Graph neural networks (GNNs) have transformed molecular property prediction — but one component has remained surprisingly underdeveloped: graph pooling, the step that compresses all atomic information into a single molecular vector. Most molecular GNNs rely on simple sum or mean pooling, ignoring decades of chemical knowledge encoded in molecular fingerprints.
Standard pooling approaches have well-documented failure modes. Sum pooling introduces size bias; mean pooling washes out discriminative contrast; max pooling suffers from outlier dominance. More sophisticated hierarchical methods (DiffPool, TopKPool) were designed for large social graphs — when applied to small molecular graphs, they often produce clusters with no biochemical meaning, and may destroy chemically important functional groups.
The deeper issue: all existing pooling methods ignore the fact that molecules are not arbitrary graphs. They are structured assemblies of well-characterized functional units — hydroxyl groups, pharmacophores, ring systems — that chemists have encoded for decades as molecular fingerprints.
FPPool uses molecular fingerprints not as input features, but as chemistry-aware guides that directly control how the molecular graph is pooled. The framework:
Each attention level produces learnable importance weights — providing built-in interpretability at atom, substructure, and fingerprint granularities, without any post-hoc explanation methods.
We evaluated FPPool on 10 MoleculeNet benchmarks and 30 activity-cliff datasets from MoleculeACE. Key findings:
Activity cliffs are pairs of structurally similar molecules with very different biological activities — the so-called "magic methyl effect" where a single methyl group can shift potency by orders of magnitude. Current GNNs fail on these because they map similar structures into overlapping feature spaces.
FPPool addresses this by forcing the model to attend to individual substructures. Visualization shows that FPPool consistently assigns highest importance weights to the precise regions where paired molecules differ, rather than the shared scaffold. This produces more meaningful structure-activity relationship explanations for medicinal chemists.
FPPool exemplifies a design philosophy with wide applicability: rather than treating neural network components as generic building blocks, encoding established domain knowledge directly into the architecture yields models that are simultaneously more performant, more interpretable, and more aligned with expert reasoning.
In molecular machine learning, decades of cheminformatics research have produced rich structural vocabularies — fingerprints, pharmacophore models, scaffold decompositions — that remain largely untapped by current deep learning architectures. FPPool demonstrates one concrete way to bridge this gap at the pooling stage.