hPP Corpus: A Tagged Biomedical Corpus for Automatic Extraction of Human Protein Phosphorylation for Understanding Cellular Functions

Protein Phosphorylation Process

Protein phosphorylation plays an essential role in cellular function and signal transduction, and receives a significant amount of attention [16]. Phosphorylation is one of the most common post-transcriptional modification of proteins where a serine/threonine/tyrosine residue is phosphorylated by a protein kinase. Phosphorylation occurs on several amino acids within a protein. Apart from the commonly observed serine, threonine, and tyrosine residues, phosphorylation has been reported to occur on phosphoserine residue in eukaryotes, and on histidine / aspartate residues in prokaryotes [17]. The biological entities related to protein phosphorylation are:

(i) Enzyme/kinase (that phosphorylates protein) (ii) Protein/substrate (that is phosphorylated) (iii) Site (the residue (serine/threonine/tyrosine) that is phosphorylated

Retrieval of Literature Dataset

PubMed is the main resource for biomedical literature. The database contains 122,250 MEDLINE abstracts related to human protein phosphorylation and literature data sampling for creating the corpus is highly complicated. However, experimentally verified information on phosphorylation is available in UnitProt Knowledgebase [1] as well as in specialized databases such as Phospho.ELM [18], PhosphoBase [19], PhosphoSitePlus [20], PhosphoNET [21] and PhosphoPOINT [22].

The literature data for corpus creation is derived from PubMed using the MEDLINE abstract references from two popular phosphorylation databases namely Phospho. ELM [18] and PhosphoSitePlus [20]. Phospho.ELM is a database of experimentally verified phosphorylation sites in eukaryotic proteins, providing UniPROT/Ensembl accession number, sequence, phosphorylated residue and its position, PubMed ID, the upstream kinase (when known), source (High- ThroughPut/Low-ThroughPut) and entry date for the various species including human. PhosphoSitePlus is a comprehensive resource on experimentally determined post-transcriptional modification of proteins including acetylation, methylation, and phosphorylation in man and mouse. The database entries include protein name, type and accession number, phosphorylation residue, PubMed ID etc. created by cell-signaling technology.

We adopted the likelihood prediction approach of GeneTag [23] to rank the MEDLINE abstracts containing information on gene/protein names i.e. substrate and kinase. High-scoring abstracts always contain substrate/kinase names when compared to the low-scoring abstracts with few or no substrate/kinase names. We randomly selected 1000 high-scoring MEDLINE abstracts as the basis for hPP corpus, because we required the corpus to contain both substrate and kinase names. A sample sentence from biomedical literature on protein phosphorylation information is illustrated in Example 1, where ERalpha is a substrate protein phosphorylated by a kinase, protein kinase A at the phosphorylation site serine-236.

Example 1: PMID: 9891036

Here we show that ERalpha is phosphorylated by protein kinase A on serine-236 within the DNA binding domain.

Text Preprocessing

Our text preprocessing is carried out in two phases. An initial preprocessing to retrieve article ID, title and abstract from MEDLINE format is carried out with an in-house Java program. We prefer MEDLINE format than XML format for data collection, because the former automatically converts the special characters such as α, β, μ etc. to their corresponding English terms i.e. alpha, beta, mew etc., thus facilitating the automatic text processing by text mining tools. The input text is split into a set of lines, where each line contains only one sentence assigned with PubMed ID. Sentences containing phosphorylation related keywords (Table 1) are more likely to contain phosphorylation information than those containing other words and are filtered out for further processing.

Corpus Annotation

We describe the corpus annotation with two key steps namely, entity annotation and event annotation. The first stage of corpus annotation is the identification of biomedical entities such as proteins, substrates, protein kinases and phosphorylation sites related to protein phosphorylation. The next stage is an extended annotation of the identified biomedical entities to recognize the pertaining phosphorylation event relationship between the entities. For instance, the annotated entities in Example 1 are not only proteins (ERalpha and protein kinase A) and phosphorylation sites (serine-236), but also the biological process “ERalpha is phosphorylated by protein kinase A on serine-236” that pertain to the proteins (Figure 1).

Click to enlarge

Automatic Annotation of Gene/Protein Names

NER approach using NAGGNER: Two of the three phosphorylation objects are proteins i.e., kinases and substrates. Hence, we used the hybrid protein name tagger called NAGGNER [24] for automatic recognition of protein names. NAGGNER consists of three major components namely (i) machine learning using conditional random fields (CRF) for initial learning and labeling, (ii) rule based tagging to improve the performance of initial tagging process, and (iii) a two stage abbreviation identification algorithm for the recognition of human gene/protein mentions.

