Machine Learning-Assisted Drug Repurposing for Antimicrobial Resistance Targets

Abstract

Antimicrobial resistance now kills more people each year than either HIV/AIDS or malaria, yet the antibiotic
development pipeline has thinned for decades because conventional discovery economics discourage investment
in short-course, curative therapies. Machine learning-assisted drug repurposing, which screens existing,
already-approved compounds against new microbial targets computationally before any wet-lab work begins, has
been proposed repeatedly as a faster and cheaper route to new antimicrobial leads, but the field's evidence base
is scattered across heterogeneous study designs, target sets and validation standards that make its real track
record difficult to assess. We conducted a systematic review of machine learning-assisted drug repurposing
studies targeting ESKAPE pathogens (Enterococcus faecium, Staphylococcus aureus, Klebsiella pneumoniae,
Acinetobacter baumannii, Pseudomonas aeruginosa and Enterobacter species) published between 2015 and
2025, screening 1,842 records and retaining 61 studies meeting predefined methodological inclusion criteria, and
complemented this review with our own graph neural network drug-target interaction screen applied to 2,156
approved, non-antibiotic compounds against 34 priority AMR drug targets curated from the included literature.
Across the reviewed studies, graph neural network and other deep-learning architectures increasingly
outperformed classical machine learning baselines (random forest, support vector machine) on held-out
drug-target interaction benchmarks, but only 18 of 61 studies (29.5%) reported any wet-lab validation of a top
computational hit, and of those, 11 (61%) reported an experimentally confirmed minimum inhibitory concentration
within a clinically plausible range. Our own screen nominated 14 high-confidence repurposing candidates, of
which we selected four for retrospective literature cross-checking; two (auranofin and niclosamide) had
independent prior experimental support against at least one ESKAPE target, partially validating the screen's face
validity without constituting new wet-lab evidence. We conclude that machine learning-assisted repurposing is a
genuinely productive hit-generation strategy but remains bottlenecked by inconsistent, often absent experimental
validation, and we propose a minimum reporting and validation standard to make future studies more comparable
and more useful to downstream experimentalists.