Computational design of CDK1 inhibitors with enhanced target affinity and drug-likeness using deep-learning framework

Abstract

Cyclin Dependent Kinase 1 (CDK1) plays a crucial role in cell cycle regulation, and dysregulation of its activity has been implicated in various cancers. Although several CDK1 inhibitors are currently in clinical trials, none have yet been approved for therapeutic use. This research utilized deep learning techniques, specifically Recurrent Neural Networks with Long Short-Term Memory (LSTM), to generate potential CDK1 inhibitors. Molecular docking, evaluation of molecular properties, and molecular dynamics simulations were conducted to identify the most promising candidates. The results showed that the generated ligands exhibited substantial improvements in target affinity and drug-likeness. Molecular docking results showed that the generated ligands had an average binding affinity of -10.65 ± 0.877 kcal/mol towards CDK1. The Quantitative Estimate of Drug-likeness (QED) values for the generated ligands averaged 0.733 ± 0.10, significantly higher than the 0.547 ± 0.15 observed for known CDK1 inhibitors (p < 0.001). Molecular dynamics simulations further confirmed the stability and favorable interactions of the selected ligands with the CDK1 complex. The identification of novel CDK1 inhibitors with improved binding affinities and drug-likeness properties could potentially fill the gap in the ongoing development of CDK inhibitors. However, it is imperative to note that extensive experimental validation is required prior to advancing these generated ligands to subsequent stages of drug development.

Keywords: CDK1 inhibitors; Cyclin dependent kinase 1 (CDK1); MD simulations; Molecular docking; Recurrent neural network.

Fig. 1

The results of a principal component analysis performed on the datasets consisting of 200 randomly selected molecules from the second-generation ligands (GEN2) and known CDK1 inhibitors (ORI), based on their physiochemical features. Additionally, further statistical analysis reveals significant differences in the mean values among various groups within the first two principal components (all p-values <0.001).

Fig. 2

The generated ligands were divided into four groups based on Tanimoto similarity using Butina cluster method. From each group, one molecule (1, 3, 5, 7) was randomly selected and listed on the left side (Generated ligands). The corresponding molecules with the highest Tanimoto similarity in CDK1 inhibitors were listed on the right side of each selected compound.

Fig. 3

The graphical representation of the QED (Quantitative Estimate of Drug-likeness) statistics distribution for CDK1 inhibitors (ORI), first-generation ligands (GEN1), and second-generation ligands (GEN2) is presented through box-plots, with the medians indicated by solid lines in the center of each box. The average QED values for these three sets of data are as follows: 0.54 for ORI, 0.66 for GEN1, and 0.73 for GEN2. Notably, the results demonstrate a significant enhancement in QED values between the ORI and GEN1 groups, as well as between the GEN1 and GEN2 groups (ORI vs. GEN1 and GEN1 vs. GEN2, with all p-values <0.001 according to t-test analysis).

Fig. 4

The distribution of molecular weight (MW) and the number of hydrogen bonds in the reference and generated ligands. (A) The graph illustrates a gradual decrease in MW across the three datasets. The average MW values are reported as 486.7, 387.5, and 368.8, respectively. Significantly different MWs are observed between the ORI and GEN1 datasets, as well as between the GEN1 and GEN2 datasets (p < 0.001, t-test). (B) The bar chart reveals that the number of hydrogen bond donors of the second-generation ligands tends to concentrate around one after two rounds of selection (>50 % of compounds), and no compound has more than four hydrogen bond donors.

Fig. 5

The distribution of binding affinity towards CDK1 is examined among reference ligands (ORI), first-generation (GEN1) ligands, and second-generation (GEN2) ligands. The average binding affinities are reported as −10.2, −10.4, and −10.7 kcal/mol, respectively. Statistical analysis using the Kruskal-Wallis test reveals significant differences in binding affinity among the three datasets (p < 0.0001).

Fig. 6

3D visualization of docked ligands interacting with CDK1, where hydrogen bonds are represented by blue solid lines, hydrophobic interactions by gray dashed lines, and π stacking interactions by green dashed lines. The figures were generated using the PLIP server [37].

Fig. 7

Prediction of the pharmacokinetics of known CDK1 inhibitors (ORI), first-generation ligands (GEN1), and second-generation ligands (GEN2).

Fig. 8

Analysis of MD trajectories from simulations of CDK1 complex. In each case, MD simulation of 100 ns was performed. (A) RMSD of the protein backbone in complex with the four selected ligands and the known CDK1 inhibitors LZ9. (B) RMSD of the five ligands.

Fig. 9

Root-mean-square fluctuation (RMSF) analysis was performed on the Cα atoms of residues belonging to the three chains (A: CDK1, B: CyclinB1, C: CKS2) in the CDK1 complex.

Fig. 10

Radius of gyration (R_g)of the CDK1 complex structure.

Fig. 11

Protein-Ligand Hydrogen Bonds and Interaction Energy Analysis. (A) The number of hydrogen bonds between CDK1 and the selected ligands is depicted, with a 50-ps running average plotted for improved clarity. In the 100 ns trajectory simulation of the Ligand1-CDK1 complex, the number of hydrogen bonds remains relatively stable, with the majority of the time exhibiting one hydrogen bond. Notably, Ligand LZ9 maintains an average hydrogen bond count of 3.3 throughout the simulation period. (B)The Ligand3-CDK1 complex demonstrates the highest stability throughout the simulation despite having an average number of hydrogen bonds less than 1. This suggests that other non-covalent interactions, such as hydrophobic and van der Waals forces, contribute significantly to the overall binding stability and affinity.