Multi-source language models fused with a dynamic routing capsule network for antimicrobial compound identification

Antibiotics are a vital class of drugs closely associated with the prevention and treatment of bacterial infections. Accurate prediction of molecular antimicrobial activity remains a key challenge in the pursuit of novel antibiotic candidates. However, laboratory-based antimicrobial compounds identification is costly, time-consuming, and prone to rediscovering known antibiotics, highlighting the urgent need for efficient and accurate computational models. Recent advances in deep learning have significantly enhanced the ability to explore chemical space and identify potential antimicrobial compounds. In this study, we propose CapMolPred, which integrates five domain-adapted chemistry language models for high-dimensional small-molecule encoding and adopts cross-attention for heterogeneous embedding alignment. It also combines a capsule network with dynamic routing and a novel asymmetric loss function to achieve superior inhibitory potency prediction against Escherichia coli, Acinetobacter baumannii, and Staphylococcus aureus compared to conventional methods. We conducted a series of ablation studies to elucidate the contributions of model design. Case studies validated the usability and effectiveness of our model and provided insights into its decision-making process, highlighting its interpretability. To facilitate accessibility, we developed an intuitive web portal (https://dmci.xmu.edu.cn/CapMolPred/indexpage.php) to disseminate this tool. Our results indicate that the approach underscores the potential of interpretable artificial intelligence methods in accelerating antibiotic discovery. Accurate prediction of molecular antimicrobial activity remains a key challenge in the pursuit of novel antibiotic candidates. Here, the authors introduce CapMolPred, a deep learning model integrating chemistry language models and a capsule network, demonstrating superior prediction of inhibitory potency against key bacteria and highlighting the potential of AI in accelerating antibiotic discovery.