Intelligent De novo Design of Multifunctional Taste Peptides via an integrated deep learning, conditional diffusion, and reinforcement learning platform.

Taste peptides from dietary proteins elicit taste sensations and have reported bioactivities, but their rational prioritization remains hampered by the limited throughput of conventional screening. Here, we describe a multi-stage computational workflow for de novo prioritization of multifunctional taste peptides, combining multi-task benchmarking, conditional diffusion sequence generation, feature-space novelty filtering, and PPO-based sequence optimization. A multi-label dataset of 4801 experimentally reported peptides was assembled from BIOPEP-UWM, MBPDB, FermFooDb, and published taste-peptide benchmark sets. The best-performing classifiers achieved area under the receiver operating characteristic curve (AUC-ROC) values of 0.923 (umami), 0.872 (bitter), 0.828 (ACE-inhibitory), and 0.807 (antioxidant), with bootstrap 95% confidence intervals estimated for each task. Conditional diffusion yielded 9401 unique candidates after validity filtering and deduplication. Novelty-aware filtering and an independently optimized PPO policy were compared using the same composite reward. The novelty score used for ranking is a Euclidean feature-space distance rather than a flow-matching likelihood. All outputs are presented as computational priorities for synthesis and experimental validation, not as measured activities.