GENTRL (Generative Tensorial Reinforcement Learning) designs molecules by combining a variational autoencoder with reinforcement learning. The autoencoder maps chemical structures into a continuous latent space, where an agent can search smoothly using gradient-based optimization. section 13.3.1
Its reward balances several requirements rather than optimizing potency alone:
- Predicted TNIK binding affinity receives heavy weight. section 13.3.1
- Selectivity, solubility, CYP inhibition, and synthetic accessibility constrain toxicity, exposure, drug interactions, instability, and practical synthesis. section 13.3.1
- Human medicinal-chemistry judgment can shift objective weights when high predicted potency produces poor drug-like properties. section 13.3.1
TNIK design without a starting ligand
During the rentosertib program, GENTRL generated INS018_055 for TNIK despite the absence of known TNIK-selective compounds, bound-ligand crystal structures, or prior structure–activity data. section 13.3.1
