Prediction of radiosensitivity and radiocurability using a novel supervised artificial neural network

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Radiotherapy efficacy varies significantly among cancer patients.
• Traditional gene-based machine learning models predict radiosensitivity but lack advanced approaches.
• Artificial neural networks (ANNs) show promise but can overfit and learn irrelevant features.

Purpose of the Study:

• To develop a novel artificial neural network with Selective Connection based on Gene Patterns (ANN-SCGP) for predicting radiosensitivity and radiocurability.
• To integrate gene pattern information into ANNs to enhance prediction accuracy and biological relevance.

Main Methods:

• Developed ANN-SCGP, utilizing gene patterns to control the first layer of weights for learning gene interaction information.
• Trained and validated ANN-SCGP on 82 cell lines and 1,101 patients across 11 pan-cancer cohorts.
• Evaluated prediction performance using metrics like root mean squared error (RMSE) and C-index.

Main Results:

• ANN-SCGP achieved the lowest RMSE for predicting survival fraction at 2 Gy.
• ANN-SCGP demonstrated high performance in predicting radiocurability, achieving top C-index values.
• The model's low-dimensional output successfully reproduced gene similarity patterns.
• Pan-cancer analysis revealed associations between immune signals, DNA damage response, and radiocurability.

Conclusions:

• ANN-SCGP exhibits superior prediction capabilities for radiosensitivity and radiocurability compared to traditional models.
• The integration of gene pattern information enhances model performance and biological interpretability.
• This study offers novel insights into predicting and understanding cancer patient response to radiotherapy.