MOFormer: Multi-Objective Antimicrobial Peptide Generation through Conditional Transformer with Multi-Fusion Descriptors

Abstract Deep learning offers considerable potential for designing peptides with desirable properties, essential for expediting drug discovery. Despite recent advances in generating antimicrobial peptides (AMPs), the simultaneous production of AMPs with multiple desired properties remains a formidable challenge. We introduce MOFormer, a multi-objective AMP design pipeline that optimizes several AMP properties concurrently. This model leverages a condition Transformer to refine the AMP sequence-property landscape for multi-objective generation, while multi-fusion descriptors improve the capture of intricate sequence details. Regularization techniques enable the model to uphold a highly structured latent space, facilitating the generation of more effective and potent new AMPs. Comparative analyses highlight that MOFormer outperforms existing methods in the dual-objective generation tasks of enhancing antimicrobial activity and minimizing hemolysis. Furthermore, we apply a Pareto-based non-dominated sorting algorithm and deep-learning surrogate models for hierarchical candidate screening. MOFormer's high efficacy is verified through two aspects: (1) elucidating candidates' structure and functions using molecular simulations and interpretable analysis, and (2) improving known AMPs towards additional dual-objective and triple-objective scenarios.