De novo design and experimental characterization of bitter peptides.

Abstract

Bitter taste is a critical quality determinant in food systems, particularly those using sustainable protein hydrolysates, where the unpredictable formation of bitter peptides severely limits consumer acceptance. Achieving predictive control over flavor chemistry requires deciphering the complex sequence-activity relationship. To address this, we integrated the generative capacity of a protein language model with BitterPep-GCN, a Graph Convolutional Network (GCN) capable of robust in silico bitter/non-bitter classification, to target the de novo design of functional bitter and non-bitter sequences. We achieved this by generating two strategic peptide libraries: a targeted tripeptide library derived from known bitter and non-bitter peptide sequences, and a set of de novo designed sequences. For the de novo designed peptides, we fine-tuned the conditional language model ZymCTRL on our curated dataset of sensory-validated bitter peptides (BPS-1000). Both libraries were subjected to classification and rigorous filtering using BitterPep-GCN to select high-confidence candidates for validation. The selected peptides were purchased and rigorously assessed for high purity. Sensory tests were conducted by an expert human panel to determine intrinsic taste quality and taste recognition thresholds. The results validated the high predictive fidelity of our pipeline: out of the 31 tested peptides, 25 were correctly classified, including 15 confirmed bitter and 10 confirmed non-bitter sequences. This study successfully demonstrates the application of machine learning frameworks in the design of bioactive peptides. It provides a set of novel taste-active peptides that can be used to accelerate the rational mitigation of off-tastes in next-generation food products.