PIGNet: A physics-informed deep learning model toward generalized DTI predictions

What is it about?

Recently, deep neural network (DNN)-based drug-target interaction (DTI) models were highlighted for their high accuracy with affordable computational costs. Yet, the models’ insufficient generalization remains a challenging problem in the practice of in-silico drug discovery. We propose two key strategies to enhance generalization in the DTI model. The first is to predict the atom-atom pairwise interactions via physics-informed equations parameterized with neural networks and provides the total binding affinity of a protein-ligand complex as their sum. We further improved the model generalization by augmenting a broader range of binding poses and ligands to training data. We validated our model, PIGNet, in the comparative assessment of scoring functions (CASF) 2016, demonstrating the outperforming docking and screening powers than previous methods. Our physics-informing strategy also enables the interpretation of predicted affinities by visualizing the contribution of ligand substructures, providing insights for further ligand optimization.

Featured Image

Photo by julien Tromeur on Unsplash

Photo by julien Tromeur on Unsplash

Why is it important?

In this work, we studied the inadequate generalization problem of deep learning models that are often encountered in real-world applications where data for training is very scarce and imbalanced. As an important practical example, we focused on drug-target interaction (DTI) models for the fast and reliable virtual screening of drug candidates. The resulting model, named PIGNet, could achieve better generalization as well as higher accuracy compared to other deep learning models. We attribute the success of our model to the following two strategies. The first one is to employ the physics-informed parameterized equations. The physics modeling acts as a proper inductive bias for the neural model, guiding the model to learn the underlying physics of the chemical interactions. We further improved the model performance by augmenting training data with protein-ligand complexes from the wider chemical and structural diversity. We analyzed the effects of the physics-informed model and the data augmentation through the ablation study and found that both contribute to the model generalization. These strategies can be readily adopted by other science fields where similar data problems are expected and the related physics is well-established. Such applications would include materials design, structural biology, and particle physics. A similar improvement in the generalization reported in the atomistic modeling of solid materials is one such example.