CROssBAR: Comprehensive Resource of Biomedical Relations with Deep Learning Applications and Knowledge Graph Representations

Systemic analysis of available large-scale biological and biomedical data is critical for developing novel and effective treatment approaches against both complex and infectious diseases. Owing to the fact that different sections of the biomedical data is produced by different organizations/institutions using various types of technologies, the data are scattered across individual computational resources, without any explicit relations/connections to each other, which greatly hinders the comprehensive multi-omics-based analysis of data. We aimed to address this issue by constructing a new biological and biomedical data resource, CROssBAR, a comprehensive system that integrates large-scale biomedical data from various resources and store them in a new NoSQL database, enrich these data with deep-learning-based prediction of relations between numerous biomedical entities, rigorously analyse the enriched data to obtain biologically meaningful modules and display them to users via easy-to-interpret, interactive and heterogenous knowledge graph (KG) representations within an open access, user-friendly and online web-service at https://crossbar.kansil.org. As a use-case study, we constructed CROssBAR COVID-19 KGs (available at: https://crossbar.kansil.org/covid_main.php) that incorporate relevant virus and host genes/proteins, interactions, pathways, phenotypes and other diseases, as well as known and completely new predicted drugs/compounds. Our COVID-19 graphs can be utilized for a systems-level evaluation of relevant virus-host protein interactions, mechanisms, phenotypic implications and potential interventions.


Main
Systemic analysis of available large-scale biological and biomedical data is critical for developing novel and effective treatment approaches. Parts of this big-data are continuously being updated and maintained by different organizations and institutions, thus, data is scattered across individual computational resources. Although these entities are biologically related and complementary to each other, the connections between datapoints at different resources are not well-established and explicit at all. In addition to the connectivity problem, another issue related to biomedical data is the incompleteness in knowledge space (e.g., possible ligands of a target biomolecule, or the disease related implications of a critical mutation). There is a clear requirement for innovative computational approaches to integrate available biomedical big-data and to complete missing information with in silico predictions, to serve the ultimate aim of proposing novel treatment options (especially) for complex diseases.
There are numerous studies in the literature that aimed to integrate the available biomedical data [1][2][3][4][5][6][7][8][9][10] . These studies provided useful tools and methods to the life-sciences research community; however, many of them miss important functionalities that prevent them from becoming widely adopted tools/services (Supplementary Information section 1).
In this project, we aimed to address the current shortcomings by developing a comprehensive open access biomedical system entitled CROssBAR via integrating various biological databases to each other, inferring the missing relations between existing data points, and constructing informative knowledge graphs based on specific biomedical components/terms such as a disease/phenotype, biological process, gene/protein and drug/compound, or specific combinations of them. To construct the CROssBAR system, we accomplished multiple sub-projects: (i) the construction of the CROssBAR database to house, and its API service to serve the integrated biomedical data, (ii) development of deeplearning based drug/compound-target protein interaction (DTI) prediction models, CROssBAR database (CROssBAR-DB) comprises carefully selected features from various data sources namely UniProt, IntAct, InterPro, Reactome, Ensembl, DrugBank, ChEMBL, 3 PubChem, KEGG, OMIM, Orphanet, Gene Ontology, Experimental Factor Ontology (EFO) and Human Phenotype Ontology (HPO). Extract-Transform-Load (ETL) pipelines were developed for heavy lifting of data from these resources by persisting specific data attributes with the implementation of logic rules. These pipelines fetch, cleanse, validate and consolidate the data, and thus, implement a multi-omics data integration approach to release a single resource based on MongoDB collections. CROssBAR-DB, which provides a broad spectrum of information such as biomolecular functions, domains, interactions, pathways, diseases, phenotypes, drugs, compounds, etc., is hosted and maintained by the EMBL-EBI. Current data statistics of the CROssBAR-DB and the database schema are shown in Fig. 1b and in Fig.   S1, respectively. CROssBAR-DB is periodically updated on demand/request basis via an automated procedure, which makes the underlying data up to date most of the time.
CROssBAR-DB can be queried via a public RESTful API at: www.ebi.ac.uk/Tools/crossbar/swagger-ui.html, which provides a multi-faceted view of the stored data through 12 endpoints (Fig. S3).
