Fig. 4. First-principles AML discovery in a virtually unlimited space.

a Median values of molecular fitness F over the prioritized N_batch molecules at the different learning steps (step 0 shows the median of the initial population N_initial). b Corresponding violin plot showing the (kernel-density estimated) distribution of molecular fitness F over the batch. These smooth kernel-density estimated distributions can slightly extend beyond the true range of F values as indicated by the explicit values marked by blue crosses. The number of queries leading to favorable and unfavorable molecules is indicated next to each violin. Due to descriptor calculation failures (see text) these numbers do not always add up to N_batch = 100. c Examples of top-performing molecules identified at various learning steps (see text for an explanation of the different color-highlighted geometric motifs). An extended list of the 4 top-performing molecules of each learning step is shown in Supplementary Fig. 10.