Traditional software runs on rules a person writes down: if a compound has this property, flag it. Machine learning works the other way around. You supply many examples — say, thousands of compounds already labeled as active or inactive against a protein — and the system searches for statistical regularities that separate the two groups. Nobody hands it the rule; the rule is inferred from the examples.
This is why the approach suits cancer drug discovery. The relationships between a molecule's structure and its behavior in the body are real but too tangled to express as clean hand-written rules. A model can still detect that certain structural features tend to co-occur with activity, and use that to score a new, never-tested compound.
Two consequences follow immediately. First, the output is a probability or a ranking, not a fact — a score of 0.9 means "similar to things that worked," not "will work." Second, the model has no access to biology beyond what the examples encode. It learns the pattern in the data it was given, including whatever errors and gaps that data contains.