NETME: on-the-fly knowledge network construction from biomedical literature

Background The rapidly increasing biological literature is a key resource to automatically extract and gain knowledge concerning biological elements and their relations. Knowledge Networks are helpful tools in the context of biological knowledge discovery and modeling. Results We introduce a novel system called NETME, which, starting from a set of full-texts obtained from PubMed, through an easy-to-use web interface, interactively extracts biological elements from ontological databases and then synthesizes a network inferring relations among such elements. The results clearly show that our tool is capable of inferring comprehensive and reliable biological networks. Supplementary Information The online version contains supplementary material available at 10.1007/s41109-021-00435-x.

Processing (Krallinger et al. 2005) tools employ information extraction methods to translate unstructured textual knowledge in a form that can be easily analyzed and used to build a functional network (i.e. a network in which the relations between two entities are not necessarily physical but can be indirect), or knowledge graphs (Szklarczyk et al. 2016;Dörpinghaus et al. 2019;Nicholson and Greene 2020). This technology allows us to infer putative relations among molecules, such as understanding how proteins interact with each other or determining which gene mutations are involved in a disease. In the context of biology and biomedicine, the Biological Expression Language (BEL) (Slater 2014), or Resource Description Framework (RDF) (McBride 2004) have been widely applied to convert a text in semantic triplets having the following form: <subject, predicate, object>. The subject and object represent biological elements, whereas the predicate represents a logical or physical relationship between them (Szklarczyk et al. 2016;Himmelstein and Baranzini 2015).
However, the implementation of biological text mining tools requires highly specialized skills in Natural Language Processing and Information Retrieval. Therefore, several ecosystems and tools have been implemented and made available to the bioscience community. Relevant tools include PubAnnotation (Kim et al. 2019), a public resource for sharing annotated biomedical texts based on the "Agile text mining" concept; PubTator (PTC) (Wei et al. 2019), a web service for viewing and retrieving bio-concept annotations (for genes/proteins, genetic variants, diseases, chemicals, species, and cell lines) in full-text biomedical articles. This latter tool annotates all PubMed abstracts and more than three million full texts. The annotations are downloadable in multiple formats (XML, JSON, and tab-delimited) through the online interface, a RESTful web service, and bulk FTP. Another interesting tool is SemRep (Rindflesch and Fiszman 2003), which extracts relationships from biomedical sentences in PubMed articles by mapping textual content to an ontology that represents its meaning. To establish the binding relation, SemRep relies on internal rules (called "indicator rules"), which map syntactic elements, such as verbs, prepositions, and nominalization, to predicates in the Semantic Network. We also mention Hetionet (Himmelstein et al. 2017), a heterogeneous network of biomedical knowledge that unifies data from a collection of several available databases and millions of publications. Also, the edges are extracted from omics-scale resources and consolidated through multiple studies or resources. Finally, in Yuan et al. (2019) authors propose a minimally supervised approach for knowledge-graph construction based on 24,687 unstructured biomedical abstracts. Authors included entity recognition, unsupervised entity and relation embedding, latent relation generation via clustering, relation refinement, and relation assignment to assign cluster-level labels. The proposed framework can extract 16,192 structured facts with high precision.
Starting from our previous work (Muscolino et al. 2021), we introduce NETME a novel web-based app (available at https:// netme. click/ website, and https:// github. com/ alemu scoli no/ netme. git github repository), which is capable of extracting knowledge from a collection of full-text documents. The tool orchestrates two different technologies: • A customized version of the entity-linker TAGME (Ferragina and Scaiella 2010) (called OntoTAGME) for extracting network nodes (i.e., genes, drugs, diseases) from a collection of full-text articles. • A software module, developed on top of SpaCy (Honnibal et al. 2020) and NLTK (Loper and Bird 2002) libraries, that derives relations (edges) between pair of nodes. Edges are weighted according to their frequency within the collection of full-texts used to create the on-fly knowledge graph.
