A deep learning framework for accurate reaction prediction and its application on high-throughput experimentation...

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Chemistry
• Machine Learning for Chemical Synthesis

Background:

• Artificial intelligence (AI) offers transformative potential in chemical synthesis.
• Current AI applications are hindered by inadequate chemical reaction representations and limited reaction data.
• Developing robust methods for reaction prediction is crucial for advancing chemical synthesis.

Purpose of the Study:

• Introduce GraphRXN, a novel graph-based reaction representation for AI-driven reaction prediction.
• Evaluate the performance of GraphRXN against existing models using public datasets.
• Assess the effectiveness of GraphRXN with experimentally generated data, including high-throughput experimentation.

Main Methods:

• Developed GraphRXN, a universal graph-based neural network framework.
• Encoded chemical reactions directly from 2D structures as input.
• Validated the model on three public chemical reaction datasets.
• Generated new reaction data through wet-lab experiments and high-throughput methods.

Main Results:

• GraphRXN achieved comparable or superior performance to baseline models on public datasets.
• The model demonstrated decent accuracy (R² of 0.712) when trained on high-throughput experimental data.
• Successfully integrated GraphRXN into a workflow combining robotics and AI for forward reaction prediction.

Conclusions:

• GraphRXN provides an effective method for representing chemical reactions for AI prediction.
• The model shows promise for advancing AI applications in chemical synthesis, particularly in forward reaction prediction.
• The integration of GraphRXN with experimental data generation highlights its practical utility in automated chemical research.