Exposing the Limitations of Molecular Machine Learning with Activity Cliffs.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Machine learning in drug discovery

Background:

• Machine learning (ML) accurately predicts molecular properties like bioactivity in drug discovery.
• Activity cliffs, structurally similar molecules with significant potency variations, are often overlooked and challenge ML model performance.

Purpose of the Study:

• To address the knowledge gap regarding best ML practices for handling activity cliffs in drug discovery.
• To benchmark various ML and deep learning methods on their ability to predict the potency of activity cliff compounds.

Main Methods:

• Benchmarking 24 ML and deep learning algorithms using curated bioactivity data from 30 macromolecular targets.
• Evaluating model performance specifically on activity cliff compounds.

Main Results:

• All tested methods exhibited challenges when predicting the properties of activity cliffs.
• ML approaches utilizing molecular descriptors outperformed complex deep learning methods in this context.
• Significant performance variations were observed across different methods and datasets.

Conclusions:

• Dedicated activity-cliff-centered metrics are crucial for ML model development and evaluation in drug discovery.
• Novel algorithms are needed to improve the prediction accuracy for activity cliffs.
• An open-access platform, MoleculeACE, has been developed to facilitate research on activity cliffs in molecular ML.