These inferred networks are handy in biomedicine, where it is essential to understand the difference between various components and mechanisms, such as genes and diseases, and their relations, such as up-regulation and binding. Therefore, the tool helps scientists fast identify reliable relations among the biological entities under investigation, based on their occurrences and mentions in PubMed 's articles. The novelties with respect our previous work (Muscolino et al. 2021) include: • The sentence's grammatical structure is extracted by Spacy linguistic annotations. Such a structure includes the word types (parts of speech) and how the words are related to each other. In the previous NETME release, the nltk bottom-up and top-down approach were employed for building the syntactic tree of each document sentence. Furthermore, the Spacy's Matcher has been used to identify verbs' passive forms. With this approach the system is now capable of properly establishing the correct edge direction. • In Muscolino et al. (2021), the proposed system was able to build a network composed of only genes, diseases, and drugs. Now, thanks to the extension we made on OntoTAGME, our new system is able to build networks composed of much more biological entities such as: genes, variants, diseases, drugs, compounds, molecular function, biological proves, pathways, enzymes, etc. • Finally, we designed and implemented a new module to handle the disambiguation among gene symbols and the acronyms of diseases or other biological elements. In fact, in many documents, the authors assign acronyms for very long biological elements that are usually equal to genes symbols.
To the authors' knowledge, NETME is the first tool that allows to interactively synthesize biological knowledge-graphs on-the-fly starting from a PubMed query. The paper is organized as follows. Section "The NETME model" introduces NETME system together with its components. Section "The annotation tool" provides the technical details of the back-end and the front-end of NETME. Section "Experimental analysis" reports two different case studies that allow evaluating NETME 's prediction qualitatively. The first one is focused on: (i) recovering known gene interactions; (ii) avoid false-negative ones. For this purpose, we selected a subset of gene-gene interactions in KEGG/REACTOME (Kanehisa and Goto 2000;Kanehisa 2019Kanehisa , 2000Fabregat et al. 2017) by making use of STRING API. More precisely, such interactions were obtained by selecting 100 random gene-gene interactions (manually curated in KEGG or REACTOME database) for each of the following STRING text-mining score intervals: 500-600,600-700, 700-800, 800-900, >= 900 . Next, we selected the first 100 pairs of non-interacting genes from the Negatome 2.0 database (Blohm et al. 2013;Smialowski et al. 2009) in order to understand if NETME can avoid false-negative interactions. The experiment yielded accuracy values from 58% when the STRING text-minig score is in [500,600] interval, to 84% when the value of such a score is higher than 900. Whereas, the second case study is focused on building a "CD147genes" interaction network through selected papers containing valuable information about CD147 gene. We compared the network returned by NETME against a manually-curated network derived from these selected papers. The experiment yielded 98% sensitivity and 100% specificity. Therefore, both experiments clearly showed the high reliability of NETME inferred networks. Moreover, we have also assessed the NETME performance for inferring "CD147-diseases" interactions by selecting 100 random interactions from DisGenNET, and the same "abstracts" used by DisGenNET for inferring these interactions. NETME detected 63 True Positive values out of 100, revealing a sensitivity of 63% Sect. "Conclusion" ends the paper and sketches future research directions.

The NETME model
A Knowledge Graph (also known as a semantic network) is a systematic way to connect information and data to knowledge. It represents a collection of interlinked descriptions of entities, real-world objects, and events, or abstract concepts, obtained from knowledge-bases such as ontologies (O 1 , O 2 , · · · , O k ) . Basically, a semantic network is defined as a graph G = (V , E) where entities are in V, and relationships in E. Each relation represents a connection between entities of one (intra-relationship) or more (inter-relationship) ontologies (Nettleton 2014). Therefore, there might exist a relation e = An ontology is a formal description of knowledge as a set of domain-based concepts in relationships among them. As a result, the ontology does not only introduce a shareable and reusable knowledge representation, but it can also provide new knowledge about the considered domain (Xiaoke and Lin 2012). NETME builds a biomedical knowledge graph starting from a set of n documents obtained through a query to the PubMed database. Papers can be sorted by relevance (default) or publication date. Users can also provide a list of PMCID/PMID or a set of PDF documents. The inferred network contains biological elements (i.e., genes, diseases, drugs, enzymes) as nodes and edges as possible relationships.
