SheddomeDB 2023: A Revision of an Ectodomain Shedding Database Based on a Comprehensive Literature Review and Online Resources

Abstract

Ectodomain shedding of membrane proteins is a proteolytic event involved in several biological phenomena, including inflammation, development, diseases, and cancer progression. Though ectodomain shedding is a post-translational modification that plays an important role in cellular regulation, this biological phenomenon is seriously underannotated in public protein databases. Given the importance of the shedding events, we conducted a comprehensive literature review for membrane protein shedding and constructed the database, SheddomeDB in 2017. In response to user feedback, novel shedding findings, more associated biomedical events, and the advance in web technology, we revised SheddomeDB to a new version, SheddomeDB 2023. The revised SheddomeDB 2023 includes 481 protein entries across seven species; all the content was manually verified and curated. The content of SheddomeDB 2023 mainly came from a comprehensive literature survey by our newly developed semiautomated screening tool. We also integrated verified and updated cleavage and secretome information from other databases into the revision. In addition, SheddomeDB 2023 features a graphical presentation of cleavage information and a user-friendly interface for searching and browsing entries in the database. This revised comprehensive database of ectodomain shedding is expected to benefit biomedical researchers across different disciplines.

Introduction

Proteolytic ectodomain shedding, or briefly, shedding, is a post-translational modification in which proteolytic cleavage takes place to release the soluble extracellular fragment of membrane-bound proteins.¹ Only a portion of membrane-bound proteins may undergo ectodomain shedding; this process is regulated by various signaling pathways.² Shedding affects various types of membrane proteins, such as cell adhesion molecules, ion channels, and receptors of signal transduction. Proteolysis of membrane proteins was initially thought to regulate their membrane abundance. However, this cellular process has been shown to regulate several biological phenomena, including development,³ inflammation,⁴ cell proliferation and migration,⁵ diseases, and cancers. In addition to regulating intracellular signaling, shedding also plays a role in intercellular signaling at the synapse in an autocrine or paracrine fashion.⁶ If a proteolytic shedding event is associated with cancer or disease, the released fragment in the body fluid might be developed as a noninvasive biomarker. In a list of 26 FDA-approved cancer biomarkers,⁷ 12 of their gene products are proved to be shed membrane proteins (Table 1).

Table 1. Gene Products of FDA-Approved Cancer Biomarkers That Undergo Ectodomain Shedding^a.

gene product	cancer
EGFR	colorectal cancer;
	non-small cell lung cancer
FGFR2	cholangiocarcinoma
FGFR3	urothelial cancer
ERBB2 (HER2)	breast cancer;
	gastric and gastroesophageal cancer
c-KIT	gastrointestinal stromal tumors
KIT	aggressive systemic mastocytosis
MET	non-small cell lung cancer
NTRK1	solid tumors with NTRK gene fusion
NTRK2	solid tumors with NTRK gene fusion
NTRK3	solid tumors with NTRK gene fusion
PDGFRB	myelodysplastic syndrome/myeloproliferative disease
PD-L1	breast cancer cervical cancer;
	esophageal squamous cell carcinoma;
	gastric or gastroesophageal junction adenocarcinoma;
	head and neck squamous cell carcinoma;
	non-small cell lung cancer;
	triple-negative Breast Carcinoma;
	urothelial carcinoma

^aThis table is a partial list from ref (7).

The ectodomain shedding events, however, are seriously underannotated in protein databases. There is one term in Gene Ontology⁸ for the concept of ectodomain shedding, GO:0006509, which stands for “membrane protein ectodomain proteolysis.” Up to October 4, 2022, only 71 mammalian membrane proteins (23 human, 26 mouse, 16 rat, and six bovine proteins) are annotated with this GO term in the UniProt⁹ database. So far, 481 protein entries have been known to undergo ectodomain shedding, according to our survey. The annotations for ectodomain shedding in UniProt only account for less than 15% of the literature evidence.

