Machine Learning May Sometimes Simply Capture Literature Popularity Trends: A Case Study of Heterocyclic...

Area of Science:

• Synthetic Chemistry
• Machine Learning
• Computational Chemistry

Background:

• Machine learning (ML) is increasingly applied to predict chemical reactivity using literature data.
• Accurate predictive models are crucial for optimizing synthetic chemistry reactions.
• The assumption is that large datasets enable robust model construction.

Purpose of the Study:

• To evaluate the efficacy of ML models in predicting optimal reaction conditions for Suzuki-Miyaura coupling.
• To determine if curated literature data is sufficient for accurate ML-driven chemical reactivity prediction.
• To identify limitations of current ML approaches in synthetic chemistry.

Main Methods:

• A curated database of over 10,000 literature examples of Suzuki-Miyaura coupling was utilized.
• Various ML models, including feed-forward and graph-convolution neural networks, were applied.
• Different chemical representations (fingerprints, descriptors, latent representations) were tested.
• Model performance was compared against naive frequency-based assignments.

Main Results:

• ML models failed to provide meaningful predictions for optimal reaction conditions (solvents, bases).
• Performance did not significantly improve across different ML architectures or chemical representations.
• ML predictions were not substantially better than random guessing based on common conditions.

Conclusions:

• Abundant, curated literature data alone is insufficient for training accurate ML models in synthetic chemistry.
• Biases in literature reporting (e.g., subjective preferences, reagent availability) and lack of negative data hinder ML performance.
• Systematic generation of reliable, standardized datasets is crucial for advancing ML in chemical reactivity prediction.