Homology modeling

Protein models and homology models were generated using Prime (Schrödinger Release 2019–4: Prime, Schrödinger, LLC, New York, NY, 2020) [14]. For protein models, crystal structures were imported into Prime by PDB ID (4FG7, 4FG8, 4FG9). Using Prime Protein Preparation Wizard, missing loops and side chains were filled in, missing hydrogens were added, and the protein underwent restrained energy minimization. Water molecules with less than 3 contacts with the protein were removed. For homology modeling of PNCK, the canonical amino acid sequence was obtained from UniProt (Q6P2M8). All homology models underwent loop refinement using serial loop sampling. For all models, extended molecular dynamic simulations (500ns-1us) were run to assess model validity and stability. All models contained co-crystallized ligands and as such, ligands were manually removed and then re-docked into the binding site to assess for model accuracy in predicting the experimentally derived binding pose and residue contacts and interactions. Homology models were assessed using SWISS-MODEL Structure Assessment tool [15,16] for structural reliability which generates a MolProbity score of global structural assessment and Ramachandran plots to ascertain allowed and disallowed dihedral angles of residues in comparison to resolved parent crystal structures. Additionally, our generated homology models were compared to those predicted by Alpha Fold via structural alignment and RMSD [17]. "[The MolProbity score] is a log-weighted combination of the clashscore, percentage Ramachandran not favored and percentage bad side-chain rotamers, giving one number that reflects the crystallographic resolution at which those values would be expected. Therefore, a structure with a numerically lower MolProbity score than its actual crystallographic resolution is, quality-wise, better than the average structure at that resolution” [16]. Residue-by-residue overlay analysis and analysis of binding site interactions between both models showed considerable overlap in structure and binding mode of ATP. Variability existed mainly in the alpha-c loop. (Fig A in S1 File).

Docking Grid Generation

Receptor docking grids were generated for virtual screening with Glide Receptor Grid Generation (Schrödinger Release 2019–1: Glide, Schrödinger, LLC, New York, NY, 2019) and OpenEye’s Make_Receptor function [18] (OpenEye Scientific Software, Santa Fe, NM. http://www.eyesopen.com.) All proteins were initially prepared using Schrödinger Protein Prep Wizard prior to grid generation. Grids were then generated by using the co-crystallized ligand as the centroid for docking similarly sized ligands. Therefore, de novo grids in the active sites did not need to be created. Docking restrictions were made in the grid to ensure certain hydrogen-bond interactions were maintained during docking, particularly interactions with the hinge residues. Default settings were used in both OpenEye and Schrodinger grid generating modalities.

Molecular Dynamics (MD) analyses

All-atom explicit solvent Molecular Dynamic (MD) simulations were run using the Desmond GPU Accelerated suite of Schrodinger (Schrödinger Release 2019–1 Desmond Molecular Dynamics System, D. E. Shaw Research, New York, NY, 2020. Maestro-Desmond Interoperability Tools, Schrödinger, New York, NY, 2020.) Simulations used the multi-sim Desmond implementation and the simulation times varied between 300ns and 3μs, with a maximum of 3000 frames and a recording interval of 1/1000th of the overall simulation time. NPT (Isothermal-Isobaric) ensemble was used for the simulation, which most accurately represents laboratory conditions at ambient temperature and pressure (300K and 1.01325 bar). System boundary distances were specified as 10.0 Å in each dimension in an orthorhombic box and each model system was relaxed before simulation for several nanoseconds. Systems were built also using Desmond with the SPC solvent model and OPLS3e forcefield with an addition of 0.15 M NaCl to balance the charges. Simulations were computed across dedicated GPU cores. Trajectories were clustered using default settings for the “Cluster Trajectory” function.

Machine learning and compound enrichment

3 unique virtual screening campaigns were used to prioritize commercially available compounds to purchase for biochemical screening: A naïve Bayesian classifiers model, a neural net model, and a Tanimoto color and shape screen. Our group has previously developed ligand-based Laplacien-modified Naïve Bayesian classifiers to predict small molecule kinase activity trained on small molecule activity data across the entire human kinome [19]. A list of homologous proteins was obtained via Clustal Omega multiple sequence alignment tool using the primary sequence of the PNCK kinase domain obtained from Uniprot [20,21]. The top 31 results had sequence identity homology which ranged from 70% to 30%, with annotated active compounds with IC50s <100nM (AURKA, AURKB, AURKC, BRSK1, CAMK2D, CAMK2G, CHEK2, DAPK3, FER, MAPKAPK2, MKNK2, MYLK, NUAK1, PDPK1, PLK4, PRKAA1, PRKACA, PRKACB, PRKACG, PRKD1, PRKD2, PRKD3, PRKG2, PRKX, RPS6KA1, RPS6KA2, RPS6KA3, RPS6KA5, SGK1, STK17A, TSSK1B). Amongst selected kinases, there were 9,693 active small molecule compounds in the training set. The model was then tested using over 7 million commercially available compounds from the eMolecules library. Compound predictions from eMolecules were then scored based on their probability of binding to said target kinase- using a score known as EstPGood. A probability cut-off of 0.1 was used to enrich the virtual screening library. Additionally, compounds with multiple target predictions were prioritized while those with only one target prediction were filtered out. Compounds with a molecular weight <250 g/mol or >700 g/mol were also filtered. In total, 49,512 unique compounds were prioritized for the first virtual screen.

