MetastaSite: Predicting metastasis to different sites using deep learning with gene expression data.

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Deep learning models offer potential for phenotype prediction from omics data.
• Interpretability of deep neural networks is crucial, especially in medical applications.
• Current methods lack comprehensive prediction of metastasis to multiple sites.

Purpose of the Study:

• To develop an interpretable deep learning framework for predicting cancer metastasis.
• To distinguish primary tumors from metastases in the brain, bone, lung, and liver.
• To identify key genes and biological functions driving metastasis prediction.

Main Methods:

• An AutoEncoder framework was used to learn gene relationships.
• DeepLIFT was applied to calculate gene importance scores.
• A multi-class deep neural network (DNN) was trained for metastasis prediction and biological function identification.

Main Results:

• The DNN model achieved prediction performances with AUC ranging from 0.93 to 0.82.
• The framework successfully predicted primary versus metastasized samples.
• Identified key genes and biological functions utilized by the model for prediction.

Conclusions:

• This is the first multi-class DNN for generic metastasis site prediction.
• The developed framework provides interpretable insights into cancer metastasis.
• The model aids in understanding biological mechanisms underlying cancer spread.