A Hybrid Explainable Multi-Omics Machine Learning Framework for Breast Cancer Mutation Prediction and Clinical Risk Stratification Using Tcga-Brca Data

Breast cancer is a leading cause of cancer-related mortality in women and detection of high-risk genomic alterations is crucial for improving breast cancer diagnosis, prognosis and personalization of breast cancer treatment. With the advent of next-generation sequencing (NGS) and large-scale cancer genomics projects, multi dimensional datasets have become available that can add to the accuracy of predictive models. So, in the present research, we present hybrid multi-omics machine learning model for predicting mutations and clinical risk stratification in breast cancer using data present in The Cancer Genome Atlas - Breast Invasive Carcinoma - TCGA-BRCA cohort. A combination of some of the somatic mutations, transcriptomic profile and clinical data into the proposed data set, to enable predictive modelling. The structured numerical data (such as SIFT score, PolyPhen score, sequencing depth, etc.) and mutation consequences of the mutations were extracted from the TCGA-BRCA mutation annotation file (MAF). The top 3,000 highly variable genes, detected in RNA sequencing data by variance-based feature selection, were used to record transcriptomic information. Some clinical data such as age, receptor status and pathological stage are made available to enhance the biological and prognostic value, Taverna said. The resulting data set was 2,764 balanced data samples and approximately 3,023 integrated predictive assets. A few supervised machine leaning models were evaluated, which include Logistic Regression, Decision Tree, K-Nearest Neighbors, Naive Bayes, Support Vector Machine, Random Forest, Gradient Boosting, XGBoost, and a Neural Network models. A stacking ensemble model was also developed that seemed to have the ability of combining the optimal of multiple classifiers was also implemented. Three types of accuracy (accuracy, area under receiver operating characteristic curve, and cross validation) were used to evaluate the model performance. Random Forest classifiers achieved the best performance with an accuracy of 75.41%, AUC value of 0.8377 and a stable performance across the five cross validation splits. The genomic/transcriptomic data features associated with the breast cancer were selected by using two XAI methods: SHAP and LIME. In the developed framework, the clinical report, molecular classification of subtypes and the mutation risk score were also produced, which were proving their validity for precision oncology. This research offers a scalable and reproducible computational platform to facilitate integration of multi-omics data and machine learning methods to aid breast cancer diagnosis, mutation interpretation and clinical decisions.