| Popis: |
Extracting cancer-related information from genomic data specially multi-source datasets has been an ever-growing challenge during the past years. The identification of subtype-specific genomic markers can lead to a sounder diagnosis and treatment. While several algorithms are proposed for feature extraction, to best of our knowledge, none of them consider between modality relations to discover modular disease associated biomarkers. In this paper, we represent an integrative deep learning approach to identify modular subtype-associated critical genes from three sets of input modalities for a better diagnosis of cancer subtypes. First, we train deep classifiers with different integration stages and distinct number of input modalities to predict cancer subtypes. Next, we use the optimized weight matrices of the classifier with the best performance to extract interactive top-ranked features among all input modalities. Lastly, we evaluate those ranks with other feature scoring methods according to their classification performance after feature extraction. Our results and analysis illustrate that the modular candidate biomarkers can be useful for cancer subtype detection. |
