PyrMol: A Knowledge-Structured Pyramid Graph Framework for Generalizable Molecular Property Prediction

Accurate prediction of molecular properties is essential for accelerating drug discovery. While graph neural networks (GNNs) have achieved impressive progress, most models rely on atomic graphs with limited chemical semantics, leading to suboptimal generalization across diverse biochemical tasks. Pretraining alleviates this challenge but incurs substantial costs and depends heavily on large chemical corpora. We propose PyrMol, a knowledge-structured pyramid representation learning frame-work that integrates multiple chemical expertise into GNNs. PyrMol constructs heterogeneous pyramid molecular graphs comprising atomic, subgraph, and molecular levels, enabling hierarchical message passing guided by functional groups, pharmacophores, and reaction-derived fragments. A multi-source knowledge fusion module adaptively aggregates heterogeneous subgraph signals, while a hierarchical contrastive strategy aligns representations across scales to enhance discrimination and consistency. Across 10 public benchmarks, PyrMol consistently outperforms 12 state-of-the-art baselines, including advanced subgraph-enhanced architectures and leading pretraining models. Moreover, incorporating pyramid graphs into existing GNNs yields universal performance gains, demonstrating the plug-and-play versatility of framework. PyrMol thus offers a principled and generalizable paradigm for chemistry-informed deep learning, highlighting the value of structured domain knowledge in molecular AI. The source code is available at https://github.com/lyj363636/PyrMol .