Bridging Molecular Fingerprints and Graph Neural Networks through Knowledge-Guided Pooling

FPPool is a chemistry-aware graph pooling method that uses molecular fingerprints as substructure discovery modules to guide hierarchical aggregation in graph neural networks. Instead of treating molecules as generic graphs, FPPool decomposes each molecular graph into biochemically meaningful substructure groups via bit-level fingerprint matching, then progressively aggregates them through a three-level attention mechanism (atom → substructure → fingerprint → molecule).

Highlights

• 🧬 Chemistry-aware pooling — Uses molecular fingerprints (PubChem, Morgan, RDKit, etc.) to decompose molecules into meaningful functional groups before aggregation
• 📊 Strong performance — Outperforms global pooling (Sum, Mean, Max, Set2Set) and hierarchical pooling (TopKPool, SAGPool, DiffPool, MinCutPool) across 40 benchmarks
• 🔍 Built-in interpretability — Multi-scale attention weights trace predictions to specific atoms, substructures, and fingerprint types without post-hoc methods
• 🔌 Drop-in replacement — Compatible with any GNN backbone (GIN, GCN, GAT, GraphSAGE) as a plug-and-play pooling module
• ⚗️ Activity cliff robustness — Detects subtle structural modifications that drive large potency shifts in activity-cliff datasets

Installation

git clone https://github.com/xiaocui3737/FPPooling.git
cd FPPooling
pip install -r requirements.txt

Dependencies

• Python ≥ 3.8
• PyTorch ≥ 1.12
• PyTorch Geometric ≥ 2.0
• RDKit ≥ 2022.03
• NumPy, Pandas, Scikit-learn

Quick Start

MoleculeNet Benchmark

# Run on BBBP (classification)
python MoleculeNet_example.py --dataset bbbp --fp_type M+Rd+Pub

# Run on ESOL (regression)
python MoleculeNet_example.py --dataset esol --fp_type M+Rd+Pub

MoleculeACE Benchmark (Activity Cliffs)

# Run on a specific target
python MoleculeACE_example.py --target ABL1 --fp_type M+Rd+Pub

Use FPPool in Your Own Pipeline

from fppcode import FPPool, compute_fingerprints

# Replace your existing pooling layer with FPPool
model = YourGNN(pooling=FPPool(
    fp_types=['Morgan', 'PubChem', 'Rdkit'],
    hidden_dim=256
))

# FPPool handles fingerprint computation and matching internally
output = model(batch)

Repository Structure

FPPooling/
├── fppcode/                    # Core implementation
│   ├── model.py                # FPPool pooling module
│   ├── fingerprints.py         # Fingerprint computation & SMARTS matching
│   ├── gnn_backbone.py         # GNN backbone (GIN/GCN/GAT)
│   └── utils.py                # Utilities
├── MoleculeNet_example.py      # MoleculeNet training & evaluation
├── MoleculeACE_example.py      # MoleculeACE training & evaluation
├── requirements.txt            # Dependencies
├── assets/                     # Figures for README
└── README.md

Method Overview

FPPool operates in three stages:

Stage 1: GNN Message Passing — Standard MPNN updates node features through neighborhood aggregation.

Stage 2: Fingerprint-Guided Substructure Matching — Multiple fingerprint types (dictionary-based, circular, path-based) decompose the molecular graph into overlapping substructure groups via bit-level SMARTS matching.

Stage 3: Hierarchical Attention Pooling

Level	Operation	Granularity
InnerPool	Attention-weighted aggregation within each substructure	Atom → Substructure
InterPool	Attention-weighted aggregation across substructures of the same FP	Substructure → Fingerprint
FinalPool	Attention-weighted aggregation across fingerprint types	Fingerprint → Molecule

Each level produces learnable importance weights, providing built-in interpretability at every granularity.

Supported Fingerprints

Type	Name	Bits	Description
Dictionary	PubChem	881	Substructure key-based
Dictionary	MACCS	166	Structural key
Dictionary	RGroup	2048	R-group fragment
Dictionary	Fragments	86	Fragment statistical
Circular	Morgan	1024	Atom environment (ECFP)
Path	RDKit	1024	Bond path-based
Other	E-State	79	Electrotopological state

Results Summary

MoleculeNet (10 datasets)

FPPool (M+Rd+Pub) achieves the best overall performance on all 10 datasets compared to 8 pooling baselines. See the paper for full results.

MoleculeACE (30 activity-cliff datasets)

FPPool achieves the best RMSE on 23/30 datasets and the lowest mean RMSE across all four target families (GPCR, NR, Kinase, Other).

Interpretability

FPPool provides multi-scale explanations without post-hoc methods:

• Inner view: Which atoms matter within each substructure?
• Inter view: Which substructures matter within each fingerprint?
• Final view: Which fingerprint types matter for this molecule?

Citation

If you find FPPool useful in your research, please cite:

@article{cui2026fppool,
  title={Bridging molecular fingerprints and graph neural networks through knowledge-guided pooling},
  author={Cui, Chao and Feng, ShiHang and Song, ChenXiao and Yu, ShunTao and Lu, SongLin and Chen, YuZong and Shen, WanXiang},
  journal={XXXX},
  year={2026}
}

License

This project is licensed under the MIT License — see LICENSE for details.

Contact

• Wanxiang Shen (corresponding) — shenwx25@zju.edu.cn
• Chao Cui — GitHub