Understanding Conformation Importance in Data-Driven Property Prediction Models.

Area of Science:

• Chemoinformatics
• Computational Chemistry
• Machine Learning

Background:

• Accurate molecular property prediction is crucial for drug discovery and materials design.
• The choice of molecular representation, particularly the inclusion of conformational information, significantly impacts prediction model performance.
• A systematic analysis of how conformational data influences property prediction models is lacking.

Purpose of the Study:

• To investigate the impact of using multiple molecular conformers on machine learning-based property prediction.
• To compare the effectiveness of 2D and 3D molecular descriptors in property prediction tasks.
• To evaluate the performance of an end-to-end model (Uni-Mol) against traditional descriptors using controlled datasets.

Main Methods:

• Development and utilization of three controlled datasets varying in scale and property type (quantum mechanical, melting point, reaction data).
• Comparison of property prediction models employing different molecular representations: 2D descriptors, 3D descriptors, and an end-to-end model (Uni-Mol).
• Evaluation of aggregation approaches for multiple conformers, including data augmentation and mean aggregation.

Main Results:

• Using all available conformers via data augmentation consistently yielded high prediction accuracy across datasets.
• The Uni-Mol model, using atomic coordinates and ground-truth conformations, significantly outperformed traditional 2D and 3D descriptors.
• Uni-Mol demonstrated high accuracy in predicting conformation-sensitive properties, though performance decreased with incorrect conformers.

Conclusions:

• Multiple conformers, treated as data augmentation, enhance machine learning property prediction accuracy.
• End-to-end models like Uni-Mol, leveraging 3D structural information, offer superior performance over traditional descriptors for molecular property prediction.
• Careful consideration of conformational data is essential for developing robust and accurate predictive models in chemoinformatics.