Fig. 9

Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction to illustrate chemical space coverage of triplicate Glide with LigPrep REINVENT-DockStream experiments against COX2 (PDB ID: 1CX2). Morgan fingerprints (radius 3, 1024 bits) for the top 1000 and 3000 compounds (based on total score) for each replicate were extracted to illustrate topology features. Chemical space coverage is stochastic in nature and thus, multiple replicates are used in practice to optimize output diversity. “Random Sampling” refers to sampling compounds from the prior generative model whose policy has not been focused on any task. a Top 1000 (based on total score) compounds generated by the agent in each triplicate experiment b Top 3000 (based on total score) compounds generated by the agent in each triplicate experiment. Some overlap between replicate experiments is observed, demonstrating steering of chemical space exploration to similar local minima