Historical trial outcomes matter most because they are the only evidence stream that records the endpoint being predicted — what actually happened to a drug in real people. Without them, a model has no ground truth to learn from.
The outcome side of the pattern
A model learns by connecting a drug's properties to what happened to it. Structure and assay data supply the properties. Trial outcomes supply the "what happened." Remove the outcome side and the connection has nothing to attach to — the model can only echo whatever assumption was used to fill the missing label.
Volume is what makes a pattern visible
Decades of testing have produced thousands of drugs that were abandoned, and each abandonment is a labeled example of failure. A pattern can only be found when there are enough examples to distinguish a real signal from coincidence. This is why the accumulated record of past trials is treated as an asset in its own right.
The condition attached to that value
Trial records help only if they are complete and comparable — the same information captured for each drug, and failures documented as carefully as successes. In practice, failed drugs are often recorded less thoroughly, which tilts what the model learns before it has made a single prediction.