Iterative machine learning-based chemical similarity search to identify novel chemical inhibitors

Abstract

Machine learning-based chemical screening has made substantial progress in recent years. However, these predictions often have low accuracy and high uncertainty when identifying new active chemical scaffolds. Hence, a high proportion of retrieved compounds are not structurally novel. In this study, we proposed a strategy to address this issue by iteratively optimizing an evolutionary chemical binding similarity (ECBS) model using experimental validation data. Various data update and model retraining schemes were tested to efficiently incorporate new experimental data into ECBS models, resulting in a fine-tuned ECBS model with improved accuracy and coverage. To demonstrate the effectiveness of our approach, we identified the novel hit molecules for the mitogen-activated protein kinase kinase 1 (MEK1). These molecules showed sub-micromolar affinity (Kd 0.1-5.3 μM) to MEKs and were distinct from previously-known MEK1 inhibitors. We also determined the binding specificity of different MEK isoforms and proposed potential docking models. Furthermore, using de novo drug design tools, we utilized one of the new MEK inhibitors to generate additional drug-like molecules with improved binding scores. This resulted in the identification of several potential MEK1 inhibitors with better binding affinity scores. Our results demonstrated the potential of this approach for identifying novel hit molecules and optimizing their binding affinities.

Keywords: Drug design; Evolutionary chemical binding similarity; Hit identification; MEK inhibitor; Virtual screening.

Fig. 1

Schematic representation of ECBS retraining procedure with new chemical pair data. A Simplified example representing evolutionary relationships of chemicals defined by their targets and family information. T1 is a target of interest. B Definition of positive and negative samples is shown with the chemical compounds in A. C Different types of new chemical pair data are shown with newly-identified active (P_new) and inactive compounds (N_new). D The overall iterative VS strategy is illustrated

Fig. 2

Two-dimensional structures of the new MEK1 inhibitors (ZINC5814210, ZINC5479148, and ZINC32911363), experimentally proven through scanELECT and KdELECT services from DiscoverX. The POC at 10 μM and Kd values are annotated (the lower values of POC and Kd indicate higher binding affinities)

Fig. 3

Results of cell viability assay with A A549 and B HT29 cell lines. To evaluate the cellular efficacy of the compounds against cancer cells, A549 lung cancer and HT29 colon cancer cells were selected to observe the relation between the cancer cell growth inhibition and MEK inhibition. PD98059 is a known selective cell permeable inhibitor for MEK1 (IC50 = 2 ~ 7 µM) and MEK2 (IC50 = 50 µM) and was used as a positive control in the present study. The two MEK inhibitors (ZINC5814210 and ZINC32911363) showed activity against A549 and HT29 cancer cells, especially at higher concentrations

Fig. 4

Molecular clustering results of all tested molecules for MEK1. The compounds (six weak or non-binders) not in C1 or C2 are not clustered

Fig. 5

Molecules that are structurally similar to ZINC5814210 were identified using multiple structural criteria, as shown by green (MACCS fingerprints), blue (Morgan fingerprints), and gray (Substructure search) box. Tanimoto Coefficient (TC) was used to define molecular similarity with MACCS or Morgan fingerprints. The known experimental affinity values (POC, IC50, or Kd) are annotated with similarity scores

Fig. 6

Molecular Docking results of ZINC5814210 for MEK1, MEK2, and MEK5. The carbon atoms of ZINC5814210 are shown as yellow sticks. The carbon atoms of residues that participate in interaction with ZINC5814210 and contribute to the hydrophobic environment around the ligand are shown as white sticks. The secondary structures of proteins are shown as cartoon. Hydrogen bond interactions and hydrophobic interactions are represented by black and orange dotted lines, respectively. X-ray crystal structures of human MEK1 and MEK2 (PDB accession codes: 3V01 and 1S9I) and Alphafold predicted structure of MEK5 were used for docking

Fig. 7

Different binding scores for designed molecules generated from ZINC5814210. A The docking conformation of 15 designed molecules were evaluated using different scoring methods (including Autodock Vina, AutoDock4, graphDelta, MM/GBSA, and MM/PBSA). Each score was min–max normalized (0–1) and averaged (Avg. Score). B Molecular Docking result of C07 in MEK1. The carbon atoms of C07 are shown in blue and the beta sheet region (Ala16-Val22, and Lys37) that contacts with C07 is shown in cyan. The X-ray crystal structure of human MEK1 (PDB ID: 3V01) was used for docking