In Fig. 1 we outline the architecture of NETME. The user provides the query terms to perform the search on PubMed, and she may directly provide PDFs or PMCIDs/PMIDs of other pertinent documents. Then NETME begins to create the network as follows: 1 First, OntoTAGME converts the full-text of the input documents into a list of entities (nodes) using literature databases and ontologies (such as GeneOntology Consortium 2004, Drug-Bank Wishart et al. 2017, DisGeNET Piñero et al. 2019, and Obofoundry Smith et al. 2007) as corpus. These entities will be the knowledge graph nodes. Note that, Obofoundry contains a several ontologies, but only the following have been currently used in our model: GO, DO, PW, BTO, PRO, AEO, PATO, CL and CLO. (Loper and Bird 2002) libraries, is executed to infer the relations among nodes entitynodes belonging to the same sentence ( S i ) or to the adjacent ones ( S i , S i+1 ) of the same document. Such relationships indicate disease treatment, genes regulations, molecular functions, gene-gene interactions, gene-disease interactions, gene-drug interactions, drug-disease interactions, disease-disease interactions and drug-drug interactions.
The final network will contain both directed and undirected edges according to the predictions made by the model. At the end of the process, the network will be rendered through Cytoscape JS. The following two subsections provide the details of these two phases.
OntoTAGME: Ontology oN Top Of TAGME TAGME Ferragina and Scaiella (2010) is a state-of-the-art entity linker for annotating Wikipedia pages mentioned in an input text. The tool searches for sequences of words (spots) that can be linked to pertinent Wikipedia pages (entities) that explain those words in that context. The use of Wikipedia as corpus allows to enrich texts with explanatory links in order to provide a structured knowledge for any unstructured fragment of the text. These links are then used for drawing a network of relationships among the extracted spots.
To mitigate ambiguity and polysemy, TAGME computes a ρ value ∈ [0, 1] for each Spot-Entity (Node) association, and keeps only those ones having the ρ value higher than an established user threshold. This value estimates the "goodness" of the annotation compared to other possible associations in the input text. A suitable use of ρ ensures the highest accordance among the extracted spots.
Due to the topics-generality of the Wikipedia corpus used by TAGME, several nonbiological spots could be extracted during the annotation procedure. To overcome this limitation, we developed a customized version of TAGME, called OntoTAGME, which makes use of several ontology and literature databases, such as: GeneOntology  (Griffith et al. 2017), andPharmGKB (Whirl-Carrillo et al. 2012). The usage of topic-specific ontology databases ensures reduced disambiguation errors and therefore yields highly reliable knowledge graphs inference.
The integration consisted of releasing a new intermediate python layer (Python Parser in Fig. 2), and a customized two-steps procedure (Wikipedia Adapter module in Fig. 2) for converting ontology databases in a wikipedia-like structure. The Python layer transforms a generic ontology or database in a list of CSV files: pages.csv, pageslink.csv and category.csv. The pages.csv stores the name of each biological element, and all possible synonyms. The pageslink.csv contains all the relationships among the nodes of the ontology. Finally, the category.csv has the type of each element extracted from the ontology or database entry (i.e Genes, Diseases, Drugs).
Next, a two-steps procedure is triggered to convert each row of the page.csv file into an XML file containing a unique ID generated by our system, the name (title), type (category) and the description (page's body) of the considered biological element. Since an element j could have several linked pages "LPs" (i.e. DOID:0002116 is a DOID:10124), or redirected pages "RPs" due to synonyms (CD147 is a synonym of BSG), the process generates a tuple uniqueID j , uniqueID k for each element k belonging to LPs, and a tuple uniqueID j , uniqueID i for each element i belonging to RPs. These tuples are then stored in the SQL files "wiki-latest-pagelinks" and "wikilatest-redirect", respectively.
Finally, the SQL and XML files are used to generate the complete OntoTAGME network. It contains 331 thousand of main nodes, 700 thousand of synonyms, and 4 million of relationships.

