Dark kinase annotation, mining, and visualization using the Protein Kinase Ontology

Data sources

The ProKinO ontology includes data obtained from curated internal sources as well as external sources. Information from internal sources include annotations of kinase sequence and structural motifs retrieved from curated multiple sequence alignments. External sources are used for information related to kinase sequence and classification (KinBase & UniProt) (Bairoch et al., 2005; Manning et al., 2002) cancer mutations (COSMIC) (Tate et al., 2018), pathways (Reactome) (Croft et al., 2011) and three dimernsioanl structure (PDB) (Berman et al., 2000). The ontology is populated and updated on a regular basis using protocols described in previous studies (Gosal, Kochut & Kannan, 2011; McSkimming et al., 2016). Here, we describe further enhancements and additions to ProKinO through integration of data on kinase expression patterns and drug interactions, as described below. In a seperate significant project, we have identified and classified nearly 30,000 pseudokinases spanning over 1,300 organisms (Kwon et al., 2019). The schematic representation of the classification of kinases into groups, families, and subfamilies was already in place (Hanks & Hunter, 1995; Manning et al., 2002). Consequently, the addition of the pseudokinases and their classification was relatively simple. However, it significantly enhanced ProKinO as a comprehensive knowledge graph representing kinase-related data. The definition and nomenclature of several kinome-wide conserved motifs were standardized based on several previously published studies which describe the kinase structural features such as subdomains (Hanks & Hunter, 1995), regulatory spine/shell (Meharena et al., 2013), and catalytic spine (Hu et al., 2015). A subset of redundant or family-specific motifs were removed in the updated ontology and motif information on some of the atypical kinases such as ALPHAK2 is not included as they cannot be reliably aligned with canonical protein kinases. Portions of the text were previously published as a part of a preprint (Soleymani et al., 2022).

Ligand interactions

Information on kinase-ligand interactions were retrieved from the Kinase-Ligand Interaction Fingerprints and Structures (KLIFS) database (Kanev et al., 2021). The KLIFS database stores detailed drug-protein kinase interaction information derived from diverse (>2,900) structures of catalytic domains of human and mouse protein kinases deposited in the Protein Data Bank. In addition, KLIFS provides an application programming interface (API) for programmatic access to data related to chemicals and structural chemogenomics (Kanev et al., 2021). However, it lacks information regarding kinase pathways or diseases which prevents the user from investigating the effect of drug-mutant protein binding on downstream pathways or diseases. KLIFS annotations, which report PDB residue positions, were converted to UniProt residue numbering using PDBrenum (Faezov & Dunbrack, 2021), then converted to prototypic Protein Kinase A (PKA) numbering using Multiply Aligned Profiles for Global Alignment of Protein Sequences (MAPGAPS) (Neuwald, 2009). Entries that could not be mapped or did not map to the kinase domain were filtered out.

Ligand responses

We have also incorporated information on kinase drug sensitivity profile in the updated ProKinO. In particular, we retrieved drug dose response data for kinase-relevant ligands/drugs from the Genomics of Drug Sensitivity in Cancer (GDSC) (Yang et al., 2013). Kinase-relevant ligands are defined based on our previous study (Huang et al., 2020), which collected 143 small-molecule protein kinase inhibitors from GDSC based on four drug-target databases: DrugBank (Wishart et al., 2018), Therapeutic Target Database (Li et al., 2018), Pharos (Nguyen et al., 2017), and the Library of Integrated Network-Based Cellular Signatures (LINCS) Data Portal (Koleti et al., 2018). GDSC provides the half-maximal inhibitory concentration values (IC50) of these 143 ligands in 988 cancer cell lines.

Ligand activities

Ligand activities were retrieved from Pharos, a flagship resource (Nguyen et al., 2017) of the National Institutes of Health (NIH) Illuminating the Druggable Genome (IDG) program that includes data on small molecules, including approved drug data and bioassay data. Based on the protein classification (Lin et al., 2017), the drug targets in Pharos include kinases, ion channels, G-protein coupled receptors (GPCRs), and others. In this phase of the project, we decided to include the data relevant to ligand binding in kinases. Pharos integrates drug-target relationships from several resources, such as ChEMBL (Bühlmann & Reymond, 2020) and DrugCentral (Avram et al., 2021).