A second machine learning effort, a neural net-based approach, was also used which compiled data from the KKB (Kinase Knowledge Base) [22] and the eMolecules library of commercially available compounds [23]. For the neural net machine learning predictions, 314 kinases with 331,985 known active compounds were used for building the model. Compounds were featurized with Deepchem and ECFP4 fingerprints with a size of 1024 bits. The dataset was split randomly into 10% test, 10% verification and 80% training. There were 4 layers in a multiclass tensorflow network with the first layer being the 1024-bit input layer, the second layer having 2000 nodes, the third layer having 500 and the final layer being the output layer of 314 kinases. Prediction results came from the eMolecules library of compounds that were then predicted for their activity in this network and stored for later retrieval. The same homologous kinases used in the Naïve Bayesian model were used to retrieve predictions from the neural net. 289,008 unique compounds were identified as predictions to be used in the virtual screen. Compounds were filtered with a probability cutoff of >0.1, compounds active toward >1 kinases, and with molecular weights ranging from 250-700g/mol.

A third approach took advantage of the binding poses of ATP in the active site of PNCK homology models. A Tanimoto shape and color screen was employed using OpeneEye’s fastROCS GPU implementation. Compounds were screened in Enamine, MolportDB and eMolecules libraries. Libraries were prepared in OpenEye first using QUACPAC which involves tautomer and protomer enumeration at physiological pH ranges. OMEGA was then used to generate up to 150 conformers for each ligand. The eMolecules dataset was filtered before the shape screen using the OpenEye “Lead-Like” filter, generating 1,169,867 unique compounds. For Enamine, the advanced HTS screening library was used which contained 97,410 unique compounds while MolportDB Screening Compound Library had 287,773 compounds. A Tanimoto combination score cutoff of 0.5 (which is the sum of a color score and shape score) was used to prioritize compounds for docking using HYBRID.

Ligand Library Preparation

Compounds selected for virtual screen were populated as SMILES in a comma-separated file along with a unique identifier or catalog number obtained from the vendor. Using Pipeline Pilot (BIOVIA Pipeline Pilot, Release 2018, San Diego: Dassault Systèmes), SMILES were canonicalized and the file was transformed to a.SDF structure file to be used as input for LigPrep (Schrödinger Release 2019–3: LigPrep, Schrödinger, LLC, New York, NY, 2019) or for OpenEye QUACPAC and OMEGA (OpenEye Scientific Software, Santa Fe, NM. http://www.eyesopen.com). For large-scale virtual screens with over 100,000 compounds, OpenEye Hybrid docking was utilized. Following enrichment by Hybrid Docking Score, smaller subsets of ligands were then used to dock using Glide (Schrodinger) for further binding pose visualization and MD simulations. Ligand libraries were obtained from Enamine’s Advanced HTS library, MolportDB and eMolecules. The Enamine library was curated to have compounds with molecular weights between 300 and 500g/mol. The MolPortDB had no molecular weight filters, with compounds ranging from 150 to 750 g/mol. The eMolecule library was filtered into screening subsets using both lead-like and drug-like filters provided by OpenEye because the library was too large to screen as a whole. For ligand preparation, default settings were used in both programs. Compounds to be used in OpenEye were prepared first using QUACPAC which involves tautomer and protomer enumeration. OMEGA was then used to generate up to 150 conformers for each ligand. For GLIDE docking, ligands were prepared using LigPrep (Schrödinger Release 2021–1: LigPrep, Schrödinger, LLC, New York, NY, 2021) which generates all conformers, tautomers and ionization states at physiologic pH ranges.

Ligand shape screen

Shape screens were conducted using fastROCS GPU (OpenEye Scientific Software, Santa Fe, NM. http://www.eyesopen.com), which can screen millions of compounds in seconds. Structure files of ATP exported from docking models of PNCK were used as input for the screens. A combined Tanimoto shape and color score was used to filter the most similar compounds in eMolecules, MolportDB and Enamine prepared compound libraries.

Ligand docking

For large scale docking studies utilizing over 100,000 structures, OpenEye’s HYBRID Dock was used. HYBRID docking was done in parallel on all PNCK grids in one job thus aggregation of scores was not necessary as HYBRID only reports the highest scoring pose per compound per grid. The top 25,000 docking poses were exported for final analysis. For smaller scale docking jobs or confirmatory docking studies prior to MD simulation, Glide Docking was used (Schrödinger Release 2019–1: Glide, Schrödinger, LLC, New York, NY, 2019). All default settings were used. Glide standard precision docking (SP) was performed using 3D ligands generated from LigPrep. Up to 25 poses were generated per ligand with the top 5 poses being selected for output. Glide scores were exported for aggregation across various grids using Pipeline Pilot.

Compound storage and cataloging

All compounds were purchased as powders in quantities less than 5mg. Compounds were dissolved in various organic solvents (Methanol, Dichloromethane, ethyl acetate) and were aliquoted out in separate vials. Solvents were evaporated using a rotovap and high vacuum and aliquoted powders were stored at -20C. Compounds were dissolved in DMSO to a concentration of 10mM prior to use in assays. Compounds were stored in DMSO at -80C no longer than 1 week. All purchased compounds were catalogued virtually using the Collaborative Drug Discovery (CDD) Vault (https://www.collaborativedrug.com/).

