Fig. 2.

Boxplots of Random Forest (RF) model quality metrics and comparison of different features. The performance indexes were the average of 50 RF models. The result indicated that MMIFs show significant improvement compared to the separated features (e.g. Checkmol, PubChem, in-housed, and ring in drugs). Although models based on MMIFs show slightly lower results compared to the model with ECFP4, MMIFs are able to explain the model results with the substructure features on compounds. Statistical analysis was performed by student T-test with p values compared to the MMIFs. Single stars denote 0.01 < p < 0.05, double stars 0.001 < p < 0.01