Problem
Searching for molecular candidates with lower predicted toxicity can be computationally expensive. Tox-NIO explores direct optimization of molecular features as an alternative to random search.
Implementation
I developed a molecular editor based on Neural Input Optimization. It uses projected gradient descent to optimize molecular features on the ZINC-250k dataset, with the aim of reducing a model’s predicted toxicity score.
Technical decisions
- Used Neural Input Optimization to modify molecular features.
- Applied projected gradient descent in place of a genetic search approach.
- Evaluated changes using predicted toxicity on ZINC-250k compounds.
Results
In my project evaluation, predicted toxicity fell from about 0.55 to 0.28, a reduction of roughly 50%. These scores come from the toxicity model used during optimization.