Figure 2.

The statistical limits of any model are derived from characteristics of the training data. The maximum of the model’s Pearson R (R_max) is dictated by the standard deviation of the individual experimental values (σ_expt) and the range of affinities (characterized by σ_data over all the values in the dataset). Clearly, low error in the experiments leads to higher R_max. The diagram above show the effect of varying σ_data while keeping the same σ_expt. They show that weak models (red) have reduced R_max because of the low σ_data that comes from tightly clustered data with few outliers. Robust models (blue) have higher R_max from larger values for σ_data that come from adding data, extending the max–min limits of the data, and/or more evenly covering the range of data.