Biomolecules

Protein folding stability

Summary

Performance in keeping small proteins folded during molecular dynamics. For each protein, an energy minimisation is run from the native (folded) reference conformation, an NVT molecular dynamics simulation at 300 K is seeded with the minimised coordinates, and the ability of the model to retain the fold is measured along the trajectory. If the energy minimisation of a protein fails or is unstable, the MD is instead run from the unminimised input structure.

The benchmark uses a small set of well-characterised proteins (chignolin, tryptophan cage, and a capped villin headpiece) with experimental reference structures. Each protein is solvated in a water box, sized to accommodate the protein, giving systems of 1068 (chignolin), 2075 (tryptophan cage), and 3441 (villin headpiece) atoms.

Metrics

  1. RMSD (chignolin)

  2. RMSD (trp-cage)

  3. RMSD (villin)

For each protein, the root mean square deviation (RMSD) of the C-alpha atoms from the native reference structure is computed for each frame of the trajectory, then averaged over the trajectory. A lower RMSD indicates the fold is retained. Each protein is reported as a separate, equally weighted metric, so a model that fails to simulate any of the proteins has no overall score.

For each protein, a line plot shows the RMSD from the reference structure along the trajectory for each model.

Benchmark speed

Multi-day: one MD simulation per protein. Likely to take several days to run on GPU.

Data availability

Input structures:

  • MLIP Audit benchmark suite, InstaDeep. Native reference structures taken from the Protein Data Bank (chignolin 1UAO, tryptophan cage 2JOF, villin headpiece).

Reference data:

  • Experimental reference structures (X-ray and NMR) from the Protein Data Bank.