Robust out-of-distribution prediction of Buchwald-Hartwig reactions

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Medicinal Chemistry

Background:

• Predictive modeling for Buchwald-Hartwig cross-coupling is limited by poor data quality and narrow chemical space coverage.
• Existing electronic laboratory notebooks and open-source high-throughput experimentation (HTE) datasets are fragmented and noisy, hindering model accuracy.
• Pharmaceutical synthesis relies heavily on efficient and predictable cross-coupling reactions.

Purpose of the Study:

• To develop a systematic framework for standardizing and integrating heterogeneous reaction datasets.
• To enhance the predictive power of machine learning models for Buchwald-Hartwig cross-coupling reactions across novel substrates and conditions.
• To enable preemptive in silico reagent screening for accelerated pharmaceutical discovery.

Main Methods:

• Systematic standardization and integration of multiple reaction datasets into a high-quality, unique-structure-per-entity dataset.
• Coupling data integration with active learning strategies to strategically expand chemical space coverage.
• Merging published Buchwald-Hartwig HTE data with new experimental results for model training and validation.

Main Results:

• Achieved a machine learning model with enhanced predictive power for novel substrates and conditions in Buchwald-Hartwig cross-coupling.
• Demonstrated improved out-of-distribution prediction capabilities compared to previous methodologies.
• Experimentally validated model-guided reagent recommendations, uncovering unexplored reactivity.

Conclusions:

• The developed framework establishes a blueprint for robust machine learning applications in synthetic chemistry.
• The approach enables effective in silico reagent screening, significantly accelerating the pharmaceutical discovery pipeline.
• Integrating diverse and standardized datasets with active learning is crucial for advancing predictive capabilities in chemical synthesis.