1. Introduction

      Existing drug-target interaction (DTI) prediction methods generally fail to generalize well to novel (unseen) proteins and drugs. In this study, we propose a protein-specific meta-learning framework ZeroBind with subgraph matching for predicting protein-drug interactions from their structures. During the meta-training process, ZeroBind formulates training a model per protein, which is also considered a learning task, and each task uses graph neural networks (GNNs) to learn the protein graph embedding and the molecular graph embedding. Inspired by the fact that molecules bind to a binding pocket in proteins instead of the whole protein, ZeroBind introduces an unsupervised subgraph information bottleneck (SIB) module to recognize the maximally informative and compressive subgraphs in protein graphs as potential binding pockets. In addition, ZeroBind trains the models of individual proteins as multiple tasks, whose importance is automatically learned with a task adaptive self-attention module to make final predictions. The results show that ZeroBind achieves superior performance on DTI prediction over existing methods, especially for those novel proteins and drugs, and performs well after fine-tuning for those proteins or drugs with a few known binding partners. The framework of ZeroBind is shown in Figure 1.

figure1
Figure 1. The framework of ZeroBind. Given the support set and query set, support loss is first calculated and utilized to update the base model to a task-specific model using the support set of each task, and then the task-specific model calculates the query loss using the query set of the task. After repeating N inner steps, all losses are weighted average and gradient descent is further performed to optimize the meta model. (b) The architecture of the base model in ZeroBind. For each task, protein graph and molecule graph are fed into a backbone GCN, respectively, and obtain the embeddings of proteins and molecules. Subsequently, a SIB module is proposed to generate the IB-subgraph of a protein as potential binding pockets. The protein subgraph embedding is concatenated with molecular embedding and they are fed into a MLP module to identify the interactions. (c) Task adaptive attention module. It takes the concatenation of protein embedding and the average of all molecule embedding in the query set as the task embedding. After using the self-attention layer to compute the weight of each task, the overall loss is averaged and incorporated into the meta-training process for updating the model parameters.

2. Input

      First, for binding predictions of proteins and drugs, please input the protein 3D structure (in PDB format) and molecule smiles.
      Automatically, ZeroBind will search the well trained protein-specific model against the protein you input. If unfortunately our data set does not contain the protein you input, ZeroBind will turn to use the meta-model for the binding prediction task.

3. Output

      We will send the results to your email when the job is finished. Results will be shown in the result page (example) when the job is finished. In addition, results can be downloaded by clicking "Download results".