Research
AI for biomedical discovery, biotechnology, and clinical translation

We develop AI methods to assist biomedical research and embed them into emerging biotechnologies, turning complex biological measurements into interpretable discovery engines. Our goal is to help researchers see biological states, mechanisms, and disease processes that were previously inaccessible, then translate those insights into clinical studies that improve human health. We move deliberately rather than chasing instant results, because careful research is what lets computation meet biology with real depth.

AI Methods
Self-supervised Representation Learning
We develop self-supervised and contrastive learning techniques for biological data where labels are scarce or noisy. Our methods learn rich representations from unlabeled or weakly labeled multi-omics and imaging data, enabling downstream analysis across modalities without expensive annotation.
Multi-modal Information Fusion
We design neural architectures that disentangle and integrate multi-modal inputs — including transcriptomics, proteomics, spatial data, imaging, and phenotypic screens — learning shared and modality-specific representations that preserve biological structure across more than two modalities simultaneously.
Causal Inference in Complex Systems
We build methods for causal reasoning over high-dimensional biological data, including gene regulatory networks, genetic association studies, and drug perturbation experiments. These frameworks go beyond correlation to uncover regulatory relationships and mechanisms that are invisible when modalities are analyzed in isolation.
AI + Biomedicine

We apply these methods across biotechnology platforms and clinical datasets, connecting molecular, cellular, tissue, and patient-level information. By integrating AI with high-throughput experiments and medical studies, we aim to generate actionable biological insight and support better diagnosis, therapy discovery, and patient care.

Multi-scale Reconstruction of Disease Phenotypes
We design scalable deep learning and generative modeling methods for high-noise, high-missingness biomedical data — including histopathology, spatial omics, and single-cell sequencing. Our goal is to reconstruct reliable disease phenotype representations from tissue to cell to molecule scale, providing tools for pathological classification and mechanistic study.
Multi-modal Pathology–Molecular Fusion for Target Discovery
We build unified representation and alignment frameworks for heterogeneous multi-modal data including pathology imaging, spatial transcriptomics, single-cell omics, and clinical phenotypes. By bridging resolution and throughput limitations of individual technologies, we provide a systematic “phenotype-to-molecule” path for disease target discovery.
Causal Reasoning for Drug Discovery
We develop methods for causal validation of therapeutic targets and transferable prediction of small-molecule function using large-scale genetic variation, drug perturbation, and high-content phenotypic data. These approaches support drug repurposing, candidate screening, and personalized treatment strategies.