Fig. 1

Illustration of idealized behavior of priors, transfer learning agents and reinforcement/staged learning agents. In all cases, the models describe the probability of sampling a given token sequence corresponding to a specific molecule (green squares), represented by a colored fill. The prior model is trained to increase probability over all drug–like molecules. A transfer learning agent built from this prior increases the likelihood on a specific region (blue, middle). In staged learning (red, right), starting from the transfer learning agent, likelihood of sampling high-scoring sequences is iteratively increased, resulting in concentration on high-scoring regions (red polygon)