Pattern templates for gene/protein name annotation: We introduced a set of pattern templates to tag the protein names (e.g. suppressor of fused (Sufu) protein in PubMed ID: 21317289) that were missed by the hybrid tagger. Our templates are specific patterns of grammatical components within the noun phrase (NP) of the input sentence [25]. We parsed the input sentence with Stanford lexical parser with grammar settings to englishPCFG module [26] to obtain a parse tree structure. We located the head-word within the top level NP. The outputs from the parser were manually reviewed for parsing accuracy. The head-word is always a noun or proper name (possibly with more than one word). We considered all the words preceding the first head-word as pre-modifiers and all the words following the first head word as post-modifiers within NP and apply the pattern templates to extract the phrase with protein names. We used the proteins dictionary from our previous work [27] to confirm whether the phrase corresponds to a protein name. The syntax tags expressed in the pattern matching templates are listed in Table 2.

Pattern 1: NP [?DT/JJ/NN NN+)] A simple NP with optional pre-modifiers such as determiner (DT), adjective (JJ), or noun (NN) followed by a head-word is the most common pattern for a protein name.

Pattern 2: NP [NP ?PP [IN NP]] NP that includes another NP followed by an optional PP is tagged as a protein name. PP is made up of a preposition (IN) such as ‘for’, ‘in’, ‘etc.’, and ‘of’, followed by another NP.

Pattern 3: NP [NP ?acronym] The pattern template is similar to Pattern 2. NP within a top level NP is followed by an optional acronym. Antiporter, antizyme, complement, exchanger, neuropeptide, oncoprotein, photoreceptor, receptor, symporter, and transporter are few important acronyms which have to be tagged as part of protein names. A set of generic names which may not be tagged individually or with preceding protein names is in given in Table 3.

Example 2: PMID: 7925415 Original sentence: An epidermal-growth-factor (EGF)- receptor preparation isolated by calmodulin-affinity chromatography from rat liver plasma membranes is able to phosphorylate calmodulin.

Tagging by NAGGNER and pattern: An <KINASE> epidermal-growth-factor (EGF)–receptor </-KINASE> preparation isolated by calmodulin-affinity chromatography from rat liver plasma membranes is able to phosphorylate <PROTEIN> calmodulin </PROTEIN> .

In Example 2, epidermal growth factor receptor is identified and tagged as protein by NAGGNER and two pattern templates 1 and 3 (Figure 2).

Click to enlarge

Protein kinase recognition: Recognition of kinases among the tagged gene/protein names is a challenging task because each protein kinase has a list of related synonyms [27]. For example, spleen tyrosine kinase has 3 related synonyms (SYK, p72-Syk, tyrosine-protein kinase SYK). We developed a specialized synonyms dictionary to distinguish human protein kinases from other human genes/proteins (Table 4).

First, we downloaded all human genes/proteins approved by Human Gene Nomenclature Committee (HGNC) using the keyword (Homo sapiens [Organism]) AND HGNC from EntrezGene (http://www.ncbi.nlm.nih.gov/gene). Next, we mapped the protein names to the synonyms collected from three resources namely (i) the synonym dictionary provided by BioCreAtIvE-II [28] contest, comprising of 32,975 entries (ii) the human protein/gene name dictionary available at NCBI (http://www.ncbi.nlm.nih.gov/gene) covering 47,177 entries and (iii) the human protein synonyms available at UniProt [1, 29] covering 14,893 entries. Finally, we generated a list of 518 human protein kinases from two datasets namely (i) human genes/protein kinases approved by HGNC from EntrezGene and (ii) protein kinase entries from two popular databases namely KinBase [30] containing 516 entries and Kinweb [31] containing 518 entries. We mapped the kinase list with the corresponding entries in the specialized synonyms dictionary. Thus, the dictionary contains (i) official symbol, (ii) gene ID (iii) all possible synonyms and (iv) Disambiguation as proteins and kinases for all HGNC approved human genes/proteins.

Automatic Recognition of Phosphorylation Sites

Recognition of phosphorylation residue/site in the unstructured text is equally challenging like protein/gene name recognition. The three amino acid (serine, threonine and tyrosine) residues related to protein phosphorylation are represented in three forms in the biomedical literature: i) full name (e.g. serine), ii) short name (e.g. Ser) and iii) acronym (e.g. S) [15, 32]. The residue is commonly represented with the location number (e.g. serine 432). We used Java Regex function with a set of pattern templates to tag all possible forms of phosphorylation site. Table 5 lists the phosphorylation site patterns and the related Regex functions.

