Building agentic AI systems that reason across language, molecules, images, and the physical world.
PhD Candidate @ University of Pennsylvania
CV last updated: Nov 2025
I am a PhD student in Bioengineering at University of Pennsylvania, advised by Prof. Cesar de la Fuente and Prof. Mark Yatskar. Previously, I received my bachelor's degree in Artificial Intelligence from Huazhong University of Science and Technology. I am fortunate to work under the supervision of Prof. Jianyang Zeng at Westlake University, Prof. Xiaohui Xie in University of California, Irvine and Prof. Dongrui Wu in Huazhong University of Science and Technology.
In the era of large language models, my research objective is to build models that can interact with and truly understand the real world—both the macroscopic world of physical objects and the microscopic 3D molecular world—so as to develop a knowledgeable, reliable AI biochemical expert capable of autonomous scientific discovery. To this end, my previous work has centered on:
If you would like to discuss my research or potential collaborations further, feel free to contact me. I'm open to connect and collaborate with each other.
The paper introduces LORE, a framework that defines how large reasoning models should scale their thinking and accuracy with problem complexity. It builds a benchmark (LORE-BENCH) to test two key properties—monotonicity and compositionality—and finds current models are mostly monotonic but not compositional; fine-tuning to enforce the compute rule improves overall reasoning performance.
ApexOracle centers on a discrete diffusion language model (DLM) that does two jobs at once: it learns chemistry-aware features of molecules and it can run the diffusion process backward to generate new antimicrobial candidates. This DLM backbone is then plugged into a multimodal architecture that also ingests pathogen genomics (Evo2) and pathogen text knowledge (Me-LLaMA), and cross-attends these with the molecule features. That multimodal fusion is what lets ApexOracle stay effective on unseen or future strains: given a new genome or description, it can both predict activity and guide the DLM to sample a molecule that should work for that specific pathogen.
Here we argue that resource-efficient hybrid RNN deep-learning models can generatively design self-assembling peptides with strong self-organization, underscoring AI's expanding role in biomaterials discovery.
This study introduces a genomics-guided pipeline to uncover a vast reservoir of class II bacteriocins from the human gut microbiome. These unmodified peptides demonstrate potent, narrow-spectrum antimicrobial activity with minimal disruption to the commensal microbiota, offering a promising avenue for precision microbiome engineering and combating antimicrobial resistance.
We introduce three novel modules for foundational vision models for efficient few-shot medical image segmentation, experiments on popular abdominal CT and MRI dataset showed average improvements of 10.21% and 1.80% in terms of DSC, respectively.
Peptide to SMILES is not the hard part anymore. PepLink goes further: it supports non-canonical amino acids, diverse cyclic peptide structures, and can even reverse-parse cyclic peptide SMILES back into amino acid sequences. Built for real peptide workflows, not just toy examples.
Getting started is easy: pip install PepLink — or try it instantly in the browser via the Web UI.
Utilizing insights from BERT and MAE, we developed a sequence classification model to accurately predict students' performance through a series of gameplay actions, earning us a prestigious bronze medal (top 5.6%).
Utilizing Euclidean Alignment and Independent Component Analysis (ICA), we generated a substantial volume of high-quality data, successfully mitigating the over-fitting issue prevalent in EEG based Brain-Computer Interfaces. Impressively, I secured the 20th position among 283 global teams predominantly comprised of master's and Ph.D. students.
By incorporating a lightweight Convolutional Block Attention Module into ArcFace and optimizing the training with challenging images isolated using MCTNN, I successfully honed the network's focus on distinguishing features such as the eyes and hairline, thereby enhancing the performance by a notable 4%.
By adeptly merging a transformer-based tracker with YOLOv5, I successfully developed a system capable of mitigating player and ball occlusion issues, enabling automatic and precise tracking of the small soccer ball at real-time FPS through a whole game.
pip install PepLink or try the Web UI and have fun.