A Comparative Study of Marginalized Graph Kernel and Message-Passing Neural Network.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Predicting molecular properties is crucial for drug discovery.
• Existing methods like directed message-passing neural networks (D-MPNN) show promise.
• There is a need for methods offering accurate uncertainty quantification.

Purpose of the Study:

• To propose and evaluate a novel hybrid kernel for molecular similarity calculations.
• To compare its performance against D-MPNN for predicting molecular properties.
• To investigate the benefits of ensemble predictions and uncertainty quantification.

Main Methods:

• A hybrid kernel combining a marginalized graph kernel (MGK) and a radial basis function (RBF) kernel was developed.
• Gaussian process models were employed with the hybrid kernel.
• Bayesian optimization was used for hyperparameter tuning.
• Performance was evaluated on 11 public datasets against D-MPNN.

Main Results:

• The hybrid kernel's performance in predicting molecular properties is comparable to D-MPNN.
• Prediction errors between the two models were correlated.
• Ensemble predictions from the hybrid kernel and D-MPNN outperformed individual models.
• Principal component analysis revealed similar molecular embeddings between the two approaches.

Conclusions:

• The hybrid kernel offers competitive performance to D-MPNN for molecular property prediction.
• D-MPNN excels in computational efficiency and scalability.
• Graph kernel models provide superior uncertainty quantification capabilities.
• Ensemble modeling enhances predictive accuracy.