Multimodal immunoprofiling of peripheral blood using foundation models of the immune system for predicting immunotherapy response and toxicity in the RADIOHEAD pan-cancer cohort.

Abstract

2533 Background: While immune checkpoint inhibition (ICI) is an emerging gold standard for cancer therapy, positive response is limited among treated patients and up to 70% experience toxicity. Early and accurate response prediction that accounts for immune-related adverse events (irAEs) is crucial for identifying patients who may benefit from ICI. We present AI-powered approaches for predicting ICI response by transcriptional and cellular profiling of blood immune cells from a pan-cancer cohort. Methods: Peripheral blood mononuclear cells (PBMC) were isolated at pre- (baseline) and early on-treatment for flow cytometry and RNA-seq profiling from the RADIOHEAD cohort (Quandt et al. 2025) receiving ICI (n=1,070). Patients were clustered based on variational autoencoder embeddings for real-world progression-free survival (rwPFS) and irAEs derived by a peripheral immune system encoder trained on the BostonGene patient database (n=45,000). The logrank test and Fisher’s exact test were used to analyze survival and compare irAE frequencies between clusters, respectively. RNA-seq trajectory features were identified using hierarchical clustering, along with elastic net-regularized and simple multivariate Cox regression models for feature selection. Results: Pre-trained immunotype models (Dyikanov et al. 2024) applied to baseline PBMC revealed that G2-primed (memory CD4+ T cell-enriched) and G5-suppressive (monocyte enriched) immunotype scores stratified patients into responders (R) and non-responders (NR) (p = 0.00001). T cell receptor (TCR) dynamics revealed a significantly greater decrease in TCR diversity in NR during treatment (p = 0.046). We discovered a baseline gene set containing immune checkpoint and cancer antigen genes as well as a longitudinal trajectory set of monocyte and myeloid cell activation markers that both stratified patients by rwPFS (p=0.006; 0.03). Trained immune system embeddings identified a novel severe-risk patient group with both a high irAE incidence (p = 0.025) and short rwPFS. This group displayed both active inflammatory and tolerance pathways that stratified patients with irAEs by rwPFS (p = 0.003). Conclusions: Using pre-trained multimodal immune system projections, we 1) independently confirmed the association of peripheral immunotypes with ICI response; and 2) identified a novel severe-risk signature from patients with high irAE incidence (≥Grade 3) and short rwPFS. We also found that greater TCR diversity and T cell differentiation were associated with ICI response, while innate myeloid activation and trafficking correlated with non-response. Our unique AI-driven analytical framework underscores the potential of peripheral blood immunoprofiling for ICI treatment selection and patient stratification in prospective trials.