S

Area of Science:

• Computational chemistry
• Chemical reaction prediction
• Machine learning in chemistry

Background:

• Predicting regioselectivity is crucial for efficient pharmaceutical synthesis.
• Current methods like DFT/TST and ML have limitations (time-consuming or data-dependent).

Purpose of the Study:

• Develop a faster and accurate method for predicting reaction regioselectivity.
• Integrate ML and DFT/TST to optimize computational resources.

Main Methods:

• A hybrid workflow combining ML and DFT calculations.
• DFT calculations are selectively triggered for low-confidence ML predictions.
• The model was trained and tested on SNAr reaction datasets (Pfizer and USPTO).

Main Results:

• Achieved 96.3% accuracy on the Pfizer dataset and 94.7% on the USPTO dataset.
• Significantly reduced computational time compared to conventional TST methods.

Conclusions:

• The hybrid ML-DFT approach offers a fast and accurate solution for regioselectivity prediction.
• This method can streamline drug discovery by minimizing unproductive synthetic efforts.