Minhuan Li(李黾奂)

I am a Flatiron Research Fellow at the Center for Computational Mathematics, Flatiron Institute. I received my PhD in Applied Physics from Harvard University in January 2025, advised by Doeke Hekstra. I also hold a S.M. in Computational Science & Engineering from Harvard, and an undergraduate degree in Physics from Fudan University.

During my PhD, I was a Fellow at the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard, and interned at D.E. Shaw Research, where I worked on jointly fitting a family of QM/ML force field models to quantum mechanical and experimental data. I am also a core member of Reciprocal Space Station, an open-source consortium for structural biology software. I care deeply about contributing to the scientific community through both computational tools and shared infrastructure.

My research builds scalable, mathematically principled methods for biomolecular dynamics. I approach this by designing an interacting ecosystem of Data, Models, and Neural Priors, drawing heavily on the biophysical principles underlying experiments such as X-ray crystallography and cryo-EM, generative modeling, and statistical-physics-inspired sampling.

In this framework, Data represents the raw experimental observables (e.g., cryo-EM micrographs, X-ray scattering patterns); Models are the unified parameterizations of the molecule (e.g., atomic coordinates); and Neural Priors are the foundation models (e.g., AlphaFold) that capture the statistical rules of biology.

Data → Model. The core inverse problem, and where I spend most of my time. I develop differentiable forward models that faithfully map molecular structures to experimental observables, robust objective functions that remain well-behaved under noise and sparsity, and efficient samplers for high-dimensional, multimodal conformational landscapes.

Prior → Model. Steering pretrained foundation models with experimental likelihoods at inference time — without retraining — so they serve as principled priors for the inverse problem above.

Data → Prior. Training biological foundation models on heterogeneous, context-rich data — integrating sequence, structure, surface chemistry, evolutionary signals, and experimental readouts — so the model learns to condition on the full biological context rather than any single modality alone.

Prior → Data. Closing the loop: using the neural prior’s uncertainty to guide experimental design and enable efficient, AI-centric data collection. This is the frontier I aim to develop alongside experimental collaborators.