Building the Virtual Cell: The Evolution from AI4Science to LLM4Science  

Abstract:

Understanding biological systems and building virtual cells revolves around four foundational stages: Represent structures in molecules and cells, Predict cell populations under new treatments, Reason & Act with evidence and tools to guide decisions, and Test & Learn from experimental outcomes. While this four-stage loop has long anchored traditional AI for Science, the gradual evolution from AI for Science to LLM for Science is reshaping how each stage is realized and connected—bridging quantitative domain modeling with prior knowledge, mechanistic reasoning, and agentic decision-making.  

 

In this talk, I will trace this progression from AI4Science to LLM4Science through our research across the four-stage virtual cell framework. Starting from cellular representation learning (scCello), I will dive into PerturbDiff for the Predict stage—a functional diffusion framework that embeds empirical cell populations into a reproducing kernel Hilbert space (RKHS) to model generative dynamics directly over probability distributions. Moving to Reason & Act, I will present CORE, which harnesses LLM in-context learning to integrate mechanistic priors with context-specific experimental observations, enabling rapid adaptation and informed intervention decisions across unseen cellular contexts. In the final part of the talk, I will discuss the emerging paradigms of LLM4Science and how they answer a central forward-looking question: how to close the Test & Learn loop with a lab-in-the-loop.

 

报告时间：2026年11月11日 14:00-15:00

报告地点: 静园六院211会议室

 

Bio：

Xinyu Yuan is a Ph.D. candidate in Computer Science at Mila – Quebec AI Institute and Université de Montréal working with Jian Tang. Xinyu went to Peking University for undergrad advised by Ming Zhang, receiving both the Outstanding Graduate and Outstanding Undergraduate Thesis awards, after winning a bronze medal in the National Olympiad in Informatics (NOI). During her Ph.D., Xinyu has published multiple first- and co-first-author papers at ICML, NeurIPS, and ICLR—including an ICML Oral (Top 2%) and a NeurIPS Spotlight (Top 2%)—and received the RBC Borealis PhD Fellowship (10 recipients across Canada annually). Xinyu has conducted research internship at Genentech’s Aviv Regev Lab, Tencent Hunyuan (Qingyun Talented Program), Amazon, and Intel AI Labs, and organizes the Multiomics Reading Group at Mila. Xinyu’s primary research interests are LLM post-training, reasoning, and agents, biological foundation models, and virtual cells, with a focus on generalizing to unseen tasks and contexts.