Madabhushi, Anant

8549 Background: Despite two decades of anti-angiogenic therapy in NSCLC, no predictive biomarker identifies which patients benefit. As VEGF-targeted combinations advance in clinical development, patient selection biomarkers remain a critical unmet need. QVT (Quantitative Vessel Tortuosity) Score is an automated imaging biomarker measuring chaotic tumor vasculature from baseline CT scans and has been shown to be associated with immune checkpoint inhibitor (ICI) outcomes. We hypothesized that QVT Score could identify patients with chaotic angiogenesis tumors and greater sensitivity to anti-angiogenic therapy. Methods: We analyzed a subset of 334 patients from the SWOG S0819 (NCT00946712) trial with treatment-naïve stage IV non-squamous NSCLC receiving carboplatin/paclitaxel with (56%) or without (44%) bevacizumab. QVT Scores were derived from baseline CTs based on radiologist-defined tumor annotations and radiomic features of tumor vasculature (e.g. vessel twisting, curvature, and branching). Overall survival (OS) was evaluated as the primary endpoint using Cox proportional hazards models with QVT × bevacizumab interaction terms. Bevacizumab was non-randomized in S0819 (physician/patient discretion): to address this, we employed doubly robust estimation combining inverse probability of treatment weighting (IPTW) with multivariable covariate adjustment for age, performance status, histology, smoking history, and stage. Results: The QVT Score × bevacizumab interaction was significant (p=0.007, doubly robust estimation) and stayed consistent across sensitivity analyses (p=0.017 multivariable; p=0.047 IPTW), suggesting differential treatment benefit based on pre-treatment vascular phenotype. QVT Score was strongly OS-associated without bevacizumab (HR=3.03, p=0.003) but not with bevacizumab (HR=1.71, p=0.079), consistent with mitigation of high-risk biology. Bevacizumab benefit increased across QVT quartiles (Table 1): patients in the highest quartile had a 60% reduction in mortality (HR=0.40, p<0.001), while patients in the lowest quartile showed no OS benefit (HR=0.89, p=0.64). Conclusions: Despite unfavorable prognosis on standard-of-care therapies (chemotherapy, ICIs), we found that patients with elevated baseline QVT Score may benefit from the addition of anti-angiogenic agents. As VEGF-targeted combinations including bispecific antibodies enter development, radiomic biomarkers may play a crucial role in enriching trial populations and enabling mechanistic longitudinal monitoring. These findings warrant prospective validation in independent randomized clinical trials. Bevacizumab treatment effect by QVT score quartile. QVT Score Quartile N Bevacizumab effect (Hazard Ratio) P Q1 84 0.89 (0.56 -1.42) 0.64 Q2 83 0.70 (0.44-1.12) 0.14 Q3 83 0.54 (0.34-0.85) 0.01 Q4 84 0.40 (0.24-0.65) 0.0003

8534 Background: Immunotherapy (IO) has transformed treatment for non–small cell lung cancer (NSCLC), but not all patients (pts) benefit, underscoring the need for predictive biomarkers to guide IO versus chemotherapy (Ch). CheckMate 227 (CM227; NCT02477826) showed a survival benefit of first-line nivolumab plus ipilimumab (nivo+ipi) over Ch in stage IV NSCLC. CD8⁺ T cells are key mediators of antitumor immunity and are associated with IO benefit. We developed an artificial intelligence (AI)–based pipeline (VIGOR-CD8) that predicts spatial CD8⁺ immune signatures from routine baseline H&E whole-slide images using histopathology foundation model embeddings (H-Optimus-0) and virtual gene expression modeling (HE2Gene), and evaluated its ability to identify pts who do and do not derive IO benefit over Ch in CM227. Methods: 1,598 pts with advanced NSCLC were analyzed, including multi-institutional retrospective cohorts (n = 487; 65 for patch-level CD8⁺ prediction and overall survival (OS), 422 for patient-level OS) and a blinded CM227 validation subset (n = 1,111). For orthogonal validation, 86,470 H&E patches were co-registered with quantitative CD8⁺ immunofluorescence. H-Optimus-0 and HE2Gene immune-related embeddings were used to train a random forest classifier to predict patch-level CD8⁺ probability. Patch-level probabilities were aggregated into a patient-level CD8⁺ signature and dichotomized into biomarker-positive (B⁺) and biomarker-negative (B⁻) groups by the training-set median. Cox models assessed the impact of VIGOR-CD8 on OS. In CM227, prognostic and predictive utility were evaluated using treatment-specific analyses; investigators were blinded to outcomes, and models were trained on independent, non-overlapping cohorts. Results: VIGOR-CD8 was associated with longer OS in the testing cohort (n = 422; HR 0.68, 95% CI 0.53–0.87, p = 0.00163) and CM227 (n = 1,111; HR 0.8, 95% CI 0.67–0.96, p = 0.016), irrespective of treatment type and PD-L1 expression. Among pts with evaluable PD-L1, B⁺ pts treated with nivo+ipi had superior OS versus Ch (n = 617; HR 0.72, 95% CI 0.58–0.90, p = 0.003), while in B⁻ pts there was no significant OS difference between treatment arms (n = 130; HR 1.18, 95% CI 0.79–1.75, p = 0.436), supporting a predictive rather than purely prognostic role for VIGOR-CD8. Conclusions: An AI-derived CD8⁺ immune signature from routine baseline H&E slides was associated with favorable OS in CM227 and predicted differential benefit from nivo+ipi versus Ch. VIGOR-CD8 may help identify advanced NSCLC pts most likely to benefit from first-line dual IO, but further validation in independent and prospective trials is warranted.

