bjornshomelab/alphafold2-fold-switching-sensitivity

AlphaFold2 Fold-Switching Sensitivity Analysis

Systematic RMSD analysis of 183 proteins from the DeepMind fold-switching benchmark, comparing AlphaFold2 predictions under baseline vs decoy input conditions, with a random perturbation control to establish a noise baseline.

Dataset Summary

This dataset contains per-protein RMSD values comparing AlphaFold2 predictions to experimentally determined structures under three conditions:

1. Baseline vs Experimental — standard AlphaFold2 inference
2. Decoy vs Experimental — AlphaFold2 with modified input sequences/templates
3. Decoy vs Baseline — divergence between the two prediction modes
4. Random Perturbation vs Experimental — experimental structures perturbed by Gaussian noise (σ = 2.0 Å on Cα atoms) as a noise control

Key Findings

Condition	Mean RMSD	Median RMSD	% > 6 Å
Random perturbation	3.44 Å	3.42 Å	0%
Decoy vs Experimental	5.05 Å	2.51 Å	24.0%
Baseline vs Experimental	20.0 Å	19.29 Å	89.6%
Decoy vs Baseline	16.47 Å	16.93 Å	41.3%

The random perturbation control establishes that ~3.5 Å RMSD represents non-meaningful structural variation. Baseline AlphaFold2 predictions deviate from experimental structures far more than this noise floor, while decoy-conditioned predictions often recover closer-to-native conformations.

Dataset Structure

protein_rmsd_analysis.csv / .parquet
├── protein_id: str                    # PDB ID + chain
├── experimental_ca: int               # Number of Cα atoms in experimental structure
├── decoy_vs_experimental_rmsd: float  # RMSD(decoy prediction, experimental)
├── baseline_vs_experimental_rmsd: float # RMSD(baseline prediction, experimental)
├── decoy_vs_baseline_rmsd: float     # RMSD(decoy, baseline)
├── random_perturbation_rmsd: float   # RMSD(perturbed experimental, original experimental)
└── fold_switching_candidate: bool      # True if any RMSD > 6 Å

Source Data

• Experimental structures: DeepMind fold-switching benchmark, AF2Rank/final2_and_debug2_inputs
• Predictions: AlphaFold2 baseline + decoy, AF2Rank/all_folds2_output/pdbs
• Original paper: DeepMind (2025). AlphaFold predictions of fold-switched conformations are driven by structure memorization. Nature Structural & Molecular Biology.

Methods

RMSD Calculation

• Alignment: Kabsch algorithm on Cα atoms
• Proteins with < 3 Cα atoms excluded
• Truncated to common length where experimental and prediction differ

Random Perturbation Control

• Each Cα atom displaced by Gaussian noise N(0, σ²=4.0 Ų)
• Establishes baseline for "non-meaningful" structural variation

Analysis Paper

Full methodology and interpretation: Revealing AlphaFold2's Uncharacterized Sensitivity Landscape: A Knowledge Graph Approach to Protein Prediction Instability (Nous Research Group, 2026-04-24).

Usage

import pandas as pd
df = pd.read_parquet("protein_rmsd_analysis.parquet")

# Baseline predictions are systematically worse than decoy
baseline_worse = df[df["baseline_vs_experimental_rmsd"] > df["decoy_vs_experimental_rmsd"]]
print(f"Baseline worse than decoy for {len(baseline_worse)}/{len(df)} proteins")

# Random perturbation establishes noise floor (~3.5 Å)
print(f"Random RMSD: {df['random_perturbation_rmsd'].median():.2f} Å")

Citation

@dataset{alphafold2_sensitivity_2026,
  title={AlphaFold2 Fold-Switching Sensitivity Analysis},
  author={Nous Research Group},
  year={2026},
  url={https://huggingface.co/datasets/bjornshomelab/alphafold2-fold-switching-sensitivity}
}

License

CC-BY-4.0