China could change drug development more than AI over the next ten years, predicts pharmaceutical researcher Jack Scannell
China could change drug development more than AI over the next ten years, predicts pharmaceutical researcher Jack Scannell
On November 11, AthenaBIO, an initiative to build open datasets on ovarian aging, published an interview with Jack Scannell. He bases his prediction on the quality, speed, and low cost of research in China, along with potential competition between Chinese and US drug regulators. He suggests assessing AI by the decisions a model helps people make during drug development.
If AI makes it cheaper to find molecules that bind to a target protein while testing their effects in cells, animals, and humans remains expensive, candidates will accumulate at the testing stage. Scannell predicts that those who conduct these expensive tests will be able to profit from the influx. Patients may benefit, while investors who funded large-scale searches for molecules may lose money.
AI protein structure prediction draws on decades of experimental data, and its predictions can be checked relatively quickly. Scannell allows that cheaper protein design could open up new forms of treatment. When researchers seek treatments for a poorly understood disease, testing a hypothesis in humans takes years. Scannell himself heads Etheros Pharmaceuticals, which develops treatments to protect nerve cells. He says AI helps him compare findings across studies and look for evidence supporting hypotheses that will then need to be tested in clinical trials.
In training data, drugs can share the label “toxic to the liver” even when the causes of harm differ: a patient's genetic predisposition, drug interactions, or accumulation of a substance because of impaired kidney function. If these conditions are not recorded, the model has a harder time predicting which molecule will be dangerous.
A practical approach is to start with the decision and work backward.Selecting successful molecules and screening out likely failures require different metrics. When choosing a few candidates from a large pool, the proportion of selected candidates that succeed matters: a mistaken selection means developers spend money testing an ineffective molecule. When screening candidates out, they need to account for the risk of discarding a useful molecule. Two models with the same overall score on a standard test may therefore perform differently when selecting molecules and screening them out.
The next molecules in a project may differ from those used to train the AI. Scannell therefore suggests recording predictions of experimental results before running the experiments:
Make predictions, run the experiments, and see what happens.
Back in July, Scannell proposed first checking how well experiments in cells and animals predict treatment outcomes in humans, then expanding the search with AI. For ovarian research, he now proposes using shared protocols and obtaining women's consent to collect biological material usually discarded during IVF, then linking it to their medical histories and pregnancy outcomes. Such a dataset would allow researchers to test associations between laboratory measures and pregnancy outcomes and use those findings to prepare data for training AI.