Tumor Immunogenicity Unlocked: Multi-omics Models Predict Immunotherapy Response and Survival

Abstract

Background: Immune Checkpoint Inhibitor (ICI) outcomes are highly heterogeneous among cancer patients. Some individuals experience remarkable treatment success and durable responses, whereas others do not show any benefit from ICI therapy. Concurrently, patients may experience severe or even potentially life-threatening adverse effects. Although currently available individual biomarkers, such as the gold-standards tumor mutation burden (TMB) or PDL1 expression, correlate with ICI outcomes, their predictive accuracy remains limited. More reliable models that predict ICI treatment outcomes with higher accuracy are urgently needed to identify patients who are likely to benefit from ICI therapy. Integrating distinct biomarkers from multiple omics layers into predictive models is a promising strategy for improving ICI outcome predictions. Methods: Various genomic and transcriptomic biomarkers that are predictive of ICI outcomes and that describe different aspects of tumor biology were derived from next-generation sequencing (NGS) data. These included tumor-intrinsic properties, such as TMB, PDL1 expression, immune cell infiltration, and known mechanisms of ICI response, as well as host factors such as HLA-related metrics. Subsequently, the biomarkers were combined into multiomics models using machine learning (ML) approaches and interpretable scoring methods to predict ICI outcomes with higher accuracy than individual biomarkers. The models were trained on cohorts of melanoma patients with annotated clinical ICI outcomes. Subsequent analyses were performed using additional datasets of various cancer types. Results: This cumulative dissertation resulted in two scientific, peer-reviewed publications. In the first study, we developed Least Absolute Shrinkage and Selection Operator (LASSO) models that integrate biomarkers derived form bulk RNA sequencing (RNA-seq) and whole exome sequencing (WES) to predict ICI outcomes. The predictions of these multiomic models outperformed those of the gold-standard ICI biomarker TMB. SHapley Additive exPlanations (SHAP) values were applied to provide explanations of the predictions. The second study focused on an interpretable multi-omics model, the Multi-Omics Tumor Immunogenicity score (MOTIscore), which extended the outcome prediction to survival analyses and investigated underlying molecular mechanics of the tumors. This MOTIscore accurately predicted ICI outcomes, and discriminated patient groups with survival benefits following ICI treatment. The MOTIscore revealed the predictive role of genes that are part of the C-X-C motif ligand pathway, which plays an important role in immune cell activation. Conclusion: This cumulative work demonstrated that integrating multi-omics data substantially improved the prediction of ICI outcomes compared to individual biomarkers. The multi-omics models developed in this study offer a promising tool for the selection of patients for treatment with ICIs. Using SHAP-based explanations and the interpretable MOTIscore model, we provided explanations for the predictions, which is a crucial prerequisite for a clinical decision-making tool. However, prospective studies are necessary to validate all the findings of our study for potential clinical applications, preferably with a larger patient cohort for which sequencing is performed.