DCA: An Interpretable Deep Learning Model for Cancer Classification and New Knowledge Discovery Using Attention Mechanism with Discriminate Feature Constraint
Jialin Zhang, Chuanyan Wu, Kai Lu, Rui Gao · 2024
With the advancement of sequencing technology, an increasing number of diverse omics data has been generated and extensively utilized in cancer research, including RNA-seq data. However, the high-dimensional characteristics of these data presents a challenge, as only a limited set of genes are implicated in the occurrence and progression of cancer. Therefore, it is crucial to develop classification algorithms that possess both interpretability and dimension reduction abilities for handling such data effectively. In this paper, we propose a deep learning strategy based on discriminative feature constraint attention mechanism (DCA), which comprises an attention module and a classification module. The attention module uses a modified SE block as the backbone, and the discriminate feature matrix calculated by Fisher score drives the updating direction of backbone. The MSE loss measures the difference between the discriminant feature matrix and backbone attention matrix, and dynamic task loss is utilized to balance the weights of the two tasks for optimal performance. We evaluated the DCA model on six bulk RNA-seq datasets from four types cancer and compared its performance with several commonly used and newly proposed classification methods. The results demonstrated that our approach achieved the optimal performance on five out of the six cancer datasets. Additionally, we conducted biological functional validation to assess the interpretability of DCA, confirming its ability to identify significant features. The proposed DCA deep learning model demonstrates its ability to extract critical information from high-dimensional omics data and to effectively predict cancer. Moreover, through biological functional validation, it has been determined that the features extracted by DCA possess biological significance.