Predictive Minisci late stage functionalization with transfer learning.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Organic Chemistry

Background:

• Structural diversification of lead molecules is crucial for drug discovery.
• Late-stage functionalizations (LSFs) enable the creation of diverse molecular products from complex intermediates.
• Predicting regioselectivity in LSFs remains a significant challenge, limiting the efficiency of drug discovery pipelines.

Purpose of the Study:

• To develop a predictive model for regioselectivity in late-stage functionalization reactions.
• To overcome data limitations in late-stage functionalization product characterization for machine learning approaches.
• To enhance the exploration of chemical space in drug discovery through accurate reactivity prediction.

Main Methods:

• Development of a hybrid approach combining a message passing neural network and 13C NMR-based transfer learning.
• Application of the model to predict atom-wise functionalization probabilities for Minisci and P450-based reactions.
• Validation of the model through retrospective analysis and prospective experimental studies.

Main Results:

• The developed model accurately predicts the regioselectivity of Minisci-type and P450 transformations.
• The approach demonstrates superior performance compared to traditional Fukui-based reactivity indices and other machine learning algorithms.
• Successful prediction of functionalization outcomes was validated experimentally, confirming the model's reliability.

Conclusions:

• The integration of message passing neural networks and 13C NMR transfer learning offers a powerful solution for predicting LSF regioselectivity.
• This method significantly advances the ability to explore chemical space and accelerate drug discovery.
• The developed approach provides a robust tool for guiding synthetic strategies in medicinal chemistry.