Efficient design of peptide-binding polymers using active learning approaches.

Area of Science:

• Chemical Engineering
• Materials Science
• Computational Chemistry

Background:

• Active learning (AL) is increasingly researched for efficient discovery of novel chemicals, materials, and polymers.
• AL enables rational design of drug delivery systems using small datasets and iterative model refinement.
• Key advantages include developing models with limited data and progressively improving predictions.

Purpose of the Study:

• To assess the applicability of AL for discovering polymeric micelle formulations for poorly soluble drugs.
• To compare various AL protocols for identifying biologically active molecules using synthetic datasets.
• To investigate AL performance based on initial training set size, task complexity, and dataset selection.

Main Methods:

• Comparison of different active learning (AL) protocols.
• Utilizing synthetic datasets for evaluating AL effectiveness in regression and classification tasks.
• Investigating the impact of initial training set characteristics on AL performance.

Main Results:

• Active learning (AL) regression modeling showed no benefit over random search.
• AL for classification tasks outperformed random selection but remained imperfect.
• The most effective AL protocol was applied to discover and validate novel polymers for asialoglycoprotein receptor (ASGPR) targeting.

Conclusions:

• Active learning (AL) demonstrates potential for accelerating the discovery of targeted drug delivery systems.
• The effectiveness of AL is task-dependent, with classification tasks showing more promise than regression.
• Experimental validation confirmed the utility of the best-performing AL approach for identifying functional polymers.