Accurate Physical Property Predictions via Deep Learning.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Machine learning for molecular property prediction

Background:

• Deep learning models have advanced drug discovery but require further accuracy improvements for molecular property predictions.
• Existing methods often rely on complex structural data, limiting their applicability.

Purpose of the Study:

• To introduce a novel deep learning architecture, the Bidirectional long short-term memory with Channel and Spatial Attention network (BCSA), for end-to-end molecular property prediction.
• To leverage SMILES tokenization and data augmentation for a data-driven approach, bypassing the need for complex spatial structures.

Main Methods:

• The BCSA model integrates Bidirectional Long Short-Term Memory (LSTM) for sequence processing with channel and spatial attention modules.
• The model utilizes data augmentation and SMILES tokenization, trained in a fully data-driven, end-to-end manner.
• Bayesian optimization was employed to further enhance the model's performance.

Main Results:

• The trained BCSA model demonstrated capability in predicting aqueous solubility.
• BCSA showed superior and competitive performance in predicting the oil-water partition coefficient compared to state-of-the-art graph models like GCN, MPNN, and AttentiveFP.

Conclusions:

• The BCSA model offers a powerful and accurate deep learning approach for molecular property prediction in drug discovery.
• This method provides a competitive alternative to existing graph-based models, particularly for predicting solubility and partition coefficients.