Biochemical PNCK Inhibitor Screen

ADP-Glo Kinase Assay (V6930, Promega) was used to screen the first round of compounds against PNCK. GST-tagged PNCK recombinant protein was purchased from Novus (NBP1-99886). Autocamtide-II (GenScript RP10271) was used as the peptide substrate. This peptide was dissolved in water and prepared immediately prior to the assay. Calcium/Calmodulin 10X (Abcam ab189137) solution was used to create a custom kinase buffer: 40mM Tris,pH 7.5; 20mM MgCl2; 0.1mg/ml BSA; 50μM DTT and Ca2+/Calmodulin solution (0.03μg/μl, Calmodulin, 1mM Tris,pH 7.3, 0.5mM CaCl2). The assay was run according to protocol using 25uM of ultra-pure ATP. A standard curve was generated per run which displayed the linear range of luminescence at various ratios of ATP to ADP. A standard curve was then done to determine the amount of protein needed to generate a signal in the linear range of the ATP curve. 100ng of PNCK and 2ng of substrate was used per well in the 96-well plate format of this assay. Compounds were screened at 10uM and incubated at RT for 1-2h. Plates were read using the Clariostar Plus. Data was normalized to a DMSO and staurosporine controls.

Cell-based PNCK inhibitor Screen

The NanoBRET Target Engagement Intracellular Kinase Assay (Promega N2520) is a cell-based binding screen which reads loss of BRET signal as a result of target binding. A PNCK Nano-luciferase fusion vector plasmid was generated custom by Promega. The K-10 tracer was used for the assay, as determined by Promega. The plasmid was maxiprepped by Genewiz for use in all future assays. HEK293 cells were seeded in a T175 flask in DMEM with 10% FBS. The following day, cells were counted to 10,000, resuspended in assay media (Optimem with no phenol red, 1% FBS). Plasmid and transfection reagent (Fugene HD, Promega E2311) were added to the mixture and cells were seeded in a white 96-well plate. After 20–24 hours, media was changed, and the assay proceeded according to protocol. K10 tracer was added at the recommended concentration of 1uM while compounds were initially screened at both 10uM and 1uM. Cells incubated with compound at 37 degrees Celsius for a maximum of 2 hours before the plate was read using Clariostar Plus. Data was normalized to Staurosporine and DMSO controls. For compounds selected for further analysis, dose response curves were generated using the same testing protocol but at 8 different concentrations: 24uM, 12uM, 6uM, 3uM, 1.5uM, .75uM 0.375uM and 0.1875uM. UM195 was tested with concentrations from 12uM to 0.09375. Similarly, compounds from the analog by catalog analysis were tested at concentrations from 24uM to 0.1875uM.

Data analysis

Data analysis of ADP Kinase Glo results were carried out per the manufacturer’s instructions. Results were represented as a percent decrease in activity. Data was normalized to staurosporine (positive control) as 100% decrease and DMSO (negative control) as a 0% decrease. Data for NanoBRET was analyzed by dividing the acceptor by the donor emission and dividing by 1000 to obtain raw millibret units (mBU). Background was subtracted using control emission averages. Data was normalized to positive and negative controls of staurosporine and DMSO. Concentration-response curves were plotted in GraphPad Prism 8 and IC50s were extrapolated using non-linear fit with concentrations transformed to logarithmic scale. Baseline was held constant and represented the average mBU of staurosporine controls while the top was constrained to represent the average mBU of DMSO controls. Error bars on dose-response curves represent standard deviations.

Building and Assessing PNCK Homology Models

No experimentally derived crystal structure exists for PNCK. However, several related structures in the CAMK family have been resolved which could be used as a structural template to generate a homology model. To determine the CAMK member with greatest homology to PNCK, the amino acid sequence of PNCK (isoform 1) was used in pBLAST (NCBI). PNCK canonical amino acid sequence was obtained from Uniprot [20] (ID: Q6P2M8) and the kinase domain is predicted to be from residues 15–270. CAMK1a was identified as the most homologous structure in the PDB with sequence identity of >70% in the kinase domain [24]. Three individual iterations of CAMK1a structures were selected as templates. All structures contained an N-terminal lobe of 5 anti-parallel beta-sheets and the canonical alpha-helix with a C-terminal lobe composed of primarily alpha helices. All native CAMKs contain a C-terminal region with auto-inhibitory and CaM binding sequences, assuming a helix-loop-helix structure [6,24–27]. CAMKIa (1–320) (PDB: 4FG9, 2.40 Å) represents a full-length version of the calmodulin kinase with the regulatory auto-inhibitory domain intact. Therefore, this kinase was experimentally shown to be inactive despite being co-crystallized with ATP in the active site. PDB structures 4FG8 (1–305, 2.20 Å) and 4FG7 (1–293, 2.70 Å) are also co-crystallized with ATP but represent two distinct forms of the kinase. 4FG8 is a truncated structure with no calmodulin binding site, thus it is in an inactive state. 4FG7 has the entire regulatory domain removed and thus represents a constitutively active form. It has been determined that the regulatory domain of CAMK1a inhibits kinase activity by interacting with the N-terminal lobe and masks the Thr177 phosphorylation site, restraining the kinase in an inactive conformation [24]. The regulatory domain, however, does not occlude nucleotide binding site, thus 4FG9 (1–320) is able to bind ATP. The activation loop still adopts an inactive conformation, with a “DFG-up” like state whereby the phenylalanine is not completely rotated out of the binding pocket, representative of a distinct inactive conformation [24,28]. 4FG8 (1–315) similarly adopts an inactive conformation, with no magnesium detected in the active site. 4FG7 (1–293) displays the structural elements of a kinase in active conformation, particularly the alpha-helix displays interactions with its conserved glutamate, forming salt bridges with ATP. This structure has a true, active, “DFG-in” conformation. Conformational analyses of the DFG structures in our homology models with KinCore confirmed 4FG7, was indeed DFG-In while 4FG8 and 4FG9 were classified as “DFG-Intermediate” [29] (Fig B in S1 File). In all three ATP-bound structures, the activation loop is largely disordered; contrasted with the Apo-structure, whereby the activation segment adopts a helical formation. The use of three unique templates of CAMK1a allowed for a broad sampling of conformational states in the virtual screen.

