Multitask machine learning models for predicting lipophilicity (logP) in the SAMPL7 challenge.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Accurate prediction of lipophilicity (logP) from molecular structures is crucial for advancing drug discovery.
• Existing methods often require refinement to improve predictive performance and generalization.

Purpose of the Study:

• To construct a novel machine learning model for accurate logP prediction.
• To enhance model performance through architectural improvements and data integration.
• To evaluate the model's efficacy in the context of the SAMPL7 challenge.

Main Methods:

• Utilized Directed-Message Passing Neural Networks (D-MPNNs) as the core architecture.
• Integrated additional datasets from ChEMBL to augment training data.
• Incorporated helper tasks, including predictions from other logP/logD models, to improve generalization.

Main Results:

• The final model achieved a Root Mean Square Error (RMSE) of 0.66 and Mean Absolute Error (MAE) of 0.48 in the SAMPL7 challenge, ranking 2nd out of 17 submissions.
• Inclusion of ChEMBL datasets improved RMSE by 0.03, and helper tasks improved RMSE by 0.04.
• Retrospective application to the SAMPL6 challenge yielded an RMSE of 0.35, which would have ranked first.

Conclusions:

• The developed D-MPNN-based model demonstrates strong predictive accuracy and generalization for logP.
• The novel approach of using predictions from other models as helper tasks shows promise for broader applications.
• Further refinements are suggested to potentially enhance the model's performance even further.