Trustworthy Deep Neural Network for Inferring Anticancer Synergistic Combinations

Area of Science:

• Computational biology
• Drug discovery
• Cancer research

Background:

• Lack of a gold standard for synergy quantification in chemotherapeutic drug combinations.
• Potential for biased identification of ineffective synergistic combinations due to neglected sensitivity.
• Need for robust predictive models for effective cancer treatment strategies.

Purpose of the Study:

• To develop SynPredict, a deep learning model for predicting drug combination synergy across five metrics and combination sensitivity.
• To evaluate the impact of multimodal data fusion (gene expression, drug chemical features) on predictive accuracy.
• To compare SynPredict's performance against state-of-the-art predictive models.

Main Methods:

• Developed a deep learning-based model, SynPredict, utilizing multimodal fusion of gene expression and drug chemical features.
• Trained and validated the model on ONEIL and ALMANAC anticancer combination datasets.
• Assessed the influence of different input data fusion architectures and training datasets on model performance.

Main Results:

• SynPredict effectively predicts synergy across five metrics and combination sensitivity.
• Training dataset characteristics had a more significant and consistent impact than input data fusion architectures.
• SynPredict demonstrated superior performance, reducing mean square error by up to 74% compared to leading models like DeepSynergy and TranSynergy.

Conclusions:

• Emphasizes the critical need to incorporate multiple synergy metrics and combined sensitivity into predictive models for cancer drug combinations.
• SynPredict offers a promising advancement in accurately identifying effective chemotherapeutic drug combinations.
• Highlights the importance of dataset selection in developing reliable synergy prediction models.