What makes a target worth drugging?

The idea had good reasons behind it. People with higher HDL cholesterol tended to have lower coronary risk. Torcetrapib blocked CETP, a protein involved in transferring cholesterol between particles in the blood, and pushed HDL in the direction researchers wanted. The hope was that the patients would benefit too.

When they did worse, even the failure was hard to interpret. Torcetrapib also raised blood pressure. Had the target been wrong, or had another effect of the molecule caused the harm? A later drug against the same target would reduce coronary events, while changing more than HDL. It took years of experiments to separate the questions that had once seemed like one.

The usual lesson drawn from HDL is that correlation does not establish causation. But a scientist choosing a target already knows that. The difficulty is deciding what to do with an incomplete, sometimes contradictory body of evidence while there is still time to choose where the program goes.

Which biology gets a chance?

It is tempting to answer that uncertainty by returning to the targets we know best. There are assays, papers and colleagues who have worked on them. An unfamiliar target asks a team to do more work just to decide whether it deserves further work.

That preference has left a mark on the research itself. In a 2018 study, Thomas Stoeger and colleagues found that much of the attention given to human genes could be explained by properties that made them easier to study and by earlier work in model organisms. Medical importance alone did not explain which genes attracted the most research.

AI can inherit that history. A model that is good at finding what has already been written may keep returning the same names, with increasingly persuasive explanations. We think its more interesting job is to make an unfamiliar target worth investigating: to predict what drugging it could do, and give scientists enough evidence to take that prediction seriously.

This is the bet behind Theorema. We want disease models to change which biology gets a drug program. Recovering familiar targets is necessary to earn trust, but a model that only confirms the existing shortlist would fall short of what we built it for.

Starting with the disease

We are introducing Causal Disease Models for 15 diseases, including Parkinson’s, MASH, heart disease and inflammatory bowel disease. Each model connects molecular activity with cellular mechanisms and disease outcomes, so a target can be examined in the context of the disease it is meant to treat.

That context matters. Lowering a protein can affect several processes, and the result that first made a target interesting may tell only part of the story. We want to understand what follows the intervention, including effects that make a promising target less attractive.

There is evidence for this work beyond the genes with the longest publication lists. Our source data include measurements from more than 100 million human cells, genetic perturbations, human genetics and clinical study results. A gene may have been measured in thousands of cells without becoming the subject of a dedicated drug discovery program.

The possibilities also extend across disease boundaries. A mechanism involved in one condition may matter in another, giving a team with an existing molecule a reason to investigate a new indication. The disease names differ; some of the biology is shared.

What would make that worth trusting?

There is an awkward problem with testing a model meant to find overlooked targets. The answers we can check tend to be the ones the field knows best. For many untested targets, we do not know whether a prediction is wrong. Treating everything outside the known successes as a failure can penalize the discoveries we are asking a model to make.

We start with answers that experiments have established. In one oncology evaluation, our models recovered relationships between cancer mutations and therapeutic targets 13.7 times more often than chance. We also asked Theorema-SL whether a cancer cell would survive the disruption of two genes together. Both genes in each test pair were held out of training. Those tests address different questions; neither can tell us that a newly proposed target will work in patients.

We also run internal benchmarks drawn from hundreds of thousands of clinical trial records, check the models for consistency and examine the evidence behind their reasoning. A strong score should survive more than one way of asking the question.

The next test comes from the scientists working on the disease. A partner can bring its own experimental data into the models and see how the evidence changes a recommendation. The result that disagrees with the model is especially useful: it gives the team something concrete to investigate before committing to the next experiment.

That requires a recommendation whose reasoning can be followed back to the evidence. We built that traceability into the models from the start, for a regulated industry where other scientists will need to examine how a decision was reached.

Julie Proft at IOCB Tech: