AI-assisted exploration of novel non-invasive biomarkers of ICI therapy response.

Abstract

e14591 Background: Immune checkpoint inhibitors (ICIs) have transformed cancer treatment, yet their efficacy varies significantly between patients, underscoring the need for accurate predictive models of ICI response. Emerging evidence highlights the gut microbiome’s critical role in modulating ICI outcomes, linking gut microbial composition to clinical responses. However, current prediction models primarily focus on genetic, transcriptomic, and immunohistochemistry-based features directly related to established ICI mechanisms. Microbiome-based biomarkers associated with underexplored aspects like the gut-immune axis remain underutilized, and efforts to develop predictive tools that generalize across patient cohorts have faced significant challenges. Methods: BluMaiden Biosciences has developed proprietary AI frameworks to address the existing gaps and identify clinically actionable, microbiome-derived features while accounting for the high inter-individual variability of gut microbiota. We applied these tools to diverse public datasets, including hepatocellular carcinoma and melanoma patients from Caucasian and Asian populations. Results: We developed a novel, cross-cohort microbial biomarker-based stratification tool, based on advanced machine learning. Through re-analysis of clinical datasets based upon original feature engineering and learning architecture design, we identified microbial biomarkers that robustly predict therapeutic outcomes in a mixed-cohort melanoma population (discovery cohort). Validation in an independent cohort demonstrated high predictive performance, with an area under the receiver operating characteristic curve (AUC) exceeding 0.82. Implementing this stratification tool can significantly enhance treatment efficacy response rates, as suggested by our predictive models, increasing from an estimated 20–50% to over 90% in stratified patient groups. These findings highlight the tool’s promise in addressing a critical unmet need in melanoma therapy, pending further validation in independent clinical trials. Conclusions: This work a) proves the existence of underexplored potential of human microbiome markers combined with custom machine learning models to dervive clinically applicable drug response predictors and b) highlights the importance of rigorous, cross-cohort validation to ensure predictor’s generalizability and robustness, paving the way for non-invasive, microbiome-based tools to guide ICI therapy in clinical practice.