Ligand-based virtual screening (LBVS) seeks strong early enrichment when searching ultra-large libraries, but practical screening often relies on 1D/2D descriptions while 3D information is expensive and uncertain due to conformer generation and alignment. We propose GNN-MA, a retrieval-style pairwise scoring model for query–candidate molecular pairs that uses molecular graphs as a unified representation. Built on intra-graph message passing, GNN-MA adds cross-graph attention to learn atom-level soft alignment that focuses on key substructures relevant to activity matching, and introduces a bond-to-atom semantic aggregation module to better exploit chemical bond cues for similarity scoring. The framework uses 2D molecular graphs derived from SMILES for retrieval-style matching and does not rely on explicit 3D conformational modeling or alignment. Experiments on DUD-E and LIT-PCBA show that GNN-MA achieves competitive overall discrimination (ROC-AUC) and, relative to its ablated variants, provides consistent gains in early-enrichment metrics (EF@1–5%) on DUD-E, while on LIT-PCBA the improvements are more target-dependent. The learned atom-level soft alignment also provides a qualitative interpretability cue in case studies. Throughput benchmarks suggest that GNN-MA is most suitable as a re-ranking/refinement model after a fast prefiltering stage.