DeepReac+

DeepReac+ predicts chemical reaction outcomes and identifies optimal reaction conditions by combining a graph-neural-network-based model (DeepReac) with deep active learning applied to 2D molecular structures.

Key Features:

• Graph-neural-network model (DeepReac): The DeepReac architecture directly processes 2D molecular structures to generate representations for reaction prediction.
• Deep active learning strategies: Integrates active learning to select informative data points and reduce the number of experiments and labeled examples required for training.
• Universality across tasks: Adapts to various prediction tasks without extensive model reconfiguration by using 2D structural inputs.
• Outcome prediction and condition optimization: Produces predictions of chemical reaction outcomes and assists in identifying optimal reaction conditions.
• Demonstrated performance: Achieves state-of-the-art results with minimal labeled data across three diverse chemical reaction datasets in multiple scenarios.

Scientific Applications:

• Reaction outcome prediction: Predicting chemical reaction outcomes from 2D molecular inputs.
• Reaction-condition optimization: Guiding selection of optimal reaction conditions while minimizing experimental runs through active learning.
• Low-data reaction modeling: Enabling accurate modeling and prediction of reactions with minimal labeled experimental data.
• AI-aided chemical synthesis: Assisting exploration and optimization of complex chemical reactions using machine-learned models and active learning.

Methodology:

Employs a graph-neural-network-based model (DeepReac) that directly processes 2D molecular structures and integrates deep active learning strategies to select informative data points and reduce required labeled experiments; validated on three diverse chemical reaction datasets.