Antibiotic resistance is a growing global health threat, with carbapenem-resistant Escherichia coli representing a critical clinical challenge. Understanding the genomic basis is classically addressed through bacterial GWAS, which identifies genomic variants statistically associated with a phenotype. However, GWAS requires large labeled datasets, struggles with bacterial genome complexity, and produces associations rather than generalizable predictive models.
In this study, we propose a few-shot learning strategy using Bacformer [1], a genomic language model generating contextualized protein embeddings, combined with a Gated Attention-based Multiple Instance Learning (MIL) framework [2], to predict meropenem resistance in E. coli from 122 whole-genome assemblies (61 resistant, 61 susceptible) collected from the BV-BRC database [3]. Crucially, the gated attention mechanism of the MIL framework provides interpretability by highlighting resistance-associated proteins, allowing the predictions to be directly compared against known GWAS findings on antibiotic resistance.
Scan the QR code to access the full poster and additional details.
[1] Wiatrak M et al. bioRxiv. 2025. doi:10.1101/2025.07.20.665723.
[2] Zhao L, Yuan L, Hao K, Wen X. Multimed Syst. 2023;29(1):275–287. doi:10.1007/s00530-022-00992-w.
[3] Olson RD et al. Nucleic Acids Res. 2023;51(D1):D678–D689. doi:10.1093/nar/gkac1003.