Abstract
Peptide therapeutics, a major class of medicines, have achieved remarkablesuccess across diseases such as diabetes and cancer, with landmark examplessuch as GLP-1 receptor agonists revolutionizing the treatment of type-2diabetes and obesity. Despite their success, designing peptides that satisfymultiple conflicting objectives, such as target binding affinity, solubility,and membrane permeability, remains a major challenge. Classical drugdevelopment and structure-based design are ineffective for such tasks, as theyfail to optimize global functional properties critical for therapeuticefficacy. Existing generative frameworks are largely limited to continuousspaces, unconditioned outputs, or single-objective guidance, making themunsuitable for discrete sequence optimization across multiple properties. Toaddress this, we present PepTune, a multi-objective discrete diffusion modelfor the simultaneous generation and optimization of therapeutic peptide SMILES.Built on the Masked Discrete Language Model (MDLM) framework, PepTune ensuresvalid peptide structures with state-dependent masking schedules andpenalty-based objectives. To guide the diffusion process, we propose a MonteCarlo Tree Search (MCTS)-based strategy that balances exploration andexploitation to iteratively refine Pareto-optimal sequences. MCTS integratesclassifier-based rewards with search-tree expansion, overcoming gradientestimation challenges and data sparsity inherent to discrete spaces. UsingPepTune, we generate diverse, chemically-modified peptides optimized formultiple therapeutic properties, including target binding affinity, membranepermeability, solubility, hemolysis, and non-fouling characteristics on variousdisease-relevant targets. In total, our results demonstrate that MCTS-guideddiscrete diffusion is a powerful and modular approach for multi-objectivesequence design in discrete state spaces.