Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
AlphaFold Database Debiasing for Robust Inverse Folding
Authors: Cheng Tan, Zhenxiao Cao, Zhangyang Gao, Siyuan Li, Yufei Huang, Stan Z. Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comparative analysis of structural feature distributions reveals that AFDB structures exhibit distinct statistical regularities, reflecting a systematic geometric bias that deviates from the conformational diversity found in experimentally determined structures from the Protein Data Bank (PDB). ... we investigated the performance of several representative inverse folding models when trained on different structural datasets and on a consistent, held-out set of experimentally determined PDB structures. ... As shown in Figure 1(b), models trained on PDB data achieved recovery rates between 34.11% and 43.76%, while those trained on AFDB structures performed markedly worse, with recovery rates ranging from 17.16% to 27.83%. |
| Researcher Affiliation | Academia | 1Shanghai AI Laboratory 2Hong Kong University of Science and Technology 3AI Lab, Research Center for Industries of the Future, Westlake University |
| Pseudocode | No | The paper describes the architecture of De SAE and the Frame Aggregation and Frame Updating modules using mathematical formulas and descriptive text (e.g., equations 3-9, 11), but it does not present a structured pseudocode or algorithm block with a clear 'Algorithm' label. |
| Open Source Code | No | We will release the data and code. |
| Open Datasets | Yes | The Alpha Fold Protein Structure Database (AFDB) [4, 5, 6] provides an unprecedented repository of structural information... experimentally determined structures from the Protein Data Bank (PDB). ... We begin by curating a paired dataset of experimentally determined structures from PDB and their corresponding AFDB data (see Appendix B). |
| Dataset Splits | Yes | We adopt the validation and test splits from CATH 4.2 [78], removing any entries with high sequence similarity with our paired dataset. ... The De SAE model is trained to learn a mapping from potentially biased predicted structures to conformations more representative of experimental observations. To facilitate this, we constructed a dataset of paired protein structures, where each pair consists of: 1. An Alpha Fold-predicted structure sourced from the AFDB [4, 1]. 2. Its corresponding experimentally determined structure from the PDB [13]. ... Following this filtering, our final benchmark test sets comprise: 893 structures from CATH 4.2, 38 from TS50, 382 from TS500, and 1575 from CATH 4.3. |
| Hardware Specification | Yes | De SAE has only 5.9M parameters and is capable of processing process about 20k AFDB structures in 3 minutes on a single NVIDIA A100 GPU. |
| Software Dependencies | No | The paper does not explicitly mention specific software dependencies with version numbers, such as Python or PyTorch versions. |
| Experiment Setup | Yes | Training spanned 60 epochs with an initial learning rate of 1 10 3, a batch size of 16, and a Cosine Annealing LR scheduler to anneal the learning rate. The SE(3) encoder comprised eight equivariant layers, and the decoder comprised six layers, each with a hidden dimensionality of 128. ... Each model was trained for 50 epochs with a learning rate of 1 10 3 and a batch size of 32. |