Ontology databases
In order to build the OntoTAGME annotation networks we used the following nine ontology and six bio-databases. DrugBank Wishart et al. (2017) contains data about drugs name, drugs synonyms, drug-drug interaction, and other comprehensive drug-target information. The database release used in our project is the v5.1 which contains 13, 367 drugs entries, including 2, 611 approved small molecule drugs, 1, 300 approved biotech (protein/ peptide) drugs, 130 nutraceuticals and over 6, 315 experimental drugs. Additionally, 5, 155 non-redundant protein (i.e. drug target/enzyme/transporter/carrier) sequences are linked to these drug entries.
HGNG (HUGO Gene Nomenclature Committee) Gray et al. (2016) assigns unique and informative gene symbols and names to human genes. Standardized HGNC approved nomenclature is used in publications and biomedical databases to remove ambiguity and facilitate communication between researchers worldwide. The last database release contains more than 40, 000 approved gene symbols of which over 19, 000 are for protein-coding genes. The HGNC also names a set of small and long non-coding RNA genes and pseudo-genes (659 since 2017). The genes are grouped on the basis of several shared characteristics such as homology, associated phenotype and encoded protein function.
Ensembl Birney (2004) contains genome annotation (i.e genes, variation, regulation and comparative genomics) across the vertebrate sub-phylum and key model organisms. This tool is also able to compute multiple alignments, predicts regulatory function and collects disease data. The last complete version of the Ensembl database has been downloaded through their FTP service, and then integrated in OntoTAGME thanks to Python Parser layer. All data in Ensembl are used in combination with those coming from HGNC to detect Genes name and symbols within a text.
Disgenet Piñero et al. (2019) contains collections of genes and variants associated with human diseases. It integrates data from scientific literature, GWAS catalogues, expert curated repositories and animal models. Additionally, several original metrics are provided to assist the prioritization of genotype-phenotype relationships. Dis-GeNET releases two types of databases, Gene-Disease Associations and Variant-Gene Associations.
CIViC Griffith et al. (2017) is an expert-crowd-sourced knowledge-base for Clinical Interpretation of Variants in Cancer describing the therapeutic, prognostic, diagnostic and predisposing relevance of inherited and somatic variants of all types. CIViC is committed to open-source code, open-access content, public application programming interfaces (APIs) and provenance of supporting evidence to allow for the transparent creation of current and accurate variant interpretations for use in cancer precision medicine.
PharmGKB Whirl-Carrillo et al. (2012) is an interactive tool for researchers investigating how genetic variation affects drug response. It displays genotype, molecular, and clinical knowledge integrated into pathway representations and Very Important Pharmacogene (VIP) summaries with links to additional external resources. A user may search and browse the knowledge-base by genes, variants, drugs, diseases, and pathways through the website: http:// www. pharm gkb. org).
OBO Foundry Smith et al. (2007) is the Open Biological and Biomedical Ontology (OBO) Foundry. It provides well-formed and scientifically accurate ontology thanks to the collaboration of ontology developers. They contribute to develop an evolving set of principles and common syntax based on ontology models that ensure the proper functioning of the system. In NETME, we use the following list of ontology: • Gene Ontology (GO) Consortium (2004)  The data relating to the number of nodes and relationships extracted from each mentioned ontology have been listed in Table 1 Network edge inference Once the network nodes have been extracted the system will annotate their position and their main characteristics within the text. We capture the significant elements in each sentence, by making use of the parts of speech (POS tags). Then through a syntactic analysis we verify the coherence of the extracted elements. Indeed, sentences have an internal organization that can be represented using a tree. Solving a syntax analysis problem for a sentence consists of looking for predefined syntactic forms which, like a tree, branch out from the single words. The main syntactic form is the sentence (S) which contains noun phrases (NP) or verb phrases (VP) that are formed by further elementary syntactic forms such as nouns (N), verbs (V), determiners (DET), etc (see Table 3). All these information will be used by the textual analysis phase to infer relations between them. A transition-based dependency parser is then used to first check the syntactic coherence and then build the syntactic tree. The dependency parser component inside the spaCy library jointly learns sentence segmentation and labelled dependency parsing. The parser uses a variant of the non-monotonic arc-eager transition-system (Honnibal and Johnson 2015), with the addition of a break transition to perform the sentence segmentation. Nivre's (2005) pseudo-projective dependency transformation is also used to allow the parser to predict non-projective parses. The parser is trained through an imitation learning objective. It follows the actions predicted by the current weights and, at each state, it determines which actions are compatible with the optimal parse that could be reached from the current state. The weights are updated in a way that the scores assigned to the set of optimal actions is increased, while scores assigned to other actions are decreased. Note that more than one action may be optimal for a given state.