As a part of the CROssBAR project, we developed two novel deep-learning-based predictive systems: DEEPScreen and MDeePred, with the aim of enriching bioactivity data by identifying unknown interactions between drugs/drug-candidate compounds and target proteins. DEEPScreen employs convolutional neural networks to process 2D structural images of drugs/compounds in 704 individually optimized high performance target-based prediction models, suited for well-studied targets 11 . MDeePred utilizes both compound and target protein features within a pairwise input hybrid deep neural network architecture to produce real valued bioactivity predictions, especially for targets with a few or no training instances 12 . We trained both systems using carefully filtered and integrated data in CROssBAR-DB, and ran our trained-models on large compound and human protein spaces to obtain comprehensive bio-interaction predictions, which are included in our knowledge graphs. We also developed an accompanying computational tool, iBioProVis, which is an unsupervised-learning-based visualization system for exploring large drug/compound-target interaction datasets in reduced dimensions 13 .
The term knowledge graph (KG) defines a specialized data representation structure, in which a collection of entities (nodes) are linked to each other (edges) in a semantic context 14 . In this study, we chose to represent heterogeneous biomedical data in KG structures. In CROssBAR knowledge graphs (CROssBAR-KG), biological components/terms (i.e., drugs/compounds, genes/proteins, bio-processes/pathways, phenotypes/diseases) are represented as nodes, and their known or predicted pairwise relationships are annotated as edges (a protein and its coding gene is treated as one merged term/entry/node). The logic behind the construction of a knowledge graph is centered around queried biological 4 components/terms, as shown in Fig. 1c with a work-flow diagram and with an example disease term query. At each step of the process, an overrepresentation-based enrichment analysis has been performed to select the terms that are significantly associated with the growing graph, and to discard the rest. This analysis comprises a series of hypergeometric tests, based on the recorded relations in the CROssBAR database. Here, we applied a layered construction approach, always taking the genes/proteins at the centre of the enrichment analysis. Finally, additional relation types are incorporated to the graph as edges between the existing nodes (e.g., drug-disease, disease-pathway and disease-HPO), to further enrich the provided information. During the construction of graphs, terms (nodes) and their pairwise relations (edges) are directly obtained from the CROssBAR-DB.
CROssBAR-KGs clearly display the direct and indirect relations between all of the terms in the graph. These intensely-processed heterogeneous biological networks are expected to aid biomedical research, especially to infer mechanisms of diseases in relation to biomolecules, systems and candidate drugs.
We developed the CROssBAR web-service (CROssBAR-WS) to make the CROssBAR-KGs available to the public in an easily interpretable and interactive way (https://crossbar.kansil.org). KGs are presented visually on web-browsers as flexible Cytoscape 15 networks. Users can create queries with biomedical terms, individually or in combination, to obtain the relevant graph. Combinatory term query is especially critical as it provides the ability to investigate the indirect biological relationships between the terms from both the same and different biomedical components. Since there are billions of different ways to query CROssBAR, it was not feasible to pre-calculate the resulting graphs; therefore, they are set to be constructed on-the-fly, in real-time. Several options are provided to users to customize the procedure both before the search, such as the UniProt databases to be used (UniProtKB/Swiss-Prot or UniProtKB/Swiss-Prot+UniProtKB/TrEMBL), taxons to be included, and the number of terms/nodes to include from each entity type (selected from enrichment score-based ranked lists). It is also possible to display the graph using a variety of layout options, including our in-house CROssBAR-layout (Fig. 2a). Saving options let users to store the graph in different formats, including json, figure-ready snapshots and protein-centric delimited data-tables. The interactive visualization also lets users prepare a custom display by relocating the nodes/edges as desired.
As a use-case, we present Coronavirus disease 2019 (COVID-19) CROssBAR-KGs (https://crossbar.kansil.org/covid_main.php). Starting from the end of 2019, the new coronavirus (SARS-CoV-2) pandemic has ravaged the entire globe and caused immeasurable damage 16 . As of July 2020, the scientific endeavour to develop effective drugs and vaccines is at peak, and systemic evaluation of the current knowledge about SARS-CoV-2 infection is 5 expected aid researchers in this struggle . To demonstrate the capabilities of CROssBAR, we   have constructed two different versions of the COVID-19 knowledge graph, (i) a large-scale   version including nearly the entirety of the related information on different CROssBARintegrated data sources, which is ideal for further network and machine learning based analysis or a detailed inspection (Fig. 2b), and (ii) a simplified version distilled to include only the most relevant genes/proteins as provided in UniProt-COVID-19 portal (https://covid-19.uniprot.org), which is ideal for fast interpretation (Fig. 2c). It is interesting to observe the indirect relations between the diseases/phenotypes in the KGs and COVID-19 over the incorporated host proteins and enriched pathways, and between COVID-19 and our in silico predicted drugs, as they may reveal further evidence to be utilized against COVID-19 (Fig.   2b,c). For this, we conducted a short literature-based validation study and found that many of these drugs have already been experimented at both preclinical and clinical stages for new COVID-19 treatments (Supplementary Information section 2).
