Shapley values

[Alex6022]

Introduced SHAP (SHapley Additive exPlanations) analysis of the surrogate model to analyze the feature importance of finished campaigns. This is especially interesting in combination with the molecular encodings that are already built-in into BayBE.

In a previous project from the AC-BO-Hackathon, different molecular encodings were previously tested to screen molecules for high corrosion inhibition. Analyzing the highly succesful MORDRED campaign with the new SHAP functionality yields the following summary plot:

Besides the measurement parameters "Time_h" and "Salt_Concentrat_M", the Mordred-specific features "SMILES_MORDRED_NdS" and "SMILES_MORDRED_nS" suggest the importance of sulphur groups for corrosion inhibition. Interestingly, this is in agreement with previous literature in the field. I hope that this new feature will be of interest for many other applications in the future.