Bayesian active learning for drug combinations.

Area of Science:

• Computational Biology
• Pharmacology
• Systems Biology

Background:

• Complex diseases involve intricate interactions of multiple factors.
• Drug combinations can improve therapy by targeting diverse factors.
• Designing effective drug combinations is challenging due to nonlinear drug interactions and high-dimensional dose spaces.

Purpose of the Study:

• To develop an efficient method for identifying optimal drug combinations in continuous, high-dimensional dose spaces.
• To overcome limitations of existing discretization-based heuristics for drug combination discovery.

Main Methods:

• Modeling biological system response using Gaussian processes (GP) for continuous drug doses.
• Employing a closed-loop experimental design guided by an expected improvement criterion.
• Utilizing a particle filter for Bayesian marginalization of GP hyperparameters.
• Applying a hybrid Monte Carlo algorithm for efficient exploration of high-dimensional search spaces.

Main Results:

• Demonstrated effectiveness on Drosophila, Herpes simplex virus type 1, and simulated apoptosis networks.
• Significantly reduced the number of required trials compared to existing methods.
• Successfully navigated high-dimensional continuous drug spaces.

Conclusions:

• The proposed Gaussian process-based approach offers a more efficient and effective strategy for drug combination optimization.
• This method advances the design of multi-drug therapies by handling continuous dose spaces and complex biological interactions.