Although COVID-19 is a respiratory disease and lung lesions have been considered the major damage caused by SARS-CoV-2, liver injury has also been reported in about one-third of hospitalized patients infected with the virus and the majority of COVID-19 patient deaths are associated with cytokine storm/release syndrome resulting in multi organ damage 17 .
Hence, with the aim of indicating the biological relevance of the information in CROssBAR-KGs, we conducted in vitro experimentation on drug treated liver cancer cell-lines and comparatively analysed the results on both COVID-19 KGs. Chloroquine (CQ) phosphate was reported to be used in treatment of COVID-19 with controversies 18 . CQ is an antiinflammatory drug that has been used in autoimmune diseases and can significantly alter the production of pro-inflammatory and anti-inflammatory cytokines. We investigated the effect CQ on normal hepatocytes like Huh7 cells and poorly differentiated Mahlavu cells.
Cells were treated with CQ and the differentially expressed gene (DEG) data were acquired from a large multiplex panel of genes using the NanoString platform (Fig. 2d). Our experimental data indicated significant alterations in JAK/STAT, PI3K, RAS, MAPK pathways involving cytokine production in liver cells (full pathway list: Table S.3). These pathways were also presented in both KGs along with additional cytokine related pathways, such as interleukin signaling, along with dense connections to other biological components in COVID-19 CROssBAR-KGs, which is an expected output considering the mode of action of CQ in COVID-19 (Fig. 2c). Additional use-cases are provided in the Supplementary Information section 3. We believe that the CROssBAR system can be utilized towards the systematic analysis of the pharmacological effects of drugs as it brings relevant pieces of biological data together, that is relevant to the user's query, which can be manually explored by the expert to build new hypotheses. 6

CROssBAR Database & API
Integrated data resources and the CROssBAR-DB CROssBAR-DB has developed its bespoke ETL pipelines in Java 8 using the Spring batch framework to structure the jobs. The latter are executed on state-of-the-art EMBL-EBI LSF clusters powered by IBM in a parallel distributed fashion to reduce the processing time. The data are finally stored in MongoDB in the form of independent data collections, thus, providing schemaless flexibility and faster development, while sustaining data relationships in the form of nested documents. The pipelines have been both unit and integration tested using Spock framework in Groovy language.
The public databases integrated in the CROssBAR system can be listed along with the type of the biomedical data it contains as follows; The statistics regarding the number of terms and annotations incorporated to CROssBAR-DB from each resource listed above is given in Fig. 1b Supplementary Information Fig. S3, where the 12 CROssBAR-DB collections are shown. Here, the collections entitled "Activities", "Assays", "Molecules" and "Targets" correspond to bioactivity, bioassay, compound and target entries in the ChEMBL database, respectively. "Drugs" corresponds to drug entries in DrugBank, "EFO disease terms" corresponds to the disease entries in the Experimental Factor Ontology, "HPO" corresponds to phenotype entries in the Human Phenotype Ontology, "Proteins" correspond to a subset of the protein entries in the UniProtKB. The remaining four collections belong to the PubChem data. There is no one-to-one correspondence between the incorporated data resources and the CROssBAR-DB collections since some of the resources had to be split to multiple collections for easier query (e.g., ChEMBL and PubChem). Also, some of the sources are directly incorporated from the UniProt database, thus, reside in the proteins collection (e.g., both terms and annotations for InterPro, Reactome, and only annotations for IntAct, OMIM and Orphanet).
