🧪 Research Experience

Washington University AIHealth, School of Medicine / CSPL

PIs: Chenyang Lu, Joanna Abraham
Feb 2023 – Dec 2024

🩺 surgVAE – Generative multi-task model for cardiac surgery outcome prediction (Lead)

• Probabilistic VAE-based framework for preoperative prediction of six critical postoperative complications in high-risk cardiac surgery patients, leveraging cross-surgery EHR (89k+ cases), disentangled latent spaces, and multi-task learning to improve both performance and interpretability on modifiable risk factors.

🌈 Deep learning for broadband spectrum reconstruction (Contributor)

• Optical conv2seq model for reconstructing broadband spectra from crystal responses, with GPT-style baselines and noise-robustness experiments.

University of Washington, Seattle

PI: Sheng Wang
Jun 2024 – Nov 2024

🧬 Slide-level pretraining for pathology WSI FMs (Co-Lead)

• Developed slide-level pretraining for pathology FMs by representing each WSI as a variable-length sequence of ViT tile embeddings and training a slide encoder with DINO/DINOv2-style self-distillation.

🩻 CT segmentation foundation models (Contributor)

• Implemented SAM, MedSAM, and SegVol pipelines for CT segmentation foundation models.

Chinese University of Hong Kong, CSE

PI: Yu Li
Jun 2022 – Dec 2024

🧫 USPNet – Unbiased signal peptide prediction with protein language models (Lead)

• Organism-agnostic SP prediction framework leveraging MSA Transformer and ESM2 with a label distribution–aware margin loss to handle severe class imbalance and domain shift; outperforms prior methods on SP classification and cleavage prediction and discovers hundreds of novel SP candidates from metagenomes.

🧬 Genome–drug foundation model for AMR prediction (Lead)

• Genome–drug paired FM that predicts AMR phenotypes from full bacterial genomes and antibiotic structures using genomic LLM pretraining, multi-modal binding, and multi-instance learning.

💊 DNA-FM + molecular-FM for antibacterial drug combination synergy (Co-Lead)

• Framework that predicts synergy of antibacterial drug combinations by combining pre-trained molecular graph models for drugs with genomic foundation models for full bacterial genomes.

🧵 LLM-based multi-agent system for AMP discovery (Co-Lead)

• LLM-based multi-agent pipeline that automates discovery of novel antimicrobial peptides from human metagenomic data by orchestrating sequence models, AMP/SP predictors, and filtering tools into a unified workflow.