Abstract 3358: Tertiary lymphoid structures (TLS) estimated by AI tools from digital H&E slides significantly enhance survival prediction in high-risk AJCC stages III/IV melanoma

Abstract Tertiary lymphoid structures (TLS) in tumors have emerged as novel disease prognostic and predictive biomarkers with immunotherapy. Recent advances in whole-slide image (WSI) digitization of pathology slides and AI tools are making in situ prediction of TLS more feasible. Methods: We applied HookNet-TLS, a deep learning-based automated TLS quantification tool, to identify and quantify TLS and germinal centers (GC) in digitized H&E slides from 376 patients (pts) with resected cutaneous melanoma enrolled in the E1609 phase 3 trial that tested adjuvant ipilimumab (IPI) vs interferon-α (IFN). We tested the prognostic value of TLS by correlating the presence of TLS and the normalized number of TLS (number of TLS divided by total area of segmented tissue) in E1609 pts with AJCC7 stages IIIB (n=180), IIIC (n=151), M1a (n=30), M1b (n=15), by stage group, sex, age, treatment and tumor ulceration, utilizing Kaplan-Meier analysis and log-rank tests. Results: TLS were present in the tumors of 207 pts (55%) and predicted better overall survival (OS) (P = 0.09), especially in those with >1 TLS (P = 0.0096). TLS density (TLS per unit area) was even more prognostic (P = 0.0036 for median cutoff; P = 0.0012 for optimal cutoff). Presence of both TLS and GC within TLS (n = 114) demonstrated significantly better OS compared to those without GC (P = 0.022). Higher TLS scores were associated with significantly better OS in IIIB (P = 0.021) and IIIC (P = 0.019) with a similar trend towards improved recurrence free survival (RFS). Associations in M1a and M1b were limited by sample size, although improvement in RFS was seen in M1a with higher TLS (P = 0.07) and presence of GC (P = 0.007). By treatment, higher TLS predicted better OS (P = 0.033) and RFS (P = 0.018) among IPI treated patients (n=246). A similar trend in OS was seen among IFN treated patients (n=130) that did not reach significance. By sex, higher TLS & presence of GC were associated with better OS (P = 0.0062 & P = 0.072) and RFS (P = 0.01 & P = 0.04), respectively in males (n=226). By age, higher TLS was associated with better OS among pts <65 (n = 310) (P = 0.0002). Similar trends in females (n = 150) and age ≥65 (n = 66) that were not significant, likely limited by sample size. Higher TLS predicted better OS and RFS in pts with non-ulcerated melanoma (n=138) (P = 0.001 & P = 0.022) but not those with ulceration (n=165). Conclusions: Automated TLS quantification using a deep learning tool (HookNet) effectively identified and characterized TLS in H&E slides where the application of TLS significantly enhanced survival prediction by AJCC stage group, sex, age, treatment and tumor ulceration. These findings highlight the potential of AI-driven approaches to standardize TLS assessment using low-cost H&E staining, with the potential to improve prognostic stratification within AJCC, that warrants further investigation and validation. Citation Format: Ahmad A. Tarhini, Tingyi Li, Ranran Tao, Sandra J. Lee, Islam Eljilany, F. Stephen Hodi, Howard Streicher, Arivarasan Karunamurthy, John M. Kirkwood, Xuefeng Wang. Tertiary lymphoid structures (TLS) estimated by AI tools from digital H&E slides significantly enhance survival prediction in high-risk AJCC stages III/IV melanoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 3358.