Protein Sampling¶
Purpose¶
This benchmark evaluates the quality and accuracy of Machine Learning Interatomic Potentials (MLIP) by analyzing the conformational sampling of amino acids in small proteins during molecular dynamics simulations. Specifically, it computes backbone Ramachandran angles (phi/psi) and side chain rotamer angles (chi1, chi2, …). The sampled probability distribution of these angles is then compared against reference data [1] and outliers are detected.
Description¶
This benchmark evaluates the conformational sampling of the protein simulations of the folding stability benchmark. The sampled probability distribution of backbone and side chain dihedrals in these simulations is compared to a reference distribution. The main metrics are the RMSD and the Hellinger distance between the sampled and reference distributions. We also compute the outliers ratio of the sampled dihedrals. An outlier is defined as a conformation that is far away from any point of the reference data.
Dataset¶
See dataset section of the folding stability benchmark.
Interpretation¶
The RMSD and the Hellinger distance are measures of the similarity between the sampled and reference distributions. The lower the value, the more similar the distributions are. The outliers ratio provides a measure of how often the MLIP samples conformations that do not appear in the reference data. The lower the value, the fewer outliers there are. An MLIP with an outlier ratio higher than 0.3 should be considered as not sampling the protein conformational space correctly.