Recognition of cancer mediating genes using MLP-SDAE model

Abstract

This article introduces a predictive deep learning model called MLP-SDAE, which combines Multilayer Perceptron (MLP) and Stacked Denoising Auto-encoder (SDAE) techniques. Our model, MLP-SDAE is trained using Stacked Denoising Auto-Encoder for feature selection, and backpropagation is employed within the MLP structure. We have incorporated dropout to enhance the model’s performance and prevent overfitting. The primary objective of the MLP-SDAE model is to identify associations among genes that have undergone significant alterations from a normal to a diseased state based on their expression behaviors. This concept allows us to predict disease-mediating genes and their altered associations. The methodology involves calculating gene-based correlation coefficients and selecting a subset of genes based on this analysis. We have demonstrated the effectiveness of our methods using four gene expression datasets related to human leukemia, lung, colon, and breast cancer. As a result, we have identified several potentially important genes, such as CACLA, HBA, IGFBP3, EFGR, TFN, TP53, LI6, and TMTC1, which may play a crucial role in developing these cancers. Furthermore, we conducted a comprehensive comparative study with other deep learning techniques, including Recurrent Neural Network (RNN), Deep Belief Network (DBN), Deep Boltzmann Machine (DBM), Auto-encoder (AE), and Denoising Auto-encoder (DAE). Our results have been validated through biochemical pathway analysis, t-tests, F-score, Gene Ontology (GO) identification, and the NCBI database. These validations demonstrate that our proposed MLP-SDAE model outperforms existing methods.

Keywords

• Stacked denoising auto-encoder
• Deep learning
• MLP
• p-value
• t-test