Expression data

An important part of our recent additions was kinase expression data. Genomic expression data (protein, RNA), as well as transcription factors and epigenomic associations, are among many facets of the data included in Pharos. Furthermore, the GDSC repository contains gene expression data (Affymetrix Human Genome U219 Array), as well. Additionally, COSMIC’s Cell Lines Project includes a significant amount of gene expression data, including kinase expression.

Dark kinases

Dark kinases were labeled based on the information from Dark Kinase Knowledgebase (Berginski et al., 2021).

Protein kinase knowledge graph: schema and data organization

The ProKinO ontology consists of classes, sub-classes, class types, relationships, relationship types, and constraints of protein kinase and related data (Fig. 2). The hierarchy connects all classes to the root, which is ProKinOEntity. Moreover, the schema defines types and constraints for the relationships. With such explicit and constrained schema, composing queries is more intuitive than conventional relational databases. In particular, to enable integrative mining of dark kinase expression data in the context of kinase sequence and structural features, we have introduced three new classes in ProKinO, the Ligand class (including its name, source, and chemical structure) and the following three related classes: (1) LigandInteraction, placed between the Ligand and (already existing) Motif classes to capture kinase-ligand binding and selectivity at the motif and residue level; (2) LigandActivity, placed between the Ligand and (already existing) Protein classes to represent kinases targeted by ligands (and drugs); and (3) LigandResponse, located between the Ligand and (already existing) Sample classes and representing ligand (and drug) sensitivity in kinases. To capture kinase expression, we added the GeneExpression relationship linking the Protein and Sample classes. The outline of the recently added classes and their relationships in ProKinO is illustrated as a UML class diagram, shown in Fig. 2.

ProKinO population

The ProKinO knowledge graph is automatically populated from several external and local data sources at regular intervals, as originally described (Gosal, Kochut & Kannan, 2011), ProKinO schema and the associated knowledge graph population software are routinely updated to incorporate additional sources of data such as pseudokinase and “dark” kinase classification and incorporating information on ligand interactions, ligand responses, ligand activities, kinase expression and associated object and datatype properties. We have been using the Protégé ontology editor for the schema creation and its subsequent modifications. The organization of the schema after these modifications is available at https://prokino.uga.edu/about.

The population software has been coded in Java and uses the Jena Framework. The population process is performed in several steps to add instances, their properties, and a combination of reading the prepared data from CSV, RDF, XML, and other file formats and accessing many remote data sources using their provided API (for example, Reactome’s REST API). Entity interconnections across data retrieved from different data sources are accomplished using UniProt identifiers, kinase names, and other accession identifiers. We modified the population software to create instances and properties for the newly added classes and relationships.

More specifically, using the KLIFS API, we retrieved the relevant kinases, ligands, and residue-level interaction data. The data was retrieved and then processed by custom Perl scripts. ProKinO ontology schema was modified, and ligands were included as new data, while interaction data (motifs) were either reconciled with the motifs already present in ProKinO or added as new, if not already there.

Similarly, the ligand response data was retrieved from GDSC and then processed by custom Perl scripts to create suitable CSV files. Additional ligands were included as new data, while the response data and the relevant samples were either reconciled with the samples already present in ProKinO or added as new, if not already there.

In order to populate the data on ligand activities, we retrieved from Pharos kinase-relevant ligands, as well as their binding data on targeted kinases, for example, IC50 values. This data was retrieved and then processed by custom Perl scripts to produce the necessary CSV files. Additional ligands, not included in the KLIFS dataset, were included as new data. All kinases targeted by ligands were already present in ProKinO, so they were reused in this step.

Data on kinase expression was first retrieved from Pharos, COSMIC, and GDSC. As before, the relevant kinases were already present in the ProKinO knowledge graph. The expression data was stored as individuals in the Expression class. Some of the relevant data about samples were already present in ProKinO, as we already had a significant amount of sample data from COSMIC. Additional samples were included as new data.

We reviewed and updated all the motifs already present in ProKinO. Furthermore, we updated the motif naming in cases where there were differences with the standard motif names.

Finally, we assembled an up-to-date list of dark kinases (Berginski et al., 2021) and added a Boolean datatype property, isDarkKinase, to identify them among all other kinases in the ProKinO knowledge graph.

