Fig. 1: The framework of PIANO.

a The graph masked self-distillation learning module for pre-training to capture intricate structural patterns indicated in the structural context of an amino acid. Both encoders and both decoders were configured with the graph transformer and GCN with ARMA filters, respectively. The mask and re-mask strategies were applied before encoders and decoders, respectively. This module was learned by minimizing the loss function which is the summation of feature mask reconstruction loss and the graph similarity comparison loss (distillation loss). b The multi-branch network module for ΔΔG prediction, the student encoder pre-trained in (a), the GCN with ARMA filters, and the CNN were employed for residue, atom, and protein sequence embeddings, respectively. c The illustration of a graph structure. d Implementation details of graph transformer. Q, K, and V are query, key, and value matrices, respectively. e Implementation details of GCN with ARMA filters. \(\hat{L}\) is the modified Laplacian matrix, W and H are learnable weight matrices. f The self-attention pooling on atom-level embeddings.