For fields such as cancer treatment, it is crucial for us to understand how genetic alterations are reflected in tissue histology. A complete understanding of both is critical for tasks such as assessing patient risk and predicting survival outcomes. My project intends to build upon our current knowledge of genotype-phenotype associations by pretraining multimodal models – incorporating both tissue images and genomic information from The Cancer Genome Atlas – in a self-supervised manner, before fine-tuning on the task of survival outcome prediction. I leverage multimodal transformers with a masked pretraining strategy to gain a cohesive genotype-phenotype understanding.

Through this SuperUROP project, I hope to apply my machine learning knowledge (particularly from 6.864 and 6.869) to an advanced research project. I’ m very excited to explore deep learning’ s applications in the medical/biological fields and to continue to hone my technical skills, while making a meaningful contribution to my group. Ultimately, I am hopeful that this project may evolve into (or at least be good experience for) a future MEng thesis.