Once OntoTAGME have extracted the set of nodes n 1 , . . . , n z from a list of N fulltext documents [p 1 , p 2 , . . . , p N ] , the edge inference module of NETME (developed on top of the Python library NLTK Loper and Bird 2002 and spaCy (Honnibal et al. 2020)) starts to establish any verbal relationships between those pairs of nodes. When two or more nodes are detected within a sentence or adjacent sentences, the syntactic analyzer extracts the parts of speech and syntactic dependencies within the sentence. For each sentence we then get a set of labelled tokens lt 1 , lt 2 . . . , lt k i . Each token is a tuple of the following form {token, POS, dependency_label} , where POS and Dependency label are valued with the data present in Table 3. Irrelevant POS are filtered out (stop-words, URLs, etc.), we keep only the useful verb forms and the nodes which correspond to the noun parts. A final pruning phase is also executed in which we use: (i) POS tag labels and dependency labels to check if the syntactic link between the verb form and the annotations is correct and consistent, as described in the Fig. 3; (ii) a dictionary of biological verb forms to check if they are pertinent. The surviving nodes and verb forms will allow to generate network edges.
In our final network, each edge e = (a, b) is weighted with three parameters: the term frequency and inverse document frequency (tf.idf ), the medium relatedness (mrho) and the biological degree (bio). More specifically, tf.idf is a measure of how much information the edge provides, namely if it is common or rare across all input documents. In formula, we compute tf .idf (e, p, P) = tf (e, p) * idf (e, P).
Where, term frequency tf (e, p) is the frequency of edge e, is defined as tf (e, p) = f e,p / e ′ ∈p f e ′ ,p , with f e,p representing the number of times that edge e occurs in paper p. The inverse document frequency idf (e, P) is a measure of how much information the edge e provides. It is defined as idf (e, P) = logN /|{p ∈ P : e ∈ p}| , where N is the number of documents analyzed by the query such that N = |P| , and |{p ∈ P : e ∈ p}| is the number of documents where the edge e appears. The parameter mrho measures the relatedness of the labels starting from the ρ value assigned by OntoTAGME to the two annotations involved, i.e. mrho(e) = ρ a * ρ b 2 . The bio-parameter is the cosine similarity (having a value ranging from 0 to 1) between the inferred relationship and a set of biological verb forms (see Table 2). Figure 4 provides an example of such an annotation.

The annotation tool
NETME is provided with a front-end developed in PHP and Javascript, in which the network rendering is performed through the CytoscapeJS library (Franz et al. 2015). Its back-end, which integrates OntoTAGME, is written in Java and communicates with both Python NLTK (Loper and Bird 2002) and SpaCy (Honnibal et al. 2020) libraries for the NLP module. PubMed search is performed with the Entrez Programming Utilities (https:// www. ncbi. nlm. nih. gov/ books/ NBK25 501/), a set of server-side programs providing a stable interface to the Entrez database and to the query system at the National Center for Biotechnology Information (NCBI).  [...]. Through OntoTAGME we detect the spots ["BSG", "VEGFA", "PGF"]. After the syntactic analysis, three noun parts are identified (the phrase spots, highlighted via orange segments): two of them ("VEGF" and "PLGF") have a joint relationship with the first ("CD147"). The verbal part is the root between the two pairs of nouns ("CD147"-"VEGF"), (CD147-"PLGF") NETME is equipped with an easy-to-use web interface providing three major functions (see Fig. 5): (i) Pubmed query-based network annotation; (ii) user-provided freetext network annotation; (iii) user-provided PDF documents network annotation.
In the query-based network annotation, the user provides a list of keywords, which are employed to run a query on PubMed, or a list of article ids. The top resulting papers are retrieved and then the network inference procedure is run. Several parameters can be set by the user (or left with default values) such as: the number of top article to retrieve from PubMed, and the criteria used to sort papers (relevance or date).