Manual Annotation Process

The common practice in the annotation process is to start with the automatically parsed and annotated text. However, since the automatic annotation could not provide 100% accurate results, we chose to perform a second level of manual annotation on the automatically tagged text. We hired three annotators with knowledge in protein phosphorylation to investigate the automatically annotated data: (i) to tag biological entities missed / incorrectly tagged by automatic tagging Example 3: PMID: 2457390 Automatic Tagging: The major site of phosphorylation by adenosine cyclic 3’, 5’ -phosphate dependent protein kinase was on the carboxy-terminal half of the molecule at <SITE> Thr-216 </SITE>. Manual Tagging: The major site of phosphorylation by <KINASE> adenosine cyclic 3’, 5’ -phosphate dependent protein kinase </KINASE> was on the carboxy-terminal half of the molecule at <SITE> Thr-216 </SITE>. (ii) to remove / re-tag the incorrectly assigned tags Example 4: PMID: 12118371 Automatic Tagging: These results suggest that phosphorylation of the <SUBSTRATE> EGF receptor </ SUBSTRATE> at <SITE> Thr669 </SITE> and <SITE> Ser671 </SITE> mediates interaction of the receptor with a specific <PROTEIN> tyrosine kinase </PROTEIN> substrate and is required for efficient ligand-induced receptor internalization. Manual Tagging: These results suggest that phosphorylation of the <SUBSTRATE> EGF receptor </SUBSTRATE> at <SITE> Thr669 </SITE> and <SITE> Ser671 </SITE> mediates interaction of the receptor with a specific tyrosine kinase substrate and is required for efficient ligand-induced receptor internalization. (iii) to re-tag the complex phosphorylation site mentions in simple machine readable form Example 5: : 2019585 Automatic Tagging: Solid phase sequencing revealed phosphorylation at serines 122 , 150 , 212 , 220 , 234 , and 315 and <SITE> threonine 159 </SITE> . Manual Tagging: Solid phase sequencing revealed phosphorylation at serines <SITE> Ser 122 </SITE>, <SITE> Ser 150 </SITE>, <SITE> Ser 212 </SITE>, <SITE> Ser 220 </SITE>, <SITE> Ser 234 </SITE>, and <SITE> Ser 315 </SITE> and <SITE> threonine 159 </SITE> . (iv) to confirm the event relationship between the related entities using a set of annotation rules. We created a set of annotation rules using the available pattern templates [15, 32], experience gained during manual annotation and experts’ knowledge for identifying sentences with positive event relationship. The entity and relationship annotations were created in parallel, partially based on the specialized synonyms dictionary for annotating entities and a set of available pattern templates for identifying positive event relationship (Table 6). The annotation of hPP corpus consumed 10 months including time spent on related software modules developments and the design of annotation scheme.

PK-phosphorylation keyword; P-substrate; K-protein kinase; S-phosphorylation site Table 6: Pattern templates and annotation rules for event relationship.

Corpus Dataset

The two popular protein phosphorylation databases namely Phospho.ELM [18] and PhosphoSitePlus [20] are utilized for collecting PubMed abstract IDs related to human protein phosphorylation. Phospho.ELM contains 2,530 PubMed IDs with 5,374 proteins and 37,144 phosphorylation sites (27,421 serine, 6,256 threonine and 3,467 tyrosine residues) related to human protein phosphorylation information. PhosphoSitePlus consists of 4,902 PubMed IDs with 1,725 proteins and 4,820 phosphorylation sites (2,826 serine, 963 threonine and 1,031 tyrosine residues) related to human protein phosphorylation information. The statistics of human related phosphorylation data available in both databases is given in Table 7.

The combined list of 6,232 PubMed IDs (without duplicate entries) from both databases is ranked with the likelihood function of GeneTag [23]. We randomly selected 1,000 abstract IDs from top 3,000 entries for downloading the abstracts from PubMed database (http://www.ncbi.nlm. nih.gov/pubmed) for corpus creation. The random selection is executed for 100 times to select the most frequently sampled abstracts. The current version of hPP corpus is a sentence-based corpus containing phosphorylation related sentences from the selected list of 1,000 MEDLINE abstracts.

