Graph-based transformer to predict the octanol-water partition coefficient.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Machine learning in pharmaceutical sciences

Background:

• Lipophilicity (logP) is crucial for drug behavior, impacting solubility, permeability, and metabolism.
• Accurate logP prediction is vital for efficient drug candidate selection.
• Graph-based deep learning models show promise for molecular property prediction.

Purpose of the Study:

• To develop and evaluate GraphormerLogP, a novel deep learning model for accurate logP prediction.
• To curate a large, diverse dataset (GLP) of over 42,000 SMILES-logP pairs for training and evaluation.
• To compare GraphormerLogP's performance against state-of-the-art methods.

Main Methods:

• Utilized a fine-tuned pre-trained GraphormerMapper model for logP prediction.
• Trained and tested the model on a newly compiled dataset (GLP) and a benchmark dataset.
• Employed graph-based deep learning to learn representations directly from molecular graphs.

Main Results:

• GraphormerLogP achieved competitive to superior predictive accuracy on both datasets.
• Attained mean absolute error (MAE) values of 0.251 on the GLP dataset and 0.269 on the benchmark dataset.
• Demonstrated the model's effectiveness compared to Random Forest, Chemprop, CheMeleon, StructGNN, and Attentive FP.

Conclusions:

• GraphormerLogP offers a high-performance solution for logP prediction in drug discovery.
• The curated GLP dataset provides a valuable resource for advancing lipophilicity prediction research.
• Graph-based deep learning, particularly with Graphormer, shows significant potential for molecular property prediction tasks.