Lead optimization runs as a repeating cycle. Chemists design a set of compounds, synthesize them, test them in the lab, and analyze the results; the analysis informs the next round of designs. Each turn of the cycle takes weeks, and the number of compounds that can be made and tested in one turn is limited by bench capacity.
AI changes the economics of the loop in two places. First, it moves part of the testing forward: property and activity models predict how a batch of proposed structures is likely to behave, so the compounds that are actually synthesized are the ones predicted to be worth the bench time. Fewer compounds are made, and a larger fraction of them are informative. Second, it moves part of the analysis forward: because the models are trained on the accumulated results of earlier rounds, each new measurement updates the predictions, so the next design round starts from a better-informed position than the last.
The loop itself does not disappear. Predictions remain estimates, and the experimental result is still the only thing that converts an estimate into knowledge. What changes is the number of turns required and the number of compounds consumed per turn — the cycle gets shorter and more informative, not unnecessary.