Mutation and expression of understudied PAK5 in human cancers

One possible way to prioritize dark kinases for functional studies is to ask the question, “which dark kinases are most mutated in human diseases, such as cancers?”. Typically answering this question would require collating and post-processing data from multiple resources such as COSMIC, Pharos, and the Dark Kinase Knowledgebase. However, with the updated Protein Kinase Ontology, these questions can be quickly answered using SPARQL. Having the “isDarkKinase” property within the Protein class and the RDF triples connecting the “Mutation”, “Sample” and “Sequence” classes, one can formulate aggregate queries requesting all dark kinases mutated in cancer samples. To avoid biases introduced by the length of protein/gene sequences (longer proteins tend to have more mutations), the query can be modified to normalize mutation counts by sequence length. Executing this modified query (Query 27, available at http://prokino.uga.edu/queries) displays the rank-ordered list of dark kinases based on mutational density. The top ten dark kinases with the highest mutational density are shown in Fig. 3A. Notably, the p21 activated kinase 5 (PAK5) is at the top of the list with a mutational density of 1.917, followed by CRK7 (1.054), PKACG (1.011), PSKH2 (1.01) TSSK1 (1.008), CK1A2 (0.991), ERK4 (0.966), DCLK3 (0.912), PKCT (0.894) and PAK3 (0.859). Having identified PAK5 as the most frequently mutated dark kinase in cancers, one can further query the ontology to explore the role of this kinase in various cancers. With the addition of the new “GeneExpression” class in ProKinO and the RDF triples connecting gene expression to the “Sample” and “Protein” classes (GeneExpression:InSample: Sample; GeneExpression:hasProtein: Protein), one can formulate queries for PAK5 expression in different samples (Fig. 3B). Rank ordering the samples based on PAK5 expression (Query 33) reveals cancer types such as adenocarcinoma (Zscore: 4,701.5) and hepatocellular carcinoma (Zscore: 2,038.3) that have previously been associated with abnormal PAK5 expression (Fang et al., 2014; Han et al., 2018; Huo et al., 2019; Zhang et al., 2017). However, the role of PAK5 in other cancer types such as acute myeloid leukemia (Zscore: 136.5) is relatively underappreciated (Quan et al., 2020). The identification of new cancer sub-types with dark kinase expression and regulation further exemplifies the use of ProKinO in knowledge discovery.

Mutational hotspots in the activation loop of PAK5

Because ProKinO encodes a wealth of information on the structural and regulatory properties of multiple kinases, it can be used to generate mechanistic predictions on cancer mutation impact. We demonstrate this for the PAK kinases by asking the question “where are PAK5 mutations located in the protein kinase domain?” Using the RDF triples connecting the “Mutation”, “Motif” and “Sequence” classes (“Mutation: LocatedIn: Motif”; “Mutation:InSequence: Sequence”), one can formulate a query (Query 28) listing mutations in different structural regions/motifs of the PAK5 kinase domain. Examination of the query results reveals that the C-terminal substrate binding lobe (C-lobe) is more frequently mutated (320 mutations) relative to the N-terminal ATP binding lobe (N-lobe: 173 mutations) (Fig. 4A). Within the C-lobe, nearly 78 mutations map to the activation loop, which is known to play a critical role in substrate recognition and activation in a diverse array of kinases (Huse & Kuriyan, 2002; Kornev & Taylor, 2015; Oruganty & Kannan, 2012). Despite the prevalence of activation loop mutations in PAK5, there is currently no information on how these mutations impact PAK5 kinase structure and function. Nonetheless, based on the evolutionary relationships captured in ProKinO (based on the alignment of human kinases to the prototypic protein kinase A), one can formulate queries mapping mutations to specific aligned positions in the shared protein kinase domain. A query listing (wild type) WT type and mutant type residues in the activation loop of PAK5 and the equivalent aligned residue positions in PKA (Query 29) provides additional context for these mutations. For example, two distinct mutations map to residue $\mathrm{P}{602}^{\mathrm{P}\mathrm{A}\mathrm{K}5}$ in the activation loop of PAK5 that structurally corresponds to a phosphorylatable residue, $\mathrm{T}{197}^{\mathrm{P}\mathrm{K}\mathrm{A}}$, in PKA (Yonemoto et al., 1993). Having this context provides a testable hypothesis that S602 mutations in PAK5 impact kinase phosphorylation and regulation. Likewise, WT residue $\mathrm{P}{607}^{\mathrm{P}\mathrm{A}\mathrm{K}5}$ is mutated in four distinct cancer samples and this position is equivalent to PKA residue $\mathrm{P}{202}^{\mathrm{P}\mathrm{K}\mathrm{A}}$, which configures the activation loop for substrate recognition (Knighton et al., 1991). Thus, mutation of this critical residue is expected to impact substrate binding and activation loop phosphorylation in PAK5. Additional insights into these mutations can also be obtained by visualizing these residues in the context of the PAK5 AlphaFold models using the ProtVista viewer described below.

