Protein phenotype prediction moves past single sites: adding epistasis and amino acid clustering tops R² in 9 of 13 datasets
Synopsis
Quantitative phenotype prediction has mostly relied on how single positions correlate with a trait, which cannot capture several residues acting together. The new scheme adds second-order epistasis and position-specific amino acid clustering on top of a Monte Carlo position search, then fits a regression from sequence distances to phenotype distances. Across 13 datasets spanning temperatures, enzyme inhibition, myoglobin concentration and rhodopsin absorption wavelength, the AC-MC-E variant using all three components reached the highest R² in 9 datasets and ranked in the top three in all 13.
Interpretation
The search for phenotype-modulating positions switches from greedy selection to Monte Carlo sampling and adds two sequence representations: epistasis, which treats two positions as one unit that counts toward distance only when both change, and clustering, which groups amino acids into classes at each position. The earlier single-site version picked positions with a greedy heuristic and compared sequence differences with phenotype differences one position at a time, so it could not express joint effects between positions. Across the 13 cases, measured by the Spearman correlation between sequence distance and phenotype distance in training, adding Monte Carlo search, epistasis and amino acid clustering made AC-MC-E the best of the five variants in both training and prediction.
Prediction first fits a linear regression from sequence distances to phenotype distances on the training set, then assigns each test sequence the phenotype value that minimizes the sum of squared differences between predicted and known phenotype distances. This distance-regression procedure follows the earlier work; what changes is that the position set fed to the regression comes from the new search and may consist of single positions or position pairs. Measured by R² between predicted and known phenotypes of test sequences, the improvement over the single-site version is statistically significant for MC-SS and largest for AC-MC-E; the literature-reported position set performs comparably to the original single-site version.
The two variants divide the work: the single-site-plus-clustering version uses on average about 5% of alignment positions, while the epistasis version predicts better but uses about 30%. A multi-objective analysis weighs the fraction of selected positions against prediction improvement; the literature position set is dominated by algorithm variants, and only AC-MC-SS and AC-MC-E remain on the Pareto front. For HIV-1 protease inhibition by FPV, AC-MC-SS reaches R²=0.60 with 10 positions and AC-MC-E reaches R²=0.70 with 40 positions, the latter spread more widely across the protein.
Perspective
The evaluation covers 13 cases: optimal growth temperature of four enzymes and three transcription factors, optimal function temperature of one enzyme, inhibition of HIV-1 protease by three antiretroviral drugs, maximal muscle myoglobin concentration in mammals, and absorption wavelength of microbial rhodopsins; all conclusions come from computational prediction on these datasets and include no experimental validation.
Amino acid clustering correlates with hydrophobicity (R² of 0.61 and 0.62 for the first principal component), but the composition of each cluster is not fully explained, and no significant correlation was found for the second component across 566 amino acid properties. Overlap with literature-annotated positions is low, with median Jaccard indices of 0.05 and 0.12, and categorical performance is near random in most cases, deviating clearly from the diagonal only for the three HIV-1 protease cases. The authors also note that the most suitable approach depends on the biological and evolutionary determinants of the property studied.