Dose-response prediction for in-vitro drug combination datasets: a probabilistic approach

Area of Science:

• Computational biology
• Pharmacology
• Machine learning

Background:

• High-throughput drug screening generates complex, often noisy, dose-response data.
• Predicting drug combination effects is crucial for developing effective combination therapies.
• Existing models struggle with sparse, heterogeneous, and uncertain experimental data.

Purpose of the Study:

• To develop a robust probabilistic framework for accurate dose-response prediction in drug combinations.
• To address challenges posed by noisy, sparse, and varying quality experimental data.
• To identify synergistic drug interactions from complex datasets.

Main Methods:

• Proposed PIICM (Probabilistic framework for Intrinsic Co-regionalization Model).
• Utilized a permutation invariant intrinsic co-regionalization model for multi-output Gaussian process regression.
• Incorporated an observation model to handle experimental uncertainty and varying data quality.

Main Results:

• PIICM accurately predicted dose-response surfaces in held-out experiments.
• The model successfully learned from sparsely observed and noisy cell-viability measurements.
• The resulting predictive function identified key features indicative of synergistic drug interactions.

Conclusions:

• PIICM offers a powerful probabilistic approach for dose-response prediction in high-throughput drug combination studies.
• The framework effectively handles experimental noise, uncertainty, and data sparsity.
• PIICM aids in discovering synergistic drug combinations, advancing precision medicine.