Insights into PAK5 ligand binding sites

With the conceptualization of new information related to kinase ligands, their mode of action and interaction with specific motifs in the kinase domain, new aggregate queries linking mutated kinases to drug sensitivity profiles, mode of action, and ligand binding sites can be performed using the updated ProKinO. For example, queries such as “list proteins and drugs or ligands interacting with the protein’s gatekeeper residue (GK.45)” (Query 31) and “list ligands targeting the Epidermal Growth Factor Receptor (EGFR) kinase and their mode of action” (Query 34) can be rapidly performed using the updated ProKinO ontology. We demonstrate the application of these new additions in the context of PAK5 by asking the question “what are the drugs targeting PAK family (PAK1-6) kinases?” Query 30 answers this question using the RDF triples connecting the “Ligand”, “Motif” and “Protein” classes (list triples) (Fig. 5). Examination of the query results indicates multiple drugs targeting PAK family kinases, including STAUROSPORINE and N2-((1R-2S)-2-AMINOCYCLOHEXYL) that bind to structurally equivalent residues/motifs in the ligand binding pocket of PAK4 and PAK5, respectively. The ligand binding sites, and associated interactions can also be visualized using the ProtVista viewer described below. Additional queries linking dark kinases to drug sensitivities, structural motifs, and pathways are listed on the ProKinO website at https://prokino.uga.edu/queries.

Visualization tools for dark kinase annotation and mining

To provide structural context for cancer mutations and to enable interactive mining of dark kinase sequence annotations in the context of 3D structures and predicted models from AlphaFold (Jumper et al., 2021; Tunyasuvunakool et al., 2021), we developed and incorporated a modified version of the ProtVista viewer in ProKinO. The viewer can be deployed for any protein kinase of interest by navigating to the Structure tab in the protein summary page and selecting either a PDB structure or AlphaFold model of interest. A snapshot of the ProtVista viewer displaying the AlphaFold model of PAK5 kinase is shown in Fig. 6. The ProtVista viewer uses an enhanced version of the Mol* viewer and the PDB web component (Watkins et al., 2017) to provide two-way interactive navigation between the 3D structure (Fig. 6A, top panel) and annotation viewer (Fig. 6A, bottom panel).

The annotation viewer consists of multiple tracks populated dynamically based on data from ProKinO and external sources such as UniProt. In addition, prediction confidence scores for AlphaFold models are displayed in the annotation viewer along with additional annotations such as conserved sequence motifs, subdomains, and structural motifs involved in kinase regulation. The annotation viewer also shows other annotations from external sources such as ligand binding sites and predicted functional sites. Users can hover over the residues on the 3D structure viewer to view the equivalent information on the annotation viewer and vice versa. For example, selecting the “activation loop” in the annotation viewer highlights the corresponding structural region in the AlphaFold model of PAK5 (Fig. 6A). Likewise, the selection of residues in the activation loop (S602 and P607) in the structure viewer highlights the annotations associated with these and interacting residues in the sequence viewer. Such interactive mining is expected to accelerate the functional characterization of dark kinases and provide new insights into disease mutations. For example, visualizing the interactions associated with S602 in the activation loop of PAK5 (Fig. 6B) indicates a hydrogen bonding interaction with R567, which is part of the conserved HRD motif (sequence annotation). Because the HRD-Arg is known to play a role in kinase regulation by stabilizing activation loop conformation (Huse & Kuriyan, 2002), it provides additional context for predicting the impact of S602 altering mutations. Likewise, examining the structural and sequence context of P607 interacting residues provides new insights into how the alteration of this residue might impact substrate binding and kinase regulation. Together, these examples highlight the value added by the ProtVista viewer in the visualization and annotation of mutations in dark kinases.