Reaction classification and yield prediction using the differential reaction fingerprint DRFP.

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Computational methods accelerate chemical research by predicting reaction outcomes.
• Deep learning-based learned fingerprints show high performance in reaction classification and yield prediction.
• Learned fingerprints have limitations, including large data requirements, inherent bias, and complex architectures.

Purpose of the Study:

• To introduce a novel, data-independent computational method for chemical reaction prediction.
• To develop a fingerprinting algorithm that overcomes the limitations of existing methods.

Main Methods:

• The differential reaction fingerprint (DRFP) algorithm was developed.
• DRFP processes reaction SMILES to generate a binary fingerprint.
• The fingerprint is based on the symmetric difference of molecular n-grams from reactants and products.

Main Results:

• DRFP outperforms DFT-based fingerprints in reaction yield prediction.
• DRFP surpasses structure-based fingerprints in reaction classification.
• DRFP achieves performance comparable to state-of-the-art learned fingerprints.

Conclusions:

• DRFP offers a powerful and efficient alternative for reaction prediction.
• The method is data-independent, reducing bias and data requirements.
• DRFP provides a valuable tool for accelerating chemical research.