At the heart of our Generative AI is optimization of rewards selected by the user, through a method known as Reinforcement Learning. This allows the generators to incrementally converge towards optimal molecules which are predicted to have good properties.
Rewards are scores that are computed on generated molecules and sent back to the generator, which learns from these scores what did and didn't work, and tries to generate new molecules with better scores.
In Makya, users can select from a variety of 2D or 3D rewards from the following:
- Tanimoto Similarity to a reference chemical space, in order to find molecules close to that space;
- Predictions from QSAR Models, trained on data uploaded by the user;
- 3D Shape and Pharmacophore Similarities to a reference ligand, whose conformer can either be uploaded by the user, or generated by Makya;
- Docking Score and Contact Score (a proprietary Iktos score akin to a protein-ligand interaction fingerprint), in order to find molecules that fit nicely in a designed protein binding pocket.
IMPORTANT NOTE: The more rewards are selected, the less the generator can explore the available chemical space and the more the generation is constrained, trying to solve multiple, potentially conflicting goals at the same time. As a result, it is important to strike a balance that allows generators to explore new ideas. If your generator fails at optimizing all the rewards you set, start a new generation focusing only on the most crucial rewards!