Integrative Genetic and Proteomic Modeling for Predicting Chronic and Treatment-Induced Autoimmunity

About This Grant

Project Summary Autoimmune diseases (AID) are complex, chronic conditions with rising global prevalence, contributing substantially to morbidity and healthcare costs. Patients frequently face diagnostic delays due to nonspecific clinical presentations, leading to irreversible tissue damage that could be prevented with earlier detection. Additionally, autoimmunity manifests as a potentially severe side effect of cancer immunotherapy, termed immune-related adverse events (irAE), which can complicate treatment and decrease patient survival rates. While polygenic risk scores (PRS) effectively identify individuals with elevated inherited risk, they fail to capture the dynamic immune states that dictate the timing of disease onset. Conversely, plasma proteomics provides real-time molecular signatures of inflammation and immune regulation but is rarely modeled jointly with genetics at a population scale. The overarching objective of this proposal is to create and interpret state-of-the-art risk prediction models for both chronic AID and irAE by integrating static genetic susceptibility with dynamic proteomic activity. Our preliminary data demonstrate that combining proteomic risk scores with PRS improves predictive performance for inflammatory bowel disease and reveals interactions between genetic load and protein levels. The central hypothesis is that integrating PRS with plasma protein levels will enhance predictive accuracy over either modality alone for multiple AID phenotypes and that these biobank-derived signatures can transfer to the context of immunotherapy toxicity. To test this, Aim 1 will utilize the UK Biobank Pharma Proteomics Project to build and benchmark machine learning models for six major chronic AIDs, ranging from interpretable regression to non-linear neural networks. Aim 2 will assess the generalizability of these models across autoimmune phenotypes by transferring them to a prospective cohort of cancer patients undergoing immune checkpoint inhibitor therapy. This aim will employ models built in a traditional supervised learning framework as well as models trained on the low-dimensional proteomic embeddings of a self- supervised masked autoencoder, with the goal of enabling robust risk stratification in a limited- sample setting. Together, these studies will bridge the domains of spontaneous and drug- induced autoimmunity. This work has the potential to improve clinical care by enabling earlier identification of high-risk individuals and defining shared and disease-specific predictive biomarkers across autoimmune conditions.