Multiomic machine learning integration of DNA and RNA features to predict immunotherapy benefit in MSS-CRC and other rare cancers.

Abstract

2531 Background: Immune checkpoint inhibitors (ICIs) have revolutionized the oncology landscape, yet remain constrained by imprecise predictive biomarkers (i.e., PD-L1, TMB, MSI), which fail to capture the complexity of the tumor microenvironment. The Immune Profile Score (IPS), an AI/machine learning (ML) driven DNA-/RNA-based molecular signature addresses this gap in translational molecular biomarkers of ICI response. IPS integrates TMB, single-gene RNA features and RNA signatures, and was independently validated for prognostic utility in > 1,500 advanced solid tumor patients (pts) treated with FDA-approved ICI. Here we evaluated the ability of IPS to accurately stratify ICI treatment outcomes in two independent cohorts representing traditionally ICI-resistant populations. Methods: From our multimodal real-world database, we used the ML-derived IPS algorithm to analyze two cohorts of high unmet need for which ICI is not approved: 1) microsatellite stable colorectal cancer (MSS CRC); and 2) rare solid cancer as defined by FDA ( < 200,000 cases/year) treated with off-label ICI. Pts were categorized as IPS-H and IPS-L using a previously independently validated and published threshold. Cox proportional hazards models were fit to demonstrate prognostic utility for real-world overall survival (rwOS). Association with time-to-next-treatment (TTNT) on prior chemotherapy (CT) was compared in the same pts to rwOS on subsequent ICI therapy to assess ICI-specific predictive value of IPS . Results: IPS-H consistently identified a subset of pts with improved clinical outcomes across both cohorts. In the MSS-CRC cohort (n = 46): IPS-H pts (6/46 = 13%) had longer rwOS than IPS-L pts (40/46 = 87%) (HR 0.22; 90% CI: 0.04-1.16). No difference was observed in TTNT between IPS-H and IPS-L for prior CT (HR 1.07; 90% CI: 0.60-1.91), while there was improvement in rwOS IPS-H vs IPS-L on subsequent ICI therapy (HR 0.21; 90% CI: 0.04-1.22). In the rare cancer cohort (n = 90): there were 26 solid tumor subtypes without an FDA-approved ICI label; carcinosarcoma (n = 19, 21%) and pancreatic ductal adenocarcinoma (n = 17, 19%) were the most commonly represented. IPS-H pts (16/90 = 18%) had longer rwOS than IPS-L pts (HR 0.26, 95% CI: 0.09-0.73). In this rare cancer cohort, IPS remained significant even when restricted to subtypes with representation from both IPS-H and IPS-L (HR = 0.18, 95% CI: 0.04-0.69). Conclusions: IPS is a novel multiomic genomic signature that identifies a subset of advanced MSS-CRC and rare solid cancer patients who may benefit from ICI therapy. By integrating multimodal genomic features, IPS emerged as a possible predictive biomarker in a population where ICI is not currently approved. IPS suggests a paradigm shift toward AI/ML-driven signatures to refine ICI candidate selection and personalize clinical decision-making in oncology.