Mode prediction results from both the testing portion of the model fitting dataset and the Briggs dataset fitted using the L2 objective function.

For each CF and AHo position in a given dataset, we determine whether the predicted mode (i.e. highest-frequency amino acid) from our best model is the same as the actual mode. Results are aggregated based on whether or not the input sequence has the correct mode. At the left side of the vertical bar (∣) is the count for the germline predicted modes (i.e. situations when the predicted amino acid mode is the naive sequence amino acid) and at the right side is the count for the non-germline predicted modes (vice-versa).