A doctor prescribes a pill for the heart and a pill for blood pressure. Individually, each helps, but together they can scramble signals in the blood and cause harm. These unwanted encounters between drugs are called interactions, and testing every pair in the lab is impossible: there are millions of them.

A computer can step in: it looks at the structure of drugs like blueprints and predicts which parts might latch onto each other. But previously, all the testing programs had different textbooks and different exams. One was praised for one thing, another for something else, and comparing them was impossible.

Now there is a common testing ground — a set of six well-known drug data collections and twenty predictor programs. Each program was run through the same tasks: find the dangerous pair, don't mistake a safe one, don't break on new data. The programs that learned from examples, like a student working through problems with answers, performed best.

This common exam revealed where programs fail most often. Now we can fix them with precision, not at random. That means in the future, doctors will have a reliable assistant that warns about a dangerous pair before the patient even swallows the pills.