Biology-Guided Representation Learning for Computational Pathology
Name
vaidya-avaidya-phd-hst-2026-thesis.pdf
Size
86.47 MB
Format
Adobe PDF
Checksum (MD5)
7cfe5212edbc242c0a2b747e822ce970
Author(s)
Vaidya, Anurag Jayant
Advisor(s)
Mahmood, Faisal
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
Cancer diagnosis has been grounded in tissue morphology for over a century, yet the molecular programs that drive disease progression, treatment response, and patient outcome are not directly visible in standard histology slides. Current foundation models for computational pathology learn representations primarily from morphology, without access to the underlying biology of mutations, gene expression, or protein states. This thesis investigates whether encoding biological structure directly into computational pathology models can produce stronger representations than those learned from morphology alone. The thesis introduces biology as a learning signal at multiple levels of the computational pathology pipeline. At the task level, structuring gene expression into biological pathways and fusing them with histology produces more interpretable and accurate survival predictions than unstructured approaches. At the representation level, a demographic bias audit reveals that encoder quality is the most consequential lever among the design choices evaluated, motivating a shift from task-specific integration of biological information toward using biology directly as a signal for representation learning for computational pathology encoders. At the pretraining level, three self-supervised methods learn slide-level representations by aligning histology with biological structure: immunohistochemistry stains for organ-specific pretraining, paired genomic and transcriptomic profiles for pan-cancer pretraining, and spatial protein markers for a new imaging modality beyond H&E histology. These results provide evidence that biological structure, whether encoded as pathways, stains, genomes, or protein markers, is an effective and transferable learning signal for computational pathology. Whether biology-guided representations can improve patient outcomes when deployed in clinical decision-making, and how such systems should be integrated into diagnostic workflows, validated across diverse populations, and regulated in practice, remain open questions that span science, engineering, and health policy.
MIT Department
Harvard-MIT Program in Health Sciences and Technology
Terms of Use
Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)
Copyright retained by author(s)
Persistent DSpace Link
