Research map / Graph learning / GNN architectures
Expressivity and Generalization: Fragment-Biases for Molecular GNNs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Fragment-WL, a WL-style expressivity hierarchy for fragment-biased molecular GNNs, and FragNet, a message-passing model over a higher-level fragment graph with a rings-paths fragmentation and ordinal fragment encoding.
Problem
Fragment-biased GNNs are empirically strong for molecules but lack theoretical expressivity analysis, and higher-order GNNs generalize poorly; the paper asks how fragment information should be encoded and how expressive these models are.
Approach
Defines Fragment-WL variants (NF-WL, FR-WL, HLG-WL) that perform WL on graphs augmented with fragment nodes and higher-level fragment edges, yielding a strict expressivity hierarchy; builds FragNet, which message-passes on the augmented graph, uses a rings-paths fragmentation that covers the whole molecule, and an ordinal (class + size) fragment embedding that generalizes to unseen fragments.
Key results
- FragNet achieves 0.0775 MAE on ZINC 10k (best among GNNs) and 0.0237 on ZINC full, and 0.2462 MAE on Peptides-struct and 0.6678 AP on Peptides-func, outperforming all GNNs in the paper's tables.
- On ZINC with 7-rings removed from training, FragNet's test MAE is 0.34 versus 0.61 for GRIT (about 1.8x lower).
- Zero-shot transfer from ZINC to QM9 penalized logP gives FragNet MAE 1.12, versus 1.22 for GRIT and 3.43 for HIMP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FragNet instead of GRIT | Peptides-struct prediction · Peptides-struct | MAE: 0.2460 ± 0.0012 -> 0.2462 ± 0.0021 | No clear effect |
| Use FragNet instead of GRIT | Peptides-func prediction · Peptides-func | AP: 0.6988 ± 0.0082 -> 0.6678 ± 0.005 | Hurt |
| Use FragNet instead of CIN++ | GNNs on Peptides-func · Peptides-func | AP: 0.6569 ± 0.0117 -> 0.6678 ± 0.005 | Helped |
| Use FragNet instead of GRIT | ZINC 10k prediction · ZINC 10k | MAE: 0.059 ± 0.002 -> 0.0775 ± 0.005 | Hurt |
| Use FragNet instead of GRIT | ZINC full prediction · ZINC full | MAE: 0.023 ± 0.001 -> 0.0237 ± 0.00 | No clear effect |
| Use FragNet instead of GRIT | ZINC 10k prediction with unseen 7-rings · ZINC 10k, unseen 7-ring test | MAE: 0.61 -> 0.34 | Helped |
| Use FragNet instead of GRIT | ZINC-full molecules grouped by rarest fragment frequency · ZINC full, rarest fragment <0.1% | MAE: 9.5 -> 5.3 | Helped |
| Use FragNet instead of GRIT | ZINC-full molecules grouped by rarest fragment frequency · ZINC full, rarest fragment <1% | MAE: 0.26 -> 0.15 | Helped |
| Use FragNet instead of GRIT | ZINC-full molecules grouped by rarest fragment frequency · ZINC full, rarest fragment <10% | MAE: 0.026 -> 0.045 | Hurt |
| Use FragNet instead of GRIT | ZINC-full molecules grouped by rarest fragment frequency · ZINC full, rarest fragment ≥10% | MAE: 0.018 -> 0.021 | Hurt |
| Use FragNet instead of GRIT | Zero-shot transfer from ZINC to QM9 · QM9 | MAE: 1.22 -> 1.12 | Helped |
Surprising findings
- GRIT performs slightly better than FragNet on molecules with frequent fragments, which the authors note as a limitation.
- Fragment-biased GNNs as a group outperform higher-order GNNs despite lower computational cost.
Limitations
- Rings-paths fragmentation and ordinal encoding are tailored to molecules and not suited to large dense graphs such as citation or social networks.
- GRIT performs slightly better on molecules with frequent fragments.
Experimental setup
Datasets: ZINC 10k ZINC full Peptides-struct Peptides-func QM9
Benchmarks: ZINC 10k ZINC full Peptides-struct Peptides-func QM9
Metrics: MAE AP recovery rate substructure count accuracy
Baselines: GRIT GPS Graphormer CIN++ HIMP GSN Autobahn GatedGCN SAN RingGNN
Builds on: Weisfeiler-Leman test HIMP CIN GSN
Closest papers by meaning
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.88)
- Improved Expressivity of Hypergraph Neural Networks through High-Dimensional Generalized Weisfeiler-Leman Algorithms (ICML 2025 · similarity 0.88)
- LLaMo: Large Language Model-based Molecular Graph Assistant (NeurIPS 2024 · similarity 0.86)
- MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation (ICLR 2025 · similarity 0.86)
- Towards Bridging Generalization and Expressivity of Graph Neural Networks (ICLR 2025 · similarity 0.86)
- B-XAIC Dataset: Benchmarking Explainable AI for Graph Neural Networks Using Chemical Data (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.