The design-make-test-learn cycle is a closed-loop drug-discovery workflow: models select compounds for synthesis and testing, experimental results update the models, and the process repeats. section 13.2.3
The cycle connects four practical stages:
- Design candidates using predictions, ensemble models, and acquisition functions that balance expected value with uncertainty. section 13.2.3
- Make prioritized compounds through experimental or robotic synthesis. section 13.2.3
- Test compounds experimentally instead of screening an entire library indiscriminately. section 13.2.3
- Learn by feeding measurements back into the models and selecting the next experiment. section 13.2.3
Industrial-scale cycle
Exscientia's Centaur Chemist combines acquisition functions with continuous robotic synthesis on a two-to-three-week cycle; four to eight cycles synthesize 350–500 compounds rather than 2,500 or more. section 13.2.3
Early cycles may favor diverse selection because too little labeled data makes predictions and uncertainty estimates unreliable. Fixed libraries also limit discovery of novel chemotypes. section 13.2.3
