Connecting molecular design with biological evidence.
We develop AI that learns molecular structure, designs for multiple goals, and plans how to refine a lead. Next, we are building AI agents that bring these tools together with experimental feedback—helping collaborators decide what to test next.
Better molecular design starts with connected evidence.
A molecule has a structure, many competing properties, and a response that depends on its biological context. Our work connects these pieces: learning useful representations, generating candidates, planning refinements, and testing biological hypotheses with collaborators.
Cell · 2026Biology-guided discovery
Design with the disease in view.
What if we searched for molecules by the disease-related gene expression changes we want to reverse?
Our collaborative GPS study predicts transcriptional responses from chemical structures, screens compounds, and optimizes leads. Experimental studies in liver cancer and pulmonary fibrosis connect computational design to biological evidence.
With Jing Xing, Bin Chen, and collaborators Jiayu Zhou, co-corresponding author
Explore the trade-offs between molecular properties.
MolSearch starts with existing molecules and searches through chemical transformations to optimize multiple objectives. Efficient search makes broader exploration practical.
Translate a design brief into progressive generation.
Chain-of-Generation breaks a text description into ordered semantic goals, guiding molecular diffusion in stages so complex requirements can be introduced progressively.
TRACE treats tool selection as a sequence of decisions. An LLM-based agent considers how each refinement affects later steps while retaining similarity to the original lead. It is our first step toward agentic drug discovery.
Together, these studies connect molecular knowledge → controlled design → sequential refinement → biological evaluation. They are complementary research methods, with opportunities to integrate them in new collaborations.
Much of this program grew from a long-standing collaboration with the Bin Chen Lab at Michigan State University, whose expertise in transcriptomics-based drug discovery connects our methods to biology and experimental validation.
02 / The next direction
An agentic AI–experiment vision
Learn from the next experiment.
More measurements help most when they answer the right questions. Building on TRACE, we envision AI agents that use our molecular models as tools, plan multi-step design, and learn from both promising and unsuccessful outcomes to guide the next round—with scientists deciding what gets made and tested.
Explore the discovery cycle
Published methods01 / 04
Generate with a purpose.
Start with a lead or a design brief. Use search and progressive generation to explore candidates under structural and property constraints.
Research building blocks: MolSearch · Chain-of-Generation · TRACE
↻
The goal: a more informative next round. Agentic planning and experimental feedback are directions for joint research. This illustration does not run a molecular model.
01Preserve the intent. Keep essential structural constraints visible.
02Inspect the decisions. Make intermediate candidates and tool choices reviewable.
03Learn from outcomes. Use failures to refine hypotheses and next experiments.
Lingxiao Li, Haobo Zhang, Bin Chen, and Jiayu Zhou.
Accepted at NeurIPS 2026 · linked preprint: 2025
04 / Work with us
Let's collaborate.
If you are working on a disease, a lead compound, or an assay and think AI might help, we would love to hear from you. We enjoy working with biologists, chemists, and drug discovery teams, and an informal chat is a fine place to start. Just email Jiayu Zhou at jiayuz@umich.edu.
Acknowledgments
This research is supported in part by the National Science Foundation under Grant IIS-2212174, the National Institute on Aging (NIA) under R01AG072449, and the National Institute of General Medical Sciences (NIGMS) under R01GM145700.
