Deriving general structure-activity/selectivity relationship patterns for different subfamilies of cyclin-dependent kinase inhibitors using machine learning methods

Abstract

Cyclin-dependent kinases (CDKs) play essential roles in regulating the cell cycle and are among the most critical targets for cancer therapy and drug discovery. The primary objective of this research is to derive general structure-activity relationship (SAR) patterns for modeling the selectivity and activity levels of CDK inhibitors using machine learning methods. To accomplish this, 8592 small molecules with different binding affinities to CDK1, CDK2, CDK4, CDK5, and CDK9 were collected from Binding DB, and a diverse set of descriptors was calculated for each molecule. The supervised Kohonen networks (SKN) and counter propagation artificial neural networks (CPANN) models were trained to predict the activity levels and therapeutic targets of the molecules. The validity of models was confirmed through tenfold cross-validation and external test sets. Using selected sets of molecular descriptors (e.g. hydrophilicity and total polar surface area) we derived activity and selectivity maps to elucidate local regions in chemical space for active and selective CDK inhibitors. The SKN models exhibited prediction accuracies ranging from 0.75 to 0.94 for the external test sets. The developed multivariate classifiers were used for ligand-based virtual screening of 2 million random molecules of the PubChem database, yielding areas under the receiver operating characteristic curves ranging from 0.72 to 1.00 for the SKN model. Considering the persistent challenge of achieving CDK selectivity, this research significantly contributes to addressing the issue and underscores the paramount importance of developing drugs with minimized side effects.

Figure 1

The Venn diagram of the VIP-selected descriptors used for developing active/inactive classifiers for five groups of CDK molecules: CDK1: blue, CDK2: crimson, CDK4: green, CDK5: yellow, CDK9: brown.

Figure 2

The density plot, box plot, and beeswarm plot (a–c) for Hy molecular descriptor for the active and inactive groups of CDK1, CDK2, and CDK5 molecules (d–f) for nRCONHR molecular descriptor for the active and inactive groups of CDK1, CDK2, and CDK5 molecules (g–i) for C-043 molecular descriptor for the active and inactive groups of CDK1, CDK4, and CDK9 molecules.

Figure 3

The density plot, box plot, and beeswarm plot (a–c) for nBnz molecular descriptor for the active and inactive groups of CDK4, and CDK5 molecules (d–f) for TPSA_(Tot) molecular descriptor for the active and inactive groups of CDK2 and CDK5 molecules.

Figure 4

One-against-one visualization of the abstract PC space made by the 31 VIP-selected molecular descriptors used for multiclass classification of active CDK molecules (Dark blue: active CDK1, magenta: active CDK2, light blue: active CDK4, gold: active CDK5, orange red: active CDK9).

Figure 5

The density plot, box plot, and beeswarm plot (a–c) for Ms (d–f) for TPSA_(Tot) (g–i) for nCbH molecular descriptor(s) within five different groups of active CDK inhibitors.

Figure 6

The distribution and coordinates of the PubChem-R and active CDK molecules within the two-dimensional subspaces created by one-by-one comparison of four VIP-selected molecular descriptors (a) TPSA_(Tot) versus T (N.. N), (b) TPSA_(Tot) versus Ms, (c) TPSA_(NO) versus T (N.. N), (d) TPSA_(NO) versus T (N.. N).