Research map / Graph learning / Graph self-supervised learning
Learning Repetition-Invariant Representations for Polymer Informatics
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 GRIN, a max-aggregation GNN aligned with maximum-spanning-tree (Prim-style) message passing plus repeat-unit augmentation (1-RU and 3-RU chains), to learn polymer representations invariant to the number of repeat units.
Problem
Graph neural networks for polymer property prediction model only a single repeating unit, so their embeddings are inconsistent across the varying numbers of repeat units that define the same polymer and fail to extrapolate to long chains.
Approach
Chain repeat units with polymerization anchor vertices to form multi-RU graphs, use max aggregation with L1 sparsity so message passing mirrors Prim's maximum-spanning-tree construction (inheriting DP-style size generalization), and augment training with 1-RU and 3-RU graphs, with a theory argument that 3 repeats is the minimal sufficient augmentation.
Key results
- On homopolymer PolyDensity, GRIN reaches R2 0.747 on Test60 vs. best baseline (BFGCN 0.631), roughly a 10-15% gain claimed in the paper's introduction.
- On copolymer IP with GIN, GRIN reaches R2 0.965 / RMSE 0.092 on Test1 and 0.962 / 0.096 on Test20, versus GIN 0.925 / 0.135 on Test1.
- GRIN achieves cosine similarity 0.999 between embeddings of repeat sizes 1 and 60 on 12,764 unlabeled polymers, vs. 0.672 for GCN and 0.903 for BFGCN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add 3-RU augmentation to the 1-RU training graphs (ablation) | GRIN-RepAug · MeltingTemp and IP, Test1 and Test60/Test20 | R2 | Helped |
| Increase augmentation pair size from {1,2} to {1,3} (ablation) | GRIN · MeltingTemp and IP | R2 | Helped |
| Increase augmentation size beyond 3 RUs (ablation) | GRIN · MeltingTemp and IP | R2 | No clear effect |
| Reduce the 3-RU augmented-sample proportion from 1:1 to 1:0.9 or 1:0.8 (ablation) | GRIN · MeltingTemp Test1 and Test60 | R2: 0.745±0.004 -> 0.736±0.005 -> 0.729±0.005 (Test1); 0.746±0.002 -> 0.734±0.001 -> 0.729±0.004 (Test60) | Hurt |
| Replace max aggregation with mean aggregation (ablation) | GRIN-RepAug · GlassTemp and MeltingTemp Test60 | R2: 0.866±0.014 -> 0.836±0.012; 0.707±0.009 -> 0.674±0.002 (GCN); 0.876±0.009 -> 0.838±0.011; 0.705±0.001 -> 0.666±0.008 (GI | Hurt |
| Replace max aggregation with sum aggregation (ablation) | GRIN-RepAug · GlassTemp and MeltingTemp Test60 | R2: 0.866±0.014 -> 0.818±0.003; 0.707±0.009 -> 0.551±0.055 (GCN); 0.876±0.009 -> 0.815±0.005; 0.705±0.001 -> 0.606±0.009 (GI | Hurt |
| Add 3-RU repeat-unit augmentation (ablation) | GCN baseline · GlassTemp Test10 | R2: 0.824±0.007 -> 0.872±0.004 | Helped |
| Add 3-RU repeat-unit augmentation (ablation) | GCN baseline · MeltingTemp Test10 | R2: 0.668±0.004 -> 0.683±0.002 | Helped |
Surprising findings
- Most size-generalization and OOD baselines (DIR, GREA, IRM, RPGNN, SSR) collapse to negative R2 on Test60, while GCN/GIN retain positive R2 on some tasks.
- DISGEN nearly fails on PolyDensity and O2Perm, attributed to its GNN-Explainer-based augmentation misprioritizing nodes.
Limitations
- Theoretical arguments rely on heuristic assumptions (e.g., L-Lipschitz contraction with L<1, sparsity heuristically encouraging MST alignment) rather than fully proven guarantees.
- Evaluation is limited to six property datasets with a single fixed architecture family (GIN/GCN) and three random seeds.
- Augmentation uses synthetic chains that approximate real polymers; the paper notes real polymers have far more repeat units than tested.
Experimental setup
Datasets: GlassTemp MeltingTemp PolyDensity O2Perm EA IP 12,764 unlabeled polymers (Polymer Info)
Benchmarks: GlassTemp MeltingTemp PolyDensity O2Perm EA IP
Metrics: R2 RMSE cosine similarity between repeat sizes
Baselines: GCN GIN DIR GREA IRM RPGNN SSR DISGEN BFGNN GAT GraphGPS
Builds on: Neural algorithmic alignment (BFGNN) GIN GCN DISGEN SizeShiftReg
Closest papers by meaning
- Improving Self-supervised Molecular Representation Learning using Persistent Homology (NeurIPS 2023 · similarity 0.85)
- N-Gram Graph: Simple Unsupervised Representation for Graphs, with Applications to Molecules (NeurIPS 2019 · similarity 0.85)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.84)
- TRIDENT: Tri-Modal Molecular Representation Learning with Taxonomic Annotations and Local Correspondence (NeurIPS 2025 · similarity 0.83)
- A Plug-and-Play Quaternion Message-Passing Module for Molecular Conformation Representation (AAAI 2024 · similarity 0.83)
- SubgDiff: A Subgraph Diffusion Model to Improve Molecular Representation Learning (NeurIPS 2024 · similarity 0.83)
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.