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

Li-Hsin Cheng, Te-Cheng Hsu, Che LinScientific 2021 · DOI

Abstract

Breast cancer is a heterogeneous disease. To guide proper treatment decisions for each patient, robust prognostic biomarkers, which allow reliable prognosis prediction, are necessary. Gene feature selection based on microarray data is an approach to discover potential biomarkers systematically. However, standard pure-statistical feature selection approaches often fail to incorporate prior biological knowledge and select genes that lack biological insights. Besides, due to the high dimensionality and low sample size properties of microarray data, selecting robust gene features is an intrinsically challenging problem. We hence combined systems biology feature selection with ensemble learning in this study, aiming to select genes with biological insights and robust prognostic predictive power. Moreover, to capture breast cancer's complex molecular processes, we adopted a multi-gene approach to predict the prognosis status using deep learning classifiers. We found that all ensemble approaches could improve feature selection robustness, wherein the hybrid ensemble approach led to the most robust result. Among all prognosis prediction models, the bimodal deep neural network (DNN) achieved the highest test performance, further verified by survival analysis. In summary, this study demonstrated the potential of combining ensemble learning and bimodal DNN in guiding precision medicine.

Introduction

Breast cancer is notoriously heterogeneous, meaning the molecular profile and clinical outcome vary significantly between patients. To guide precision medicine, clinicians require reliable prognostic biomarkers—specific genes that can predict whether a patient is likely to have a favorable or unfavorable outcome. However, identifying these biomarkers from microarray gene expression data is plagued by the "curse of dimensionality": datasets typically contain tens of thousands of gene features but only a few hundred patient samples.

Purely statistical feature selection often results in "unstable" gene lists, where minor changes in the data lead to completely different sets of genes, many of which lack clear biological relevance. This study addresses these challenges by bridging the gap between raw statistics and biological reality, creating a more robust and interpretable prediction model through a combination of systems biology and deep learning.

Methodology

Two-Stage Workflow

The researchers proposed a two-stage pipeline to improve both the selection of biomarkers and the accuracy of prognosis prediction.

  1. Feature Selection: Instead of relying solely on differential expression, the study utilized an unsupervised systems biology feature selector. This tool maps genes onto known interaction networks (using the BioGrid repository) to identify how gene interactions differ across prognosis groups. To solve the instability problem, three ensemble approaches were tested: data perturbation (subsampling), function perturbation (grouping patients by biological markers like ER or HER2), and a hybrid ensemble combining both.
  2. Prognosis Prediction: A Bimodal Deep Neural Network (DNN) was developed. Unlike standard neural networks that process a single flat vector, this bimodal structure uses two separate sub-networks to process genomic features and clinical data (e.g., age, tumor size) independently before merging them for a final prediction.

Datasets

The study utilized the METABRIC dataset, the largest open-access breast cancer cohort. From an initial 1,980 samples, a labeled set of 582 samples (categorized into "good" or "poor" prognosis based on 5-year disease-specific survival) and an unlabeled set of 1,282 samples were used for model training and feature selection.

Results

Robustness of Feature Selection

The study found that "function perturbation"—using different biological lenses—was more effective at improving robustness than simple data subsampling. The hybrid approach, which integrated both strategies, yielded the most stable and high-performing gene list. The model successfully narrowed thousands of candidates down to a highly stable 16-gene signature, including EGFR, ESR1, BTRC, and PLK1.

Comparison of ensemble feature selection methods showing AUC performance and stability across different perturbation strategies.

Predictive Performance and Survival Analysis

The bimodal DNN achieved an AUC of 0.7836, significantly outperforming traditional machine learning models like Support Vector Machines (SVM) and Random Forests (RF). Furthermore, the model's Concordance Index (CI) of 0.6683 was notably higher than the traditional Cox proportional hazards model (0.6251).

Survival analysis confirmed the clinical utility of the model. It successfully stratified patients into two distinct groups: at the 5-year mark, the "good prognosis" group had a 68% survival rate, compared to only 24% in the "poor prognosis" group.

Kaplan-Meier survival curves demonstrating the model's ability to stratify patients into distinct good and poor prognosis groups over a 60-month period.

Discussion and Conclusion

This research highlights that multi-gene approaches are superior to single biomarkers for capturing the complex biological distortions leading to cancer progression. By processing clinical and genomic data through separate subnetworks, the bimodal DNN learns unique features from each heterogeneous data type, leading to superior synergy.

While the hybrid ensemble is computationally expensive, it provides the most robust results. For resource-constrained settings, the authors recommend function perturbation as an efficient alternative. Future work will involve validating these 16 genes using RNA-Seq data and conducting wet-lab investigations to confirm the roles of under-studied genes like BTRC as potential therapeutic targets.

Summary updated

Abstract

Breast cancer is a heterogeneous disease. To guide proper treatment decisions for each patient, robust prognostic biomarkers, which allow reliable prognosis prediction, are necessary. Gene feature selection based on microarray data is an approach to discover potential biomarkers systematically. However, standard pure-statistical feature selection approaches often fail to incorporate prior biological knowledge and select genes that lack biological insights. Besides, due to the high dimensionality and low sample size properties of microarray data, selecting robust gene features is an intrinsically challenging problem. We hence combined systems biology feature selection with ensemble learning in this study, aiming to select genes with biological insights and robust prognostic predictive power. Moreover, to capture breast cancer's complex molecular processes, we adopted a multi-gene approach to predict the prognosis status using deep learning classifiers. We found that all ensemble approaches could improve feature selection robustness, wherein the hybrid ensemble approach led to the most robust result. Among all prognosis prediction models, the bimodal deep neural network (DNN) achieved the highest test performance, further verified by survival analysis. In summary, this study demonstrated the potential of combining ensemble learning and bimodal DNN in guiding precision medicine.
BibTeX
@inproceedings{cheng2021integrating,
  author = {Li-Hsin Cheng and Te-Cheng Hsu and Che Lin},
  title = {{Integrating ensemble systems biology feature selection and bimodal deep neural network for breast cancer prognosis prediction}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-92864-y},
}
Figure