Fig. 2

Heatmap visualization of LSTM softmax probability outputs on a known PIK3CA active. We used a known PIK3CA active, “CC1(C(= O)N2CCCI(N3CCc4c(-c5cnc(N)nc5)nc(N5CCOCC5)nc43)C2)CCCO1” with 10.2 pKd activity recorded for PIK3CA. The given SMILES is cut off at the first 40 tokens for the visualization. a, b The tokens on the x-axis are the input SMILES on the time steps, and the token on the y-axis corresponds to each output of the softmax layer of the prior generator and the agent generator, respectively. The highlighted cells indicate the correct next tokens to be sampled to obtain the given sequence. a The prior model's conditional likelihood heatmap. b The conditional likelihood heatmap of the LOGICS agent from the PIK3CA experiment. c Side-by-side comparison of the correct next token output probability of the prior and the agent and the log2 fold change of the agent likelihood over the prior likelihood. The fold change is calculated by \({log}_{2}{G}_{\varphi }({x}_{t}|{x}_{0:t-1})/{G}_{\theta }({x}_{t}|{x}_{0:t-1})\)