Foldseek, with the longest representative used as a qARP structure for functional prediction. Cluster assignation Before passing into one-class SVM models for functional prediction, qARP structures were aligned to all ARP structures of the ARG class, with Foldseek under default settings. This was performed so qARP structures could be assigned to 20%, 30%, or 50% ARP structure clusters (cluster levels) for a particular ARG class, based on their maximum seqID to an ARP structure(s) (Table 3). An approach which scaled with the homology of the qARP to known ARGs was deemed necessary to maintain algorithmic precision of functional predictions of qARPs at a range of seqIDs. Table 3Assignation of qARP structures to high-lDDT ARP structure cluster levelsBest seqID with ARP structure(s) (%)Assigned to cluster level (seqID)0–2520% > 25–5030% > 5050% Assignation of qARP structures to high-lDDT ARP structure cluster levels As an exception, if the ARP structure with the highest seqID to the qARP did not belong to a valid cluster (≥ 3 members) at the assigned cluster level (Table 3), the qARP was dropped to the next lower cluster level. This was repeated until a valid cluster containing that ARP structure was found. In this case, the qARP was compared only against the specific cluster containing that ARP structure, not all clusters at the lower level. For example, qARP Aph01 had a maximum seqID of 52.4% to APH(6)-Ic, assigning it to 50% clusters. APH(6)-Ic was a singleton at this level, so Aph01 was dropped to 30% clusters. Again, APH(6)-Ic did not form a valid cluster at 30% seqID, so Aph01 was dropped to the 20% cluster. At this level, APH(6)-Ic belonged to a valid cluster, and Aph01 was predicted by ARG-PASS using only this cluster. The number of qARP structures assigned to each seqID cluster level are presented in Additional file 1: Table S9, and qARP structures are available from high-lDDT ARP structure databases at https://zenodo.org/records/21209970. Clustering of class B3 MBLs omitted for functional prediction of potential MBLs downloaded from the AFDB No qARP structures downloaded directly from the AFDB with less than 20% seqID to known class B3 MBLs were predicted functional by ARG-PASS using standard clustering. We therefore conducted an additional analysis in which prior 20% seqID clustering was omitted specifically for class B3 MBLs, instead creating a single MSTA and corresponding high-lDDT ARP structures across all class B3 MBL structures. qARP structures downloaded directly from the AFDB were then compared against both these high-lDDT class B3 MBL structures and the 20% seqID cluster of high-lDDT class B1/B2 MBLs. We chose not to integrate the omission of prior 20% seqID clustering for divergent potential MBLs into the ARG-PASS pipeline to retain the high precision of the algorithm. Nonetheless, this step demonstrates how reducing the clustering thresholds can relax ARG-PASS to identify more distant homologues, with the consideration that false positives may increase. Functional prediction Each qARP structure was aligned using Foldseek to databases of clustered high-lDDT ARP structures, to obtain an x,y distribution of seqIDs (x) vs TM-scores (y) for the qARP structure (also scaled using scikit-learn’s ‘StandardScaler’). As with the training data, this was computed using Foldseek’s TM-align mode, specified with –alignment-type 1, which uses the Needleman-Wunsch (NW) algorithm [58] with affine gap costs. The NW algorithm was crucial for appropriate global alignment of shorter high-lDDT ARP structures to full-length qARP structures, in contrast to Foldseek’s default alignment mode which uses the Smith-Waterman algorithm and is more suited to local alignments [82]. Additionally, Foldseeks’ --exhaustive-search was specified and the high-lDDT ARP structure (subject) length was used to normalise TM-scores and calculate seqIDs. Let f(x) be the decision function learned by the one-class SVM from the training data, where x represents a 2D feature vector (i.e., seqID and TM-score). For a given qARP structure constituting multiple points {xqARP(1),xqARP(2),...,xqARP(n)}, the qARP is classified as functional if any of its points lie on or within the DB: That is, the qARP is considered functional if at least one of its associated feature points has a decision score equal to or greater than zero, indicating it lies within the DB learned from the training data. Otherwise, the qARP is classified as non-functional. This criterion reflects the default decision threshold of the one-class SVM and assumes that any point outside the DB (i.e., f(xqARP(i))<0) is an outlier with respect to the training distribution. Manual adjustment of Foldseek seqID calculations Due to the alignment of shorter high-lDDT ARP structures against full-length qARP structures using Foldseek TM-align, seqID calculations were underestimated due to alignment gaps not being properly accounted for in the aligned region of the qARP structure. Manual adjustment of seqIDs was required by incorporating an alignment length of the high-lDDT ARP structure (subject length). Validation of ARG-PASS on an existing protein structure-based method of predicting ARGs Validation of ARG-PASS was performed on proteins previously experimentally verified by an existing computational method also based on protein structure used to identify ARGs, called pairwise comparative modelling (PCM) [75]. PCM predicted ARGs from a 3.9 million gene catalogue of the human intestinal microbiota. Referring to Supplementary Table 1 of the PCM paper, we collected protein sequences from the project website (https://mgps.eu/Mustard/index.php?id=accueil), and focussed on experimentally tested candidates with less than 95% seqID to known ARGs (n = 59). Protein structures encoded by sequences were downloaded directly from the AFDB or predicted by AF2/AF3, confirmed to be absent from the training distribution, and functionally predicted by ARG-PASS. Computational predictions and the functional synthesis results from the PCM paper are presented in Additional file 1: Table S5 and summary statistics of precision, specificity, and sensitivity for each ARG class are presented in Additional file 1: Table S6. Structure files of PCM candidates are available at https://zenodo.org/records/21209970. Validation of ARG-PASS on a functional metagenomics dataset The sensitivity of ARG-PASS was evaluated using ResFinderFG v2.0, a curated database of ARGs identified by functional metagenomics from environmental and animal-associated microbiomes [30]. Protein sequences encoded by ARGs were downloaded from the associated GitHub repository (https://github.com/RemiGSC/ResFinder_FG_Construction) and deduplicated by clustering at 95% seqID using CD-HIT [43]. Sequences were then searched against the custom CARD database used to create ARG-PASS one-class SVM models using Diamond (--max-target-seqs 1 -e 10 --ultra-sensitive --query-cover 70 --subject-cover 70), and those with > 95% seqID to known ARGs were excluded as these are represented in the training data, which resulted in 1153 protein sequences. To avoid class imbalance, up to 30 sequences were sampled per ARG class included in the ARG-PASS method and available from the 1153 protein sequences. For classes with fewer than 30 available sequences, all were retained: AAC(3') (n = 21), AAC(6’) (n = 28), ANT (n = 4), APH (n = 9), class B1/B2 β-lactamases (n = 8), PBP (n = 4), and TRPP (n = 15). For classes with more than 30 sequences, six sequences were randomly sampled from each of five seqID bins (0–35%, 35–50%, 50–65%, 65–80%, 80–95%) to ensure representation at different seqIDs. If a bin contained fewer than six sequences, no additional sequences were sampled from other bins to avoid over-representing particular seqID ranges. The overall sampling yielded a final dataset of 204 protein sequences. Protein structures encoded by protein sequences were predicted using the AF3 web server and analysed by ARG-PASS. Sensitivity results per ARG class across stratified seqID bins are presented in Additional file 1: Table S7 and structure files of ResfinderFG v2.0 candidates are available at https://zenodo.org/records/21209970. Signatures of mobility near functional qARPs To infer if functionally verified qARPs from HMP reference strains co-localised with MGEs, MGE-associated sequences (transposon, plasmid, phage, and integron-associated sequences) were queried 10 ORFs up and downstream of qARPs by alignment against the MobileOG-db [12] using Diamond with a minimum seqID of 80% (over 90% query and subject length). In addition, the Galaxy server of PlasmidHunter [91] was used to infer if contigs containing qARPs had a chromosomal or plasmid origin. Experimental verification qARPs from HMP reference strains were selected for experimental verification to ensure representation across all ARG classes predicted functional by ARG-PASS. All qARPs were tested from ARG classes with a low amount of functional predictions (i.e., ≤ 2). For ARG classes with higher numbers of functional predictions, the majority of Dfr candidates were selected; however, only one PBP candidate was selected due to the practical constraints of assembling and cloning longer membrane-bound proteins. Additionally, five qARPs predicted non-functional by ARG-PASS (AAC, class B1 β-lactamase, and Sul) were selected to assess specificity. The Minimal Antibiotic Resistance Platform includes vectors grouped by a kanamycin (pGDP1 and pGDP2) or ampicillin (pGDP3 and pGDP4) selective marker, and by promoter strength, with the stronger β-lactamase bla promoter (pGDP1 and pGDP3) and the weaker lac promoter (pGDP2 and pGDP4) [18]. DNA sequences of selected qARPs, positive controls not already available in the Minimal Antibiotic Resistance Platform (i.e., dfrK (NCBI accession CBL80435.1) and sul3 (NCBI accession ACJ63260.1)), and control folA from E.coli MG1655 (NCBI accession NP_414590.1), were codon optimised for E. coli and synthesised by IDT. Note that no positive control was included for PBPs, as multiple attempts were not successful in assembling a β-lactam-resistant PBP3 [102] into pGDP1 or pGDP2. With the exception of bla04, we opted to assemble potential qARP β-lactamases into the pGDP2 series with the weaker lac promoter to limit potential protein misfolding arising from expression by the stronger bla promoter in pGDP1. The remaining qARPs from other classes were assembled into the pGDP3 vector. Synthesised gblocks were assembled into pGDP vectors using the NEBuilder® HiFi DNA Assembly (New England Biolabs), transformed into E. coli MM294 by the chemical transformation method, and grown in Lysogeny Broth (LB) for 1 h at 37 °C with agitation. Colonies were grown overnight (o/n) on LB and agar plates containing 100 µg/mL of ampicillin or 50 µg/mL of