Each homology model was created using PRIME and included the entire protein with ATP in the binding site (See Methods) (Fig 1A–1C). Loops were extensively refined, and the structure went through multiple rounds of energy minimization to produce the most stable model. To determine if the model is valid, the co-crystallized ATP was removed from the binding site and was docked back in using Glide. A model was deemed valid if it accurately predicted the pose and binding mode which was present in the experimentally derived structure within 1–2 Angstroms. The PDB structures all have a conserved interaction with the adenine amino group and a glutamate or valine in the hinge. Additionally, there are hydrogen bonds between the catalytic lysine (Lys44) to the gamma phosphate group and hydroxyl groups on the sugar. For PNCK, the predicted hinge residues included Val93 and Glu91 with the gatekeeper residue being Met90 and the catalytic lysine at Lys44. Models were finally evaluated for stability through extended all-atom explicit water molecular dynamic (MD) simulations (Fig C in S1 File). MD simulations were run using Desmond (GPU implementation) for 500ns (see Methods). An analysis of the RMSD of the C-alpha backbone as a function of time was used to assess overall structure integrity; as the change of RMSD approaches zero, the structure “converges” to a local stable conformation. Each PNCK model converges at about 150-300ns to an average RMSD of about 3–5 angstroms (S3 Fig). Models were additionally assessed using SWISS-MODEL Structure Assessment to ensure MolProbity scores at or less than the resolution of the parent structures used to generate the homomlogy models and Ramachandran allowed residue percentages greater than 90% (Fig D in S1 File and Table 1). Finally, structures were compared to those predicted by Alpha Fold via structural alignment and RMSD (Table 2). To further increase the conformational space sampled during docking, the trajectories of the 500ns molecular dynamic simulations were clustered and 5 structurally diverse poses for each of the three homology models were extracted (Fig 1D). Thus, in total, 15 unique representations of PNCK were used in parallel for the virtual screening docking campaign.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. PNCK Homology Models with ATP Docked.

ATP is docked in the active site of each model (displayed in yellow) with the hinge and gatekeeper residues colored in red. Compound UM195 is also overlayed in the active site of each model (displayed in blue). A) 4FG7 Homology model with no CAM-binding site or auto-inhibitory domain. B) 4FG8 Homology Model with CAM-binding site (Green helix) and no autoinhibitory domain. C) 4FG9 Homology model with full length PNCK including autoinhibitory domain and CAM-binding site (Green helix). D) Overlay of 5 representative poses extracted from the 500ns molecular dynamics (MD) simulation trajectory of the 4FG9 homology model.

https://doi.org/10.1371/journal.pcbi.1010263.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. SwissModel Structure Assessment.

https://doi.org/10.1371/journal.pcbi.1010263.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Structural Alignment of Homology Models using AlphaFold.

https://doi.org/10.1371/journal.pcbi.1010263.t002

Curating compound libraries for scalable Ligand-based Virtual screening

In total, three unique virtual screening campaigns were used to prioritize commercially available compounds for in vitro screening. Each combined a scalable ligand-based approach with docking and molecular dynamics. Specifically, we used Naïve Bayesian classification models, a multi-task deep neural network model [23], and 3D shape-based screening (Fig 2) (see Methods).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Workflow of AI-Assisted PNCK Chemical Probe Study.

Compounds from commercially available sources totaled over 7 million (from eMolecules, Enamine and MolDB). Enrichment via machine learning and shape screening led to over 358K compounds to be docked. Using the top docking scores and clustering scaffolds which scored similarly, 64 compounds were selected to use in chemical screens. 16/64 compounds were considered “hits” (>50% inhibition at 10 micromolar in the NanoBRET screen) and 5 of those were selected to further analysis and improvement by SAR. The cycle of docking, MD and cell-based screening was repeated until compound binding was improved both in silico and in vitro.

https://doi.org/10.1371/journal.pcbi.1010263.g002

Our group has previously developed ligand-based Laplacien-modified Naïve Bayesian classifiers to predict small molecule kinase activity. These models have also been used to predict novel small molecule inhibitors of EGFR and BRD4 [19,30]. The predictors were trained on small molecule activity data across the entire human kinome. Although there are no published selective compounds that target PNCK, similarity amongst kinase binding sites and cross-kinase activity of many kinase inhibitors facilitate the enrichment of PNCK inhibitors using predictive models of the most similar kinases. A list of homologous kinase domains was obtained by multiple sequence alignment with Clustal Omega using the primary sequence of PNCK kinase domain obtained from Uniprot [20,21] (See Methods). Most of the homologous kinases with sufficient small molecule data for high-quality predictors belonged to the CAMK family, including DAPK3 and BRSK1; other kinases included PRKACA, AURKA and AURKB. The model was then employed toward a dataset with over 7 million commercially available compounds from the eMolecules library. In total, 44,736 unique compounds were prioritized for the first virtual screen.