In addition to UniProt, the association between ectodomain shedding and certain physiological or biological phenomena can be found in published research. However, this evidence is distributed across the vast amount of biomedical literature and is challenging to retrieve. Furthermore, membrane proteins that were proved to undergo proteolytic shedding can be found in secretome databases, such as HCSD,¹⁰ Sys-BodyFluid,¹¹ and SPD.¹² Since membrane protein shedding can be treated as an unconventional pathway of protein secretion,¹³ if a shedding event accompanies a particular disease, the released fragment in body fluids might be detected with the secretome analyses of the disease. Unfortunately, ectodomain shedding is not clearly annotated in secretome databases, making it difficult to determine if secretion occurs through alternative pathways, such as extracellular vesicles or exosomes. According to all those above-mentioned, there is a need for a repository of ectodomain shedding.

Given the importance of ectodomain shedding in the biomedical research community, we constructed and announced SheddomeDB in 2017.¹⁰⁰ We conducted a comprehensive literature review for research articles relevant to ectodomain shedding. In addition to manual review, we also integrated into SheddomeDB 2017 the shedding information from proteolytic cleavage databases (HPRD,¹⁴ PMAP,¹⁵ and MEROPS¹⁶) and secretome databases (HCSD,¹⁰ LOCATE,¹⁷ Sys-BodyFluid,¹¹ and SPD¹²). To the best of our knowledge, SheddomeDB 2017 is the most comprehensive online resource for ectodomain shedding. In response to user feedback, novel shedding findings, studies of large-scale membrane protein cleavage site identification,^18,19 more associated biomedical events, and the advance in web technology, we revised the SheddomeDB 2017 to a new version.

In this study, we introduce SheddomeDB 2023, which integrates machine learning into our literature screening workflow. Unlike the previous version of SheddomeDB, which relied exclusively on manual review for database curation, the current approach seeks to decrease the burden of literature screening while enhancing the accuracy of shedding event identification and updating existing entries. Additionally, we have improved the web interface to augment data accessibility and usability.

Materials and Methods

Overall workflow for the construction of SheddomeDB 2023 is shown in Figure 1. Similar to SheddomeDB 2017, we first reviewed several online databases to retrieve verified and updated information on membrane protein secretion and proteolytic cleavage. Second, we performed an initial literature search to select articles relevant to ectodomain shedding as thoroughly as possible. The abstracts of the retrieved candidate articles were subjected to a document classifier to evaluate the degree to which the candidate articles were relevant to ectodomain shedding. To accelerate the process of manually reviewing the candidate articles, we prioritized the candidate articles based on the evaluation scores. Finally, the shedding information was extracted from the manually verified relevant articles and deposited into SheddomeDB 2023.

Figure 1.

Workflow to update SheddomeDB. We first check the update of information about secretome and proteolytic cleavage from online databases. Most online databases are obsolete, but we keep their records in SheddomeDB 2023 if their PMID indices remain available. The motifs of cleavage sites are mapped to the latest protein sequence provided by UniProt. The second part is to retrieve new shedding events. In this part, we first used the review result of the initial search of SheddomeDB 2017 to train a document classifier to predict the degree to which a query abstract is relevant to ectodomain shedding. Second, we applied the same search strategy to perform the initial search of SheddomeDB 2023 to obtain candidate abstracts. The candidate abstracts were subjected to the document classifier to get their relevance scores. Based on the scores, we prioritized the candidate abstracts to produce a ranking for the downstream manual review. Finally, the shedding information was extracted from the manually verified relevant articles and deposited into SheddomeDB 2023.

