Leveraging Graph Attention Mechanisms to Create an Explainable Multi-Function Machine Learning Model
Date
ORCID
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Purpose:
Identifying target-specific ligands is a difficult task, especially in cases where receptors display high structural similarity. Such is the case for metabotropic glutamate receptor subtype 2 (mGlu2) and metabotropic glutamate receptor subtype 3 (mGlu3), which are prime targets for various neurological treatments. However, signal transduction through these two receptors often yields opposing physiological function and differentially affect pathologies.
Methods:
Understanding the need to differentiate ligands based on their binding to mGlu2 and mGlu3, we employed a machine learning (ML) approach. The ML model performed three distinct tasks and leveraged transfer learning to inform each subsequent task. Task 1: Simple Classification was performed, as the ML model predicted if the ligands displayed selectivity for the mGlu2 or mGlu3 class. Task 2: Regression was performed, as the ML model estimated the IC50 values of individual input ligands. The classification weights from Task 1 were broadcasted into the attention layers of the ML model for Task 2, serving as a starting point. Task 3: Classification was performed, as the ML model sought to determine if a ligand displayed low or high potency for the target class. Classification weights and regression weights from previous tasks were broadcasted into the model.
Results:
The model yielded greater than 99% accuracy in the selectivity classification task, while also delivering satisfactory performance when predicting potency (72.80% error). The model yielded 83% accuracy in correctly identifying high potency mGlu2 ligands, as high potency mGlu2 compounds. Meanwhile, the algorithm displayed 75% accuracy in correctly identifying high potency mGlu3 ligands, as high potency mGlu3 compounds.
Conclusions:
This approach allows for prediction of multiple target properties using a single model. With access to other high-quality datasets, this model has the potential to apply to other ligand classes of interest, posing great potential for drug repurposing studies.