An interpretable machine learning model for selectivity of small molecules against homologous protein family - supplementary material

  

Supplementary Table 1: Hyperparameters values of five machine learning models trained to predict the selectivity of small molecules against JAK family of proteins.

 Supplementary Table 2: Hyperparameters values of five machine learning models trained to predict the selectivity of small molecules against DRD family of proteins 

  

Supplementary Table 3: Rp of Fragments that are preferred for active small molecules of JAK1, JAK2, JAK3 and TYK2.

  

Supplementary Table 4: Rn of fragments that are preferred for inactive small molecules of JAK1, JAK2, JAK3 and TYK2

  

Supplementary Table 5: Rp of fragments preferred for active small molecules of DRD1, DRD2, DRD3, DRD4, and DRD5

  

Supplementary Table 6: Rn of fragments that are preferred for inactive small molecules of DRD1, DRD2, DRD3, DRD4, and DRD5.

 Supplementary Figure 1: The input and output representation used for A. XGBoost and B. Other models 

 Supplementary Figure 2: The fragments that are predominantly present in the inactives of JAK1, JAK2, JAK3 and TYK2 obtained from XGBoost, followed by SHAP method 

 Supplementary Figure 3.  The prevalent fragments in actives of  DRD1, DRD2, DRD3, DRD4, and DRD5 proteins. The ratio of positives (Rp) value for each of the fragments is also provided. 

  

Supplementary Figure 4: Molecules containing fragments which provide selectivity to JAK1, JAK2 and JAK3 inhibitors. For each example, the selectivity feature is highlighted in red and the corresponding Rp value is mapped from the supplementary Table 3.