Online Databases

Like SheddomeDB 2017, we collected secretion and proteolytic cleavage information from other online databases. For each secreted membrane protein recorded in the secretome databases HCSD, LOCATE, SPD, or Sys-BodyFluid, we collected its cell type or the body fluid in which the protein was released. For each shed membrane protein recorded in the proteolytic cleavage databases HPRD, MEROPS, or PMAP-Substrate, we collected its cleavage site sequence, involved protease, and the experiment analysis for the detection of its proteolytic cleavage. Currently, only MEROPS is available and maintained for all the above-mentioned databases. Though databases other than MEROPS are obsolete, the PubMed indices extracted from these databases are kept in SheddomeDB 2023 if the indexed articles can be accessed. Following the release of SheddomeDB 2017, two large-scale studies were published, utilizing mass spectrometry to identify membrane protein cleavage sites. Amy M. Weeks et al. employed subtiligase N-terminomics to reveal novel N-termini cleavage events of cell surface proteins, creating the ASCENT database.¹⁸ Kazuya Tsumagari et al. used TMT-labeling N-terminomics and C-terminomics to identify membrane cleavage events across 10 cell lines.¹⁹ To enhance our database, we integrated cleavage site data from these studies into our existing membrane entries, referencing the studies in the “PMID” column. For ASCENT, we noted it in the “Database” column of the Proteolytic Cleavage table, and for the other study, we indicated it as “PAPER” in the same column.

Search Strategy for Literature Evidence

In SheddomeDB 2017, shedding OR proteolytic OR cleavage OR protease OR released OR “soluble form” was defined as the search strategy for the PubMed database. The search results subsequently underwent manual review. To check if there were new findings of ectodomain shedding after the announcement of SheddomeDB 2017, we applied the same strategy to search relevant articles indexed by PubMed between September 3, 2016, and January 1, 2022. We retrieved 376,150 abstracts from PubMed for the downstream analysis.

Document Classification and Literature Prioritization

To accelerate the manual review process, we utilized the review result of the initial search of SheddomeDB 2017 for training a document classifier. After a manual screening, we used 428 relevant and 677 irrelevant article abstracts for our classifier training. We combined the text of the title and abstract for each article in the downstream analysis.

We preprocessed the text to remove redundant information and noise in the data. We transformed words to lowercase, removed punctuations and stop words such as “a,” “the,” “this,” and “that,” and reduced words to their original forms using the Porter Stemmer algorithm. After word stemming, we tokenized the text by whitespaces to separate individual word stems. We used the preprocessing tools implemented in the Python gensim package to conduct the aforementioned transformation.

During the feature selection process, we constructed a dictionary that contained the 500 most frequently used word stems in all transformed abstracts based on their document frequency. We believe this allows for better generalization to other abstracts that need to be screened. We subsequently applied one-hot encoding to represent each abstract as a binary vector of length 500 to indicate which word stems in the dictionary appeared in the abstract.

We selected a random forest classifier as the classification algorithm because it can handle the class imbalance problem and learn the nonlinear relationship in our data. We used the default settings in the implementation of the Python scikit-learn package, with 100 trees in the forest, Gini impurity as the criterion for splitting, and nonspecified maximum depth. Stratified 10-fold cross-validation was performed to evaluate the reliability and validity of our classifier. We achieved a recall of 95.6 ± 2.6%, precision of 95.8 ± 3.1%, F1-score of 95.7 ± 2.4%, specificity of 97.3 ± 2.1%, and overall accuracy of 96.7 ± 1.9% on the validation data. After ensuring the stability of the results, we trained a classifier with all 1105 binary vectors and their relevance labels. The final classifier was used to infer the relevance score for 376,150 abstracts to be screened.

Rather than simply categorizing all abstracts into relevant or irrelevant articles, the classifier generated a relevance score for each prediction, indicating the confidence that the query abstract is relevant to ectodomain shedding. Based on the relevance scores, we prioritized the 376,150 abstracts and produced a ranking. Our manual review proceeded in the ranking order and stopped once 100 consecutive reviewed abstracts were all irrelevant to ectodomain shedding. After confirming the relevance of the abstracts to ectodomain shedding, we manually downloaded and reviewed the full-text articles. To include an article for creating new entries, it must meet the following criteria: (1) Confirm molecular weight changes before and after cleavage. (2) Establish a link between the molecular weight change and sheddase activity. (3) Explore the biological regulation or significance of the cleavage event. Once a study satisfied these requirements, we manually extracted the shedding information and incorporated it into SheddomeDB 2023.

