GAMS: Graph-based Autoregressive Molecular Generation System for Drug Discovery

The Project

This project is about developing an AI-powered system that can automatically generate new, drug-like molecules to accelerate the early stages of drug discovery. Inspired by my grandfather’s illness and the delays in finding effective medication, I wanted to create a tool that could help scientists explore chemical space faster and more efficiently. The core of the project is a decoder-only transformer model combined with a Dynamic Guided Decoding (DGD) strategy, which uses real-time feedback to guide molecule generation toward better drug-likeness, lower toxicity, and easier synthesis. I also used a graph-based molecular optimization framework to analyze and select high-potential compounds. One of the most difficult parts was designing a system that could balance multiple properties while keeping the molecules chemically valid. After testing and evaluating the results, I found that the model could consistently generate high-quality candidates, showing strong potential for real-world applications in pharmaceutical research.

View this project on YouTube