Interpretable attention-based deep learning ensemble for personalized ovarian cancer treatment without manual...

Area of Science:

• Oncology
• Biomedical Engineering
• Computational Biology

Background:

• Bevacizumab, a VEGF-targeting therapy, is a key treatment for ovarian cancer.
• Predicting patient response to bevacizumab is crucial for optimizing therapy selection.
• Angiogenesis-related proteins are potential biomarkers for bevacizumab efficacy.

Purpose of the Study:

• To develop an interpretable, annotation-free deep learning model for predicting bevacizumab therapeutic effect in ovarian cancer.
• To investigate the predictive value of angiogenesis-related proteins (VEGF, Angiopoietin 2, Pyruvate kinase isoform M2) for bevacizumab response.
• To identify optimal protein biomarkers for selecting patients likely to benefit from bevacizumab therapy.

Main Methods:

• Utilized immunohistochemical whole slide images of tissue microarrays from ovarian cancer patients.
• Developed an attention-based deep learning ensemble framework for prediction.
• Evaluated model performance using five-fold cross-validation and statistical survival analyses (Kaplan-Meier, Cox proportional hazards).

Main Results:

• The ensemble model using Pyruvate kinase isoform M2 and Angiopoietin 2 achieved high performance metrics (F-score 0.99, accuracy 0.99, AUC 1.00).
• The model successfully identified patients with a predictive therapeutic sensitive group, correlating with low cancer recurrence (p<0.001).
• Cox proportional hazards model confirmed the predictive significance of the ensemble model (p=0.012).

Conclusions:

• The proposed deep learning ensemble model effectively predicts bevacizumab therapeutic response in ovarian cancer.
• Protein expressions of Pyruvate kinase isoform M2 and Angiopoietin 2 are significant predictors of bevacizumab efficacy.
• This model can aid in personalized treatment planning for bevacizumab-targeted therapy in ovarian cancer patients.