In this relatively simple, 3-layer deep model developed from 370 compounds, the core node (1) has almost 50-50 distribution of inactives and actives (189 inactives and 181 actives). The Saagar feature SGR10013, which represents the number of C atoms in a molecule, splits the core node into children nodes (2) and (3) with 213 (57.6%) and 157 (42.4%) compounds, respectively. Node 2 (SGR10013 <25) has mostly inactives (175/213 or 82.2% inactives), while node 3 (SGR10013 >=25) has mostly actives (143/157 or 91.1% actives). Thus, one can infer that larger molecules with 25 or more C atoms are likely to be VDR agonists. This structure-based reasoning would have correctly predicted Elocalcitol and Elgocalciferol as VDR agonists, though these two compounds were not part of the 370-compound training set in this example.

However, one must be careful about predictions from such simple, global rules. What if our query compound were a straight chain hydrocarbon with 30 C atoms? This model would falsely predict it to be a VDR agonist. Therefore, before accepting such a prediction, the overall “similarity” of the query molecule with the compounds in the node 3 must be determined as an applicability domain check.

Furthermore, individual predictions can be refined by considering multiple features. For example, in node 2, all compound with <25 C atoms are not inactive; in fact, 38 of the 213 compounds (~18%) are active. Two additional Saagar features, SGR10382 and SGR10829, can identify 17 of those 38 actives in node 11 with 100% purity. A representative compound, CHEMBL 116438, is shown, with SGR10382 highlighted in blue and SGR10829 highlighted in red.

Saagar-based RP models can also be used to help design compounds with desired properties. For example, if one were looking for compounds that are not agonists of VDR, the starting point could be a compound with <25 C atoms and fewer than 4 occurrences of the group with O connected to the middle C of a 3-C chain (SGR10382), because 90% (170/189) of such compounds are inactive in the VDR agonist assay (node 4).

Though the VDR model shown is relatively simple, it demonstrates how the recursive partitioning models built with Saagar features provide chemistry-based reasoning that can be used to predict biological activity or design compounds with desired properties, with confidence assigned based on prediction accuracies at each node. Notably, the Saagar set is extensible and because Saagar substructures and motifs are coded in SMARTS, they can be processed independent of any third-party software.