Date: 7 October 2026 (Wednesday)
Time: 10:00 am – 11:00 am
Venue: Lecture Theatre 2, G/F, William M.W. Mong Block, 21 Sassoon Road 
Host: Professor Rio Sugimura

Dr. Jin Liu is an Associate Professor and Presidential Scholar at the School of Data Science, The Chinese University of Hong Kong, Shenzhen. He is an elected member of the International Statistical Institute (ISI), an Associate Editor of JASA, and serves on professional committees of the Chinese Mathematical Society. He is also an Adjunct Associate Professor at Duke-NUS Medical School, National University of Singapore. His research focuses on artificial intelligence and statistical computing for statistical genetics and genomics, including single-cell and spatial transcriptomics integration, foundation models for cells and genes, multimodal models of transcriptomes and tissue graphs, and reasoning- and agent-based downstream inference. His work has appeared in Cell, Nature Communications, Genome Biology and other leading journals. He has led four Singapore AcRF Tier 2 grants and projects funded by NSFC.

Existing gene-embedding approaches in genomics often overlook intergene relationships. CoxFormer integrates literature-derived gene knowledge with co-expression networks from bulk tissues and large-scale single-cell atlases, learning 512-dimensional representations for 32,016 human genes. These embeddings support histology-based expression imputation, chromatin-accessibility-based gene activity prediction, subcellular super-resolution inference, and pathological region detection. Meanwhile, abundant spatial transcriptomic data enable molecular inference from H&E images. We propose HistAgent, a unified framework combining a vision-omics foundation model with spatial AI modules. Trained on 2.23 million paired H&E–spatial transcriptomics spots, the model generates relevance-ranked molecular readouts from local H&E images and their surrounding context. Spatial AI agents organize these outputs into evidence cards containing inferred cell composition, functional programs, and spatial context, enabling question-driven, multi-round analysis.

All are welcome.