A Pan-Cancer Classification Method Based on Multi-omics Data Integration and Sample Similarity Network

Cancer poses an important threat to human life and health. Patients can effectively extend their lives if cancer is detected early. The use of computer technology can provide assistance in cancer diagnosis. Aiming at the problem of cancer diagnosis, a pan-cancer classification technique is proposed. The provided method, KPGAT_NN, based on multi-omics data feature extraction and the sample similarity network. In addition, the similarity between cancer samples can also be informative for classification. However, directly merging multi-omics data will raise the data dimension while decreasing the signal-to-noise ratio, resulting in an imbalanced dataset, which makes it difficult to classify cancer. In order to deal with the problem, the data balancing technique integrating Borderline Smote and Tomek Links was applied, and then KPCA and PCA methods were integrated to extract omics data features and aggregate them as cancer features. The sample similarity network was constructed using the Jaccard similarity coefficient. Subsequently, the Graph Attention Network was utilized to acquire novel representations of cancer features. Finally, the data was fed into a Deep Neural Network for malignancy classification. The proposed method outperforms several other classification algorithms in terms of evaluation indicators.