Dear Patrick,

In this benchmark, since it is about regression modeling,
it might be quite interesting to see at least one or two unfolded-counted fingerprints:
e.g. ECFP or atom-pairs.

I might send a PR about this, if I manage to wrap my head
around your piece of code.

Regards,
F.

xgboost might be good to give you a competition winning model; but it's way too slow (at least to my taste)
to get a baseline regressor.

I might contribute a random forest regressor to replace it.

That would also make training models two times faster.

Problem might be: you compute mean and stddev over the 10 folds for each model.
Maybe, you will not trust mean and stddev computed over just 5 folds.

This being said, maybe you want to do iterated testing and do those stats over something like
50 repeats (e.g. we train/test on random partitions of 80%/20% and we repeat this 50
times for each model).

Dear Pat,

You might be interested in such a script:

#!/bin/bash                                                                     
                                                                                
pip3 install tmap mhfp xgboost seaborn jupyter                                  
                                                                                
# # extract pure Python code                                                    
# jupyter nbconvert --to python benchmark_map4.ipynb

I did not probably cover everything needed, but at least it allows to run the benchmark on my computer.

Regards,
Francois.