Atomic Watchtower: A Multi-Modal Machine Learning Framework for Swift Antimicrobial Resistance Prediction Using Whole Genome Sequencing and SMILES-Based Molecular Mapping

Abstract

AMR (antimicrobial resistance), estimated to become deadlier than cancer, could cause up to 10 million deaths a year by 2050 if left unchecked. It is one of the biggest global health threats of the 21st century. Standard susceptibility tests usually take 48-72 hours which creates a diagnostic gap where empirical, broad-spectrum antibiotics are often used. In this paper, we present the Atomic Watchtower, a multi-modal machine learning framework that incorporates Whole Genome Sequencing (WGS) k-mer features alongside drug-structure data encoded as SMILES. Our framework poses the resistance prediction task as a structural-matching problem between pathogen genotype and drug chemistry. The Minimal Inhibitory Concentration (MIC) value is predicted by a Random Forest Regressor. Moreover, drug-target binding behaviour is predicted by a DeepPurpose-based interaction model. The two-tier architecture is used to predict anti-SARS-CoV-2 drug candidates. In a review cohort of 491 Indian clinical isolates from the Indian_AMR_Multimodal_Dataset, the framework achieved a median AUROC of 0.90 for the AV-Susceptible antibiotics classes and 0.99 classification accuracy for ciprofloxacin. A federated-learning simulation using the Flower framework is presented in this study as a proof of concept for cross-institutional training that can preserve privacy. The findings are from retrospective, single-dataset, simulation-based evaluation and should be interpreted as preliminary; prospective, multi-centre validation are needed for clinical use. A prototype Streamlit dashboard demonstrates how the outputs of the framework could be displayed to clinicians.