Multi-armed bandit approach for multi-omics integration

As the omics technology has developed, it has resulted in the accessibility of large datasets with multi-omics information. Holistic analysis of biological processes require integration of different omics modalities. Several recent works suggest multi-omics integration outperforms the use of single omics for the task of disease subtyping. However, most of these works propose a black box integration of multi-omics through data or model ensembles. In contrast, our method searches for function transformations applied directly on single-omics to derive a mathematical expression for multi-omics integration. This allows for reproduciblility and quick validation of the results. We employ multi-armed bandits where 10 candidate functions compete with each other to be used for the transformation of each omics. Experiments with different exploration strategies namely $\epsilon$-Greedy and Upper Confidence Bound were investigated. Additionally, reward was defined as the prediction score of logistic regression model on unseen data. Based on cumulative reward per candidate over multiple trials, a winner was chosen for the ultimate formulation. We perform the task of cancer classification on liver data and provide comparisons between single omics, randomly integrated and optimally integrated multi-omics data. Further, we investigate the efficacy of the generated empirical formula on other cancer datasets viz. breast cancer, colon cancer and pancreas cancer. Experiments affirm the efficacy of the proposed method for optimally integrating multi-omics data.