PharmaLLaMA is a state-of-the-art generative AI model designed to create novel drug-like molecules in SELFIES format, optimized for specific properties. The model will attempt to optimize an input molecule, described as a Self-Referencing Embedded String (SELFIES) format, towards desired chemical properties outputting a novel optimized molecule. In this case study, we used JAK2 inhibition as an optimization parameter. Built on Meta's LLaMA transformer architecture, the model integrates Proximal Policy Optimization (PPO) to refine molecular generation, advancing the field of de novo drug design.
This project demonstrates how transformer-based architectures and reinforcement learning can be combined to surpass traditional graph-based models and even some transformer based models that were previously reported for molecular generation in producing highly valid, novel, and diverse molecules suitable for experimental validation.