Abstract
Antimicrobial resistance poses a major global threat, driven by diminishing efficacy of current treatments and limited new therapies. Combination therapy with existing drugs offers a promising solution, yet current empirical screening methods are expensive and often lead to suboptimal efficacy and inadvertent toxicity. We introduce CALMA, a computational framework that quantitatively analyzes the potency-toxicity landscape of multi-drug combinations. Integrating genome-scale metabolic modeling with a neural network that reflects metabolic subsystems, CALMA enhances interpretability and prioritizes pathways influencing drug interactions. The incorporation of metabolic architecture in the neural network leads to over 92% reduction in model parameters, enabling it to learn generalizable mechanistic signals and reducing the experimental search space of optimal combinations by 97%. CALMA identified promising antimicrobial combinations against Escherichia coli and Mycobacterium tuberculosis that were antagonistic for kidney and liver toxicity and uncovered the nucleotide salvage pathway as a selective influencer of toxicity, which was validated in vitro. Mining of health records of over 400,000 patients showed reduced frequency of kidney side-effects in patients taking a vancomycin combination identified by CALMA. CALMA provides a rational, mechanistic approach to streamline combination treatment design.