Amino acid size, charge, hydropathy indices and matrices for protein structure analysis

Prediction of protein folding and specific interactions from only the sequence (ab initio) is a major challenge in bioinformatics. It is believed that such prediction will prove possible if Anfinsen's thermodynamic principle is correct for all kinds of proteins, and all the information necessary to form a concrete 3D structure is indeed present in the sequence.

We indexed the 200 possible amino acid pairs for their compatibility regarding the three major physicochemical properties--size, charge and hydrophobicity--and constructed Size, Charge and Hydropathy Compatibility Indices and Matrices (SCI & SCM, CCI & CCM, and HCI & HCM). Each index characterized the expected strength of interaction (compatibility) of two amino acids by numbers from 1 (not compatible) to 20 (highly compatible). We found statistically significant positive correlations between these indices and the propensity for amino acid co-locations in real protein structures (a sample containing total 34630 co-locations in 80 different protein structures): for HCI: p < 0.01, n = 400 in 10 subgroups; for SCI p < 1.3E-08, n = 400 in 10 subgroups; for CCI: p < 0.01, n = 175). Size compatibility between residues (well known to exist in nucleic acids) is a novel observation for proteins. Regression analyzes indicated at least 7 well distinguished clusters regarding size compatibility and 5 clusters of charge compatibility. We tried to predict or reconstruct simple 2D representations of 3D structures from the sequence using these matrices by applying a dot plot-like method. The location and pattern of the most compatible subsequences was very similar or identical when the three fundamentally different matrices were used, which indicates the consistency of physicochemical compatibility. However, it was not sufficient to choose one preferred configuration between the many possible predicted options.

Indexing of amino acids for major physico-chemical properties is a powerful approach to understanding and assisting protein design. However, it is probably insufficient itself for complete ab initio structure prediction.