靜態結構恢復不等於動態穩定:以動力學評測蛋白質基序支架設計
Static Recovery Is Not Dynamic Stability: Dynamics-Aware Benchmarking of Protein Motif Scaffolding
研究團隊指出,靜態基序支架評測不足以判斷蛋白質設計的動態穩定性:55 組靜態評測成功的設計與問題配對中,只有 31 組通過分子動力學驗證,僅 8 組通過所有檢查。團隊推出開放研究資源 BILBY,包括 220 項基於 PROSITE 連結實驗結構的 BILBY-PRO 基準,以及評估結構漂移、基序接觸保留和構象系綜保真度的 BILBY-MD 框架。
Can we trust a protein design just because its predicted structure looks correct?
In our NeurIPS 2026 Evaluations & Datasets paper, we show why static motif-scaffolding benchmarks tell only part of the story. Of 55 statically successful design–problem pairs, only 31 passed our molecular dynamics validation criteria, and just 8 passed every check.
We introduce BILBY, two complementary open research resources:
🧬 BILBY-PRO: 220 systematically derived functional motif-scaffolding benchmarks from PROSITE-linked experimental structures.
🔬 BILBY-MD: A reusable molecular dynamics framework for assessing structural drift, motif-contact preservation, and conformational ensemble fidelity against experimental references.
Our MD study evaluates designs from the original MotifBench benchmark, independently of the new BILBY-PRO dataset.
We hope BILBY will help make evaluation of generative protein-design methods more rigorous, reproducible, and physically meaningful.
Paper: https://arxiv.org/abs/2611.07723
Dataset: https://huggingface.co/datasets/HFMI-PROFOUND/BILBY-PRO
MD framework: https://github.com/PROFOUND-HFMI/BILBY-MD
Feedback and discussions are very welcome!
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co