In the user-provided free-text network annotation, users provide a free text which is then input to the network inference procedure.   [...]. Through OntoTAGME we detect the spots ["BSG", "VEGFA", "PGF"], and after the syntactic analysis and noise reduction steps, we detect two valid edges: ["BSG", "regulate", "VEGFA"] and ["BSG", "regulate", "PGF"]. Note that "regulate" is a biological verb forms and it has bio parameter set to 0  In the user-provided PDF documents network annotation, users give a set of PDF documents which are then input to the network inference procedure.
The result of the network inference procedure is a direct graph (network) which shows all inference details in three main tables containing: the list of extracted papers, the list of annotations, and the list of edges together with their weight.
The user can then click on a node of the network to view all incoming and outgoing connections, or she can click on an edge to display its type and the verbal relation between the nodes it connects.

Experimental analysis
To analyze the reliability of NETME knowledge graphs, we performed two case studies. The first one aims at providing a comprehensive analysis of NETME performance by checking its ability to predict known relations between genes drawn from Kyoto Encyclopedia of Genes and Genomes -KEGG (Kanehisa and Goto 2000;Kanehisa 2019Kanehisa , 2000 or REACTOME (Croft et al. 2010;Joshi-Tope 2004;Croft et al. 2013) pathways and, on the other hand, its ability to avoid inferring false connections between proteins by using the Negatome 2.0 database (Blohm et al. 2013;Smialowski et al. 2009). The second case study is more specific and focuses on building a network based on some selected publications that contain valuable information specific to the CD147 gene. Such a network is then compared against a manually-curated one derived from the same papers by a bio-expert. In both cases, the performance of NETME has been measured in terms of a precision/recall curve.

Case study 1
The first case study focuses on assessing NETME performance through its capability to recover known gene interactions. For this purpose, we selected a subset of gene-gene interactions from KEGG/REACTOME by making use of STRING API. More precisely, such interactions were obtained by selecting 100 random gene-gene interactions for each of the following STRING text-mining score intervals: 500-600, 600-700, 700-800, 800-900, ≥ 900 (listed in Additional files 1, 2, 3, 4, 5, respectively). These interactions form the true-positive set.
Next, we selected 100 random pairs of non-interacting genes from the Negatome 2.0 database as a true-negative set (listed in Table 5). For each interacting gene-pairs, we queried NETME with the papers used by STRING to infer the interactions. On the other hand, to annotate non-interacting genes, we queried NETME with the pair of genes of interest, selecting the top 20 papers from PubMed. Accuracy, sensitivity, specificity and PPV values, detected by NETME, are listed in Table 4 The results clearly show that NETME produces reliable results when the annotations are performed on top of relevant literature (STRING text-mining score higher than 700). On the other hand, when the STRING text-mining score is lower than 700, the NETME performances degrade in accordance with STRING predicted confidence as highlighted by their score . The reason behind such a behaviour is due: (i) not enough literature about these interactions; (ii) the interactions have been inferred by human curators as a combination of other interactions occurring in the text. Furthermore, when the text-mining score is small, STRING predictions could be wrong. In fact, as reported in Szklarczyk et al. (2016), a score of 500 would indicate that roughly every second term of an interaction might be erroneous (i.e., a false positive). Therefore, the computed value of accuracy, sensitivity, specificity and PPV could be incorrect.

Case study 2
Many tools (Alaimo et al. 2020) and computational models rely on existing network databases, such as KEGG (Kanehisa and Goto 2000;Kanehisa 2019Kanehisa , 2000 and Reactome (Croft et al. 2010;Joshi-Tope 2004;Croft et al. 2013). However, despite the enormous amount of available data, these databases are still incomplete and therefore have partial information (Menche et al. 2015). As an example, KEGG includes approximately one-third of the known genes. In this case study, we have chosen CD147, also known as Basigin (BSG) or EMMPRIN, as a starting point for the gene-gene interactions network construction. This gene represents an example of a biological element that should be supplemented to the KEGG network since it is not currently described in their pathways. Among the bibliography consulted to build the network manually, we have carefully selected 11 papers containing a significant amount of helpful information for our purpose. On the other hand, in this case study, we have also assessed the capabilities of NETME in inferring CD147-diseases relations. For this purpose we selected 100 random interactions from DisGenNET (Piñero et al. 2019), as well as the same abstracts used by DisGenNET for inferring such interactions (listed in Additional file 6).
