(a) The architecture of the proposed APN, where we plot a two-way 2-shot task from Tox21. APN is optimized over a set of training tasks. Within each training task |$T_{t}$|⁠, the support set is used to obtain the prototypes for each class and the query set is used to optimize the parameters of the moleclue encoder and AGDA module. A query molecule |$x_{t}$| is represented as |$\mathbf{z}^{\prime}$| by the moleclue encoder and AGDA module, which is used to compare the similarity with prototypes for the final prediction. (b) The attribute extractor. The attribute extractor can not only extract three types of fingerprint attributes from 14 molecular fingerprints, including single fingerprint attributes, dual fingerprint attributes, and triplet fingerprint attributes, but also extract deep attributes through self-supervised learning methods, where DFP1, DFP2...DFPn are deep fingerprints generated by self-supervised learning methods 1 to |$n$|⁠. (c) The overall framework of the proposed AGDA. All nodes representations of a molecule sequentially pass a attributes-guided local-attention module and a attributes-guided global-attention module to obtain the final attributes-refined molecular representation.