Multi-Objective Peptide Design via Token-Aligned Preference Optimization

Protein language models have recently shown promise for de novo protein and peptide design, but they lack mechanisms for controllable optimization of functional properties. This limitation is particularly critical in therapeutic peptide discovery, where candidates must simultaneously satisfy multiple, often conflicting, biochemical constraints. We present a token-aligned preference optimization framework that adapts a pretrained protein language model using pairwise sequence preferences conditioned on property-specific control tokens. By learning from comparative feedback rather than scalar rewards, our approach enables multi-objective control and shows generalization to novel property combinations. As a case study, we apply the method to antimicrobial peptide (AMP) design, a clinically relevant but challenging testbed. Our approach achieves substantial improvements in jointly satisfying biochemical constraints, demonstrating the potential of preference alignment for controllable peptide design.