GAMB-GNN: Graph Neural Networks learning from gene structure relations and Markov Blanket ranking for cancer classification in microarray data

Abstract

Microarray data plays a significant role in the classification and prediction of cancer, which has become an essential area of research in computational biology and bioinformatics. Several machine learning methods have been applied to help classify and predict cancer using microarray data. However, previous methods have mainly leveraged the expression level of a subset of genes as a sample feature. These methods lack the utilization of the structure relation information among genes. At the same time, there is a lack of analysis of redundancy among features in the existing gene selection algorithms. This paper proposes a Graph Neural Networks model with the Markov Blanket ranking method for cancer classification (GAMB-GNN) in microarray data to combat these challenges. The problem of cancer classification in the microarray data is considered by gene attributes and multi-type relation networks among genes in our model. The proposed GAMB-GNN obtains the gene scores and ranking according to the relevance of genes and cancer by the gene ranking algorithm based on Markov Blanket and then selects a top fraction of the ranking list to construct the multi-type gene relations graph. Further, GAMB-GNN employs a graph classifier with a relation attention mechanism and node aggregation pooling to learn the multiple relations structure and weight to obtain a better feature representation of the sample. The proposed model is validated through experiments in six public microarray datasets. The experiment results show that the accuracy and f1-score of our method on the cancer classification task on all tested microarray datasets improved by 3.98%–24.36% and 4.22%–31.93% on average over the baseline and state-of-the-art methods.