Validate network predictions

[MihailBogojeski]

Included notebook and python script for evaluating how model predictions are distributed per residue. The purpose is to see whether the model tends to predict the similar affinity values for the same residue type regardless of the neighboring environment.

The residue-wise affinity predictions are collected for all samples in the dataset, and the means and standard deviations are printed for each dataset. If the means are similar among different datasets with very low deviation, then the model tends to predict the same values regardless of the environment.

An additional analysis is provided correlating the model RMSE across different different datasets, with the deviation from the training mean. This analysis is intended to show whether the model does better than the mean predictor, and whether the error is dependent on the deviation from the mean.

[moritzschaefer]

Let's leave this PR open, so we can share/discuss code here on Marco's next/new model?