Fig. 1. The framework of ZeroBind.

a Network-based negative sampling strategy. The bipartite network consisting of drugs and protein targets: The square nodes represent the protein nodes and the circle nodes represent the molecule nodes, and there are only edges between different types of nodes, representing the corresponding drug-target interaction. Solid lines represent existing drug-target interactions (DTIs) and dotted lines represent the generated negative interactions with the shortest path distance ≥ 7. b The positive ratio of the training set before the network-based negative sampling strategy. c The positive ratio of the training set after the network-based negative sampling strategy. d Given the support set and query set, ${\mathcal{L}}_{support}$ is first calculated and utilized to update the base model with parameter $\mathit{\theta }$ to a task-specific model with parameter ${\mathit{\theta }}^{\prime }$ using the support set of each task, and then the task-specific model calculates the ${\mathcal{L}}_{query}$ using the query set of the task. After repeating N inner steps, all losses are weighted average by ${\left\{{\mathit{\omega }}_{{\mathbf{T}}_{\mathbf{i}}}\right\}}_{\mathbf{i}\mathit{=}\mathbf{1}}^{\mathbf{N}}$, and gradient descent is further performed to optimize the meta model. e The architecture of the base model in ZeroBind. For each task, the protein graph and the molecule graph are fed into a backbone graph convolutional network (GCN) with parameters ${\mathit{\theta }}_{\mathbf{P}}$ and ${\mathit{\theta }}_{\mathbf{M}}$, respectively, to obtain their embeddings. Subsequently, a SIB module is proposed to generate the IB-subgraph of a protein as potential binding pockets in a weakly supervised way. The protein subgraph embedding is concatenated with the molecular embedding and they are fed into a Multilayer Perceptron (MLP) module to identify the interactions. f Task adaptive attention module. It takes the concatenation of the protein embedding ${\mathbf{G}}_{\mathbf{p}\mathit{,}\mathbf{k}}$ and the average of all molecule embeddings ${\left\{{\mathbf{G}}_{\mathbf{m}\mathit{,}\mathbf{k}}^{\mathbf{i}}\right\}}_{\mathbf{i}\mathit{=}\mathbf{1}}^{\mathbf{m}}$ in the query set as the task embedding. After using the self-attention layer to compute the weight of each task, denoted as ${\left\{{\mathit{\eta }}_{{\mathbf{T}}_{\mathbf{i}}}\right\}}_{\mathbf{i}\mathit{=}\mathbf{1}}^{\mathbf{N}}$, the overall loss is averaged and incorporated into the meta-training process for updating the model parameters. Source data are provided as a Source Data file.