Approach
How Topologix predicts drug resistance
Topologix is developing models that predict whether a mutation in a drug target will make a given compound lose potency. This page describes what the models use, what they produce, and how we check that the predictions are real.
What the platform will predict
Input: a target protein sequence, the amino-acid substitutions of interest, and a set of compound structures.
Output: for every mutation–compound pair, a probability that the compound loses at least tenfold potency against the mutant compared with the wild-type protein, and a ranking across pairs and compounds.
What the models look at
Each mutation–compound pair is described by three groups of features. The first describes the substitution itself: the change in size, charge, polarity and related properties between the original and the new amino acid. The second is evolutionary: a protein language model, trained on natural sequences, estimates how acceptable the new residue is at that exact position. The third describes the compound’s structure. A machine-learning model combines the three into one score.
We describe the method at this level publicly. Design partners can discuss the details with us directly.
How we evaluate
Published resistance and binding predictors often look strong on random train–test splits and weaken sharply once similar proteins or compounds are kept apart. A random split lets a model score well by memorising which protein or chemical series a data point came from, without learning anything about the mutation. We design our evaluations so that this cannot happen.
- Paired measurements. Every mutant measurement is paired with a wild-type measurement of the same compound, against the same protein, in the same paper, with the same assay readout. This removes much of the noise that comes from comparing labs.
- Whole chemical series held out. Compounds are grouped into series by structural similarity, and each series is held out entirely during testing.
- Controls built in. Every run includes a model that sees only protein identity, which must score at chance, and a run with shuffled labels, which must fall to chance.
- Pre-registration. Comparisons and pass marks are written down before any model is trained.
Early results
On a ChEMBL-derived corpus of 10,983 mutant and wild-type activity pairs across 215 target proteins, our first model reaches AUROC 0.827 ± 0.021 on compounds from 1,672 chemical series, each held out of training in turn, under a pre-registered evaluation. The protein-identity control scores 0.50. These results cover new compounds against targets in the training data, most of them kinases and HIV-1 reverse transcriptase.
What we are building next
We are adding structural information, such as each residue’s distance to the bound ligand and its solvent exposure, and physics-based estimates of binding energy. The aim is to extend the models to targets with little resistance data, and to bacterial and viral targets beyond the current corpus.
Become a design partner
We are working with a small number of drug-discovery teams and CROs to decide what Topologix should do first. If resistance has cost you a programme, or you are trying to anticipate it in one, we would like to hear how you handle it today and what a useful prediction would look like.
Design partners can also run a blind test on their own data: send structures and mutations, keep the measurements, and score our predictions yourself. We are happy to sign an NDA first.