An integrated deep learning and dynamic programming method for predicting tumor suppressor genes, oncogenes, and...

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Genomics

Background:

• Uncontrolled tumor growth stems from mutations in proto-oncogenes (ONGO) and loss of tumor suppressor genes (TSG).
• Cancer's heterogeneity necessitates computational approaches to identify gene functionality for targeted drug development.
• Existing methods for classifying gene roles in cancer drivers have limitations.

Purpose of the Study:

• To propose and validate a novel classification method for predicting gene functionality as ONGO or TSG.
• To leverage 3D protein structural information and deep convolutional neural networks (DCNNs) for gene classification.
• To compare two distinct approaches for processing structural data and assess their impact on classification accuracy.

Main Methods:

• Feature map extraction from 3D protein structures (Protein Data Bank - PDB).
• Application of a deep convolutional neural network (DCNN) for initial functional classification.
• A dynamic programming-based method for final prediction of primary structure functionality.
• Validation using the Cancer Gene Census (COSMIC) database.

Main Results:

• Approach 1 (pooled structures) achieved a DCNN AUROC of 0.978 and a final classification AUROC of 0.989.
• Approach 2 (separated structures) achieved a DCNN AUROC of 0.765 and a final classification AUROC of 0.879.
• Both approaches surpassed the current state-of-the-art AUROC of 0.924 for ONGO vs TSG classification.

Conclusions:

• The proposed deep learning method effectively classifies gene functionality relevant to cancer drivers.
• The strategy of pooling 3D protein structures (Approach 1) demonstrated superior performance compared to separating them (Approach 2).
• These findings support further application of deep learning models for predicting human gene functionality in cancer.