Machine learning (ML)-based quantification of tumor-infiltrating lymphocytes (TIL) and clinical outcomes of patients with melanoma treated with immune-checkpoint inhibitors (ICI).

Abstract

9560 Background: TIL quantification has shown promising prognostic and predictive impact in various tumors treated with ICI. More recently, TIL therapy has become an emerging treatment agent for ICI-refractory melanoma. In this work, we studied the effect of intrinsic TILs on clinical outcomes of patients with melanoma treated with ICI and quantified its utility as a biomarker in combination with tumor mutational burden (TMB). Methods: We applied a previously developed ML model to process digital whole-slide images of hematoxylin and eosin slides to quantify TIL density (TIL/mm2) in 182 metastatic melanoma samples treated with ICI at the Dana-Farber Cancer Institute. All samples underwent next-generation targeted-panel sequencing and had available TMB data (mutations/megabase). Overall survival (OS) was calculated from the data of ICI initiation to the date of death. TTF was calculated from the data of ICI initiation to the date of next line treatment or death. Alive patients were censored at the date of last follow-up. Multivariable cox regression was used to estimate adjusted hazard ratios (adj-HR) and p-values (adj-p), with TMB, prior lines of treatment and regimen type (single vs. combination ICI) used as covariates. TMB-high (TMB-H) group was defined as samples with TMB≥10. TIL-high (TIL-H) group was defined as samples with TIL density higher than the median. Results: Of 182 patients, 60% (110/182) were male and the median age at ICI start was 65.5 years. 63% (115/182) of patients received single-agent ICI (anti-PD1 or anti-CTLA4). Of 182 tumor specimens assessed for TILs, 47 (25.8%) were obtained from the primary site. There was no difference in TIL density (medians 583 vs. 817 TILs/mm2, p=0.57; 709 TIL/mm2 for full cohort) or TMB (medians 11.4 vs 12.2 mutations/megabase) between primary and metastatic samples, respectively. Moreover, there was no correlation between TIL density and TMB (pearson’s correlation coefficient = 0.92, p=1) across all samples. Compared to patients in the TIL-low (TIL-L) category, patients in TIL-H had significantly longer median TTF (33.9 [15.9-NR] versus 13.0 [7.8-18.6] months, adj-HR: 0.53 (95% CI: 0.36-0.78); adj-P=0.0012), while there was no difference in OS (P=0.57). In fact, patients in the TMB-H/TIL-H group had the longest TTF, followed by TMB-H/TIL-L, TMB-L/TIL-H then TMB-L/TIL-L (table). Conclusions: Our results support the potential use of ML-based TIL scoring as a novel and independent biomarker to predict time to failure on ICI. There is ongoing effort to validate our findings in an external dataset. [Table: see text]