Finding the most potent compounds using active learning on molecular pairs.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Active learning accelerates molecular optimization but struggles in early stages with limited data, potentially yielding analogs with low scaffold diversity.
• Exploitative active learning can lead to models that overfit to initial data, limiting exploration of novel chemical space.

Purpose of the Study:

• To introduce ActiveDelta, an adaptive active learning approach using paired molecular representations to predict improvements and guide data acquisition.
• To evaluate ActiveDelta's effectiveness in enhancing both potency and chemical diversity of identified molecular inhibitors.

Main Methods:

• ActiveDelta was applied to graph-based deep (Chemprop) and tree-based (XGBoost) models for active learning.
• Performance was assessed on 99 K_i benchmarking datasets, comparing ActiveDelta against standard active learning methods (Chemprop, XGBoost, Random Forest).
• Chemical diversity was evaluated using Murcko scaffolds.

Main Results:

• ActiveDelta implementations significantly outperformed standard active learning in identifying more potent inhibitors.
• The approach successfully identified molecular inhibitors with greater chemical diversity based on Murcko scaffolds.
• Deep learning models (Chemprop) trained with ActiveDelta-selected data showed improved accuracy in simulated time-split test sets.

Conclusions:

• ActiveDelta represents a significant advancement for active learning strategies, particularly in low-data scenarios.
• Molecular pairing approaches, like ActiveDelta, can accelerate the identification of potent and diverse drug candidates.
• This method holds substantial potential for improving hit identification against critical drug targets.