Fig. 2

Dataset extraction process for a single BP \(\mathscr {S}\). (1) The BP’s model is downloaded from the Biomodels repository. (2) The BP is converted to ODEs and simulated to steady state multiple times. (3) The sensitivity is computed from the simulations’ results for each possible input/output pair in the BP. (4) The BP-related protein interaction graph \(\mathcal {G}_{\mathscr {S}}\) is built by retrieving the BioGRID interactions among the proteins in the BP; note that multiple species can be mapped to the same protein (orange arrows), e.g. \(s_1, s_2\) to \(u_1\), and a single species can be mapped to multiple proteins, e.g. \(s_5\) to \(u_2\) and \(u_4\). (5) The DyPPIN dataset consists of the DPs for each I/O species pair that are mapped to each protein-protein pair. (6) Each DyPPIN data sample induces a graph \(\mathcal {G}_{\mathscr {S}}^{in,out}\) for the DGN training: the skeleton conveys the PPIN subgraph topology, while the node features \(\mathscr {X}_{\mathscr {S}}^{in,out}\) represent whether the node is the input \(u_{in}\) or the output \(u_{out}\). Optionally, the node features can be augmented with protein embeddings from UniProt (7). The graphs \(\mathcal {G}_{\mathscr {S}}^{in,out}\) and their sensitivity labels \(y_{\mathscr {S}}^{in,out}\) are used as input and the target variables of our graph classifier