Active learning approaches in molecule pKi prediction.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Identifying potent drug candidates requires extensive screening.
• Assaying vast chemical libraries is resource-intensive.
• Active learning (AL) offers a strategy to optimize candidate selection and minimize experimental assays.

Purpose of the Study:

• To benchmark various AL strategies for drug discovery.
• To identify an optimal AL approach for high model performance and efficient molecule selection.
• To develop a unified AL strategy balancing exploration and exploitation.

Main Methods:

• Simulated active learning (AL) workflow using virtual experiments.
• Leveraged ChEMBL datasets with known molecular bioactivity values.
• Proposed and evaluated a hybrid AL selection strategy with adjustable parameters (n and c).

Main Results:

• In classification tasks, exploration and hybrid strategies (c<1 for n=1, c≤0.2 for n=2) built high-performance models with minimal data.
• For property selection, exploitation and hybrid strategies (c≥1 for n=1, c≥0.7 for n=2) were efficient.
• The hybrid strategy with c=0.7 balanced both model performance and property selection effectively.

Conclusions:

• Active learning significantly reduces the number of assays needed in early drug design.
• The proposed hybrid AL strategy offers adaptable and efficient molecule selection.
• This approach enhances the identification of drug candidates with desired properties and predictive model accuracy.