CD147 is a transmembrane glycoprotein of the immunoglobulin superfamily, expressed in many tissues and cells, which is known to participate in several high biological and clinical relevance processes and is a crucial molecule in the pathogenesis of several human diseases (Xiong et al. 2014). Recently Wang et al. (2020) discovered an interaction between host cell receptor CD147 and SARS-CoV-2 spike protein, together with Angiotensin-Converting Enzyme 2 (ACE2), as an entry point for SARS-CoV-2.
In this direction, CD147 is an example of how a missing crucial gene within a biological network can compromise scientists' efforts to understand certain molecular phenomena. In literature, there are many valuable tools (Himmelstein et al. 2017;Himmelstein and Baranzini 2015) to integrate the missing information into bio-databases, such as KEGG. However, the most reliable approach in terms of accuracy and updated information remains the manual curation of such networks through careful and time-consuming literature analysis. On the other hand, a manually constructed network provides partial information due to the limited number of articles that a scientist could read. Our second case study affords this issue by providing a practical example of how NETME can create valuable networks by analyzing quickly and automatically larger sets of publications. The set of 11 selected papers, described in Fig. 7a, was analyzed by a bio-expert to derive a CD147-genes interactions network manually. This process resulted in 50 genes and 64 interactions, as shown in Fig. 7a. Next, by using the same set of papers, we run NETME with no upstream filter. The automatically generated network consisted of 86 genes and 139 relationships between them (see Fig. 7a, b). As the manually curated network consists of genes and proteins, only elements from these two categories were selected for the evaluation. This was performed by considering edges with the lowest "bio" score for each node pair. Qualitatively, this network includes most of the interconnections mentioned in the papers, thus providing a reliable and comprehensive overview of the molecular function of Basigin. Quantitatively, NETME achieved an accuracy of 98.99%, a sensitivity of 100%, a specificity of 98.98%, and a positive predicted value of 46.32%. Figure 6a-c depicts the precision/recall curve (AUC 0.997), the sensitivity/specificity curve and the True positive rate/False Positive Rate one. The construction of the curves considered all possible gene-pairs and their edges.
Finally, we queried NETME with the selected 100 random CD147-diseases interactions in DisGenNET, selecting the same PubMed abstract used by DisGenNET for inferring those interactions. NETME detected 63 True Positive values out of 100, revealing a sensitivity of 63% It is essential to stress that NETME allows us to extract a satisfactory and valid amount of information in a few minutes, compared to a manual search that may take days or weeks. We also believe that this case study is significant because, in the evaluation, we considered not only the presence of a link between two nodes but even more closely the type of edge, hence the adequacy and specificity of the annotated edge in its biological context.

