Fig. 6.

Comparison of the chemical space of the LIGAND set and generated molecules. Shown are all known ligands (orange) and desired molecules (black). Moreover shown are generated molecules by DrugEx v1 (A, E, I, M, blue), v2 (B, F, J, N, red), ORGANIC (C, G, K, O, green) and REINVENT (D, H, L, P, purple). A distinction can be made between the multi-target case (A–H) and target specific case (I–P). Additionally a distinction can be made between PF scheme based scoring (A–D and I–L) and WS scheme based scoring (E–H and M–P). Chemical space is represented by the first two components in t-SNE with ECFP6 descriptors of molecules. Similar to our previous work it can be seen that DrugEx better covers the whole chemical space of the input data. In particular in the multi-target case with a Pareto optimization based scoring function (E–H) the improved coverage in all sections, including isolated active ligands, becomes clear