Table 2. Maximum J(m) scores averaged over 10 runs, the number of molecules evaluated per run, and the required CPU time per run. See the text for an explanation of the methods. Results for the non-GB methods are taken from the study of Yang et al.2 where the number of molecules evaluated per run is estimated based on the average number of molecules generated per minute and the CPU time.
Method | Average J(m) | No. molecules | CPU time |
GB-GA (50%) | 6.8 ± 0.7 | 1000 | 30 seconds |
GB-GA (1%) | 7.4 ± 0.9 | 1000 | 30 seconds |
GB-GM-MCTS (62%) | 2.6 ± 0.6 | 1000 | 90 seconds |
GB-GM-MCTS (80%) | 3.4 ± 0.6 | 1000 | 90 seconds |
GB-GM-MCTS (80%) | 4.3 ± 0.6 | 5000 | 9 minutes |
ChemTS | 4.9 ± 0.5 | ∼5000 | 2 hours |
ChemTS | 5.6 ± 0.5 | ∼20 000 | 8 hours |
RNN + BO | 4.5 ± 0.2 | ∼4000 | 8 hours |
Only RNN | 4.8 ± 0.2 | ∼20 000 | 8 hours |
CVAE + BO | 0.0 ± 0.9 | ∼100 | 8 hours |
GVAE + BO | 0.2 ± 1.3 | ∼1000 | 8 hours |