Fig. 1: UniPMT framework.

The P–M–T, P–M and P–T relationships are first represented as a graph, where the initial embedding of P and T is learned via ESM19, and that of M is its pseudo-sequence3. Then, a GNN is applied to learn the embeddings of each input node. Finally, a DMF-based learning strategy is applied to unify the binding prediction tasks for P–M–T, P–M and P–T. w and y denote the weights and prediction scores, respectively.