AttenGene: A Deep Learning Model for Gene Selection in PDAC Classification Using Autoencoder and Attention Mechanism for Precision Oncology

Abstract

Pancreatic Ductal Adenocarcinoma (PDAC) ranks among the most severe and fatal types of cancer that are commonly found at an advanced stage, thus accounting for its low survival rate. The problem of early diagnosis remains, and the existing diagnostic devices are inaccurate and not effective. This research proposal is a prototype of a new deep learning (DL) architecture, AttenGene, that would deal with these problems by adding an autoencoder to learn unsupervised features and a self-attention mechanism to select sparse genes. The proposed model is capable of dealing with the high-dimensionality of the gene expression data, and the number of features can be decreased without compromising the classification performance. AttenGene may classify better and be interpretable than conventional classifiers, including XGBoost and AE + CNN, because of its smaller number of biologically meaningful genes. The second factor that will make the model useful in the clinical environment is that the model is simple and convenient, and provides information on the possible biomarkers that may be used to diagnose and treat PDAC. By being the first model that can combine both model performance with interpretability, AttenGene stands as an important milestone in the field of precision oncology, not only in PDAC but also in other cancers where the choice and classification of genes play a vital role.