Abstract A051: Multi-omic explainable machine learning improves cancer treatment outcome prediction

Abstract Background: Advancements in multi-omics data integration and explainable Machine Learning (ML) have shown promise in precision oncology. Multi-omic data used to train ML models may include genomics, transcriptomics and histopathology to characterize cancer cells and the tumor microenvironment (TME). Explainability methods, such as SHAP, have enabled researchers and clinicians to unravel the decision-making rationale of ML models predicting cancer progression and treatment response. We developed an explainable ML framework that incorporates multi-omic features of cancer and the TME. This framework was applied to predict patient response to neoadjuvant chemotherapy (NAC) in breast cancer and immune checkpoint inhibitor (ICI) in melanoma. Methods: For breast cancer, we used the cohort from Sammut et al. [1] (n=157 training, n=75 test). For melanoma, we assembled a cohort comprising 229 patients (n=138 training, n=53 test cutaneous, n=38 test non-cutaneous) from five independent studies. We improved the performance of the ensemble ML models in Sammut et al. [1] by implementing a shared-learning architecture to enable component models to influence each other as training progresses. We applied this ensemble (Ens:LR+RF+SVM) to predict NAC response in breast cancer and ICI response in melanoma by integrating clinical, DNA sequencing, RNA sequencing and histopathology (only for breast cancer) data. For melanoma, we also trained three single ML models (LR, RF, and SVM) and another ensemble (Ens:LR+RF), and introduced a novel dual utility of SHAP for feature-selection during training and biomarker threshold identification during validation. Results: The ensemble model trained on the multi-omic breast cancer features achieved ROC-AUC of 0.88 and showed a potential 25% reduction in false positives (i.e., incorrect predictions of good response) compared to its predecessor from Sammut et al. [1]. In the melanoma cohort, the Ens:LR+RF model achieved ROC-AOC of 0.77 but was outperformed by the RF model, ROC-AUC 0.78. SHAP revealed unique interactions between each ML model and the feature space, resulting in distinct training feature sets per model. During validation, the intersection between feature values and SHAP scores revealed numerical thresholds underpinning good versus poor responses of clinically meaningful biomarkers such as neoantigen load (>2.25 good, <2.25 poor, values in log10 scale). Across these two studies, we developed and open-sourced a scalable and versatile ML workflow (xML-workFLow) for rapid experimentation in biomedical research. Conclusions: This work showcases the potential of multi-omics explainable ML in advancing precision oncology to improve treatment outcome prediction. With further experimental validation, the use of explainable ML to determine numerical thresholds could guide the development of companion diagnostics and inform combination therapeutic strategies. References: 1. Sammut, S.J., et al., Multi-omic machine learning predictor of breast cancer therapy response. Nature, 2022. 601(7894): p. 623-629. Citation Format: Khoa A. Tran, Venkateswar Addala, Lambros T. Koufariotis, Jia Zhang, Scott Wood, Conrad Leonard, Lotte L. Hoeijmakers, Christian U. Blank, Mireia Crispin-Ortuzar, Amy McCart. Reed, Po-ling Inglis, Sunil R. Lakhani, Elizabeth D. Williams, John V. Pearson, Olga Kondrashova, Nicola Waddell. Multi-omic explainable machine learning improves cancer treatment outcome prediction [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A051.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.