It is possible to obtain cross-collection relational data (i.e., integrated relational data from multiple collections) by writing programmatic queries and submitting them to the API, as it is applied in CROssBAR-WS to construct the knowledge graphs. However, it is not possible to obtain this complex relational data in a single query using the Swagger graphical user interface. Currently, CROssBAR knowledge graphs do not include PubChem data due to both elevated computational demand (the sizes of PubChem collections are large) and high redundancy (a large portion of bioactivity data points in PubChem and ChEMBL databases are shared). However, it is possible to query the CROssBAR-DB using the provided API 8 service, to obtain data entries from PubChem database collections.
The database and API construction work has been handled by the Protein Function Development (UniProt database) team at EMBL-EBI, utilizing their expertise in biological database development and maintenance together with the available strong computational infrastructure, the team managed to build a huge but stable resource. The professional service providing approach applied by the team allowed the proper and constant maintenance of both the database and the entire CROssBAR system.

Deep-learning-based predictors and dataset construction
The the current knowledge corresponds to less than 0.001% of the whole compound-target space 19 . The high rate of missing DTI data negatively impacts the integrated biomedical resources as well. In the CROssBAR project, we aimed to address this issue by producing machine learning based DTI predictions and incorporating these predictions to the CROssBAR resource. The studies specifically about the development of these tools have already been published or under review 11,12 , however, we used our tools to produce DTI predictions to be incorporated in our knowledge graphs in the framework of this study.

Bioactivity dataset construction
One critical topic in developing DTI prediction models is the source dataset to be used in system training procedures. It is especially critical to construct large-scale DTI datasets to train deep-learning models. To address this issue, we prepared a DTI dataset from the ChEMBL database that is suitable for training machine learning systems, with standardized filtering operations on targets, compounds and bioactivities. The dataset is periodically updated with each ChEMBL database release. We employed this dataset for the training and validation of the deep-learning based DTI prediction models we developed in the framework of the CROssBAR project, and also as the source dataset for drug/compound-target 9 interaction space visualization (the methods are described below). It can also be used for developing new DTI prediction models. The current version of the bioactivity dataset (ChEMBL v27) is available for public use in: https://github.com/cansyl/CROssBAR/blob/master/CROssBAR_DB_API/ChEMBL27_preproc essed_activities_sp_b_pchembl.zip. Details regarding the dataset can be found in our recent article 11 .
Deep learning base predictor 1 -DEEPScreen DEEPScreen was the first DTI prediction system that we developed in this endeavour.
DEEPScreen is a high-performance drug-target interaction predictor that utilizes deep convolutional neural networks and 2-D structural compound representations (i.e., simple images) to predict their activity against intended target proteins. DEEPScreen system is composed of 704 target protein specific prediction models, each independently trained using experimental bioactivity measurements against many drug candidate small molecules, and optimized according to the binding properties of the target proteins. The main novelty of DEEPScreen is employing readily available 2-D structural representations of compounds at the input level instead of conventional drug/compounds descriptors (e.g., molecular fingerprints) that display limited performance. DEEPScreen produces binary predictions, meaning that a compound is either predicted as active or inactive against a target protein.
During the development of this method, we also carried out cell-based in vitro wet-lab experiments on computationally generated DTI predictions, with the purposes of both validating the accuracy of the prediction models, and for gaining biological insight in the framework of health and disease, especially to contribute to the understanding of processes active in different cancer subtypes. DEEPScreen can be used for the fast screening of the chemogenomic space, to provide completely new DTIs that can later be investigated experimentally in the fields of drug discovery and repurposing 11

Construction of knowledge graphs
In CROssBAR, the data is stored in a non-relational database (MongoDB), as separate collections for easy maintenance and fast querying. As a result, the database itself is not a knowledge graph. Instead, biologically relevant small-scale knowledge graphs are constructed on-the-fly, triggered by users' queries with a single or multiple term(s) such as the names or ids of genes/proteins, diseases/phenotypes, compounds/drugs and/or pathways/biological processes of interest.
In CROssBAR knowledge graphs, biological entities are represented as vertices/nodes.
Relations between different types of biological entities are expressed by the edges of the graph. Edge types vary according to the defined relations. For a relation between; (i) two proteins, the edge is labelled as "interacts_with", (ii) for a gene/protein and a disease, the edge label is "related_to", (iii) for a drug/compound and a protein, the edge label is "targets", (iv) for a gene/protein and a pathway, the edge label is "involved_in", (v) for a gene/protein and a phenotype term, the edge label is "associated_with", (vi) for a drug and a disease, the edge label is "indicates", (vii) for a disease and a pathway, the edge label is "modulates", and (viii) for a disease and a phenotype term, the edge label is "associated_with".
