AMR-Fold

Overview This record contains the AMR-Fold protein sequence dataset, a curated collection of antibiotic resistance genes (ARGs) and non-ARG bacterial proteins designed for training and benchmarking machine-learning models for antibiotic resistance prediction. The dataset underpins the AMR-Fold architecture, which combines protein language model (PLM) embeddings and inferred structural tokens (3Di + confidence) with a Transformer encoder and attention regularisation. In total, the dataset comprises 54,044 unique protein sequences with per-sequence labels for: Binary ARG vs non-ARG (bin ∈ {0, 1}) Antibiotic class / mechanism, including multiple ARG classes (e.g. β-lactam, tetracycline, aminoglycoside, etc.) and a dedicated non_ARG class for negatives. Protein lengths range from 31 to 2,890 amino acids (mean ≈ 351 aa). Data construction Positive (ARG) sequences Positives are derived from the MCT-ARG collection [1] (~27k proteins), which aggregates curated ARG sequences from multiple sources (e.g. Arg-annot, CARD, MEGARes, DeepARG, NCBI RefSeq ARG annotations). Sequences were unified, deduplicated, and assigned: An internal identifier Seq_x (e.g. Seq_1 … Seq_27022) An antibiotic class label (e.g. beta_lactam, tetracycline, etc.). 13 antibiotic classes + others (rare, e.g. tunicamycin, nucleoside, etc ) were considered. A binary label bin = 1 for ARG. Negative (non-ARG) sequences Negatives were sampled from UniProtKB/Swiss-Prot bacterial proteins. A multi-step filtering pipeline was applied to obtain high-confidence non-ARGs: Keyword filtering – sequences with annotations containing resistance-related cues (e.g. “resistance”, “beta-lactamase”, “multidrug efflux”, and related terms) were removed. Homology-based filtering – remaining sequences were queried against the ARG database using DIAMOND BLASTP [2], and any sequence with significant similarity to known ARGs (e.g. identity/coverage thresholds; stringent e-value cutoff) was excluded. Functional filtering – remaining candidates were screened with eggNOG-mapper; proteins with resistance-related functional annotations were excluded. Final leakage-control gate (cluster-level) – as a final safeguard, the combined ARG + candidate non-ARG set was clustered with CD-HIT [3] at 90% identity, and any candidate non-ARG sequence that co-clustered with an ARG sequence was removed. This guarantees that no non-ARG in the final dataset is ≥90% identical to any ARG sequence at the clustering threshold used for split construction. From the resulting high-confidence candidate pool, a subset was selected to yield 27,022 non-ARG sequences with bin = 0 and class = non_ARG. Label construction The antibiotic resistance class labels were derived from the original source metadata and subsequently harmonised into a unified classification scheme. The final AMR-Fold dataset contains 15 classes in total, including 14 ARG-related classes and one dedicated non_ARG category. The retained ARG classes were: aminoglycoside bacitracin beta_lactam fosfomycin glycopeptide LSAP-like multidrug others peptide phenicol quinolone rifamycin sulfonamide tetracycline To reduce label fragmentation and improve class consistency, several mechanistic or closely related subclasses were merged into broader parent categories: carbapenem and cephalosporin → beta_lactam fluoroquinolone → quinolone rifampin → rifamycin chloramphenicol → phenicol polymyxin → peptide The LSAP-like category was constructed as a pooled low-frequency resistance class comprising mainly lincosamide-, streptogramin-, and macrolide-related ARG subclasses, together with closely related rare subclasses that individually lacked sufficient representation for robust model training. Very rare or heterogeneous resistance categories were grouped into the others class. These included: aminocoumarin bleomycin elfamycin ethambutol fosmidomycin fusidic_acid isoniazid kasugamycin mupirocin nitrofurantoin nitroimidazole nucleoside oxazolidinone pleuromutilin puromycin pyrazinamide qa_compound streptothricin tetracenomycin thiostrepton triclosan trimethoprim tunicamycin and residual ambiguous or poorly defined labels Additionally, records whose annotations specifically indicated qacH were forcibly assigned to the others category due to their highly sparse representation and ambiguous resistance context. Final merged dataset Positives and negatives were combined into a single balanced dataset: 27,022 ARG sequences 27,022 non-ARG sequences 54,044 sequences in total (DB.faa.gz) All sequences were renamed to a canonical ID scheme (Seq_x), while preserving the original source header in a companion TSV file. Train/validation/test splits To avoid train–test leakage via very close homologs, splitting was performed at the cluster level: The full dataset (DB.faa.gz) was clustered with CD-HIT at 90% sequence identity to define sequence families (clusters). Clusters were assigned to splits with an 80/10/10 proportion: Training: DB_train.