An uncertainty-based interpretable deep learning framework for predicting breast cancer outcome.

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Accurate breast cancer outcome prediction is crucial for effective treatment selection and patient survival.
• Deep learning methods show promise but often lack interpretability.
• Existing models face challenges in integrating biological knowledge and patient heterogeneity.

Purpose of the Study:

• To develop a novel, interpretable deep learning model for breast cancer outcome prediction.
• To improve prediction accuracy by incorporating biological pathway knowledge and patient heterogeneity.
• To identify novel genes associated with breast cancer prognosis.

Main Methods:

• Proposed a multitask deep neural network named UISNet.
• Implemented an uncertainty-based integrated gradients algorithm for feature interpretability.
• Integrated prior biological pathway knowledge and patient heterogeneity information.

Main Results:

• UISNet achieved superior performance on seven public breast cancer datasets (average C-index = 0.691) compared to state-of-the-art methods (average C-index = 0.650).
• Identified 20 genes associated with breast cancer, including 11 previously known and 9 novel findings.
• Demonstrated robustness and accuracy in breast cancer outcome prediction.

Conclusions:

• UISNet is an accurate and robust method for predicting breast cancer outcomes.
• The model serves as an effective tool for identifying novel breast cancer-associated genes.
• Open-source code is available for the UISNet method.