For other models, we used default parameters set in the scikit-learn library.
For the selected set of models we used model averaging to address model uncertainty64.
We also compared the performance in terms of computing time for all the models we used, reported in column 3 of Table 3.
In order to identify a common transcriptional response across models, we used a vote counting method as previously described for large-scale meta-analysis15.
One issue with the nonlinear models we used in this study is estimating the asymptote.
To compare the models, we used Akaike Information Criterion (AIC) and Bayesian Information Criterion BICC).
For all models, we used an unstructured working correlation matrix and a robust estimator covariance matrix.
To construct k-nearest neighbor models we used weka.classifiers.lazy.Ibk.Ibk
For testing the different hypotheses represented by the models, we used the likelihood ratio test (LRT).
For the tiger and clouded leopard models, we used two different exemplars.
In all models, we used additive SNP effects with no interactions.
