Biomarker Discovery for Early Prediction of Autoimmunity in Immunotherapy Patients Through Deep Immune Profiling and Temporal Graph Convolutional Networks

Using temporal graph machine learning and deep immune profiling to predict which cancer immunotherapy patients will develop autoimmune adverse events before they occur.
Completed Other Biomarker Discovery Data-Driven & Quantitative Early Disease Detection Experimental Platforms & Models Human Cohorts Immune Profiling Machine Learning & AI Precision Medicine Real-world Evidence Translational & Clinical

Funding Year: 2023

Project Overview

Immune-related adverse events (irAEs) are serious, sometimes life-threatening autoimmune toxicities that arise during cancer immunotherapy, and early identification of at-risk patients could enable prompt prevention or treatment. This project built a prospective irAE cohort of 340 patients on checkpoint inhibitor drugs with deep immune profiling and rich clinical data, alongside a discovery cohort of over 8,000 patients in PennChart. Temporal graph machine learning — which handles irregular structure and complex time dependencies in clinical data — is being applied to predict whether irAEs will occur. An EHR phenotyping algorithm for detecting irAEs has been implemented to support larger AI/ML cohort construction, with refinement and automated acquisition of an augmented cohort ongoing as the basis for an R01 grant proposal.

Impact & Innovation

Predicting autoimmune toxicity before immunotherapy causes harm.

 

By combining the largest known irAE cohort with complete deep immune profiling with temporal graph ML, this project builds a predictive framework for identifying cancer patients at risk of autoimmune adverse events — before symptoms appear.

  • Constructs the largest existing irAE cohort with complete deep immune profiling across multiple time points, establishing a foundational dataset for irAE biomarker discovery and AI model training
  • Generates open-source code under MIT license on GitHub, with AI models in follow-up research to be assessed for commercialization potential through the Penn Center for Innovation
  • Advances the Consortium’s Integrated Data and Discovery Platforms pillar by developing a replicable EHR phenotyping and temporal graph ML framework applicable to irAE prediction and broader autoimmune surveillance
Research Approach

A framework designed for discovery

This project combines prospective deep immune profiling, large-scale EHR data, and temporal graph machine learning to develop and validate predictive biomarkers for immune-related adverse events in cancer immunotherapy patients. The work integrates cohort construction, algorithm development, and AI model training toward R01-level grant proposals and clinical translation.

Prospective collection and deep immune profiling of blood samples from 340 checkpoint inhibitor patients at multiple time points; construction of a discovery cohort of 8,000+ patients in PennChart with all structured clinical observations; implementation and refinement of an EHR phenotyping algorithm to detect irAEs; and development and training of temporal graph machine learning models to predict irAE occurrence from high-dimensional immune and clinical data.

Prospective irAE cohort of 340 patients with deep immune profiling and rich clinical data; PennChart EHR discovery cohort of 8,000+ checkpoint inhibitor patients; structured clinical observations and checkpoint inhibitor drug data from Penn Medicine; and temporal graph ML model training datasets incorporating immune profiling across multiple complex time points.

Development of a validated temporal graph ML model for early irAE prediction, an EHR phenotyping algorithm for large-scale irAE detection, and a deep immune profiling dataset enabling mechanistic discovery. Findings will support NIH R01 grant proposals and continued Colton funding, with AI models assessed for commercialization through the Penn Center for Innovation.

Investigators & Institutions

Powering the science

Principal Investigators

Joseph Romano, PhD, Colton Consortium Member

Assistant Professor, Department of Biostatistics, Epidemiology, and Informatics (Informatics), Perelman School of Medicine, University of Pennsylvania

Sokratis Apostolidis, MD, PhD, Colton Consortium Member

Assistant Professor, Department of Medicine (Rheumatology), Perelman School of Medicine, University of Pennsylvania

Colton Centers & Institutions

Colton Center for Autoimmunity at the University of Pennsylvania

Perelman School of Medicine

Research Outputs

From insight to impact

Publications

Chimeric antigen receptor T cells against the IGHV4-34 B cell receptor specifically eliminate neoplastic and autoimmune B cells

Science Translational Medicine
Cohen, IJ; Bochi-Layec, AC; Lemoine, J; Jenks, S; Bayat, P; Kim, KH; Zhao, H; Ugwuanyi, O; Stella, F; Ghilardi, G; Gabrielli, G; McCuaig, S; Iatrou, A; Vlachonikola, E; Karipidou, M; Bouziani, E; Espie, D; Ramasubramanian, R; Agathangelidis, A; Bhosale, A; Paruzzo, L; Medico, G; Kolar, B; Bugrovsky, R; Guruprasad, P; Wang, LP; Harris, J; Arons, E; Zhang, Y; Pajarillo, R; Kreiger, PA; Day, CP; Sahinalp, SC; Wu, CH; Santi, A; Fulmer, B; Cases, M; Palmer, MB; Porazzi, P; Wherry, EJ; Kreitman, RJ; Tiacci, E; Apostolidis, SA; Behrens, EM; Bhoj, V; Sanz, I; Inghirami, G; Schuster, SJ; Ghia, P; Stamatopoulos, K; Ruella, M February 2026
Adaptive ImmunityAnimal ModelsAutoantibodiesB Cell BiologyBiological & MechanisticExperimental Platforms & ModelsIn Vitro ModelsPrecision MedicineTherapeutic DevelopmentTranslational & ClinicalOtherSystemic DiseasesSystemic Lupus Erythematosus (SLE)University of Pennsylvania

Additional Outputs

All code developed as part of this research is open source and freely available on GitHub under the MIT license. AI models in follow-up research to be assessed for commercialization potential through the Penn Center for Innovation (PCI).

rAE EHR phenotyping algorithm paper; clinical applications paper on AI-identified mechanisms of irAEs; paper on immune mechanisms of irAEs (submission planned Fall 2025).