In our characterization of how UniRep contributes to the prediction of double-mutations when trained solely on single mutations, we find that they are often out-performed by one-hot representations. At first glance, this would suggest that learned sequence representations, like UniRep, offer negligible benefits over simple sequence representations like one-hot encodings; however, when applying predictive models to contexts where a large number of consecutive mutants are introduced, the benefit of deep-learned representations is likely to become more pronounced. As has been described in comparative analyses of the role that sequence representations play in protein fitness prediction models, linear models offer adequate performance on fitness landscapes where epistatic effects are limited; the effects of consecutive mutations trend with the additive changes in fitness predicted by the summed effects of their constituent single amino acid mutations \cite{Yang2019}. This trend is clearly illustrated in [supplementary figure X], where the one-hot trained model predictions are nearly identical to predicted additive changes in fitness (calculated using Equation 2). Furthermore, the large degree of similarity between one-hot predictions and additive predictions suggests that the double-mutant data are sampled from a region of considerably low epistatic effects, causing the double-mutant fitness values to trend with the sum of their single-mutant constituents. Following this observation, we decided to derive the conditions under which one-hot representations of protein sequences would lead to predictions that resemble additive fitness changes for consecutive mutations, when training a linear model on single-mutant data (Supplementary Discussion 2). It is expected that one-hot encodings offer adequate performance in regions of minimal epistatic effects, and when a predictive model is only considering simple mutation cases, with low diversity of interaction types, such as the combined effect of two co-expressed mutations. However, the margin of error for any model grows exponentially with each consecutive round of mutagenesis, as each additional mutation brings with it a range of new potential interactions between neighboring residues; this is likely where deep-learned representations such as UniRep show their true value, having been trained to identify patterns of favorable or unfavorable local amino acid arrangements. On top of UniRep’s ability to embed general information regarding amino acid arrangement patterns, it is likely that an evotuned representation would provide further value by embedding taxonomically-specific information, such as where specific residue arrangements should be located within a full amino acid sequence, for a given family of protein, such that the desired tertiary structures will assemble.