Research Areas

Our lab sits at the intersection of Artificial Intelligence and Precision Medicine. We develop multi-modal foundation models that bridge the gap between standard clinical imaging and high-dimensional molecular data, transforming how we understand and treat complex diseases.

Core Research Pillars

·       Pathology Foundation Models: We build pathology vision and vision-language foundation models for precision oncology (Nature 2025, JCO 2025). These models can improve the diagnosis and prediction of treatment outcomes across various cancers.

·       Virtual Spatial Proteomics: We’ve developed AI models that generate spatial proteomics profiles and predict the expression of 40+ biomarkers directly from standard H&E-stained slides (Nature Medicine 2026). This technology significantly improves immunotherapy response and survival prediction for patients with lung cancer.

·       Single-Cell Spatial Mapping: We investigate advanced methods for the single-cell characterization of tumor microenvironment, using AI to decode the spatial relationships between cells that drive disease progression (Nature Communications 2025).

Ongoing Research Projects

We are actively expanding our research into the next generation of biomedical AI:

·       Multi-Modal Foundation Models for Spatial Omics: Integrating histopathology with spatial transcriptomics and proteomics to create a holistic view of human biology.

·       Virtual Cells & World Models: Building generative world models of cellular and tissue microenvironments to simulate disease progression and therapeutic perturbations in silico.

·       AI Agents: Developing AI agents capable of autonomous reasoning across multi-modal biomedical data for clinical decision-making and scientific discovery.

Imaging AI for Precision Oncology

Previously we worked on deep learning and radiomics analysis for clinical outcome prediction in various cancer types (Radiology 2016, 2017, 2018; Annals of Surgery 2020; Nature Communications 2021; Lancet Digital Health 2022). In a Nature Machine Intelligence 2011 study, we developed a unifying radiological tumor classification system that demonstrates prognostic relevance and therapeutic implications across multiple cancers.

We’ve also explored the biological basis of clinically relevant radiological phenotypes. In recent studies, we developed a noninvasive imaging approach to evaluate tumor immune and stromal microenvironment, which could further predict immunotherapy response and survival outcomes (Annals of Oncology 2020; Lancet Digital Health 2021; Nature Communications 2023; Cell Reports Medicine 2023).