Fig. 3

(A-C) The properties distribution of the 5000 valid molecules generated by the modules “Physicochemical properties” (A, B), “Pharmacophore features” (C), and “Bayesian sampling” (C) are compared with those generated by the unconditional generation task. D The scaffold of erlotinib was selected as the input molecule for the “Structure features” module to perform fragment-based design. Annotation: MW1-LOGP: the MW and logP range were set to 150–400 and -2–6, and no optimization for QED and SA; MW-LOGP: the MW and logP range were set to 200–600 and 0–4, and no optimization for QED and SA; QED1-SA: the MW and logP range were set to 200–600 and -2–6 and optimization for QED was turned on; QED-SA1: the MW and logP range were set to 200–600 and -2–6, and optimization for SA was turned on