LipiDisease

1 Introduction

Lipids are a diverse group of organic compounds having characteristic solubilities in nonpolar solvents. They exhibit essential roles in cellular assembly, signaling and energy storage. Over the past few years, there has been an increasing interest in untargeted lipidomics, largely driven by the technological advancements in lipid profiling and functional analysis improving, thus, our understanding of lipid biology. The lipidomic analysis encompasses lipid quantification and identification, allowing both insights into changes into the structural heterogeneity and into individual lipid levels in a biospecimen. It is evident that the lipidome is altered during disease development and comparative lipidomics could identify lipid biomarkers for disease prognosis or even intervention (Ghosh and Nishtala, 2017^; Lydic and Goo, 2018). For this reason, it is important to develop methods to identify lipids that could be predictive of disease type, progression and risk using the study of lipidomics data.

One of the ways to identify disease-relevant biomarkers is to associate them with diseases using pre-existing literature data. Enrichment analysis is a computational procedure to assist this task in an automated and exhaustive manner. Enrichment analysis is performed to associate a set of biomolecules, for example, derived from a high-throughput experiment, with a biological trait. The analysis compares the association of biomolecules with a background set to identify significantly associated candidates with biological traits. The background association is derived from biological information and is typically obtained from experimental datasets and information databases. There is a plethora of such resources available for genomics and proteomics data that perform pathway enrichment, disease association and more (Alfoldi and Lindblad-Toh, 2013^; Conesa et al., 2016^; Fontaine and Andrade-Navarro, 2016^; Schmidt et al., 2014). Such databases and tools are severely lacking in the case of lipids because of the stronger focus of the omics field on DNA, RNA and protein biology (Stephenson et al., 2017), and consequently, the enrichment analysis of lipids is limited to their categorization and pathway identification (Acevedo et al., 2018^; Clair et al., 2019^; Martin et al., 2013^; Molenaar et al., 2019), while their functional enrichment is lacking. While much information on metabolite association with diseases can be accessed and retrieved from databases such as MarkerDB (Wishart et al., 2021) and Human Metabolome Database (Wishart et al., 2007), a dedicated database for lipids is missing. LipidPedia is a resource specialized in associating lipids with biomedical information (Kuo and Tseng, 2018). It utilizes full-text mining focused on individual lipids, but lacks statistical filtering and does not allow manual inspection of the underlying basis for disease associations. Furthermore, currently, there is no way to analyze a set of lipids and consider their cumulative association, which is more relevant in human diseases.

The LipiDisease web server aims to address this by providing a web app performing disease enrichment analysis using a set of lipids and ranking them for prioritizing the diseases. It utilizes the PubMed database listing more than 26 million biomedical records with their manual associations with chemicals and diseases. We expect that the web server will contribute to a better understanding of lipid-disease associations and will inspire the development of many such tools expanding our knowledge about lipids beyond membranes and signaling.

2 Implementation

LipiDisease is built around the biomedical literature from the PubMed database. With a focus on avoiding false positive, instead of full-text mining, it utilizes the manual annotations of the PubMed records using the Medical Subject Headings thesaurus MeSH. Disease terms were extracted from the branch ‘C’ of the MeSH. The list of lipids and corresponding PubChem Compound ID numbers (CIDs) were obtained from Lipid Maps (Fahy et al., 2009). Articles corresponding to the lipid PubChem CIDs were obtained from PubMed using NCBI’s Entrez Programming Utilities. The entire data was locally stored in a MySQL database (details in Supplementary Information).

The interactive web app was built using the Shiny package in R (R Core Team, 2020). JavaScript functionalities were implemented using the shinyjs package. The required input for the web server is either a set of PubChem CIDs for lipids or MeSH Unique IDs for diseases depending on the analysis. There are the following four types of analysis:

• Lipid-set enrichment;

• Lipids to Diseases;

• Diseases to Lipids;

• Lipid-set enrichment (with fold changes).

The Lipid-set enrichment option considers a list of lipids. These lipids are considered collectively for disease enrichment analysis. Lipids to Diseases and Diseases to Lipids consider individual lipids and diseases, respectively, and identify associations using individual entries; using either of these two analyses is close to browsing our underlying database of associations. Lipid-set enrichment (with fold changes) is similar to Lipid-set enrichment with the additional consideration of lipid-level statistics (e.g. fold changes for the lipid expression between two biological conditions). To understand the functionality of the web server, users can use the test cases by clicking on the ‘example’ link next to every analysis type.

An analysis is done in seconds using the pre-indexed data. The output is represented in table and plot form, which can be downloaded locally for further exploration. The output table can be downloaded in TSV format (entries separated by TAB) and can be accessed using text or spreadsheet programs. The table provides hyperlinks to corresponding lipid PubChem CIDs, disease MeSH terms and PubMed records.

The predicted enrichment is derived from the manual annotations of the PubMed records with the lipid PubChem CIDs using a computational procedure employed before for literature mining of gene to disease associations (see Supplementary Information for details; Fontaine and Andrade-Navarro, 2016). In short, lipid enrichment is performed by identifying the over-representation of PubMed articles using one-tailed Fisher’s exact test. The results are then ranked according to false discovery rate calculated by Benjamini and Hochberg method (Benjamini and Hochberg, 1995). The type of association (for example, positive or negative) cannot be derived from this. However, the links from the results to the underlying data that was used to derive the associations, particularly the connections to the relevant literature through PubMed records, facilitate post-analysis of the results.

3 Conclusion

An online web server, LipiDisease, was constructed to predict significant associations between lipids and diseases using literature mining. LipiDisease is the first tool performing disease enrichment using a set of lipids and providing a ranked list of statistically significant associations. The tool provides a user-friendly web interface and outputs results in tabular as well as graphical form, with links to the underlying data used in the literature mining procedure, making it easier to generate hypotheses about the involvement of lipids in disease.