Major Projects
Artificial Intelligence:
Computer vision for biomedical imaging
Deep learning enables automatic analysis of many imaging modalities for disease diagnoses and surgery intervention. We have developed pre-trained foundation models for optical coherence tomography (OCT) [Cui et al 2025] and improved tissue identification using 3D OCT [Ly et al 2025]. Our deep learning models have been used to automate renal carcinoma biopsy [Wang et al 2024], epidural anesthesia [Wang et al 2024] and percutaneous nephrostomy [Wang et al 2021].
Large language models for health behavior intervention and data analytics
Large language models (LLMs) can be used to power virtual counselors (chatbots) for health behavior intervention. We have adapted LLMs for message generation [Calle et al 2024] and smoking cessation counseling [Liu et al 2025]. We have also fine-tuned LLMs for unified modeling language (UML) code generation from diagram images [Bates et al 2025] and for automatic matching of patients to clinical trials.
Methodology research in machine learning
We improved the estimation of test performance with scarce data using nested cross-validation, automated hyperparameter optimization, and high-performance computing [Calle et al 2025]. We developed an interpretable neural network architecture [Adrien and Pan 2022] and a pan-disease multi-tasking learning framework [Adrien and Pan 2023].
Bioinformatics:
Prebiotics and probiotics in humane gut microbiota
Human microbiomes play important roles in our health. While prebiotics and probiotics have been shown to improve many health outcomes, their mechanisms are not known. We studied the metabolic interactions within microbial communities [Zhang et al 2022] and developed metabolic models using omics data constraints [Wang et al 2024]. Our discoveries will enable the development of prebiotics and probiotics with enhanced health benefits.
Parallel algorithms for proteomic stable isotope probing
Proteomic stable isotope probing are developed in our lab to track the flow of carbon and nitrogen from a substrate to its consumers in a complex community. We developed Sipros to achieve more accurate detection of isotopically labeled peptides from shotgun proteomics data [Xiong et al 2024]. Proteomic stable isotope probing has been used to track nutrient flows in marine communities [Kieft et al 2021] and rhizosphere soil communities [Li et al 2019]