AI identifies bitter peptides: What this means for plant protein

In short: researchers from the Leibniz Institute for Food System Biology at the Technical University of Munich have developed an AI pipeline that predicts short protein blocks as bitter or not bitter and new sequences A trained sensor panel confirmed 25 out of 31 verified predictions. This is a useful tool for taste research, but not yet a process that automatically eliminates bitterness in plant protein products.

Why protein products can taste bitter

Proteins consist of chains of amino acids. If these chains are split up during fermentation, maturation or enzyme processing, shorter pieces are created, so-called peptides. Some of them activate bitter receptors. This can be desirable, inconspicuous or disturbing for matured cheeses, kefir, protein hydrolysates and vegetable protein powders.

The problem is complex: not only a single amino acid decides on the taste. Order, length, spatial properties and concentration play together. Even small changes can turn a bitter into a hardly bitter sequence. Manufacturers often test many process variants and formulations before a product is sensoryly balanced.

How the two AI components work together

The team combined a protein language model with BitterPep-GCN, a graph-based neural network. The language model produced new peptide sequences, while BitterPep-GCN sorted them according to the probability for bitter or non- bitter taste. The basis were experimentally tested data from the open BPS-1000 database.

This database comprises 570 as bitter and 423 as non- bitter peptides; if available information is also stored in taste thresholds and activated bitter receptors. A curated part has been used for the generative model, which is important because older publications may contain contradictory taste indications for individual peptides.

The model produced 161 new sequences. For the laboratory test, the researchers selected particularly promising candidates, had them synthesized and tested purity and identity. Only then did a trained panel taste the solutions in specified concentrations.

What 25 correct predictions out of 31 mean

Out of 31 tested peptides, 15 predicted bitter and ten predicted non bitter sequences were confirmed accordingly. Four further results were not clear according to the previously defined panel boundaries. The procedure thus achieved around 80 percent correct classifications over the entire test selection.

This is an experimental validation, but not a general guarantee of accuracy for each food. The 31 samples were selected from a much larger set of models. In addition, pure peptide solutions were evaluated. In a finished product, fat, sweetness, acidity, salt, roasted aromas, texture and other ingredients influence perception.

The study shows above all that the search space can be reduced. Instead of synthesizing and tasting thousands of possible sequences indiscriminately, laboratories can select candidates more specifically. This potentially saves time and material in an early research phase.

How the findings could support milder products

In the long term, the predictions could help to test process steps: Which enzymes split a protein at which locations? Which fermentation conditions promote certain peptides? Which raw-material fraction contains particularly many bitter sequences? Questions like this concern plant proteins as well as dairy and fermented products.

A possible goal would be to choose production conditions in such a way that less problematic peptides arise. Another possibility is to recognize particularly bitter fractions early and to specifically balance formulations. However, this does not automatically mean "Clean Label". Whether fewer flavours, sweeteners or other flavoring ingredients are needed, must be tested separately for each product.

What consumers can take from the research today

Research does not change any list of ingredients in the supermarket. Those who compare vegetable protein products can continue to pay attention to a short, understandable list of ingredients, but should not derive a general quality evaluation from it. Processing is not automatically bad, and a natural raw material does not taste automatically mild.

In the own kitchen, bitterness and strong plant notes can often be balanced with roasting, acidity, salt, sweetness and fat. In the case of pulses, the right preparation is also crucial. You can find ideas among pulses and at fast dinners. However, a research result on synthesized peptides is no reason to avoid individual foods.

Limitations of the study

• Small validation group: 31 selected peptides were tested, not all 161 generated sequences and not the entire possible peptide space.
• Simplified test matrix: Pure solutions represent the interactions of a complex food only limited.
• Trained panel: The results describe controlled sensor technology, not a representative consumer survey.
• No health assessment: Bitterness does not automatically say anything about harmfulness or a health benefit.
• No finished production recipe: The model identifies candidates; safe and economical processes must be developed separately.

Why the topic still matters

Taste is a central hurdle for new protein sources. If a model is more reliable in identifying which small pieces of protein taste disturbingly bitter, it can better prioritize the complex search in the laboratory. It is particularly valuable that the predictions were not only evaluated mathematically, but were checked with purified samples and human sensors.

The open database is also important. The BPS-1000 database and the published workflows allow other groups to understand results and test models with additional peptides. Only such independent tests show how robust the process is across different raw materials and applications.

Conclusion: A useful filter, not yet a flavour switch

The AI pipeline has correctly distinguished a large part of bitter and non- bitter peptides in a specifically selected test series. This brings closer to a more systematic development of flavours for fermented and protein-rich foods.

Nevertheless, research is removed from a mild plant protein at the touch of a button. Formulation, manufacturing process, concentration and food matrix remain crucial. For now, the model is a research filter that helps select experiments more effectively. It does not replace sensory analysis, process development or consumer testing.

Sources and publication status

• Leibniz-LSB@TUM: AI as a Tool in Flavour Research: De Novo Design of Bitter Peptides, published on 22 September 2026.
• Steuer et al.: De novo design and experimental characterization of bitter peptides, npj Science of Food, published on June 25, 2026.
• Leibniz-LSB@TUM: Bitter Peptide Space BPS-1000, open database of experimentally tested peptides.
• PubMed: Bibliographic entry, abstract and funding information for the primary study. (Bibliographical entry, abstract and funding information for the primary study)
• DiPizio Lab: Published records and model workflows.

Author: Kochzauber editorial team. Information last updated on 23 September 2026. This article explains a research tool and makes no health or product claims. The header image is an editorial symbolic image.