Research map / Probabilistic and scientific modeling / Molecules and proteins
Protein Structure Tokenization via Geometric Byte Pair Encoding
Michael Sun, Weize Yuan, Gang Liu, Wojciech Matusik, Marinka Zitnik
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add glue-angle optimization (differentiable inverse kinematics) (ablation) | GeoBPE · Reconstruction, |V|=600 (train) | RMSD: 4.39 -> 1.66 | Improved |
| add glue-angle optimization (differentiable inverse kinematics) (ablation) | GeoBPE · Reconstruction, |V|=600 (train) | LDDT: 0.69 -> 0.73 | Improved |
| fit vocabulary on task-specific data instead of pretraining data (ablation) | GeoBPE · Functional site prediction (average, 8 tasks, 15 splits) | AUROC (%): 80.20 -> 80.23 | No clear effect |
| fit vocabulary on 1% of pretraining data instead of full (ablation) | GeoBPE · Functional site prediction (average, 8 tasks, 15 splits) | AUROC (%): 80.20 -> 80.12 | No clear effect |
| fit vocabulary on 1% of pretraining data instead of full (ablation) | GeoBPE · OOD reconstruction (CAMEO, CASP14) | RMSD: 1.77 -> 1.66 (CAMEO, |V|=600) | Mixed |
| increase downstream training data from 20% to 100% (ablation) | GeoBPE · Physicochemical property prediction (average, flexibility) | Spearman rho (%): 38.24 -> 45.26 | Improved |
| increase downstream training data from 20% to 100% (ablation) | GeoBPE · Structure property prediction (Homo, average) | Macro F1 (%): 35.92 -> 52.21 | Improved |
| increase downstream training data from 20% to 100% (ablation) | GeoBPE · Functional site prediction (average, 8 tasks, 15 splits) | AUROC (%): 79.19 -> 80.20 | No clear effect |
| increase M_max from 5000 to 20000 motif occurrences (ablation) | GeoBPE · Reconstruction (|V|=6000) | RMSD: 1.40 vs 1.39 | No clear effect |
| use GeoBPE vocabulary features instead of ESM3 VQ tokens | ESM3 · Structure property prediction (Homo, average) | Macro F1 (%): 36.44 -> 52.21 | Improved |
| use GeoBPE instead of ProteinMPNN on flexibility prediction | ProteinMPNN · Physicochemical property prediction (average, flexibility) | Spearman rho (%): 54.41 -> 45.26 | Hurt |
| use GeoBPE instead of ProteinMPNN on repeat motif prediction (Fold) | ProteinMPNN · Rep (Fold split) | AUROC (%): 77.63 -> 56.44 | Hurt |
Closest papers by meaning
- Protein Structure Tokenization: Benchmarking and New Recipe (ICML 2025 · similarity 0.90)
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.88)
- DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules (NeurIPS 2025 · similarity 0.87)
- Beyond Sequence: Impact of Geometric Context for RNA Property Prediction (ICLR 2025 · similarity 0.87)
- Atom-level Protein Representation Learning Improves Protein Structure Prediction (NeurIPS 2026 · similarity 0.87)
- LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling (NeurIPS 2026 · similarity 0.87)
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.