Corpus Annotation

The main objective of hPP corpus is to provide a gold- standard data for developing and evaluating text mining systems to extract the phosphorylation event information existing between substrate, kinase and phosphorylation site. This focus influences the entity annotation to tag the entities which are relevant to the phosphorylation information Tables 8&9. For instance, proteins relevant to phosphorylation event are tagged specifically as substrate or kinase. Additionally, the corpus tags the non-relevant protein entities to provide related information such as PPI (Figure 3).

The event annotation captures a stated relationship between phosphorylation related biomedical entities identified during entity annotation. The entity and event annotation together explicitly reflect the phosphorylation information stated in a sentence in a more structured way. The entity mentions in an event annotation vary diversely across the corpus sentences. Figure 3 presents a complete event annotation where all three entities (substrate, kinase and site) together provide a phosphorylation event. However, the corpus includes a wide range of various combinations of entities representing phosphorylation event. For instance, the kinase responsible for phosphorylation event is not available in Example 4 and the information on phosphorylated site alone is provided in Example 5 without any substrate and kinase information. Likewise, the phosphorylation related keyword is absent and the equivalent keyword indicating the phosphorylation event is ‘substrate’ as shown in Example 6.

Example 6: PMID: 2843348 <SUBSTRATE> Calmodulin </SUBSTRATE > is a substrate for the <KINASE> insulin receptor kinase </KINASE>.

Example 7: PMID: 15849194 Tagged sentence: <KINASE> p21 -activated protein kinases </KINASE> (<KINASE> Paks </KINASE>) are serine/threonine protein kinases that phosphorylate <KINASE> Raf-1 </KINASE> at <SITE> Ser-338 </SITE> and <SITE> Ser-339 </SITE>.

A very closely related biological process appears in combination with protein phosphorylation is protein- protein interaction. We incorporated the annotation on protein-protein interaction information together with phosphorylation event annotation to explore the correlation between protein-protein interaction and phosphorylation using information retrieval systems (Figure 3). On the other hand, gold standard corpora (AIMED and HPRD50) available to evaluate protein-protein interaction extraction systems contain very few annotations on protein phosphorylation.

We considered the synonym and abbreviation related to a protein name are annotated as equal. For instance, the kinase names, casein kinase 1 epsilon and CK1 epsilon are annotated as same entry (Figure 3). Identification of negative event information is an important key to avoid the extraction of false phosphorylation information by a text mining system. We annotated the sentences with negative phosphorylation information within tags <NEGATIVE PHOSPHORYLATION> and </NEGATIVE> as shown in Example 8.

Example 8: PMID: 11855836 <NEGATIVE PHOSPHORYLATION> Our results suggest that <KINASE> Akt </KINASE> is a negative regulator of <SUBSTRATE> FANCA </SUBSTRATE> phosphorylation. </ NEGATIVE>.

Corpus Statistics

The basic statistics of hPP corpus is listed in Table 8. The information on the number of annotated entities and phosphorylation information are listed in Table 9. Among 5,264 annotated protein entities, 2,480 are identified as substrates and 2,075 are identified as kinases. The phosphorylation information annotation identified 2,829 phosphorylation information in 2,381 sentences from 1,000 MEDLINE abstracts. Among these, 576 information include any one phosphorylation related entity, 1,605 include two entities and 648 include three entities. Most of the phosphorylation information 1,605 (56.7%) in hPP corpus are contributed by two entities patterns (substrate and kinase, substrate and site, kinase and site). The number of phosphorylation information by one entity pattern 576 (20.4%) and three entities patterns 648 (22.9%) are almost the same.

Comparison with Available Corpora

A public resource of existing biomedical corpora and benchmarks is maintained at http://www2.informatik. hu-berlin.de/~hakenber/links/benchmarks.html. The collection comprises of 39 corpora which are primarily intended for biomedical NLP. While ten corpora among the collection are available for NER evaluation, GENIA corpus [33] and JNLPBA [11] are specifically available for human genes/proteins. However, the protein annotation in the available corpora is general and does not distinguish protein kinases from other proteins. In phosphorylation event extraction, distinguishing protein kinases from other proteins is mandatory.

