Scaling up drug combination surface prediction.

Area of Science:

• Computational biology
• Pharmacology
• Machine learning

Background:

• Drug combinations are crucial for treating complex diseases like advanced cancers.
• Synergistic drug combinations offer enhanced efficacy and reduced toxicity compared to monotherapy.
• Current drug combination screening is costly and time-consuming, necessitating efficient predictive models.

Purpose of the Study:

• To develop and evaluate an improved machine learning model (comboKR 2.0) for predicting full drug combination dose-response surfaces.
• To address limitations of existing scalar-valued prediction methods by adopting a functional output approach.
• To enhance the prioritization of potential synergistic drug combinations for experimental validation.

Main Methods:

• Implemented a scaled-up formulation of the comboKR method, incorporating novel modeling choices for response surfaces.
• Developed a projected gradient descent method to solve the pre-image problem in functional output prediction.
• Utilized input-output kernel regression and functional modeling of response surfaces.

Main Results:

• comboKR 2.0 demonstrated robust predictive performance across three real-world datasets, including scenarios with unseen drugs or cell lines.
• The functional output prediction approach outperformed traditional synergy score prediction methods.
• The projected gradient descent method effectively addressed the pre-image problem.

Conclusions:

• Functional output prediction of drug combination dose-response surfaces offers a more relevant and powerful approach than synergy scoring.
• The enhanced comboKR 2.0 model provides a reliable tool for prioritizing drug combinations in cancer research.
• This methodology can accelerate the discovery of effective combination therapies for complex diseases.