Short-Term Scientific Mission

A Multitask Model as Fitness Function for a Generative Model for Drug Design

Main theme: Single- and Multiobjective Optimisation
Grantee: Selina Hesse, Philipps-University Marburg, Marburg, Germany
Host: Nicolas Moitessier, McGill University, Montreal, Canada

Start date: 2024-07-15
End date: 2024-09-30
Awarded: 2024-07-03
Report approved: 2024-10-28

Description

Limited data availability challenges molecular property prediction through machine learning. While multitask learning offers a solution by leveraging property correlations, such as solubility’s dependence on acidity (pKa) and hydrophobicity (logP), current approaches often lack effective integration of shared learning. Graph neural networks (GNNs) are particularly suited to molecular representation, treating atoms as nodes and bonds as edges. This STSM focuses on developing a multitask GNN model trained on diverse datasets, leveraging transfer learning to improve predictions of key properties like hydrophobicity and solubility. The ultimate goal is to integrate these predictors into generative models for drug discovery applications.

Achievements

During the STSM, the focus was on predicting multiple drug discovery properties using a multitask Graph Neural Network (GNN), where shared learning across related properties like logS, logP, pKa, logD, BBB permeability, and CACO-2 permeability was emphasized. The work demonstrated the practical application of multiobjective optimization in drug discovery and provided a benchmark for comparing future methods. Key contributions included bridging pharmacology with computer-aided approaches and integrating machine learning into drug discovery. While the work advanced optimization techniques, extending the model to new properties requires further work in ensuring the robustness of predictions. The STSM established a foundation for future publications to address these issues.