PPI relation corpora such as AIMED [12] and HPRD50 [13] consists of a fewer number of annotated sentences related to protein phosphorylation. The corpora available for protein phosphorylation information extraction are the feature evidence resource of iProLink [14], PTM phosphorylation extraction corpus (https://research.bioinformatics.udel. edu/iprolink/corpora.php) and another corpus for protein phosphorylation (http://pir.georgetown.edu/pirwww/ iprolink/ftcorpora.shtml). All the corpora are from PIR. The iProLink corpus was originally developed for evaluating a phosphorylation information extraction tool called RLIMS-P and consists of 59 MEDLINE abstracts related to protein phosphorylation. Later, PTM phosphorylation extraction corpus was released by PIR for evaluating RLIMS-P v.2 and consists of 150 MEDLINE abstracts and 105 full-length articles. In addition to iProLink, PIR has developed five literature corpora of evidence tagging for protein functional features related to post-transcriptional modification of proteins (acetylation, glycosylation, methylation, phosphorylation and hydroxylation). Each corpus consists of two types of dataset, one with abstracts and the other with full-length articles. The corpus for protein phosphorylation consists of 186 abstracts and 76 full-length articles (http://pir.georgetown. edu/pirwww/iprolink/ftcorpora.shtml) with annotations on feature lines having phosphorylation information. BioNLP shared task corpus is an event extraction corpus that includes nine different events and phosphorylation is one of the events [34, 35]. Many text mining systems for extracting phosphorylation event extraction were evaluated using this corpus [36, 37]. However, phosphorylation event extraction and phosphorylation information extraction are not the same [38]. The biological processes such as gene expression, transcription, and localization are events and their extraction from published literature is termed as event extraction. The relationship between the entities or the events themselves is defined as information extraction [39]. hPP corpus is meant for phosphorylation information extraction.

There are several aspects that make hPP corpus unique from the available corpora on NER and protein phosphorylation: (i) The corpus provides a distinct annotation of biomedical named entities (substrate, kinase and phosphorylation site) related to protein phosphorylation. Such annotation is useful to evaluate the system on identifying and distinguishing kinases and substrates, though both are proteins in general. Identification of phosphorylation related entities such as kinase, substrate and site is a prerequisite for extracting phosphorylation information by a text mining system [38]. The existing corpora do not highlight the entities within the sentences conveying phosphorylation information. (ii) The corpus further provides annotations of gene/protein names not involved in the phosphorylation information given. Though many corpora are available to evaluate systems on gene or protein mentions, none is available to distinguish kinases and substrates involved in protein phosphorylation from other proteins. (iii) The current version of the corpus is a dataset of sentences with annotation on phosphorylation information, in addition to NER. (iv) As the name indicates, hPP corpus is specifically developed for evaluating text mining system related to human protein phosphorylation. The two available corpora from PIR are general corpora (not specific to human).

Application of hPP Corpus

We developed a text mining system for extracting protein phosphorylation information from PubMed articles [38]. Our hybrid approach includes NLP (Natural Language Processing) parsing for identifying phosphorylation information patterns and Support Vector Machines (SVM), a machine algorithm for classifying the extracted phosphorylation information as true or false. We used annotations from 300 PubMed abstracts of hPP corpus to train and test our proposed system: 214 sentences from 200 PubMed abstracts as training data and 207 sentences from 100 PubMed abstracts as test data. We also used the existing corpora PLC and iProLink for validating the performance of our system. hPP corpus provided insights on protein phosphorylation patterns specific to human. Such patterns were absent in the existing corpora. Thus, a corpus specific to human protein phosphorylation is highly preferred.

Phosphorylation in Cellular Functions

Oxidative phosphorylation is the fourth step in cellular respiration that produces most of the energy in the form of adenosine triphosphate [40]. Phosphorylation is an important regulatory mechanism in cell. It regulates protein functions and cell signaling by causing conformational changes in substrate protein. The protein kinase catalyze the conformational changes in substrate protein and induces cellular transduction signaling. The hyperactivity, malfunction, or overexpression of cellular transduction signaling is found in several diseases including cancer [41]. Phosphorylation is necessary for cells to respond to stress or external stimuli that leads to cell death [42]. Thus, phosphorylation is important for various cellular activities and hPP corpus is useful for evaluating text mining systems meant for the extraction of human protein phosphorylation information from PubMed abstracts. Conclusion We have developed hPP corpus with linguistically rich entity and event annotations to facilitate text mining researches related to human protein phosphorylation. To our best knowledge this is the first and foremost annotated corpus available for human protein phosphorylation. The entity annotation includes three biomedical entities namely substrate protein, protein kinase and phosphorylation site related to protein phosphorylation. The event annotation provides the phosphorylation relationship between the annotated entities. The corpus is freely available for download and usage. We believe that hPP corpus will be used by many biomedical researchers to explore literature information on human protein phosphorylation for understanding cellular functions.