The incorporation of pathway-related information (both signalling and metabolic pathways) in CROssBAR is done based on a membership-based approach, where pathways are expressed on the graph as single nodes and the nodes of those member proteins are connected to them via edges. This approach leaves out the detailed reaction-based mechanistic information provided in pathway databases such as Reactome and KEGG pathways; however, the inclusion of this information via applying a pathway resource styled network approach would prevent the generation of large heterogeneous networks composed of tens of different pathways and other components. Nevertheless, it is possible to explore these pathways in detail using the provided links, which takes the user to the corresponding page on that pathway database. Both Reactome and KEGG pathways provide the same type of biological information at the level of large-scale biological processes; however, Reactome also divides these processes into sub-pathways, whereas KEGG only provides the pathway information at a generic level. In CROssBAR, due to the way the overrepresentation analysis is done, only specific sub-pathways are incorporated from the Reactome database, in most cases. As a result, pathway information in the knowledge graphs is displayed at different levels of specificity, and thus, not redundant.
A simplified form of the knowledge graph construction work-flow is displayed in Fig. 1c. In would not be feasible. To address this problem, we applied a multi-staged overrepresentation-based enrichment analysis process during the construction of graphs. In this analysis, we calculate an independent enrichment score for each biological entity in the database (i.e., a disease, phenotype, drug, compound, gene/protein or pathway), to be considered as its relevance to the graph that is being constructed. The calculation of enrichment score and its statistical significance is done using a modified version of the hypergeometric test for overrepresentation 20 , which also corresponds to a one-tailed Fisher's exact test, and it is based on the statistics of the relations/connections with gene/protein nodes. For example, the enrichment score (ED,W) and its significance (SD,W), in terms of p-value, for a disease term D, for graph W is calculated as follows: (1)

13
(2) where ED,W is the enrichment score calculated for the disease term D for graph W; mD 2 represent the square of the number of genes/proteins in graph W that are associated with disease D; nW represents the total number of gene/protein nodes in graph W; MD is the total number of genes/proteins (not necessarily in graph W) that is associated with disease D; and N represents the total number of reviewed human gene/protein entries (i.e., UniProtKB/Swiss-Prot entries) in the CROssBAR database that is annotated with any disease entry. SD,W represents the significance (p-value) for the disease term D for graph W calculated in the hypergeometric test.
While constructing the graph, an enrichment score is calculated for each disease entry in the CROssBAR database and these scores are used to rank disease entries according to their biological relevance to graph W. A cut-off value k is employed to include the top k relevant disease entries to graph W. The default value for k is 10, which means that only top-10 relevant diseases will be included in the graph. Apart from diseases, the same methodology is used to filter out the nodes of neighbouring genes/proteins, phenotypes, drugs, compounds and pathways. Significance values are not directly used in the filtering operation, since the main objective here is not only including significantly over-represented terms, but just reducing the number of nodes in the graph by filtering out the ones that are least relevant. In the traditional way of calculating an enrichment score, mD is without square. The reason behind taking the square of mD here is to break the tie between the scores of terms in favour of the one with a higher mD value.
Formalizing the graph construction around gene/protein entries During the construction of a knowledge graph, first, the gene/protein entries that are directly connected to the query term (i.e., core proteins) are fetched, such as the member genes/proteins of the queried signalling pathway. After that, neighbouring/interacting genes/proteins are added to the graph by calculating enrichment scores for each interacting protein, using the equation above, and filtering out based on the selected cut-off value. This is followed by the enrichment-based filtering and addition of other entity types; however, this time, both core and neighbouring genes/proteins are taken into consideration together to calculate the enrichment scores. If the user starts a heterogenous search that contains multiple terms from different entity types, both core and neighbouring genes/proteins are independently collected for each non-protein query term, queried gene/protein entries are added to this list (if there is any), and the entity collection process is continued using the union of these genes/proteins (Fig. 1c). This approach enables the exploration of direct and indirect relations between the queried terms. Bioactive compound and bioactivity selection procedure Small molecule compounds are selected and incorporated to KGs based on their reported bioactivities against target proteins. In a KG, a compound is represented as a node and a bioactivity is represented as an edge between a compound node and a gene/protein node.