Web Programming and Data Handling

SheddomeDB 2023 was built as an online database. We applied the model-view-template design pattern to develop our website with the Python Django framework. To make our website adapt to various sizes of display devices, our website templates were created using Bootstrap to feature responsive web design (RWD). HTTPS is supported with the SSL certificate from OpenSSL. SQLite handles the data storage of tabular data. A fuzzy search for protein names is done with the Python fuzzywuzzy package. Cleavage information and domain topology are presented graphically with lollipop plots using Python seaborn and matplotlib packages.

Results

Entries of Ectodomain Shedding

The content of SheddomeDB 2023 was primarily summarized from 556 articles studying shedding events. We finally have 481 protein entries across seven species (Table 2). Each entry of SheddomeDB 2023 annotates a protein substrate that comes from a specific organism; we used the UniProt ID of the substrate to index the entry uniquely. Please note that entries of a protein substrate from different organisms are considered different. We also show the difference in the number of entries between SheddomeDB 2017 and SheddomeDB 2023 in Table 2.

Table 2. Difference in the Number of Entries and PubMed Unique Identifiers (PMIDs) across Seven Species between SheddomeDB 2017 and SheddomeDB 2023.

organism	entries (% of total)		PMIDs (% of total)
	2017	2023	2023
Homo sapiens (human)	325 (70.2%)	336 (69.9%)	386 (69.4%)
Mus musculus (mouse)	112 (24.2%)	117 (24.3%)	136 (24.5%)
Rattus norvegicus (rat)	21 (4.5%)	22 (4.6%)	27 (4.9%)
Gallus gallus (chicken)	3 (0.6%)	3 (0.6%)	4 (0.7%)
Bos taurus (bovine)	1 (0.2%)	1 (0.2%)	1 (0.2%)
Oryctolagus cuniculus (rabbit)	1 (0.2%)	1 (0.2%)	1 (0.2%)
Drosophila melanogaster (fruit fly)	0 (0.2%)	1 (0.2%)	1 (0.2%)
total	463 (100%)	481 (100%)	556 (100%)

Information in each protein entry is categorized into the following five sections (Figure 2): UniProt Summary, Biomedical Literature Report of Shedding, Topological Domains equipped with a Cleavage Site Diagram, Secretome Detection, and the Proteolytic Cleavage. The content of UniProt Summary comes from UniProt; this section provides information on protein identity. Section Biomedical Literature Report of Shedding lists the PMID and summary of the research articles that offer the shedding evidence. A membrane protein may have different ways of shedding, depending on the cellular conditions and regulations. Thus, an entry may contain multiple PMIDs, each of which studies a specific shedding event. The information in the Topological Domains section is also retrieved from UniProt; this section provides an overview of the topological structure. In addition, this section is equipped with a lollipop plot that visually displays information about the cleavage site (Figure 3). The information in the Secretion Detection section is collected from online secretome databases. This section records the location where the query protein is released and literature references. According to the secretion information, the query protein exhibits its potential role for physiological and pathological biomarkers. The information in the section Proteolytic Cleavage comes from the online proteolytic cleavage databases. The proteolytic shedding information is experimentally validated and annotated in different proteolytic databases. In addition to the tabular representation, the information in this section is also used to make the lollipop visualization of cleavage sites in the section Cleavage Site Diagram.

Figure 2.

Features in SheddomeDB 2023. Each entry in SheddomeDB 2023 has the following five categories of information: UniProt Summary, Literature Summary, Topological domains, Proteolytic Cleavage, and Secretome Detection. The snapshot is the partial result of entry P12830.

Figure 3.

Lollipop plot for the protein cleavage sites. The horizontal stem (black) of the lollipop plot is the full length of the protein sequence. Cleavage sites are shown with different endpoints of stems according to their evidence source. Topological domains such as transmembrane domains (blue) and extracellular domains (green) are also shown in colored boxes.

