AI model predicts and designs bitter peptide sequences for food research

Researchers from the Leibniz Institute for Food Systems Biology at the Technical University of Munich have developed an AI-driven method to predict and design peptide bitterness. By combining a protein language model with a graph convolutional network, the system can identify structural features that cause off-tastes in protein-based products.

In a blind test involving a sensory panel, the AI correctly predicted the bitterness of 25 out of 31 de-novo constructed sequences, achieving an accuracy rate of approximately 80 percent. Beyond mere classification, the system is capable of generating entirely new peptide sequences designed to exhibit specific bitter profiles.

This technology aims to improve the consumer acceptance of plant-based protein alternatives and fermented foods by making bitterness more measurable and controllable during the molecular modeling stage, rather than relying on post-production recipe adjustments.