We start the bioactive compound collection procedure with a set of target gene/protein entries at hand (gathered in a previous step of the KG construction process), and obtain the compounds that are reported to be bio-actively interacting with these proteins, as their target biomolecules. Despite having a simple logic, this procedure is extremely complex due to practical reasons. Since there are more than 15 million bioactivity measurements in the ChEMBL database (v26), we rigorously filter these data points with the aim of providing only the most relevant bioactivity/compound information in CROssBAR-KGs. Since CROssBAR is a gene/protein centric system, we first filter out the data points where the target is not a single protein. We also set an organism filter for the targets, where the default selection is human. Additionally, we filter out bioactivities if their standard (activity) type is not one of these: IC50, EC50, AC50, XC50, Ki, Kd, potency; since these standard types provide roughly comparable measures of half-maximal response. Furthermore, we eliminated data points 15 without a pChEMBL value, which standardizes the above-mentioned standard types under one value in the negative logarithmic scale. Bioactivity data points with an assigned pChEMBL value have usually received additional curation, and thus, they are more reliable.
Finally, with the aim of only taking the data points at the active binding range (i.e., high affinities between the ligand and the target) we discard the data points with a pChEMBL value less than 5 (i.e., XC50 > 10 µM). Despite these filtering operations, we still usually end up with tens of thousands of compounds before the compound enrichment analysis, which significantly increases the KG construction run time. Exploiting the fact that it is a better choice to include a compound with higher binding affinity compared to a compound with a lower binding affinity for the same target protein, we set the pChEMBL value cut off value to 8, at the beginning of the compound collection procedure. Then, we reduce the cut off value and re-run the query if the total number of gathered compound entries is less than 1000 in the first run. We iteratively do this procedure until we obtain at least 1000 compound entries. Similarly, if the number of returned compounds is more than 2500 in the first run, we further increase the cut off iteratively until we obtain less than 2500 compound entries.
This number (i.e., 1000 to 2500) is still much higher than the number of compounds we incorporate to a KG, which is between 0 and 50; however, we aim to enter the enrichment analysis with a high amount to be able to select the compounds that are interacting with multiple proteins in the network, not just one. Another reason is to be able to select diverse compounds, in terms of their scaffolds/structures, which is explained below (under the compound clustering sub-section).

Compound clustering
There are more than one million compound entries in the ChEMBL database, most of which have bioactivity data points against target biomolecules. Since it is not feasible to include each and every bioactive compound node in a KG (otherwise the graph would be extremely crowded), only the most overrepresented compounds are tried to be incorporated. We observed that some of the compounds with the same (or a very similar) enrichment score are also structurally very similar to each other. These are mostly molecules with matching scaffolds, which are screened against the same target and produced similar results in the same bioassay. Since their enrichment scores are similar as well, they are either selected or discarded together. To provide a better selection of compounds in the graph, we incorporated a structural property-based filtering in the enrichment analysis. The aim here is to select overrepresented compounds that are as diverse from each other as possible in terms of molecular structures, so that users will be provided with a variety of ligands for the target proteins in the graph. To achieve this, we calculated the pairwise molecular similarities between all compounds in CROssBAR-DB using circular fingerprints (ECFP4) and the Tanimoto coefficient. After that, we clustered the compounds based on a predefined similarity cut-off value of 0.5, meaning that each cluster is composed of compounds that are at least 50% similar to each other. The cluster information is pre-calculated and recorded on our server. Each time a knowledge graph is being constructed, enrichment score ranked compounds are checked one by one in terms of their cluster membership and if there already is a compound from the same cluster in the graph, the compound in turn is discarded (i.e., not incorporated to the graph). The same clustering-based selection approach is applied to incorporate compounds that are computationally predicted to interact with the proteins in the graph.