Challenges in the Literature Survey of Shedding Events

An initial step to constructing our database is to perform a comprehensive literature survey. Nevertheless, this task is challenging because we have no consensus way to describe ectodomain shedding in the literature nor a standard combination of keywords for searching for shedding events. A common approach to this problem is preparing a set of keyword combinations to search candidate articles that are relevant to ectodomain shedding as thoroughly as possible, followed by a labor-intensive and time-consuming manual review to identify articles pertinent to shedding events; some of the identified articles even do not have the words “ectodomain” and “shedding” in their abstracts. We applied this approach to build SheddomeDB 2017, and this step took years.

To make the manual review more efficient, we developed a machine-learning model to evaluate the degree to which an article is relevant to ectodomain shedding. In building SheddomeDB 2023, based on the evaluation score, we suggested an order and an early-stop mechanism for reviewing the candidate articles. This significantly reduced the review time from years to months. Over 376,000 candidate articles were retrieved during searching for novel substrates of proteolytic shedding. We identified 17 novel entries (Table 3) across different model organisms. For example, one of the targets for cancer immunotherapy, PD-L1, was first reported to be a shed substrate of breast cancer cell.²⁰ In addition, we also collected the updated events for original entries.

Table 3. Novel Entries Identified with the Assistance of Machine Learning.

entry	organism	gene names (primary)
Q9NZQ7	Homo sapiens (human)	CD274
P14209	Homo sapiens (human)	CD99
P54760	Homo sapiens (human)	EPHB4
Q8NBJ4	Homo sapiens (human)	GOLM1
P05981	Homo sapiens (human)	HPN
Q6UXK5	Homo sapiens (human)	LRRN1
P23515	Homo sapiens (human)	OMG
Q13018	Homo sapiens (human)	PLA2R1
Q9NT68	Homo sapiens (human)	TENM2
P43489	Homo sapiens (human)	TNFRSF4
P12804	Mus musculus (mouse)	Fgl2
P51655	Mus musculus (mouse)	Gpc4
Q9JHJ8	Mus musculus (mouse)	Icoslg
Q9QUM0	Mus musculus (mouse)	Itga2b
Q8BYM5	Mus musculus (mouse)	Nlgn3
D3ZGQ0	Rattus norvegicus (rat)	Il17rd
Q9VIC6	Drosophila melanogaster (fruit fly)	Nlg3

Redesign the Web Interface

In response to user feedback and to provide users with a better experience, we redesigned the interface of our SheddomeDB 2023. We applied the Bootstrap framework to support responsive web design to display our web pages appropriately within devices of different sizes. We also provided a new page for queries. Queries can be done by gene names, protein names, or UniProt IDs, and fuzzy search was fully supported. In addition, we provided browse mode, in which all 481 entries are listed for browse and access. Moreover, we used lollipop plots to show the cleavage sites of membrane proteins instead of a long text sequence of peptides used in SheddomeDB 2017 for better visualization (Figure 3). Lollipop plots are commonly used to represent biological sequences visually, which can give a quick glance at the location and fragment of interest.

Discussion

In the current work, we reconstructed the SheddomeDB database and updated the content with the help of machine learning. This revised online resource is supposed to be more user-friendly than it was and benefits researchers across different disciplines in the biomedical research communities.

In addition to potential disease biomarkers, ectodomain shedding might be the treatment target. In recent years, COVID-19 has been a pandemic overwhelming the entire world. Coronavirus enters the host cell through the ACE2 receptor of the cell membrane.²¹ Hence, if the activity of ACE2 is diminished on the cell membrane, it might vastly reduce the possibility of viral entry. ACE2 undergoes ectodomain shedding;²² researchers might develop a treatment strategy based on this molecular mechanism.^23,24

Clarification for Our Inclusion of Secretome Data

We need to clarify our intention to include secretome data from other databases. It is important to note that detecting a membrane protein in a secretome analysis does not necessarily indicate its release via proteolytic ectodomain shedding. Membrane proteins can also be secreted through extracellular vesicles and exosomes and subsequently released into extracellular environments such as body fluids. As explained in the “Materials and Methods” section, we did not include entries solely based on secretome data. Our objective in including secretome data was to gain insights into potential sources of noninvasive biomarkers identified in body fluids or cell secretions. Our previous study did not explain this, and we hereby declare it.

