CDCR-Rank: a computational model for predicting drug combination dose response using ranking-based optimization.

Area of Science:

• Computational biology
• Drug discovery
• Cancer therapy

Background:

• Predicting synergistic drug combinations is crucial for cancer therapy but is hindered by complex interactions and numerous possibilities.
• Existing methods face challenges in accurately ranking potential drug combinations.

Purpose of the Study:

• To develop a novel ranking-based approach for predicting synergistic drug combinations.
• To improve the accuracy and robustness of drug synergy prediction for cancer therapies.

Main Methods:

• Utilized a pre-trained CDCR model to predict absolute synergy scores.
• Developed a new architecture for direct comparison of drug combination lists.
• Implemented a custom loss function combining mean squared error and ranking loss (uRank loss).
• Employed a one-dimensional convolutional neural network (1D-CNN) for drug representations from SMILES strings and a sinusoidal encoder for dose-response curves.

Main Results:

• CDCR-Rank outperformed state-of-the-art methods like comboFM and comboLTR on the NCI-ALMANAC benchmark.
• Demonstrated superior performance in predicting synergy for novel drug pairs, even without monotherapy data.
• Ablation studies confirmed the effectiveness of the 1D-CNN, sinusoidal encoder, and uRank loss.

Conclusions:

• The CDCR-Rank framework significantly enhances accuracy and robustness in predicting drug synergy.
• This approach can accelerate the identification of promising drug combinations for experimental validation.
• The findings have the potential to expedite the development of effective multi-drug cancer therapies.