Pairwise Neural Networks for Ranking Molecular Structures Based on Properties.

Area of Science:

• Computational Chemistry and Materials Science
• Machine Learning in Scientific Discovery

Background:

• Traditional molecular discovery relies on slow, expensive quantum-chemical calculations or experimental screening.
• Machine learning offers accelerated molecular property prediction directly from structures.
• Ranking molecular structures can be sufficient for screening, bypassing exact property value determination.

Purpose of the Study:

• To develop and evaluate a deep learning model for ranking molecular structures.
• To compare the performance of a Siamese network with pairwise learning against standard pointwise regression.
• To assess the model's robustness using the Uni-Mol molecular representation.

Main Methods:

• Developed a deep learning model utilizing a Siamese network architecture.
• Employed pairwise learning to train the model for molecular structure ranking.
• Evaluated performance on QM7x and QO2Mol datasets, comparing against pointwise regression.

Main Results:

• The Siamese network with pairwise learning outperformed standard pointwise regression for predicting absolute energetic properties (e.g., total and orbital energies).
• Pointwise regression remained effective for derived properties (e.g., HOMO-LUMO gap) and non-energy properties (e.g., dipole moment).
• The pairwise learning-to-rank approach consistently outperformed pointwise regression across various Uni-Mol model sizes (V1 and V2).

Conclusions:

• Pairwise learning with Siamese networks is a robust and effective method for ranking molecular structures, outperforming pointwise regression for specific property predictions.
• This approach accelerates molecular screening and discovery in fields like energy storage, catalysis, and drug development.
• The framework demonstrates consistent superiority even with advanced pretrained Transformer backbones like Uni-Mol.