AI as a Tool in Flavor Research: De Novo Design of Bitter Peptides

Bitter-tasting peptides are produced during the enzymatic or chemical breakdown of proteins and pose a particular challenge in the production of fermented foods or protein hydrolysates. At the same time, they can possess physiological properties and, for example, contribute to the regulation of hunger and satiety.

“To make plant-based protein sources more attractive for food production and to use them more sustainably, we need to better understand which peptides taste bitter and what structural features characterize them. AI-based methods can also make an important contribution here,” says Prof. Dr. Antonella Di Pizio. She is Head of the Molecular Modeling research group at the Leibniz Institute for Food Systems Biology at the Technical University of Munich (Leibniz-LSB@TUM) and Professor of Chemoinformatics and Protein Modelling at the TUM School of Life Sciences. In addition to her research group at Leibniz-LSB@TUM, researchers from TUM and Pompeu Fabra University in Barcelona were also involved in the current study.

Combination of a Language Model and an Artificial Neural Network

To develop such an AI-supported bioinformatics method, the team led by Antonella Di Pizio combined a protein language model—which the team had previously trained using approximately 500 known bitter-tasting peptides—with the BitterPep-GCN prediction model they had recently developed. This is a so-called Graph Convolutional Network (GCN), a specialized form of artificial neural networks used to analyze structured data.

Based on this, the researchers first generated 161 new peptide sequences that had not yet been experimentally characterized. They then identified those candidates that, according to the model predictions, were highly likely to taste bitter or non-bitter.

Predictions Largely Confirmed

They had the most promising of these peptides synthesized and then tasted by a trained sensory panel. In most cases, the AI predictions were confirmed: Of the 31 peptides tested, the test subjects correctly classified 25 as bitter or non-bitter. In addition, the research team identified numerous previously unknown bitter and non-bitter-tasting peptides.

“Our results show that not only can the bitterness of peptides be predicted, but that our new AI-based method can also be used to specifically design new bitter-tasting peptides,” says Alexandra Steuer, first author of the study and a doctoral student in Antonella Di Pizio’s Molecular Modeling research group. “This brings us significantly closer to the goal of proactively controlling taste characteristics,” adds Antonella Di Pizio.

The scientist emphasizes, that the research is ready to be implemented in application frameworks: “In the long term, these new findings could help to specifically control the formation of bitter-tasting peptides during food production. This would be particularly relevant for plant-based, protein-rich foods, whose acceptance often suffers due to undesirable flavor notes.”

Publication:

Steuer A, Ferri F, Eckrich L, Heidenkampf J, Mittermeier-Kleßinger VK, Schaefer S, Behrens M, Ferruz N, Dawid C, Di Pizio A. De novo design and experimental characterization of bitter peptides. NPJ Sci Food. June 25, 2026;10(1):200. doi: 10.1038/s41538-026-00942-0. https://doi.org/10.1038/s41538-026-00942-0

More Information:

Information about bitter-tasting peptides used for training comes from the Bitter Peptide Space (BPS)-1000 database. www.leibniz-lsb.de/en/databases/bitter-peptide-space-bps-1000-database

The publication “Srivastava P et al. Bitter peptide prediction using graph neural networks. J Cheminform. (2024). https://doi.org/10.1186/s13321-024-00909-x” introduces BitterPep-GCN, a feature-independent graph convolution network for predicting bitter-tasting peptides.

This news item from the TUM School of Life Sciences is based on a press release from Gisela Olias from Leibniz-LSB@TUM.