ZJU scientists develop multi-fidelity modeling for pKa prediction

Acid-base dissociation constant (pK_a), a key physicochemical parameter describing the ionization ability of molecules, plays a vital role in chemical, environmental, agricultural, and medical sciences. In drug discovery, pK_a determines the ionization states of drug-like molecules in the physiological environment, thus affecting a wide range of biopharmaceutical properties.

Graph Neural Networks (GNN) are a set of deep learning (DL) algorithms that can learn representation automatically from graph-structured data for node-, subgraph- or graph-level tasks, and have achieved tremendous success in molecular property predictions. However, due to the complexity of pK_a, the applications of GNN in pK_a predictions are relatively rare, even though they show immense promise.

The research team led by Prof. HOU Tingjun and Hsieh Chang-yu at the Zhejiang University College of Pharmaceutical Sciences developed MF-SuP-pK_a (multi-fidelity modeling with subgraph pooling for pKa prediction), a novel pK_a prediction model that utilizes subgraph pooling, multi-fidelity learning and data augmentation. Their research findings are published in an article titled “MF-SuP-pK_a: Multi-fidelity modeling with subgraph pooling mechanism for pKa prediction” in the journal Acta Pharmaceutica Sinica B.

Graphical abstract

In this model, the researchers drew up a knowledge-aware subgraph pooling strategy to capture the local and global environments around the ionization sites for micro-pK_a prediction. To address the scarcity of accurate pKa data, low-fidelity data (computational pK_a) was employed to fit high-fidelity data (experimental pK_a) through transfer learning. The final MF-SuP-pK_a model was constructed by pre-training on the augmented ChEMBL data set and fine-tuning on the DataWarrior data set. Extensive evaluation of the DataWarrior data set and three benchmark data sets reveals that MF-SuP-pK_a outperformed other state-of-the-art pKa prediction models while calling for much less high-fidelity training data. Compared with Attentive FP, MF-SuP-pK_a improved its accuracy rate by 23.83% and 20.12% in terms of mean absolute error (MAE) on the acidic and basic sets, respectively. 

Overall, this model is expected to serve as a powerful tool for the prediction of such biopharmaceutical properties as absorption, distribution, metabolism, excretion, and toxicity.