Global exchange or readout capacity? A controlled decomposition of the virtual node in molecular graph neural networks

Abstract A virtual node is one artificial vertex joined to every atom of a molecular graph. It is normally justified as a channel for global information exchange, yet adding it also gives the prediction head a second graph-level vector, and the usual ablation changes both at once. The two were separated here with a factorial design whose control runs the virtual node and reads its state but never broadcasts it to the atoms, so that both arms see the same learned second channel. Six arms of an edge-aware graph isomorphism network were trained with 30 seeds on four binary problems: honey bee toxicity under a Max-Min and a temporal partition, a twelve-assay Tox21 panel and OGB MolHIV. The decision threshold was declared before training. On MolHIV, the problem with the largest evaluation set, the whole virtual node gain came from the second readout channel (+0.104 MCC without exchange, +0.068 with it), and exchange itself lowered MCC by 0.036; both effects held in an independent repeat of the run. On the three smaller problems neither factor survived correction, and mean pooling alone reproduced the gain of a sum-and-mean readout on ApisTox Max-Min. A conventional 0.5 threshold reversed the sign of 9 of 24 contrasts.