This project develops a computational framework for analyzing multi-modal spatial proteomics data from 350 lymphoma patients, integrating H&E images, IHC images with 10 protein markers, Hyperion proteomics data, and mutation profiles. To address missing modalities and high-dimensional inputs, we extend scalable graph transformers with a two-stage contrastive learning approach: intra-modality representation learning followed by cross-modality alignment. An attention pooling layer will generate slide-level embeddings, enabling both spot- and slide-level predictions, such as protein marker inference from H&E and mutation prediction from multi-modal data. Beyond predictive tasks, the model aims to support patient stratification and spatial biomarker discovery.

I am participating in SuperUROP because I want to gain research experience in machine learning and computational biology, enabling me to contribute to real-world biomedical challenges and
develop skills that are essential for graduate school and an academic career.