Shujin Li

Research Interests

My research is centered at the intersection of artificial intelligence and medicinal chemistry, aiming to revolutionize the drug discovery pipeline. I focus on developing novel deep learning frameworks that can transform drug design from a process of serendipitous screening to one of rational, goal-directed engineering. My work bridges the gap between de novo molecular design and practical chemical synthesis, ensuring that AI-generated candidates are not only potent but also synthetically accessible. My research pursues two primary, synergistic directions:

1. Structure-Based Generative Molecular Design

A critical challenge in drug discovery is the efficient exploration of the vast chemical space to identify molecules that bind effectively and selectively to a specific protein target. My research addresses this by developing advanced generative models conditioned on the 3D structure of the protein’s binding pocket. My work involves: Developing and implementing state-of-the-art generative architectures (e.g., Diffusion Models, Equivariant Graph Neural Networks, and Transformers) to generate novel 3D molecules with high geometric and chemical complementarity to the target. Integrating multi-objective optimization algorithms, such as reinforcement learning and genetic algorithms, to steer the generation process towards candidates that simultaneously satisfy multiple critical criteria: high binding affinity (potency), selectivity against off-targets, and desirable ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) properties. Creating a closed-loop “design-predict-optimize” cycle that rapidly iterates through thousands of potential compounds, significantly accelerating the hit-to-lead optimization phase and increasing the probability of success for preclinical candidates.

2. Large Language Models for Drug Design

I harness the power of Large Language Models (LLMs) to unify the diverse “languages” of drug discovery—from protein sequences and chemical notations (SMILES) to natural language in scientific literature. My work focuses on developing instruction-tuned LLMs capable of multi-property optimization. This approach enables the generation of novel molecules that simultaneously satisfy complex design objectives (e.g., high binding affinity, favorable ADMET profile, and synthetic accessibility) specified through natural language prompts, effectively bridging the gap between a medicinal chemist’s design hypothesis and data-driven molecular creation.