Machine Learning–Based Insights into Circulating Autoantibody Dynamics and Treatment Outcomes in Patients with NSCLC Receiving Immune Checkpoint Inhibitors

Feifei Wei¹Hiroyuki Takeda²Koichi Azuma³Yoshiro Nakahara⁴Yuka Igarashi⁴Kenta Murotani³Haruhiro Saito⁴Shuji Murakami⁴Tetsuro Kondo⁴Taku Kouro⁴Hidetomo Himuro⁴Kayoko Tsuji⁴Mitsuru Komahashi⁵Tatsuya Sawasaki^2*Tetsuro Sasada^4*

• ¹Kanagawa Cancer Center Research Institute, Yokohama, Japan
• ²Ehime University, Matsuyama, Japan
• ³Kurume University School of Medicine, kurume, Japan
• ⁴Kanagawa Cancer Center, yokohama, Japan
• ⁵Nihon University School of Medicine, tokyo, Japan

Introduction Immune checkpoint inhibitors (ICIs) targeting the programmed death-1/ligand-1 (PD-1/PD-L1) axis have significantly improved treatment outcomes in non-small cell lung cancer (NSCLC); however, challenges remain owing to the limited durability of therapeutic responses and the occurrence of immune-related adverse events (irAEs). This study aimed to characterize dynamic changes in the circulating autoantibody (CAAB) profile during ICI treatment and explore their association with treatment outcomes in patients with NSCLC. Methods A panel of 59 CAABs showing substantial treatment-related changes was initially identified using AlphaScreen assays in a primary screening of five patients who developed ir-pneumonitis. These CAABs were subsequently profiled in paired pre-and post-treatment plasma samples obtained from 179 patients with NSCLC treated with anti-PD-1/PD-L1 therapy at two Japanese centers. Associations between CAAB dynamics and clinical parameters—including baseline characteristics, treatment regimens, and treatment outcomes (irAEs, ir-pneumonitis, response, progression-free survival [PFS], and overall survival [OS])—were evaluated using permutational multivariate analysis of variance and univariate binary logistic and Cox regression, elastic net regularization regression, and random forest regression. Results Using permutational multivariate analysis of variance and univariate binary logistic/Cox regression, we comprehensively assessed the global associations between CAAB dynamics and eight clinical parameters, including background factors (PD-L1 expression and treatment line), treatment regimens (chemotherapy exposure), and treatment outcomes (irAE occurrence, ir-pneumonitis development, RECIST-assessed response, PFS, and OS), indicating that chemotherapy exposure was the only significant and strong factor influencing CAAB dynamics. In patients receiving ICI monotherapy, univariate logistic or Cox regression analyses were performed to identify individual CAABs significantly associated with each outcome, highlighting both shared and distinct immunological features underlying different clinical endpoints. Through machine learning-based evaluation of the predictive potential of CAAB dynamics for five treatment outcomes across the overall cohort and six subgroups defined by three stratification variables, four optimized CAAB signatures with robust predictive performance for ICI treatment outcomes were established. Conclusions These findings suggest the involvement of distinct immune pathways in therapeutic benefits and toxicity. Collectively, our results provide mechanistic insights into ICI-induced humoral immune regulation, highlight the potential utility of CAABs as biomarkers to enhance benefit-to-risk assessment, and guide the development of personalized immunotherapy strategies for NSCLC.