Teaching a computer to predict bitterness

Bitterness is one of the five basic tastes, and probably the most protective. It is a warning: many spoiled or poisonous foods taste bitter, and recoiling from them is a defence. The same reflex causes trouble elsewhere. Many medicines taste bitter, which makes it harder to stick to a treatment, and some healthy foods are left on the plate for the same reason.

A family of 25 sensors

We detect bitterness through a family of 25 receptors called TAS2Rs, proteins that sit in the membrane of taste cells. They are not only in the mouth: they are also found in many other tissues, where they seem to have other jobs, such as helping to regulate inflammation and the immune defences of the airways.

These receptors do not share the work evenly. Some are generalists that respond to many, very different molecules; others react to only a handful. Bitter compounds themselves are just as varied, from alkaloids and polyphenols to peptides and salts. Finding out which receptor responds to which molecule means testing pairs one by one in the lab, which is slow and expensive. So the available data are limited, and much of what is known has been collected in a single database, BitterDB.

Letting the model show its reasoning

We gathered every in-vitro verified pair we could find: more than 300 bitter molecules and nearly 4,000 molecule–receptor pairs, covering 22 of the 25 human receptors (the other three have no known activator). Each pair was labelled as “binds” or “does not bind”, and we trained two kinds of models on it.

The first is a classic machine-learning model built from decision trees. It describes each molecule with a short list of computed chemical properties. The second is a graph neural network, which reads the molecule directly as a network of atoms and bonds. On data it had never seen, the tree model reached an ROC AUC of 0.92 and the network 0.88, on a scale where 1 is perfect and 0.5 is guessing.

A prediction alone is not very useful if nobody can see where it comes from, so both models were built to be explainable. The tree model can report which features weighed most. One finding was almost reassuring: it leaned heavily on which receptor was involved, because the two most promiscuous receptors, TAS2R14 and TAS2R46, bind so many molecules that pairing with them pushes the answer towards “yes”.

The graph model gives a more visual explanation. For strychnine, a famously bitter alkaloid, paired with TAS2R46 (the receptor drawn on the taste receptors page of this site), it highlighted the region around one nitrogen atom as the part pushing towards “binds”. Experimental structures had shown that this same nitrogen forms a hydrogen bond inside the receptor. The model had no information about the receptor’s structure. When we removed two carbon atoms next to it, the predicted probability dropped.

Why it matters

The models are not perfect. The data contain about five times more “no” pairs than “yes” pairs, so positive predictions are less precise. The tools also report how similar a new molecule is to the ones they learned from, as a guide to how far to trust the answer. Still, they can help choose which experiments are worth doing, and point to the parts of a molecule that a chemist might change: to make bitter medicines and foods easier to accept, or to design molecules that target one receptor in particular.

This work grew out of the EU project VIRTUOUS, alongside two related studies: one showing that TAS2R46 moves in a more coordinated way when strychnine is bound, and one that searched protein databases for pockets shaped like its binding site and found 145 other proteins with similar sites.

Read the paper: Ferri, Cannariato, Pallante, Zizzi, Miceli, Deriu, Journal of Molecular Graphics and Modelling 142, 109187 (2026). doi:10.1016/j.jmgm.2025.109187

Related: Cannariato et al., Frontiers in Molecular Biosciences 11, 1473675 (2024). doi:10.3389/fmolb.2024.1473675

Related: Pallante et al., Scientific Reports 14, 6296 (2024). doi:10.1038/s41598-024-56893-7