One interest of our lab is predicting where evolution will go, in the genotype space, and which paths evolution can take to improve fitness. To this end, fitness landscapes are a useful concept for understanding the predictability of evolution since they relate the genotype to the reproductive success of an individual. Empirical fitness landscapes are especially important because they allow us to compare real data with the theoretical work that has accumulated over more than 80 years of research.

The latest collaboration between the Bank and Jensen labs examines an unprecedentedly large empirical fitness landscape composed of amino-acid changing mutations in the heat shock protein Hsp90 in yeast. Our collaborator Ryan Hietpas generated and screened mutants that are up to six amino-acids away from the parental genotype, allowing the authors to look at the interactions between mutations (epistasis) and compare subsets of the data (local landscapes) to the entire, global landscape. Their main findings are:

1. The landscape is dominated by epistasis. Most of the landscape experiences antagonistic epistasis, where mutations that are neutral or beneficial when alone have a detrimental effect when they are found together. In contrast, the global fitness peak of the landscape is the result of synergistic epistasis, where mutations that are neutral or slightly beneficial when alone yield high reproductive success in combination.
2. Studying potential adaptive walks on the landscape shows that although the global peak can theoretically be reached most of the time, adaptation may stall at an intermediate fitness peak, which is reached with high probability from the parental genotype.
3. Analyses of the entire set of mutations by comparing estimated landscape statistics to expectations from theoretical landscape models suggest that the landscape is heterogeneous and its topography is globally hard to predict.

This work highlights the dual nature of epistasis: a landscape with no interactions between mutations is just a single global peak (globally predictable), and evolution will always end up at the top but it is impossible to know what paths it will take (locally unpredictable). Conversely, a landscape dominated by interactions has few viable paths (locally predictable) but each one leads to a different peak (globally unpredictable).

The empirical fitness landscape studied here suggests that our current theoretical understanding of fitness landscapes is not sufficient to describe empirical landscapes. More generally, this work draws attention to the importance of comparing and informing theoretical models with empirical data, and proposes approaches to improve this process in the future. The paper is available online in PNAS or as a preprint.