Date: Friday, May 9, 2025
Time: 11:00 AM – 12:00 PM
Room: Building 9‑354, MIT
CSB Ph.D. Candidate: Ifrah Tariq
Supervisor: Ernest Fraenkel (BE)
TDC Members: Caroline Uhler (chair), Nir Hacohen, Manolis Kellis
Title: Biologically Interpretable Representation Learning for Mechanistic Insights into Cancer Immunotherapy Resistance
Abstract: Immune checkpoint inhibitors (ICIs) have revolutionized cancer treatment, yet therapeutic resistance remains a formidable challenge across diverse tumor types. This thesis presents a multi-faceted investigation into the molecular underpinnings of immunotherapy response and resistance, integrating functional genomics, multi-omic data processing, and interpretable machine learning. First, we detail genome-scale perturbation screens that identify genetic determinants of tumor sensitivity to natural killer (NK) cells, revealing key immune-evasive mechanisms. We then outline the process of aggregating and harmonizing data from multiple clinical studies investigating ICI resistance. Finally, we present the Biologically Disentangled Variational Autoencoder (BDVAE), an interpretable deep learning model that integrates RNA-seq and whole-exome sequencing data to uncover mechanisms of response and resistance to immune checkpoint blockade across multiple cancers. The model leverages a pathway-informed architecture to learn biologically meaningful latent dimensions. The BDVAE accurately predicts clinical outcomes and disentangles immune, tumor-intrinsic, metabolic, and neuroimmune programs. It also identifies an intermediate tumor subgroup with distinct survival patterns, offering insights beyond conventional response labels.