Conclusions
In this paper, we have introduced NETME system to infer on-the-fly knowledge-graphs from a collection of either full-text papers obtained from PubMed or user-provided ones. It has been implemented upon a customized version of TAGME, called OntoTAGME, in connection to a syntactic analysis module developed on top of the Python NLTK and SpaCy libraries. Our results clearly show that NETME allows extracting reliable knowledge graphs in a few minutes or hours compared to a manual search that could take several days or weeks. The completeness of the extracted knowledge increases when the documents used by NETME comprehensively describe the desired topic under study. To evaluate NETME, we performed two case studies. The first one tested the ability of NETME in recovering relationships between genes. The experiment yielded accuracy ranging from 58%, when using low reliable relations (i.e. False Positives) from STRING, Fig. 7 a Depicts the pathway constructed by hand from the selected papers Kong et al. 2014;Ke et al. 2012;Grass and Toole 2016;Xiong et al. 2014;Rucci et al. 2010;Ding et al. 2017;Ulrich and Pillat 2020;Wang et al. 2014;Kong et al. 2014;Kirk et al. 2000), with CD147(BSG) as the central node. b Shows the molecular mechanisms summarised in the knowledge network developed by NETME in accordance with the same papers used in a NETME shows that CD147 is a potent inducer of metalloproteinases (MMPs) such as MMP2, MMP14 and MMP9 as reported in Xiong et al. (2014); Rucci et al. (2010);Ding et al. (2017). Furthermore, the overexpression of CD147, which results in increased phosphorylation of PI3K(PIK3CA), Akt(AKT1), leads to the secretion of vascular endothelial growth factor (VEGFA) in several biological contexts such as KSHV infection Xiong et al. (2014); Rucci et al. (2010). In addition to its ability to induce MMPs, CD147 regulates spermatogenesis, lymphocyte reactivity and MCT system, in particular MCT1 and MCT4 (MCTS1 and SLC16A4) expression (Xiong et al. 2014;Kirk et al. 2000). Our results also show that CD147 can increase the expression of ATP-binding cassette transporter G2 (ABCG2) protein, regulating its function as a drug transporter, as mentioned by Xiong et al. for MCF-7 cells (Xiong et al. 2014). NETME identifies also BSG as an upstream activator of STAT3, highlighting its involvement in tumor development in agreement with the literature . As summarized by our knowledge network, CD147 is regulated by various inflammatory mediators, such as RANKL (TNFSF11), denoting its involvement in inflammatory processes (Grass and Toole 2016;Rucci et al. 2010). Among the potential activators of BSG, NETME also find the transcription factor c-Myc (MYC) (Kong et al. 2014) (See figure on next page.) Table 4 Metrics on NETME 's ability to predict known interactions (from KEGG/Reactome) and noninteractions (from Negatome 2.0) between genes   Table 5 List of the first 100 pairs of non-interacting genes from the Negatome 2.0 database.The column "SOURCE" indicates the starting gene, instead the column "TARGET" indicates the gene to which the action of the source gene is directed  CD274  ICOS  RIPK1  TNFRSF10A   CD3G  ZAP70  RIPK1  TNFRSF10B   CD74  NOTCH1  SFN  TSC1   CDKN1B  TSC1  SH3KBP1  TNFRSF14   CSF2  IL3RA  SMAD1  ANAPC10   CTNNB1  HSP90AA1  SMAD4  ANAPC10   CTNNB1  DDIT3  SOCS3  JAK2   CTNND1  IL2  STIM1  TRPC6   CTNND1  APC  TANK  RBCK1   CTNND1  CTNNA1  TBC1D7  TSC2   CTNND1  CTNND1  TFDP1  CDK2   CTNND1  CTNNB1  TFDP1  CCNA1   DKK1  WNT1  TICAM1  TLR4   DKK1  SOST  TJAP1  F11R   DVL1  TSC1  TJAP1  CLDN1   EIF3I  ACVR2A  TJAP1  TJP1   EIF3I  ACVR1  TNF  EGFR   EIF3I  TGFBR1  TRADD  TNFRSF10A   EP300  CD44  TRADD  TNFRSF10B   ERBB2  PIK3R2  TRAF6  IRF3   ETS1  CREBBP  TSC1  CDKN1B   FOXO1  TSC1  VAV1  SHC1   GRAP2  SOS1  VEGFB  KDR   GRAP2  CBL  VEGFB  FLT4   HDAC2  RELA  VEGFC  FLT1   HIPK2  MDM2  VIPR2  RAMP1   HSPA4  BAX  VIPR2  RAMP2   IGF2  IGF1R  VIPR2  RAMP3   IL15  IL2RA  VWF  F8   IL1A  EGFR  YWHAB  TSC1 to 84% when such STRING relations are very reliable. At the same time, the second case study tested the ability of NETME in integrating knowledge about genes starting from a selected set of papers. The experiment yielded 98% sensitivity and 100% specificity. Therefore, both experiments clearly showed the high reliability of NETME 's inferred networks. Future work will include: (i) the construction of knowledge-graphs from all the openaccess papers stored in PubMed Central; (ii) the integration of all Obofoundry ontology within OntoTAGME; (iii) the design of a more effective algorithm to select the pertinent papers on which NETME has to be applied (Ponza et al. 2019(Ponza et al. , 2020; and finally, add a methodology that allows to extract context-based relationships