* Validation: DB_val.* Test: DB_test.* All sequences from a given CD-HIT cluster reside entirely in one split, so no cluster is shared across train/val/test (no leakage across partitions). Stratified assignment and rare-stratum policy: Cluster assignment was stratified to maintain similar label proportions across splits. In addition to stratification by binary label (ARG vs non-ARG), a rare-stratum policy was applied to improve development-time representativeness: strata represented by only one cluster were assigned to training; strata represented by two clusters were assigned to training and validation; strata with three or more clusters were split approximately 80/10/10. This preserves strict cluster-level separation while avoiding validation blind spots for small families. All sequences from a given CD-HIT cluster reside entirely in one split, so no cluster is shared across train/val/test (no leakage across partitions). As a result, splits are approximately balanced between ARG and non-ARG labels in each partition, while maintaining strict cluster-level separation. Files and formats Typical contents of this record include: FASTA files DB.faa.gz – full dataset (54,044 proteins) with headers as Seq_x. DB_train.faa.gz, DB_val.faa.gz, DB_test.faa.gz – split FASTA files. Annotation tables (TSV) DB.tsv – master annotation for all sequences with columns: ids – original protein header (source database + functional annotation) Ids – canonical ID (Seq_x) class – antibiotic class label (e.g. beta_lactam, tetracycline, non_ARG, …) bin – binary label (1 = ARG, 0 = non_ARG) DB_train.tsv, DB_val.tsv, DB_test.tsv – annotations for each split (same columns as above). Features Feature matrices are provided in this record at features.tar.gz and follow the convention: {Seq_x}.plm.npy – ProstT5 [4] per-residue embeddings, shape (L, 1024). {Seq_x}.3di_tokens.npy – per-residue 3Di token indices [5], shape (L,). {Seq_x}.3di_conf.npy – per-residue confidence scores, shape (L,). Intended use and limitations This dataset is intended for: Training and evaluating sequence-based and structure-informed models for antibiotic resistance prediction. Benchmarking binary ARG vs non-ARG classification and multi-class antibiotic class assignment. Exploring attention mechanisms and interpretability (e.g. attention-based localisation of resistance determinants). The sequences are compiled and derived from publicly available resources (e.g. curated ARG databases and UniProtKB/Swiss-Prot). Users should consult and cite the original databases and publications where appropriate, in addition to citing this dataset record. The dataset is not exhaustive for all possible ARGs or bacterial proteins and should be used as a curated benchmark rather than a complete catalogue. REFERENCES [1] He, L., Li, H., Qi, R., Zou, Q., & Wang, Y. (2025). MCT-ARG: Identification and classification of antibiotic resistance genes based on a multi-channel Transformer model. Science of the Total Environment, 1006, 180848. https://doi.org/10.1016/j.scitotenv.2025.180848. GitHub: https://github.com/nanbei45/MCT-ARG/tree/master. [2] Buchfink, B., Xie, C., & Huson, D. H. (2015). Fast and sensitive protein alignment using DIAMOND. Nature Methods, 12(1), 59–60. https://doi.org/10.1038/nmeth.3176. GitHub: https://github.com/bbuchfink/diamond. [3] Li, W., & Godzik, A. (2006). Cd-hit: a fast program for clustering and comparing large sets of protein or nucleotide sequences. Bioinformatics, 22(13), 1658–1659. CD-HIT package and documentation: http://www.bioinformatics.org/cd-hit/. [4] Bernhofer, M., Heinzinger, M., Teufel, A. I., Reeb, J., & Rost, B. (2024). Bilingual language model for protein sequence and structure. NAR Genomics and Bioinformatics, 6(4), lqae150. https://doi.org/10.1093/nargab/lqae150. Model page: https://huggingface.co/Rostlab/ProstT5.. [5] van Kempen, M., Kim, S. S., Tumescheit, C., Mirdita, M., Söding, J., & Steinegger, M. (2023). Fast and accurate protein structure search with Foldseek. Nature Biotechnology, 41, 496–502. https://doi.org/10.1038/s41587-023-01773-0.. GitHub: https://github.com/steineggerlab/foldseek.