Explainable Graph Neural Networks with Data Augmentation for Predicting pK_a of C–H Acids

Abstract

The pK a of C–H acids is an important parameter in the fields of organic synthesis, drug discovery, and materials science. However, the prediction of pK a is still a great challenge due to the limit of experimental data and the lack of chemical insight. Here, a new model for predicting the pK a values of C–H acids is proposed on the basis of graph neural networks (GNNs) and data augmentation. A message passing unit (MPU) was used to extract the topological and target-related information from the molecular graph data, and a readout layer was utilized to retrieve the information on the ionization site C atom. The retrieved information then was adopted to predict pK a by a fully connected network. Furthermore, to increase the diversity of the training data, a knowledge-infused data augmentation technique was established by replacing the H atoms in a molecule with substituents exhibiting different electronic effects. The MPU was pretrained with the augmented data. The efficacy of data augmentation was confirmed by visualizing the distribution of compounds with different substituents and by classifying compounds. The explainability of the model was studied by examining the change of pK a values when a specific atom was masked. This explainability was used to identify the key substituents for pK a. The model was evaluated on two data sets from the iBonD database. Dataset1 includes the experimental pK a values of C–H acids measured in DMSO, while dataset2 comprises the pK a values measured in water. The results show that the knowledge-infused data augmentation technique greatly improves the predictive accuracy of the model, especially when the number of samples is small.

AttenGpKa: A Universal Predictor of Solvation Acidity Using Graph Neural Network and Molecular Topology

“…The AD of the AttenGpKa model is defined based on the similarity of molecules in the training set, as determined by Euclidean distance using Morgan fingerprints . Specific details can be found in our previous work . By calculating the distance matrix of compounds in the training set, the average Euclidean distance d̅ and the standard deviation σ of these distances are obtained, which are 6.44 and 2.95, respectively.…”

“…31 Specific details can be found in our previous work. 30 By calculating the distance matrix of compounds in the training set, the average Euclidean distance d̅ and the standard deviation σ of these distances are obtained, which are 6.44 and 2.95, respectively. The distance threshold D T for describing the AD is calculated using D T = d̅ + Z*σ, where Z is an empirical parameter representing the significance level.…”

“…29 In our previous work, we established the EEGpKa model based on message passing neural networks (MPNN). 30 This model introduces a knowledge-infused data augmentation technique, improving the prediction accuracy of pK a for small sample data sets of C−H acids in both water and DMSO solvents.…”

“…MF-SuP-p K a is a GNN model integrated with chemical knowledge, using strategies such as knowledge-guided data augmentation and knowledge-aware subgraph pooling to achieve accurate prediction of p K a in aqueous solutions . In our previous work, we established the EEGpKa model based on message passing neural networks (MPNN) . This model introduces a knowledge-infused data augmentation technique, improving the prediction accuracy of p K a for small sample data sets of C–H acids in both water and DMSO solvents.…”

Editorial: Machine Learning in Bio-cheminformatics

“…Many papers on diverse directions are also gathered into this collection to describe the application of ML in cheminformatic studies. A series of articles in this collection focused on the prediction of the physicochemical characteristics of chemical compounds, and these characteristics included: temperature-dependent viscosity, solvation Gibbs energies, p K a, metal coordination geometry, binding energy, and electronic property. − Another set of papers focused on molecular generation and design by introducing software/tool, , developing transformer-based new algorithms, and optimizing molecule via molecular scaffold decoration . The remaining tested the performance of ChatGPT in chemical generation and similarity indexing …”