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

Comprehensive Summary

The leading cause of cancer mortality is metastatic spread, comprising 90% of cancer deaths. Predicting if metastasis will occur and where it will occur is even harder due to a multitude of possible metastatic sites, metastasis to many organs simultaneously, and limited data on metastatic events. This research aimed to predict specific distant metastatic sites based on patterns in gene expression from a primary tumor. Wang et al. computer perturbed correlations between miRNAs and competing endogenous RNAs to measure their correlation when the tumor sample is normal. This allows for dysregulation of biological interactions to be studied, which deeply underlies metastasis. The machine-learning model predicted distant metastasis of tumors and made site-specific predictions for bone, liver, and lung metastasis. The model was found to have an accuracy rate of over 92% across all 3 site-specific models. In comparison with other models, such as MetastaSite and GCNN-Kirchhoff, the perturbed multilayer perceptron-based approach outperformed them all. This model also demonstrated that clustering tumors based on perturbed miRNA-mRNA correlations are aligned with metastatic outcomes, suggesting these may also be used as biomarkers.

Outcomes and Implications

This paper demonstrated that a perturbed model can provide specific predictions on primary tumor metastasis location and probability of occurring; this can be used by clinicians to decide on organ screening and surveillance in addition to targeted imaging for high risk patients. Moreover, this study demonstrates disrupted gene regulatory interactions, which can be used as biomarkers of metastasis as well as potential therapeutic targets. This accurate site prediction can influence choice of therapies, development of clinical trials targeting specific organ metastasis, and risk-informed counseling for patients and families.