More recently, we developed a kinome-wide multi-task deep neural network classifier trained on aggregated curated data from ChEMBL [31] and the KKB (Kinase Knowledge Base) [22,23]. Overall, 331,985 known active compounds across 314 kinases were used for building this model. Predictions were made for the entire eMolecules library. Similar to the previous approach, compounds predicted active for homologous kinases were prioritized (See Methods). Aggregating the predictions, a total of 289,008 unique compounds were prioritized. PRKACA was the kinase that was most represented, with 13,780 compounds predicted active.

A third approach took advantage of the three-dimensional predicted binding poses of ATP in each of the active sites of the PNCK homology models. A Tanimoto shape and color screen was employed using OpeneEye’s fastROCS (GPU implementation). Compounds from the Enamine, MolDB and eMolecules libraries were screened after generating 3D conformers. ROCS (Rapid Overlay of Chemical Structures) is a powerful method for shape similarity and pharmacophore screening which uses Gaussian overlay functions to measure both shape and “color” similarity between compounds. ROCS utilizes both Tanimoto and Tvsersky functions to score overlap similarities. Volume overlap generates a shape score while color Tanimoto scores are derived from scoring similarities between electrostatic forces for functional atoms or chemical groups. The eMolecules dataset was filtered before the shape screen using the OpenEye “Lead-Like” filter, generating 1,169,867 unique compounds. For Enamine, the advanced HTS screening library was used which contained 97,410 unique compounds while MolportDB had 281,987 compounds. A Tanimoto combination score cutoff of 0.5 (which is the sum of a color score and shape score), was used to prioritize compounds for docking using HYBRID. 163,493 compounds with Tanimoto scores above 0.5 were docked. The top 25,000 scoring compounds from each dataset were then evaluated further.

Virtual screening campaign 1: Naïve Bayesian Classifiers and Docking

Compounds prioritized using the Laplacien modified Naïve Bayes models were prepared for docking using Schrodinger LigPrep. Compounds were then docked using Glide [32] into each of the 15 docking grids generated from the 3 homology models. Docking scores were then aggregated across all 15 receptors (See Methods). The top Glide Scores and mean Glide scores for each compound were used to rank and prioritize. As a reference, ATP received a glide average glide score of -8.59 kcal/mol. Docking scores for the machine-learning predicted compounds had an average of -5.52 kcal/mol and a standard deviation of -0.83. The top 1% of docking scores scored below (better than) -8.26 kcal/mol (Fig 3). Compounds were first clustered using extended connectivity fingerprint methods with a radius of 6 (ECFP_6). Clusters were then ranked using the average glide score and top clusters were evaluated individually based on synthetic tractability of the scaffold, likely binding pose, and ability for derivatization. One to two compounds from each cluster were then selected for purchase to test in in in vitro screening. In total, 25 compounds were purchased from the eMolecules catalogue (S1 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Distribution of Docking Scores in Different Machine Learning Models.

A) The naïve bayesian model and neural net model were both used to screen for predictions of inhibitory compounds towards kinases homologous to PNCK in the eMolecules library. The distribution of docking scores was very similar suggesting the neural net did not offer many new predictions. This was confirmed by manual inspection of clustered scaffolds B) vROCs shape screening and HYBRID docking using OpenEYE was performed using three small molecule libraries. The composition of each library offered much more diversity and subsequently was reflected in the distribution of docking scores, with eMolecules compounds scoring the best.

https://doi.org/10.1371/journal.pcbi.1010263.g003

Among compounds selected for biochemical screening, the docking scores ranged from -7.37 kcal/mol (UM_027) to -10.99 kcal/mol (UM_039) and molecular weights ranged from 315 (UM_019) to 410 (UM_027) g/mol. All compounds passed The REOS filter, (rapid elimination of swill) and PAIN filter screening (pan-assay interference) except for UM_036, as the presence of a 1,3-diazinane-2,4,6-trione may react with reagents used in such screens. Compounds UM_027, UM_029, UM_031, UM_034, and UM_039, all had at least one “Rule of Three” violation, while none of the compounds were predicted to be exceptionally cell-permeable (Exceptional Cell permeability has a QPPCaCo score >500) [33]. In fact, 4 compounds are predicted to be considered poorly cell permeable (QPPCaCo <25) (UM_021, UM_031, UM_034 and UM_039) from QikProp analysis (S2 Table). Thus, compounds from this cohort were more “drug like” (MW <500, cLogP <5, Molar Refractivity between 40–130) than “lead like”.

Virtual screen campaign 2: MTDNN (Multi-task Deep Neural Net) and Docking

As described above, MTDNN predictions were obtained using the same homologous kinases as campaign and 289,008 unique compounds were identified as predictions to be used in the virtual screen. Compounds were filtered with a probability cutoff of >0.5, activity to >1 kinase and with molecular weights ranging from 250-700g/mol leading to a final docking library of 39,787 compounds. Although PRKACA was the most overrepresented kinase, distributions of docking scores amongst the various kinase predictions did not vary significantly. The overall distribution of docking scores was slightly different compared to the naïve Bayesian classifier model, with the average docking scores shifting to -4.70 kcal/mol, with the best docking score being -9.69 kcal/mol (Fig 3). In fact, many of the same scaffolds as in campaign 1 were similarly identified in this method. For example, the cyclopropyl-pyrazole-carbonyl moiety in UM_019 and UM_034 was present in many of the top clusters of the MTDNN campaign. Additionally, the thieno [2,3-d] pyrimidine scaffold found in UM_027, UM_028, UM_033 and UM_035 was also highly represented in these predictions campaign. Therefore, this method did not greatly increase the diversity of scaffolds, likely due to the fact that compounds were screened again the same eMolecules library and the two machine learning methods did not identity significantly different compounds. Therefore, no compounds identified from this method were purchased for further analysis in biochemical or cell-based screens.

