PAIRMAP: A Unified Geometry-Aware Pairwise-Map Framework for Molecular Representation Learning

Molecular representation learning is fundamental to drug discovery, yet existing methods have key limitations: they often miss the pairwise interactions that determine molecular properties, rely on fixed geometric basis functions unable to adapt across tasks, and lack explicit physical constraints. To address this, we propose PAIRMAP, a unified framework centered on geometry-informed atom pair embeddings. PAIRMAP introduces three synergistic innovations: (1) learnable geometric encoding with task-adaptive basis functions, (2) geometric triangular attention incorporating physical constraints such as the triangle inequality and cosine law as learnable biases, and (3) hierarchical pair pooling to aggregate interactions across chemically meaningful distance scales. The E(3)-equivariant framework is broadly applicable to 2D graphs, 3D conformations, and protein-ligand complexes. On 20 benchmarks, PAIRMAP achieves state-of-the-art performance, including an 80.08% average ROC-AUC on MoleculeNet classification, a 29.9% MAE improvement on the QM8 quantum benchmark, and a 1.140 kcal/mol RMSE in binding affinity prediction. Under stringent protein sequence identity constraints (≤30%) on the ATOM3D LBA benchmark, PAIRMAP achieves the lowest RMSE among all methods (1.308 kcal/mol) and substantially outperforms conventional geometric GNNs in Pearson correlation, confirming that its physical constraints generalize to structurally novel proteins. Crucially, PAIRMAP outperforms large-scale pretrained models without requiring expensive conformational pretraining.