Longitudinal immune programs and association with toxicity burden and antitumor response during immunotherapy.

Abstract

2541 Background: Immune-related adverse events (irAEs) and tumor responses often co-occur during immune checkpoint inhibitor (ICI) therapy. We investigated whether blood immune programs associated with toxicity can be temporally and biologically dissociated from response programs. Methods: ICI-naïve patients with advanced solid tumors (n=37; enriched for melanoma and NSCLC) were profiled at baseline (T0), early on-treatment (T1; week 4), and later timepoints (T2–T3); patients with irAEs had an additional sample at onset (Ttox) before immunosuppression. PBMCs were analyzed by multiparameter flow cytometry and plasma cytokines by 48-plex multiplex assay (log2). We compared timepoints and delta windows (ΔT1–ΔT3), controlled multiple testing with Benjamini–Hochberg (q<0.1), and evaluated baseline predictors from T0 alongside landmark Cox models from T1 for ΔT1 predictors to avoid immortal-time bias; baseline multiple-irAE signatures were tested in penalized multivariable models with clinical covariates. Results: irAEs occurred in 20/37 (54%) patients (grade ≥3: 8/37, 22%); multiple irAEs occurred in 14/37 (38%). Objective response occurred in 15/37 (41%), and 10/15 (67%) responders developed irAEs; median time to first irAE was 86 days (IQR 63–108). Baseline multiple-irAE susceptibility centered on Tfh states (higher Tfh1/Tfh17 PD1+ICOS− and lower Tfh1 PD1−ICOS+), and a MultiTox signature predicted multiple irAEs (OR 5.46; p=0.033). Early dynamics strengthened toxicity prediction: ΔT1 Tfh2 PD1+ICOS− decreased in AnyTox/MultipleTox and predicted subsequent irAEs (HR 0.41; p=0.0019), whereas ΔT1 PDGF-BB (log2) increased risk (HR 2.25; p=0.006). Approaching onset, activated regulatory compartments (activated Treg and Tfr) contracted, followed at Ttox by a surge in IFN-inducible CXCR3 chemokines (CXCL9/MIG p=3.1×10⁻⁵; CXCL10/IP-10 p=0.0021), consistent with a Th1/IFN axis. In contrast, response-associated programs emerged later and reflected a Tfh/B cell–activated CD8 axis (memory B at T1 p=0.036; ΔT2 Tfh2 PD1−ICOS− q<0.1, p=0.0010; ΔT2 eosinophils q<0.1, p=0.0071) and remained independently associated with response (p≈0.02–0.03). Conclusions: Toxicity—particularly high-burden toxicity—appears to reflect a baseline susceptibility that is amplified by early on-treatment immune trajectories and culminates in a chemokine-rich onset state. In contrast, response-associated programs emerge later and are at least partly dissociable from high-burden toxicity. Distinct Tfh states—especially early Tfh2 dynamics—may support timing-informed monitoring and improved benefit–risk stratification during ICI therapy.