Virtual screen campaign 3: Tanimoto ROCS Shape Screen and Docking

By far, the greatest and most diverse docking results came from a third approach which simply consisted of finding molecules that are of similar shape, size, and electronic configuration as our computational model of ATP docked in the PNCK active site. Using fastROCS GPU of OpenEye, millions of compounds from eMolecules, Enamine and MolDB were queried against the 15 poses of ATP obtained from the molecular dynamic trajectories of the 3 PNCK homology models (See Methods). All conformers and ionization states generated from the compounds with Tanimoto combo scores >0.5 were then docked across all 15 receptors using OpenEye Hybrid and the top 25,000 scoring representations across all docking models were selected for further evaluation. Compounds were clustered using ECFP6 fingerprints as described before and top scoring clusters were evaluated for synthetic tractability and binding mode. As with the Naïve-Bayesian campaign, representative structures with unique scaffolds from the top scoring clusters were purchased for analysis in in vitro screens. Significant differences were noted in distribution of Tanimoto and docking scores amongst the three databases. Compounds from the eMolecules database had an average HYBRID docking score of -9.9 kcal/mol, followed by MolPortDB with an average of -8.56 kcal/mol and Enamine with an average of -7.92 kcal/mol (Fig 3).

When comparing the three homology models for PNCK, docking results from the Tanimoto screen using HYBRID show a strong preference for the 4FG9 grid (the inactive form of PNCK), with the best docking scores from these compounds coming from poses of the auto-inhibited state. Conversely, 42% of compounds selected from the Naïve Bayesian cohort (UM_019-UM_039) scored best with GLIDE in the 4FG7, constitutively active PNCK model.

Target engagement assays

The 25 Compounds prioritized by Naive Bayesian classifiers and docking (Campaign 1) were initially tested in a cell-free biochemical kinase activity screen, ADP Kinase Glo (See Methods). Six compounds were identified as hits (>50% inhibition of kinase activity at 10uM): UM_026, UM_029, UM_032, UM_035, UM_038 and UM_039 (Fig E in S1 File). However, it was determined that PNCK is not a highly active kinase and likely sub-optimal activity without phosphorylation at Thr171 (equivalent of Thr177 on CAMK1a) [34,35]. Thus, significant amounts of PNCK (>100ng per reaction) were needed to obtain signals in the linear range of the standard curve to extrapolate kinase inhibitory activity. Therefore, NanoBRET target engagement assay was used for future screens, including re-testing all compounds tested in the ADP-Glo Kinase Screen. NanoBRET is more sensitive than the ADP-Glo and additionally selects for compounds that are cell permeable (see Methods).

From the shape screen and docking campaign, compounds UM_194 to UM_244 were prioritized for testing in the NanoBRET target engagement assay. All compounds chosen for binding assays were analyzed for their putative binding mode by GLIDE docking and molecular dynamics (MD) simulations using the homology model for which it scored the highest. 55 compounds were initially tested in the NanoBRET assay, including compounds UM_019-UM_039 from the first campaign. Compounds UM_192, UM_193 UM_242, UM_243 and UM_244 were not cell permeable while UM_234 and UM_245 had poor aqueous solubility, which was properly indicated by QikProp predictions. Additionally, UM_234 and UM_243 were flagged by the PAINS filter due to the presence of a thiophene structure and a catechol structure, respectively. As such, these compounds were not subject to further analyses. Compounds from each campaign were similar in physiochemical properties but compounds from the Tanimoto shape screen were significantly more cell-permeable (Table 3). The NanoBRET screen was performed using two concentrations for compounds: 10μM and 1μM. Data was normalized to DMSO and staurosporine controls (Fig 4). Hits were defined as having greater than 50% inhibitory activity at 10μM. This assay confirmed actives and inactives that were identified via ADP kinase GLO with the exception of UM_037. UM_037 was only shown to inhibit ~31% of PNCK activity at 10uM in the ADP GLO assay. However, in the NanoBRET screen, UM_037 was shown to be a potent binder at 10uM with 56% inhibition. Conversely, top compounds in the ADP GLO assay failed in NanoBRET, likely due to poor cell-permeability. In particular, UM_039 had 91% inhibition of kinase activity in the biochemical screen with negligible activity in the cell-based screen. QPPCaco predictors from QikProp (Schrodinger) score UM_039 at 8.374, which is predicts poor cell permeability.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Physiochemical Properties of Compounds From Each Virtual Campaign.

https://doi.org/10.1371/journal.pcbi.1010263.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Cell-Based Screening of Prioritized Compounds.

All compounds from the naïve Bayesian model in addition to compounds identified in the vROCs shape screen were tested in the cell-based target engagement assay, NanoBRET. Compounds were tested at concentrations of 10μM and 1μM. Activity was reported as normalized percent inhibition using DMSO and staurosporine as controls. Pan-CAMK inhibitor KN-62 was also used as a reference. Hits were identified as compounds with at least 50% inhibition at 10uM.

https://doi.org/10.1371/journal.pcbi.1010263.g004

The top compounds chosen for follow-up validation in concentration-response curves were selected based on normalized percent inhibition at 10μM and 1μM, synthetic tractability and kinase inhibitor scaffold diversity or novelty (UM_037, UM_195, UM_210, UM_213 & UM_228). Concentration-response curves were generated for each compound using the NanoBRET system at 8 concentrations ranging from 24μM to 0.175μM. Nonlinear curve fitting after log-transforming of the concentrations allowed for calculation of IC50s (Figs 5 and 6).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Dose-Response Curves for Top 5 Hit Compounds.

