Research map / Label-efficient and robust learning / Recognition and long-tail
RFL: Simplifying Chemical Structure Recognition with Ring-Free Language
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 Ring-Free Language (RFL), which decomposes molecular graphs into a skeleton, ring structures and branch information, and a Molecular Skeleton Decoder (MSD) that predicts them hierarchically for end-to-end optical chemical structure recognition.
Problem
End-to-end optical chemical structure recognition struggles with complex 2D molecular structures (rings and multiple branches) when they are serialized into 1D markup such as SSML via depth-first traversal.
Approach
Equivalently decouple a molecule into a ring-free skeleton (rings collapsed into SuperAtoms/SuperBonds), individual ring structures, and branch connection info, and decode them hierarchically with a skeleton generation module plus a branch classification module.
Key results
- MSD-DenseWAP reaches 64.92% EM and 73.15% Struct-EM on EDU-CHEMC vs 61.35% EM for DenseWAP baseline.
- MSD-RCGD reaches 65.39% EM on EDU-CHEMC vs 62.86% for RCGD and 95.23% EM on Mini-CASIA-CSDB vs 95.01%.
- Ablation on EDU-CHEMC: removing both MSD and [conn] gives 38.70% EM, while full system gives 64.96% EM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply RFL and MSD to RCGD | RCGD · EDU-CHEMC | EM (%): 62.86 -> 65.39 | Helped |
| Apply RFL and MSD to RCGD | RCGD · EDU-CHEMC | Struct-EM (%): 71.88 -> 73.26 | Helped |
| Apply RFL and MSD to RCGD | RCGD · Mini-CASIA-CSDB | EM (%): 95.01 -> 95.23 | Helped |
| Apply RFL and MSD to RCGD | RCGD · Mini-CASIA-CSDB | Struct-EM (%): 95.38 -> 95.58 | Helped |
| Add MSD (ablation) | MSD-DenseWAP · EDU-CHEMC | EM: 38.70 -> 52.76; 44.02 -> 64.96 | Helped |
| Add MSD (ablation) | MSD-DenseWAP · EDU-CHEMC | Struct-EM: 49.45 -> 58.58; 55.77 -> 73.15 | Helped |
| Add [conn] (ablation) | MSD-DenseWAP · EDU-CHEMC | EM: 38.70 -> 44.02; 52.76 -> 64.96 | Helped |
| Add [conn] (ablation) | MSD-DenseWAP · EDU-CHEMC | Struct-EM: 49.45 -> 55.77; 58.58 -> 73.15 | Helped |
Surprising findings
- Gains on the printed Mini-CASIA-CSDB are small (MSD-RCGD only +0.22 EM), which the authors attribute to fewer complex molecules in that dataset.
Limitations
- Gains on printed data are marginal, attributed to low proportion of complex molecules.
- Relies on a hand-designed decomposition and [conn] token; evaluated only on two datasets plus a self-built complexity set.
Experimental setup
Datasets: EDU-CHEMC Mini-CASIA-CSDB ChEMBL-derived complexity dataset
Benchmarks: EDU-CHEMC Mini-CASIA-CSDB
Metrics: EM Struct-EM
Baselines: Imago WYGIWYS BTTR ABM CoMER DenseWAP RCGD
Builds on: DenseWAP RCGD DenseNet GRU attention decoder
Closest papers by meaning
- Atom-Level Optical Chemical Structure Recognition with Limited Supervision (CVPR 2024 · similarity 0.89)
- MarkushGrapher: Joint Visual and Textual Recognition of Markush Structures (CVPR 2025 · similarity 0.88)
- TinyChemVL: Advancing Chemical Vision-Language Models via Efficient Visual Token Reduction and Complex Reaction Tasks (AAAI 2026 · similarity 0.86)
- ChemVLM: Exploring the Power of Multimodal Large Language Models in Chemistry Area (AAAI 2025 · similarity 0.85)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.84)
- RingFormer: A Ring-Enhanced Graph Transformer for Organic Solar Cell Property Prediction (AAAI 2025 · similarity 0.84)
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.