A perturbed multilayer perceptron approach to predicting distant metastatic sites of cancer patients

Cancer metastasis accounts for about 90% of cancer-related mortality, but is difficult to predict. In particular, distant metastasis is more difficult to predict by a learning method than lymph node metastasis due to the limited amount of data available for training a model and the inherent complexity of distant metastasis. Predicting distant metastatic sites from a primary cancer is even more difficult than predicting whether or not distant metastasis will occur. We developed a deep learning model called a perturbed multilayer perceptron (PMLP) to predict distant metastatic sites using expression levels of competing endogenous RNAs and their correlations at the primary site of cancer samples. In independent testing of PMLP on datasets which were not used in training, it showed high predictive performance (average AUC of 0.99, accuracy above 96%, and F1 scores above 0.91) in all metastatic sites. In comparison of the model with other state-of-the-art methods, our model showed a better performance. This model along with the explanation functionality of its prediction results can be used as useful aids to predict potential distant metastatic sites from gene expressions at the primary sites of cancer. To the best of our knowledge, this is the first study to employ PMLP combined with ceRNA correlation changes (ΔSCCs) for predicting specific distant metastatic sites, showing superior predictive performance with model interpretability.