Harnessing Interpretable Deep Learning to Predict Resistance in Klebsiella pneumoniae

Abstract

Antimicrobial resistance constitutes an escalating global health threat, complicating therapeutic management and increasing morbidity and mortality. Deep learning approaches have emerged as promising tools for bacterial profiling based on omics data, particularly for predicting antimicrobial susceptibility from genomic information. This task relies on identifying genomic signatures associated with resistance mechanisms. Here, DeepMDC is introduced as a deep learning architecture designed for bacterial profiling using whole-genome data. Given that precise annotation at the gene or mutation level is often costly and ambiguous, phenotypic classification is formulated as a Multiple Instance Learning (MIL) problem, in which each genome is represented as a bag of instances with a single associated label. The core of DeepMDC is a Modern Hopfield Network that processes all open reading frames (ORFs), including small ones, derived from genomic data. A key feature of the architecture is its interpretability, enabled by attention mechanisms that facilitate biological insight and hypothesis generation. The model was evaluated against Klebsiella pneumoniae and four clinically relevant antibiotics (meropenem, cefepime, ceftazidime, and gentamicin), achieving strong performance in several metrics. Notably, genes associated with resistance consistently received high attention scores during inference, which validates the architecture and eventually may generate new hypotheses.