Following the finalization of the compound nodes, we check whether some of these compounds correspond to drugs that are already incorporated to the KG (since ChEMBL also contains bioactivity measurements belonging to approved or investigational drugs), using the identifier mapping between ChEMBL and DrugBank databases. When a positive case is detected, we merge these two nodes and set the node type as a drug, since drugs are As a result, predicted relations comprise the least reliable part of the DTI information provided in CROssBAR-KGs. With the aim of transmitting this evidence-based relation confidence information to users, we used edge labels in the KGs. In terms of visualization, these labels are encoded on the graphs as colours, such that green colour corresponds to DTIs obtained from approved or investigational drugs, blue colour corresponds to experimental bioactivity measures obtained from ChEMBL, and the red colour corresponds to computationally predicted interactions. During the generation of KGs, if a specific relation is obtained from multiple sources (e.g., when the same relation is reported both in DrugBank and in ChEMBL) the edge label of the more reliable relation is incorporated. To accomplish this task, KG construction process comprises an edge label update procedure.
Another process we applied at this step is the edge addition. Some drugs possess bioactivity data points in ChEMBL in addition to their approved targets in DrugBank. To detect this, we first do a mapping between ChEMBL compound entries and DrugBank drug entries, to find the equivalent ChEMBL entry for each drug. After that, we identify the reported ChEMBL bioactivities between that compound and all of the proteins presented in the KG. For those relations, which were not already incorporated into the KG via DrugBank, we added blue coloured edges. The same procedure is applied for adding red coloured edges to the drugs and compounds that have extra computationally predicted target interactions. A user query initiates the entity (node) gathering procedure first from the related database collections, using the CROssBAR RESTful API. Together with the entities that match the search term, the information regarding the related/connected entities are obtained from the corresponding collection. After that, the next database collection is queried with the terms gathered at the previous step. The order of the API queries follows the logic defined for the construction of the KGs, as given in Fig. 1c (simplified version), and in Fig. S2. (full version). Following the initiation of a query, the growing knowledge graph is displayed on the web-browser in real time (using CytoScape Web), starting from the collection and filtering of core and neighbouring genes/proteins. The process is continued with the collection, filtering and addition of phenotype/pathway/disease terms, drugs, bioactive compounds and predicted interacting compounds to the KG (as nodes), together with their relations with gene/protein nodes (as edges). The construction process is finished with the addition of respective edges between non-protein nodes.

CROssBAR Web-Service, User Interface and Layout
An important subject in graph/network visualization is the layout. In CROssBAR-WS, we incorporated the standard layouts of CytoScape Web, such as circle, cose, grid and concentric. However, none of these layouts were sufficient for communicating highly heterogeneous graphs with 7 different types of nodes and 9 different types of edges. To address this problem, we developed the CROssBAR layout, in which biological terms (nodes) from a specific biomedical component (e.g., diseases, pathways, ...) are placed on circular points within a fixed radius. With the aim of preventing overlapping nodes, the radius of each circle is selected as a different value. Curved edge style (i.e., unbundled-bezier) is applied to reduce the amount of edge crossing. More information regarding the usage of CROssBAR-WS and its user interface can be found at https://crossbar.kansil.org/tutorial.php.
CROssBAR web-service queries run in linear time and the actual duration of the process is correlated with the total number of core genes/proteins obtained within the query, together with the annotation volume of these genes/proteins. Highly studied genes/proteins usually have high number of associations, which in turn, extends the actual query runtime in practice. According to our tests, most of the queries with disease, gene/protein, drug and compound terms (in terms of both single and combinatory term searches) take between 1 and 3 minutes to complete (from job submission to the display of the whole graph). However, most pathway and some phenotypic term queries take longer, especially when the number of directly associated genes/proteins is over one hundred. With the aim of creating a better user experience, we applied a procedure in which the collected nodes and their edges are instantaneously and interactively displayed on the screen, before the end of the job. This way, users do not have to wait for the whole job to be finished before starting to explore the KGs.
The entire CROssBAR web-service including the web-site and the underlying API queries can be found in the GitHub repository of the project (https://github.com/cansyl/CROssBAR).

Generation of COVID-19 KGs
We constructed two versions of COVID-19 knowledge graph. First, the large-scale version that includes nearly the whole of the COVID-19 related information recently accumulated in 19 the scientific literature, organized and presented in an interpretable way. Second, the simplified version, which is suitable for quick exploration. The aim behind constructing the simplified version was that the large-scale KG is not easily explorable visually due to its huge size. Since most of the COVID-19 related data has still not been integrated into the regular releases of biological databases, the data could not be pulled to the CROssBAR database at the time of writing this manuscript. As a result, we had to make manual interventions to obtain the data from CROssBAR data resources. We applied the same knowledge graph construction methodology incorporated in CROssBAR; however, we also conducted manual curation to a certain extent, since, unlike the rest of the CROssBAR data, COVID-19 related information has not been extensively curated yet. We saved the pre-constructed KGs, which are directly accessible and viewable through the links given on our web-service (https://crossbar.kansil.org/covid_main.php). It is also important to note that, due to the integrated data resources, CROssBAR heavily contains rare and complex disease data, and mostly leaves infectious diseases out. Nevertheless, the constructed COVID-19 graphs provided rich biomedical information. Below, we describe the methodology followed for the generation of CROssBAR COVID-19 KGs.
