Fig. 7
From: Memory-assisted reinforcement learning for diverse molecular de novo design
ECFP6 analog generation during reinforcement learning with experience replay and different bucket sizes. In all figures the DRD2 QSAR model was used. (a), (c) and (e) display experiments with experience replay. (b), (d) and (f) experiments without experience replay. a, b show the number of generated ECFP6 analogs. Compounds with a prediction score of at least 0.7 and Tanimoto similarity (count-based ECFP6) to the nearest neighbor of known actives of at least 0.4 were considered ECFP6 analogs. c, d show the number of unique BM scaffolds of the generated ECFP6 analogs. e, f show the number of unique carbon skeletons of the generated ECFP6 analogs. In all panels the RL without a memory (blue line) is not affected by the bucket size as this parameter is not present