I build generative models for scientific data, mostly single-cell biology. My current work asks how to generate and forecast cells under conditions a model has never seen: multi-condition generation with gene-block attention (SAVE, ICLR 2026), causal representation learning for integrating heterogeneous scRNA-seq, and diffusion-based imputation that recovers spatial context from dissociated cells (stDiff).
What these share is a preference for structure over scale. Genes come in coordinated blocks, cells sit in spatial neighbourhoods, and measurements unfold in time — I would rather put that structure into the architecture than throw a bigger model at it. That is also why I wrote a survey on deep learning for gene regulatory network inference before building yet another network.
I started out on the physical side: short-time forecasting of radiation-belt electron flux with stacked ensembles, and hexagon-based node deployment for wireless sensor networks — both during my undergraduate years at Nanchang University. I completed my M.S. at Fudan University, where I moved from forecasting physical fields to modelling cellular ones.
Education
-
Fudan UniversityComputer Science and TechnologyM.S.2023 – 2026 -
Nanchang UniversityComputer Science and TechnologyB.S.2019 – 2023
News
Selected Publications (view all )
SAVE: A Generalizable Framework for Multi-Condition Single-Cell Generation with Gene Block Attention
Jiahao Li, Jiayi Dong, Peng Ye, Xiaochi Zhou, Haohai Lu, Fei Wang
International Conference on Learning Representations (ICLR), 2026
A generative framework for multi-condition single-cell generation that attends over gene blocks, so the model conditions on coordinated gene programs rather than individual genes.
SAVE: A Generalizable Framework for Multi-Condition Single-Cell Generation with Gene Block Attention
Jiahao Li, Jiayi Dong, Peng Ye, Xiaochi Zhou, Haohai Lu, Fei Wang
International Conference on Learning Representations (ICLR), 2026
A generative framework for multi-condition single-cell generation that attends over gene blocks, so the model conditions on coordinated gene programs rather than individual genes.