kanamycin (selection). Transformants were recovered and grown o/n in LB with selection. Plasmids were then extracted using the Monarch Plasmid DNA miniprep kit (New England Biolabs) and sequenced on a minION Flongle Flowcell (Oxford Nanopore Technologies) to confirm an error-free assembly. Confirmed plasmids were then re-transformed via electroporation into E. coli BW25113 (ΔbamBΔtolC), with the same subsequent steps as the chemical transformation method and stored at −80 °C in glycerol. The Minimal Antibiotic Resistance Platform was a generous gift from Prof. Gerard Wright of McMaster University (also available as an Addgene kit #1000000143). MIC measurement Broth microdilution (BM) Determination of MICs by broth microdilution (BM) was measured in accordance with CLSI M07 [17]. Cultures were grown overnight in MHII broth for 18 h at 37 °C, serially diluted, and applied in duplicate to microdilution plates containing MHII broth and appropriate antibiotics ranging around CLSI resistant breakpoints [16]. An approximate final concentration of 5 × 104 colony forming units (CFU) was present in each well. Following incubation for 18 h at 37 °C, MICs were recorded as the antibiotic concentration that no visible growth was present, which was also confirmed by measuring the optical density at 600 nm (OD600) of the well compared to wells without inoculation. All MICs were measured alongside previously transformed E. coli BW25113 (ΔbamBΔtolC) strains containing an empty pGDP vector (negative control), and a pGDP vector containing a known ARG in the respective class of the qARP (positive control). A qARP was considered a functional ARG if its MIC was at least two-fold higher than that of the negative control in both duplicates, in accordance with CLSI guidelines. Spot plating assay (SPA) The spot plating assay (SPA) was developed in-house in an attempt to increase through-put and automation of MIC measurements. Briefly, 2 µL of culture was collected by a Rotor HDA (Singer Instruments) from each well of a 96-well plate containing a strain for MIC testing, previously grown for 18 h at 37 °C in 200 µL LB, without selection. The 2 µL of culture was then spotted onto PlusPlates (Singer Instruments) containing MHII agar or MHII agarose (for TMP/SMX and TMP testing) and concentrations of appropriate antibiotics at CLSI breakpoints, in the same relative position as they were collected from the 96-well plate. After 18 h at 37 °C, plates were visually inspected, with an MIC recorded as the concentration for which no visible growth was present. A qARP was considered a functional ARG if its MIC was at least two-fold higher than that of the negative control. Phylogenetics To investigate evolutionary relationships between functionally confirmed qARPs in this study and ARPs encoded by closely related ARGs, FoldMason under default settings was used to create an MSTA. Maximum likelihood was applied as a search criterion in IQ-TREE v.3.0.0 using the best-fit model automatically selected by ModelFinder [59]. An Ultrafast Bootstrap analysis was performed with 1000 ML replicates and an approximate likelihood ratio test search with 10,000 replicates. Phylogenetic trees were visualised and edited with FigTree v.1.4.4 (available http://tree.bio.ed.ac.uk/software/figtree/). Supplementary Information Additional file 1: Supplementary Materials. Additional file 2. Additional file 3. Additional file 4. Acknowledgements The authors would like to thank Professor Jean-Philippe Côté and Lya Blais (Department of Biology, Université de Sherbrooke, QC, Canada) for MIC testing using the Rotor HDA. This research was enabled in part by computational support provided by Calcul Québec (calculquebec.ca) and the Digital Research Alliance of Canada (alliance​can​.ca). Authors’ contributions LB and LPH conceived and designed the study. LB and EBL performed PCR reactions, cloning assemblies, transformations, and functional verification of candidate genes. FG performed Nanopore sequencing and provided experimental guidance. LPH and SR contributed to conceptual analysis and provided experimental guidance. LB designed the computational method, performed bioinformatic analyses, and drafted the manuscript. All authors reviewed the final manuscript. Funding This work was funded by grants from Humans & the Microbiome Program Catalyst (CIFAR), Génome Québec, and Fonds de Recherche du Québec—Santé (grant: 349522). Data availability ARP and high-lDDT ARP structure databases, qARP structures analysed in this study, and reproducibility pipelines to use ARG-PASS are available for download at https://zenodo.org/records/21209970. An overview of the ARG-PASS method and code generated is also available at https://github.com/btroppo/ARG-PASS. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Sébastien Rodrigue, Email: sebastien.rodrigue@usherbrooke.ca. Louis-Patrick Haraoui, Email: louis.patrick.haraoui@usherbrooke.ca. References Associated Data Supplementary Materials Additional file 1: Supplementary Materials. Additional file 2. Additional file 3. Additional file 4. Data Availability Statement ARP and high-lDDT ARP structure databases, qARP structures analysed in this study, and reproducibility pipelines to use ARG-PASS are available for download at https://zenodo.org/records/21209970. An overview of the ARG-PASS method and code generated is also available at https://github.com/btroppo/ARG-PASS.