A) UM_37, B) UM_195, C) UM_210, D) UM_213, E) UM_228.

https://doi.org/10.1371/journal.pcbi.1010263.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. IC50s of Top 5 Hit Compounds.

https://doi.org/10.1371/journal.pcbi.1010263.g006

Physiochemical properties for compounds were predicted from 3D structures in Schrodinger using the QikProp tool. QLogPo/Pw is the predicted octanol/water partition coefficient with higher, positive values representing increased hydrophobicity. QPPCaco is the predicted apparent Caco-2 cell permeability in nm/sec. Caco-2 cells are a model for the gut-blood barrier. QikProp predictions are for non-active transport; <25 is considered poor while >500 is great. SASA is the total solvent accessible surface area (SASA) in square angstroms using a probe with a 1.4 Å radius. PSA (polar surface area) represents the Van der Waals surface area of polar nitrogen and oxygen atoms.

In conclusion, one top scaffold was identified from Naïve-Bayesian Classifier / docking campaign while the other 4 were identified through the Tanimoto nucleotide ligand shape and color screen & docking.

Structure Activity Relationship (SAR)–Analogue by catalogue

Each hit compound was studied again extensively to develop the most accurate binding hypothesis based on classical binding poses kinase inhibitors. Using the most likely predictive pose from initial docking studies (Fig 7) and MD simulations (Fig F-Z in S1 File), analogs were identified to potentially maximize hydrogen bonding interactions in the binding pocket, particularly at the hinge residues and to gain evidence for its true binding mode. Parent hit compounds were queried in ChemSPACE to find analogs among billions of chemicals in “REAL space” that had variation at one or more R-group sites (Fig F-Z in S1 File) [36]. Analogs were then docked in the PNCK models to analyze any improvement in docking scores based off additional hydrogen bond and hydrophobic interactions at the hinge, gatekeeper, or catalytic lysine residues. Analog compounds with improvement in docking scores or with modifications congruent with a structure-activity relationship hypothesis were ordered for synthesis on demand. Several custom analogs were ordered for each initial hit (limited by commercial availability) and were tested again in the NanoBRET screen at 1μM and 10μM (Fig 8). Compounds with at least 50% activity at 10μM were analyzed further with concentration-response curves to generate IC50 values (Fig 9). These data then allowed for preliminary structure-activity relationships for each of the scaffolds.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. 2D Representation of Predicted Binding Poses of Hit Compounds.

A) UM_37, B) UM_195, C) UM_210, D) UM_213, E) UM_228 (Add info about hinge resides).

https://doi.org/10.1371/journal.pcbi.1010263.g007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Cell-Based Screening of Analog Compounds.

Analogs of the 5 hit compounds were identified from ChemSpace and prioritized via docking and SAR hypothesis. Compounds were tested using the NanoBRET assay at concentrations of 10 μM and 1 μM.

https://doi.org/10.1371/journal.pcbi.1010263.g008

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Dose Response Curves of Top Analog Compounds.

A) UM_251, B) UM_252, C) UM_254, D) UM_259, E) UM_277, F) UM_280.

https://doi.org/10.1371/journal.pcbi.1010263.g009

UM_228, (6R)-2-(5-chlorothiophene-2-amido)-6-methyl-4,5,6,7-tetrahydro-1-benzothiophene-3-carboxamide: Molecular modeling of UM_228 led to several binding hypotheses. As we prioritized poses with hydrogen bonds at the hinge, the predicted pose with the chlorine atom solvent exposed, the thiophene group deep in the binding pocket and the amide group positioned for hydrogen bonding with hinge residues Glu91 and Val93 was used for selecting analogs. Additional poses reversed the orientation by 180 degrees with the methyl group being solvent exposed and the chlorine participating in bonding interactions at the catalytic lysine. In selecting analogs, various parts of the molecule were considered to determine which groups of the scaffold were important for binding, determined by analyzing docking studies and MD simulations (Fig F-H in S1 File). The methyl group was removed in one analog or replaced with bulkier groups to maximize hydrophobic interactions in the hinge. Chlorine was removed or was replaced with various hydrogen bond donors or acceptors such as a nitro group or various heterocycles. Docking studies of the analogs supported the hypothesis that the chlorine-substituted thiophene was likely situated towards the hinge, as replacing this with a purine group as in UM_251 greatly improved docking scores with sustained interactions with Val93 (Fig I in S1 File). Molecular dynamics analyses of UM_251 in the 4FG9 model demonstrates various hydrogen bond interactions along with significant mobility of the auto-inhibitory domain to create a new binding pocket, with new interactions gained at Serine156 (Fig J in S1 File). In UM_252, thiophene in UM_228 is replaced with a fluorophenyl-pyrazole. Binding hypotheses situate the fluorophenyl group in the hinge with sustained hydrophobic interactions during molecular dynamic simulations while the pyrazole and amide groups participated in multiple hydrogen bonds and water bridges to Val93 and Ser156; more than the parent compound 228 (Fig K in S1 File). This was supported via NanoBRET assays with analogs UM_251 and UM_252 improving from the IC50 of UM_228 by one order of magnitude (Fig 8).

