九色导航

---

Research Interests:

1. Language Models for Genomic and Clinical Discovery
My research focuses on designing domain-specific large language models (LLMs) to uncover biological mechanisms from genomic sequences. I develop models such as DYNA, a disease-specific LLM for variant pathogenicity, and Lingo, which generalizes to unseen genomic variants and supports clinical discovery in cardiomyopathy and arrhythmia. These models are trained on biological sequences and evaluated for generalization, interpretability, and impact in clinical decision-making.

2. Secure and Robust Biomedical Foundation Models
I study the robustness and privacy of biomedical foundation models like ESM and DNABERT-2. This includes adversarial testing using mutated or perturbed genomic inputs and investigating techniques for privacy-preserving learning, such as representation anonymization and controlled exposure of biological information. This research has implications for both personalized medicine and federated biomedical model deployment.

3. Graph Neural Networks and Knowledge Graph Reasoning
I integrate knowledge graphs with neural networks to support reasoning tasks in biomedical contexts. Current projects include using GNN for Alzheimer's disease knowledge graph completion and employing retrieval-augmented reasoning with tools like COLBERT and Weaviate to improve question answering and knowledge extraction over structured biological data.

4. Cross-modal Alignment and Adaptation

I explore theoretical and empirical frameworks that unify low-rank adaptation (LoRA) and geometric alignment techniques to improve domain adaptation across biology and language. This enables transferring knowledge from protein or DNA embeddings into language domains and vice versa, with applications in zero-shot inference and surgical activity recognition.

5. AI for Surgery and Rehabilitation Robotics
In collaboration with clinical partners, I design predictive models to anticipate surgical gestures and subphases and personalize rehabilitation robotics. I develop unified frameworks that treat surgeons and patients as agentic systems, combining sequence modeling (e.g., BLIP-2, JIGSAWS) with agent modeling and intent prediction to enhance clinical workflows.