Finding hidden patterns in graphs like finding objects in images?

Subgraph detection—finding where query patterns appear in larger graphs—is computationally hard. GraphDETR borrows the object-detection framework DETR for graphs: a GNN encodes the target graph, then learned query vectors pinpoint all pattern matches in a single forward pass via transformer decoding. Unlike traditional combinatorial solvers limited to exact matches and small patterns, GraphDETR handles approximate matching and scales to 1000-node graphs with 50-node patterns. On real molecular data (ChEMBL), it achieves 91.2% AP detecting functional groups.