Construction of the large-scale COVID-19 graph started with acquiring the related EFO disease term named: "COVID-19" (id: MONDO:0100096). We also incorporated the disease term for "Severe acute respiratory syndrome" (id: EFO:0000694) (the original SARS) into the graph since SARS is better annotated compared to COVID-19. The full-scale COVID-19 KG construction is continued as follows:

COVID-19 related genes/proteins and PPIs
We obtained related genes/proteins and their interactions from the IntAct database's COVID-19 dataset. Unlike a genetic disease, human genes/proteins represent only a portion of an infectious disease. Due to this, we aimed to incorporate SARS-CoV and SARS-CoV-2 genes/proteins, as well as, the host genes/proteins into the graph. Without any filtering, the KG contained 1,172 gene/protein and metabolite nodes from various organisms and 2,214 edges. Due to the high number of genes/proteins in the KG, there was a risk of incorporating non-specific/irrelevant terms from the other biological components at later steps. To address this risk, we applied several filtering operations on this data. First, we eliminated all non-gene/protein nodes and we discarded the genes/proteins if the corresponding organism is not human or SARS-CoV/SARS-CoV-2. Second, we eliminated the protein entries that are not reviewed (i.e., not from UniProtKB/Swiss-Prot) except SARS-CoV-2 ORF10 (accession: A0A663DJA2), which currently is an unreviewed protein entry in UniProtKB/TrEMBL. We also filtered out a portion of the host genes/proteins using 20 interaction-based information, according to their confidence scores reported in IntAct. We discarded the edges between host proteins and SARS-CoV and/or SARS-CoV-2 proteins if the confidence score was less than 0.35. We also discarded the edges between host proteins in the KG (i.e., neighbouring proteins) if their interaction confidence score is less than 0.6. We removed the disconnected components made up of host proteins, which were formed due to the edge filtering operation. Orthology relations between SARS-CoV and SARS-CoV-2 genes/proteins were annotated with "is ortholog of" edge type. The subunits of large protein complexes such as the NSPs of replicase polyprotein 1ab of SARS-CoV/SARS-CoV-2 were mapped to their corresponding protein complex nodes with "is subunit of" edge label.
After these operations, the finalized number of genes/proteins/subunits is 539 ( Pathways of COVID-19 related host genes/proteins Signaling and metabolic pathway information was taken from Reactome (via CROssBAR database) and KEGG pathways data sources. The most relevant pathways were determined by the overrepresentation analysis and mapped to the related genes/proteins in the KG.

COVID-19 related phenotypic implications
The resource for the phenotype terms is the Human Phenotype Ontology (HPO) database.
For each phenotype term that is associated with at least one gene in the KG according to HPO data, we calculated an enrichment score and p-value via overrepresentation analysis.
From the score-ranked HPO term list we selected phenotype terms that are not in a close parent-child relationship with each other in the HPO direct acyclic graph. HPO also has a curated list of SARS related phenotype terms. These terms were also added into the network and mapped to "COVID-19" and "Severe acute respiratory syndrome" disease

Competing Interests statement
The authors declare no competing interests.

Additional information
Supplementary information is available for this paper.  genes/proteins that are associated with the query disease (i.e., core genes/proteins), (3) collects additional genes/proteins (i.e., first-neighbours) using PPIs of core genes/proteins, (4) identifies biological processes (pathways), of which these genes/proteins (core+neighbouring) are members, (5) gathers phenotypic terms (HPO) associated with the whole gene/protein set, (6) obtains known drugs and drug candidate compounds targeting these genes/proteins, together with our deeplearning based interaction predictions, and (7) revisits the disease collection to make another query with all collected genes/proteins, to obtain the disease entries that have similar implications as the query disease.