Background/Purpose: Immune checkpoint inhibitors (ICIs), while revolutionary in cancer treatment, may induce severe or even fatal immune-related adverse event (irAE). The mechanisms underlying irAE development remain unclear, and early identification of predictive biomarkers for irAE is urgently needed to improve risk stratification and clinical management. However, most current studies on irAE-related potential biomarkers predominantly conduct cross-sectional analyses at a single time point, leaving a critical cap in tracking dynamic changes predicting irAE onset.

Methods: This prospective study enrolled 60 lung cancer patients receiving ICIs therapy at Peking Union Medical College Hospital from January 2022 to September 2024. Additionally, 23 irAE patients with various cancers were included at the time of irAE onset as an observational cohort. Serum samples collected longitudinally at baseline (pre-treatment), 1-3 weeks, 3-6 months post-treatment and the onset of irAE after starting ICIs were analyzed by Olink proximity extension assay technology. Advanced analytical approaches included time-series clustering, linear mixed-effects models, and differential expression analysis across multiple time points. Predictive performance of candidate biomarkers was assessed using receiver operating characteristic (ROC) analysis and support vector machine (SVM) classifiers integrating temporal features.

Results: In the prospective cohort, 26 out of 60 patients (43.3%) developed irAE within 6 months after ICIs initiation. Longitudinal proteomic analysis revealed distinct temporal profiles between irAE and non-irAE groups, with six proteins (IL10, NT-3, CCL23, FGF-19, LIF-R and uPA) exhibiting significantly different trajectories (Fig.1). At baseline, 10 serum proteins (e.g., CXCL1, CXCL6, CXCL11, IL6 and HGF) were significantly depressed in irAE patients, with moderate predictive accuracy for irAE occurrence (AUC > 0.6). Besides, patients with new-onset irAE exhibited significantly elevated CCL25, IL4, IL10 and IL20, and decreased IL-15RA and CD8A l compared to non-irAE patients. Longitudinal fold-change analyses from baseline to irAE onset further identified 24 significantly differential proteins, with CXCL9, IL33, IL-18R1, FGF-19 and FGF-23 showing strong predictive power (AUC > 0.8). Furthermore, machine learning models based on SVM integrating temporal features achieved high accuracy in predicting irAE development (AUC > 0.9) (Fig.2), suggesting that these protein-based models provide robust predictive power for irAE occurrence.

Conclusion: Our study provides valuable insights into the dynamic temporal immunological patterns associated with the development of irAE. The identification of temporally regulated biomarkers and the development of predictive models based on machine learning offer promising tools for early risk prediction, diagnostic stratification, and longitudinal monitoring of therapeutic response. These findings advance our understanding of irAE pathogenesis and may inform future biomarker-driven approaches to individualized patient management and targeted therapeutic interventions.

Fig.1. Longitudinal serum proteomic clustering analysis in irAE and non-irAE patients. Mfuzz analysis of differential proteins expression between irAE and non-irAE groups and heatmap of Z-score–normalized expression profiles across three time points (T0, T1, T2) for proteins grouped in each cluster. Linear mixed-effects regression analysis of 6 significant proteins for the three time points in two groups. irAE, immune-related adverse event; NPX, normalized protein expression; T0, before the initiation of ICIs treatment; T1, 1-3 weeks after treatment; T2, 3-6 months after treatment. The y axis represents the Olink assay value (NPX). P-values were calculated using linear mixed-effects models.

Fig.2. The composite biomarkers performance for predicting irAE in train and valid set. ROC analysis showed the composite model performance of ten differentially expressed proteins at baseline (A), six differentially expressed proteins at irAE onset (B) and top six differentially expressed proteins at the onset of irAE compare to at baseline (C) in train and test set. irAE, immune-related adverse event; ROC, Receiver operating characteristic; AUC, area under the curve.

« Back to ACR Convergence 2025

ACR Meeting Abstracts - https://acrabstracts.org/abstract/longitudinal-proteomic-signatures-for-prediction-of-immune-related-adverse-events-during-immune-checkpoint-inhibitors-therapy-a-machine-learning-guided-prospective-biomarker-discovery-study/