Reproduce “AlphaFold Database Debiasing for Robust Inverse Folding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Cheng Tan, Zhenxiao Cao, Zhangyang Gao, Siyuan Li, Yufei Huang, Stan Z. Li
The research from Westlake University introduces the Debiasing Structure AutoEncoder (DeSAE) to mitigate systematic geometric biases present in AlphaFold-predicted protein structures. Training inverse folding models on DeSAE-debiased AlphaFold Database (AFDB) structures led to an improvement of up to 18.52 percentage points in sequence recovery rates for models like PiFold, making AFDB data more suitable for sensitive design tasks.
Key insights
- AlphaFold-predicted structures exhibit systematic "geometric idealization" and "distributional shift," characterized by tighter, less diverse distributions of dihedral angles and significantly lower variance in backbone bond lengths compared to experimental PDB structures.
- Even subtle, systematic geometric biases in training data can drastically impair the generalization capabilities and performance of deep learning models in fine-grained tasks like inverse folding.
- Debiasing predicted structures can unlock the potential of large-scale computational databases like AFDB for advanced protein engineering applications, bridging the gap between predicted and experimentally observed conformational landscapes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.