UM_213,3-[2-(2,3-dihydro-1H-indol-1-yl)-2-oxoethoxy]-6H,7H,8H,9H,10H,11H-cyclohepta [c]chromen-6-one: All 3 docking models oriented UM_213 into the binding pocket of PNCK with the pyrone carbonyl hydrogen bonding with Val93 and the cycloheptane ring pointing towards either the solvent or towards to the gatekeeper, Met90. Additionally, there were hydrogen bonds between Lys44 and Ser156 with the amide nitrogen and oxygen. Molecular dynamic simulations supported hypothesis that these hydrogen bond interactions are indeed sustained; and in this pose, UM_213 was very stable with an average RMSD of <1 Å (Fig L-N in S1 File). Analogs were designed to assess the importance of the large hydrophobic cycloheptane group and to optimize hydrogen bonding interactions at the opposite end of the molecule, changing the indole group for an oxo-piperazine (UM_259) or an imidazolidine-2,4-dione (UM_258) (Fig L in S1 File). Docking of analogs supported the hypothesis that the chromenone-cycloheptanyl group fits in the binding pocket by the hinge and gatekeeper while increased hydrogen bonding interactions could be observed via Val93, Lys44 and Ser156 with additional bonds at Asp136 in the 4FG9 model for UM_259 (Fig O,P in S1 File). UM_259 was experimentally determined to have a higher affinity for PNCK than the parent, UM_213.

UM_195, 5-[5-(naphthalen-1-yl)-1,3,4-oxadiazol-2-yl]sulfanylpyrazine-2-carbonitrile: While UM_195 was one of the most potent ligands as determined via the NanoBRET assay, modeling of UM_195 proved to be the most difficult. Kinase inhibitors classically bind with strong hydrogen bond interactions at the hinge. However, UM_195 docking poses in all 15 grids did not predict any such interactions. The naphthalene group was predicted to be deep in the binding pocket with hydrophobic interactions at the gatekeeper residue (Fig 7). Some hydrogen bonding to the oxadiazol, nitrile or pyrazine groups may occur with the catalytic lyisine. Molecular dynamics interactions shifted the compound into the binding pocket so that the nitrogen atoms on the oxadiazol interacted with Val93 of the hinge, leaving the nitrile group solvent exposed (Fig Q,R in S1 File). Analogs replacing the naphthalene group with heterocycles containing hydrogen bond donors and acceptors were virtually screened. Additionally, the nitrile group was replaced with various groups to improve hydrogen bonding interactions with Ser156, Lys44, Ser27 (Fig S in S1 File). Docking of analogs, however, supported the initial binding hypothesis. Analogs tested (UM_242, UM_244, UM_245, UM_253, UM_254) represented these changes. UM_254 proved to be the most potent analog (which substituted the terminal nitrile group with a methoxy-ester) with an IC50 of 2.68uM. However, molecular dynamics analysis of UM_254 in any of the three models did not predict stable binding in the active site.

UM210, 6-chloro-2-[(quinolin-2-ylsulfanyl)methyl]-3,4-dihydroquinazolin-4-one: Initial molecular modeling predicted that UM_210 binds to PNCK with hydrogen bonding interactions to the hinge residue, Val93, via the chlorine atom on the fused pyrimidine heterocycle (Fig T,U in S1 File). It was considered more likely that the chlorine residue is solvent exposed and the heterocycle oxygen forming a hydrogen bond with either Lys44 or Val93. Indeed, molecular dynamics simulations of UM_210 in both the 4FG8 and 4FG9 models situated the compound in the active site with conserved hinge interactions at the pyrimidine scaffold, sustaining contacts with Val93 and Glu91. Analogs were selected for testing to assess the likely binding mode of the compound (Fig V in S1 File). For example, the quinoline ring was changed to an indazole or pyridine group. Additionally, the chlorine was removed to determine if this atom in this position contributed significantly to binding. Docking of the analogs, however, supported another binding hypothesis whereby the quinoline ring was situated deep in the binding pocket with the benzene ring extending into the gatekeeper region while the nitrogen on the quinoline participated in hydrogen bonding with hinge residues. Analogs tested, UM_278, UM_279 and UM_280 did not demonstrate significantly improved binding in NanoBRET assays.

UM_037, 1-(3-5H-[1,2,4] triazino [5,6-b]indol-3-ylsulfanylpropyl)-2,3-dihydro-1H-1,3-benzodiazol-2-one: UM_037 was the highest scoring compound in every docking model of PNCK. It was initially screened in ADP-KinaseGLO where it had minimal inhibitory activity. However, it showed significant binding in NanoBRET with an IC50 of 8.79 uM. UM_037 contains two fused heterocycle moieties, both of which are common structures among hinge-interacting kinase inhibitors. The benzimidazolinone is predicted to be situated towards the hinge with interactions between the nitrogen NH and carbonyl oxygen with the Val93 backbone. Molecular dynamics also suggested additional interactions of the benzimidazolinone with Ser156 while the triazino group had hydrogen bonds with Lys44 and Ser156 and the indole group interacted with Glu140 (Fig W,X in S1 File). Additionally, there is likely pi-stacking interactions with these groups with Phe158. One model did orient the molecule in the opposite direction, with the indole group making interactions at the hinge. To determine the proper pose, analogs were designed that altered the triazino/indol group, including adding or removing bulky groups or adding carbonyl groups to alter hydrogen bonding (Fig Y in S1 File). Molecular modeling of analogs supported the initial binding hypothesis whereby substituting the triazino-indol group for more flexible groups which retained hydrogen bonding ability resulted in better docking results. Despite improvement in docking scores, the analogs were not significantly better inhibitors of PNCK in vitro.