Removal of ShedP and Functional Categories

SheddomeDB 2017 features a tool, ShedP, to predict the probability of a membrane protein being shed. However, we do not provide ShedP in SheddomeDB 2023 because Zhongbo Cao et al. released a prediction model, DeepSMP,²⁵ which uses a bidirectional LSTM-attention model to predict protein shedding and outperform ShedP, so we currently suggest researchers use DeepSMP if they intend to predict protein shedding. As more shedding events are discovered, we might work on developing a new version of ShedP and integrate it into our future release.

In SheddomeDB 2017, we defined 14 functional categories based on the annotation of entries and provided a pie chart to show the relative proportion of these categories. We also offered a “Browse by Categories” mode for browsing protein entries based on their functional categories. Given that one entry might be moonlighted and belong to multiple functional categories, which is prone to confusion, we thought that “functional categories” may not be a clear concept, so we decided to remove the pie chart and the “Browse by Categories” mode in SheddomeDB 2023.

Literature Review Based on Machine Learning

The literature review is the most challenging process in constructing SheddomeDB 2023. Though various text mining algorithms and tools are available online, none fulfill our requirement for building SheddomeDB. Thus, we developed an in-house literature review tool using machine learning techniques. A limitation of our tool is that only abstracts are considered; for those articles that only mention protein substrates in the text but not the abstract, our tools will fail to identify this substrate. For example, when reviewing a previous study,²⁶ our tool failed to identify Neuroligin-1 and Neuroligin-2 as shed membrane proteins because their shedding information is described in the main text but not mentioned in the abstract. Fortunately, the shedding information of the two proteins is recorded by another review article that was regarded as relevant to ectodomain shedding by our tool.⁶ Our future work might focus on filtering full texts available in PMC collection to eliminate the potential risk of missing targets.

Our methodology comprises three stages: converting input text into a 500-dimensional numeric vector, predicting its association with an ectodomain shedding event, and ranking input texts based on relevance. It is important to note that the word count of the input text solely impacts the numeric vector conversion time, not the other two stages. The average word count for abstracts (approximately 140 words) and full texts (approximately 5500 words) in our sample implies that processing full-text articles takes roughly 40 times longer than processing abstracts. Utilizing a MacBook with 16 GB memory and an M1 chip, 376,150 abstracts were processed in 1 min and 34 s; thus, processing the same number of full-text documents would take approximately 60 min. We do not directly use full-text articles for analysis due to challenges in automatically downloading large amounts of copyrighted documents and the laborious process of handling PDF files. Upon confirming the abstract’s relevance, we manually download the full text to analyze the relevant literature comprehensively. We believe that our current approach strikes a balance between accuracy and efficiency.

Glossary

Abbreviations

PD-L1

programmed cell death 1 ligand 1

Data Availability Statement

Database URL: https://bal.lab.nycu.edu.tw/sheddomedb/

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jproteome.3c00001.

• Table S1: The list of the identified 481 shed membrane protein entries (PDF)

Author Contributions

W.-Y.H. conceived the study, conducted the data analysis, constructed the website, and wrote the paper. K.-P.W. conceived the study and wrote the paper. All authors read and approved the manuscript. The manuscript was written through contributions of all authors. All authors have given approval to the final version of the manuscript.

This work and the publication cost of this manuscript was financially supported by the Ministry of Science and Technology, Taiwan under the contract number of MOST 110-2314-B-A49A-506-MY3.

The authors declare no competing financial interest.