Limits of Prediction for Machine Learning in Drug Discovery.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Machine learning

Background:

• Machine learning (ML) is increasingly used in drug discovery for property optimization.
• Understanding the limitations of ML extrapolation is crucial for reliable predictions.
• Previous analyses of ML extrapolation effectiveness in drug discovery are limited.

Purpose of the Study:

• To systematically evaluate the extrapolation capabilities of six ML algorithms.
• To assess performance across various molecular properties and dataset configurations.
• To identify optimal ML approaches for extrapolation in drug discovery.

Main Methods:

• Tested six ML algorithms on 243 datasets.
• Evaluated extrapolation performance using molecular weight, cLogP, and sp3-atom counts.
• Compared extrapolation with shuffled data (interpolation) versus sorted data (high-to-low and reverse).

Main Results:

• Extrapolation using sorted data yielded significantly higher prediction errors compared to shuffled data.
• Linear ML methods demonstrated superior performance for extrapolation tasks.
• Performance varied based on the specific ML algorithm and dataset properties.

Conclusions:

• The method of data preparation (shuffled vs. sorted) critically impacts ML extrapolation accuracy.
• Linear models are more reliable for extrapolation in drug discovery contexts.
• Further research should focus on developing robust extrapolation strategies for ML in drug design.