Fitness Extraction from Genotype Correlations in Time Series

We introduce a simple method for computing fitness from genotype frequency time series data. Unlike many existing methods which make assumptions on the order of epistasis present in the fitness landscape, our approach uses time-dependent genotype correlations to estimate fitnesses on landscapes possessing arbitrarily high order of epistasis. Starting by time averaging the diffusion limit of population genetics, we show that fitnesses can be estimated from time-averaged frequency correlations between genotypes only using singular value decomposition. We validate the approach on a variety of microscopic models and find that even in the presence of noisy frequency trajectories, the method is able to estimate fitnesses with high accuracy for variants ranging over several orders of magnitude in terms of average frequency.