Research map / Probabilistic and scientific modeling / Molecules and proteins
Ensembits: an alphabet of protein conformational ensembles
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.
Introduces Ensembits, a residual-VQ tokenizer over sets of protein conformational frames that uses permutation-invariant encoding and single-frame-to-token distillation to produce dynamics-aware discrete tokens.
Problem
Existing protein structure tokenizers capture only static local geometry and miss correlated motions and alternative conformations present in protein ensembles; tokenizing ensemble dynamics is unaddressed.
Approach
Tokenize per-residue multisets of SE(3)-invariant descriptors computed with frame-dependent kNN neighbors, using a PerceiverIO-style permutation-invariant set encoder, a residual VQ-VAE with Hungarian-matched reconstruction, and a distillation loss that aligns single-frame embeddings with full-ensemble embeddings (SFTD).
Key results
- Ensembits P=full reaches RMSF Spearman 0.604 on mdCATH-div, beating the strongest baseline ProtProfileMD_K (0.579) by ~2.5 points.
- Ensembits P=1 (single frame) achieves RMSF Spearman 0.458 on mdCATH-div, the best single-frame tokenizer, ahead of AminoAseed (0.408).
- On ProteinGym zero-shot fitness (96 assays), Ensembits P=1 blended with ESM2 gives mean Spearman 0.518 vs 0.484 for ESM2 alone, the best blend among tokenizers.
Surprising findings
- Basis-reparameterization from AminoAseed was relatively ineffective in this setting, attributed to small data scale.
- Standalone Ensembits P=1 trails ESM3struct on ProteinGym (0.381 vs 0.384 mean Spearman), though the blend with ESM2 is best.
Limitations
- Tokenizer quality is bounded by the coverage and fidelity of MD trajectories, which are costly to generate.
- Augmentation with generated ensembles (AFSample2) is only a probe and quality depends on generator maturity.
- Standalone functional performance does not consistently beat ESM3struct (e.g., EC and GO on the sequence split).
Experimental setup
Datasets: mdCATH MISATO ProteinGym ProteinShake
Benchmarks: mdCATH-div RMSF MISATO RMSF ProteinGym ProteinShake EC ProteinShake GO ProteinShake binding site ProteinShake binding affinity
Metrics: Spearman correlation ANOVA AUROC AP skill score
Baselines: AA 3Di_tokens Random AminoAseed ProToken ESM3struct ProtProfileMD_K Vote_3Di ESM2-650M
Builds on: VQ-VAE Residual VQ-VAE Foldseek 3Di PerceiverIO ESM3
Closest papers by meaning
- Protein Structure Tokenization: Benchmarking and New Recipe (ICML 2025 · similarity 0.88)
- Predicting directional flexibility in proteins (NeurIPS 2026 · similarity 0.87)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.86)
- LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling (NeurIPS 2026 · similarity 0.86)
- Structure Language Models for Protein Conformation Generation (ICLR 2025 · similarity 0.86)
- Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression (NeurIPS 2025 · 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.