Integrating ensemble systems biology feature selection and bimodal deep neural network for breast cancer prognosis prediction

Abstract

Motivation:Breast cancer is a heterogeneous disease. In order to guide proper treatment decisions for each individual patient, there is an urgent need for robust prognostic biomarkers that allow reliable prognosis prediction. Gene feature selection on microarray data is an approach to systematically discover potential biomarkers. However, common pure-statistical feature selection approaches often fail to incorporate prior biological knowledge and thus tend to select genes that lack biological insights. In addition, due to the high dimensionality and low sample size properties of microarray data, selecting robust gene features is an intrinsically challenging problem. We therefore combined systems biology feature selection with ensemble learning in this study, aiming to address the above challenges and select genes with biological insights, as well as robust prognostic predictive power. Moreover, in order to capture the complex molecular processes of breast cancer, where multiple disease-contributing genes may exist and interact, we adopted a multi-gene approach to predict the prognosis status using machine learning classifiers. Results:We systematically evaluated three different ensemble approaches that all improved the original systems biology feature selector. We found that compared to the most popular data-perturbation approach, function perturbation can produce significant improvement with just a few ensembles. Among all, the hybrid ensemble approach led to the most robust feature selection result, and the identified genes were shown to be highly involved in pathways, such as ubiquitination and cell cycle. Final prognosis prediction models were constructed using the identified genes and clinical information as input features. Among all models, bimodal deep neural network (DNN) achieved the highest AUC (area under receiver operating characteristic curve) in test performance evaluation, where subsequent survival analysis also verified its ability to differentiate patients with different prognosis statuses. In summary, the study demonstrated the potential of ensemble learning to improve gene feature selection robustness, as well as the potential of bimodal DNN in providing reliable prognosis prediction and guiding precision medicine.

Training with Small Medical Data: Robust Bayesian Neural Networks for Colon Cancer Overall Survival Prediction

“…They were then trimmed with statistical model order selection techniques to remove false-positive connections. Prognosis relevance values (PRVs) were calculated from built networks to identify the final gene feature subset as the prognostic biomarkers for training classifiers [5], [4].…”

“…• Concat: an NN classifier fed with V , a direct concatenation of X and C (early fusion). • Bimodal: a bimodal NN classifier (late fusion) [5], [4].…”

“…Our previous research utilized systems biology approaches to identify differentially enriched pathways. Prognostic biomarkers with biological insights were selected by constructing gene interaction networks (GINs) and calculating the prognostic relevance values (PRVs) from microarray data for non-small cell lung cancer (NSCLC) and breast cancer [4], [5]. We can select a set of prognostic biomarkers for colon cancer patients through the same systems biology feature selector and used deep learning models to predict their prognostic condition.…”