Can a language model tell how long a drug stays bound?
Predicting how long a drug stays on its target usually calls for a 3D structure. We asked whether models that read sequences and chemical strings can capture some of it anyway.
Most medicines work by attaching to a protein in the body and changing what it does. For a long time, the main question asked of a new drug candidate has been how tightly it binds: its affinity. That is a fair question, but it describes a balance, a snapshot of how much drug is bound once everything has settled.
A living body is never quite settled. Drugs arrive, get diluted, get cleared. So a second question matters too: once the drug has found its target, how long does it stay there?
Holding on
Every bound drug eventually lets go. How fast that happens is described by the dissociation rate constant, koff. A small koff means the drug lets go slowly and stays bound for a long time, its residence time. The link between koff, residence time and how well a drug works in the body is well established.
Predicting koff, though, is still a hard problem. It depends on the moment-to-moment interaction between the drug and the pocket it sits in, the process of letting go, and that cannot simply be read off from measurements made at equilibrium. Approaches that model this process usually start from a detailed 3D picture of the drug inside its protein.
The question we asked
Language models trained on protein sequences and on chemical structures have learned a great deal about both, without ever being shown a 3D complex. That made us curious: can some useful information about kinetics be extracted from these models alone, with no structure of the drug in its pocket?
To find out, we built a model that sees only two strings of text:
- the protein, written as its sequence of amino acids
- the drug, written as a SMILES string, a single line of characters that encodes how its atoms are connected
Each string is read by a pretrained language representation, one for proteins and one for molecules, and a transformer on top learns to predict koff. The title of the preprint, Kinetics Beyond Pose, refers to what is missing: the pose, the exact position and orientation of the drug in its pocket.
A first answer, with limits
On a chemically and biologically diverse dataset, spanning several protein families, the model performed competitively in cross-validation and on a held-out test set, and it behaved robustly on the protein families best represented in the data.
We read this as an encouraging sign, not as proof that the approach works everywhere. The model was tested on data of the same kind it learned from, and it did best where examples were most numerous. How it would fare on protein families it has rarely or never seen, and how much of what it learns reflects the physics of unbinding rather than patterns in the data, are open questions. This work is also a preprint: it is public but has not yet been through peer review.
Why the question is worth asking
If sequence and chemistry alone carry part of the kinetic signal, residence time could be considered earlier in the design of new medicines, including for targets whose 3D structure with a drug is not available. Whether that holds up is what the next round of tests, by us and by others, will tell.
Read the preprint: Zizzi, Cicero, Gentile, Deriu, ChemRxiv (2026). doi:10.26434/chemrxiv.15003308/v1