Background The landmark study of durvalumab as consolidation therapy in NSCLC patients (PACIFIC trial) demonstrated significantly longer progression-free survival (PFS) in patients with locally advanced, unresectable non-small cell lung cancer (NSCLC) treated with durvalumab (immunotherapy, IO) therapy after chemoradiotherapy (CRT). In clinical practice in the USA, durvalumab continues to be used in patients across all levels of programmed cell death ligand-1 (PD-L1) expression. While immune therapies have shown promise in several cancers, some patients either do not respond to the therapy or have cancer recurrence after an initial response. It is not clear so far who will benefit of this therapy or what the mechanisms behind treatment failure are. Methods A total of 133 patients with unresectable stage III NSCLC who underwent durvalumab after CRT or CRT alone were included. Patients treated with durvalumab IO after CRT were randomly split into training (D1=59) and test (D2=59) sets and the remaining 15 patients treated with CRT alone were grouped in D3. Radiomic textural patterns from within and around the target nodules were extracted. A radiomic risk score (RRS) was built and was used to predict PFS and overall survival (OS). Patients were divided into high-risk and low-risk groups based on median RRS. Results RRS was found to be significantly associated with PFS in D1 (HR=2.67, 95% CI 1.85 to 4.13, p<0.05, C-index=0.78) and D2 (HR=2.56, 95% CI 1.63 to 4, p<0.05, C-index=0.73). Similarly, RRS was associated with OS in D1 (HR=1.89, 95% CI 1.3 to 2.75, p<0.05, C-index=0.67) and D2 (HR=2.14, 95% CI 1.28 to 3.6, p<0.05, C-index=0.69), respectively. RRS was found to be significantly associated with PFS in high PD-L1 (HR=3.01, 95% CI 1.41 to 6.45, p=0.0044) and low PD-L1 (HR=2.74, 95% CI 1.8 to 4.14, p=1.77e-06) groups. Moreover, RRS was not significantly associated with OS in the high PD-L1 group (HR=2.08, 95% CI 0.98 to 4.4, p=0.054) but was significantly associated with OS in the low PD-L1 group (HR=1.61, 95% CI 1.14 to 2.28, p=0.0062). In addition, RRS was significantly associated with PFS (HR=2.77, 95% CI 1.17 to 6.52, p=0.019, C-index=0.77) and OS (HR=2.62, 95% CI 1.25 to 5.51, p=0.01, C-index=0.77) in D3, respectively. Conclusions Tumor radiomics of pretreatment CT images from patients with stage III unresectable NSCLC were prognostic of PFS and OS to CRT followed by durvalumab IO and CRT alone.

The aim of this study is to evaluate whether NIS radiomics can distinguish lung adenocarcinomas from granulomas on non-contrast CT scans, and also to improve the performance of Lung-RADS by reclassifying benign nodules that were initially assessed as suspicious. The screening or standard diagnostic non-contrast CT scans of 362 patients was divided into training (St, N = 145), validation (Sv, N = 145), and independent validation (Siv, N = 62) sets from different institutions. Nodules were identified and manually segmented on CT images by a radiologist. A series of 264 features relating to the edge sharpness transition from the inside to the outside of the nodule were extracted. The top 10 features were used to train a linear discriminant analysis (LDA) machine learning classifier on St. In conjunction with the LDA classifier, NIS radiomics classified nodules with an AUC of 0.82 ± 0.04, 0.77, and 0.71 respectively on St, Sv, and Siv. We evaluated the ability of the NIS classifier to determine the proportion of the patients in Sv that were identified initially as suspicious by Lung-RADS but were reclassified as benign by applying the NIS scores. The NIS classifier was able to correctly reclassify 46% of those lesions that were actually benign but deemed suspicious by Lung-RADS alone on Sv.