SUBA4: the interactive data analysis centre for Arabidopsis subcellular protein locations

Database structure and interface

SUBA4 data are stored in a MySQL relational database (MySQL 5.6), operating on a UNIX-based system (RHEL6) using Apache 2.2.15. Extra relational semantics were added using the Python 3.5 and SQLAlchemy library. The SUBA4 search portal is a web-browser based GUI (Graphical User Interface) written in dynamic HTML using Asynchronous JavaScript + XML (AJAX) to interact with the SUBA4 server.

The SUBA4 search portal can be used without prior knowledge of SQL and allows the construction of complex queries by selecting parameters in a point and click manner from the tabs in the search menu. The data can be accessed via http://SUBA.live and is compatible with most current browsers including Windows Explorer, Safari, Firefox and Chrome. The search window contains five query tabs that separate filter options into categories for ease of use. Each tab contains pre-formulated queries in full text with pull down menus and simple text boxes. The query window allows AND, OR and bracketing to create complex Boolean queries.

Query results can be accessed via the results tabs in a tabular format with a default row format containing protein identifiers and summaries of localization data. Each column contains a pull down menu for customizing the information shown. The results from a query may be downloaded as a csv tabular file for further analysis. Each protein hit in the results view is hyperlinked to a factsheet that provides further information including sequence, hydropathy profile, protein properties, annotations, coordinates of gene, coexpression or PPI partners and localization study description. Protein annotations are displayed as hyperlinks that provide rapid access to related resources on the same protein in UniProt, AmiGO, EMBL, Ensembl Plants/Gramene, KEGG, MSU and others.

Arabidopsis proteins and annotations

SUBA4 data adds to the SUBA3 datasets and are based on the non-redundant Arabidopsis protein set obtained from The Arabidopsis Information Resources (TAIR release 10), and the Arabidopsis mitochondrial (117) and chloroplast (87) open reading frame (ORF) sets obtained from GenBank Y08501 and AP000423, respectively. SUBA4 contains 35 388 distinct protein sequences. The primary attributes of molecular weight, average hydropathicity and isoelectric point were described previously (6) and functional assignments (4) for each Arabidopsis locus were retained and updated as required for TAIR10. An additional 1,150,204 annotations (Supplementary Table S1) were integrated into SUBA4 for each AGI as a searchable format. The annotations were obtained from the data repository of Ensemble Plants release 31 (ftp://ftp.ensemblgenomes.org/pub/release-31/), Uniprot (http://www.uniprot.org/) and TAIR (https://www.arabidopsis.org/download/index.jsp). Individual databases were directly linked to the display on the SUBA4 output.

Experimental localization data sets

Experimental subcellular localizations of proteins were sourced and collated using the methods previously described (6). In short, localizations by MS were obtained by searching PubMed (http://www.ncbi.nlm.nih.gov/pubmed) for the term ‘Arabidopsis AND (proteome OR proteomics) and by florescent protein tagging for the term ‘Arabidopsis AND (CFP OR GFP OR RFP OR YFP). AGIs, aliases and subcellular locations were extracted from text or supplementary data. In addition to the previously described 11 subcellular locations, 37 location subcategories were added, increasing the subcellular and suborganellar resolution of the data as a whole (Table 1).

PPI subcellular localizations were obtained by searching PubMed (http://www.ncbi.nlm.nih.gov/pubmed) for the search term ‘Arabidopsis AND (interact OR interaction OR protein–protein)’. The retrieved literature was manually searched for BiFC and co-localizations. PPI experimentation and subcellular locations were extracted from this literature into the 49 categories (48 locations + "unclear"). SUBA4 contains 122 PPI localization studies, collating 683 pairs observed in 858 localizations covering 29 of location categories. Overall, Arabidopsis PPI datasets of 16 252 protein pairs were obtained from the IntAct database (http://www.ebi.ac.uk/intact/) and 8990 PPI protein pairs were manually curated as previously described (6). In total, 21 495 unique protein pairs were integrated in SUBA comprising 26 327 experimental individual verifications.

Arabidopsis gene coexpression pairs, mutual rank and average correlation coefficient data were sourced from the ATTED-II database (http://atted.jp) as described (7,17). Coexpression between identical AGIs (self) were omitted from the list.

The subcellular targeting predictions for 22 bioinformatic programs were described in SUBA3 (6). The SUBAcon consensus was retrained as previously described (7) using updated experimental data and the updated Arabidopsis SUbcellular REference standard (ASURE).

SUBA4 tools

The BLAST, Coexpression Adjacency Tool (CAT), PPI Adjacency Tool (PAT) and Multiple Marker Abundance Profiling (MMAP) tool were developed to aid the SUBA4 user. The BLAST tool was build using a standard BLASTX method (18). The score shown is −log₂(E) where E = pval × Neff is the P-value times the effective search space size. The score increases with sequence match. Users can enter full or parts of protein sequences, where protein sequences with gaps are treated as separate fragments that are matched against Arabidopsis proteins independently. For the CAT and PAT tools, the coexpression data and PPI data sets were linked to SUBAcon calls. The SUBAcon calls of each AGI partner was joined and categorized into proximity relationships according to their biological interpretation (Supplementary Table S2). The subcellular locations, proximity relationships, mutual rank and average correlation coefficient data were linked to the tool interface as a searchable parameter.

The MMAP tool will be described in detail elsewhere (personal communications, H.T. Parsons et al.). In short, spectral counts collated as described previously (19,20) were sourced from the MASCP gator (http://www.masc-proteomics.org) for each AGI. The spectral counts of all experimentation were normalized using the Normalized Spectral Abundance Factor (NSAF) method (21). Before linking to the SUBAcon output, the NSAF values were normalized and scaled to the total sum to generate Gator scores. Using a marker list of >17 000 high confidence classification proteins, their Gator score and location call, calculations were performed to estimate the organellar protein abundance as a relative fraction of overall expected protein abundance in a cell. The distribution of organellar proteins (SUBAcon) and organellar protein abundance (MMAP) are displayed as easy-to-interpret bar and pie charts.

Expanded datasets and annotations

A categorized search interface for quick access to varied filter options

The integration of new parameters, data sets and annotations allows more query options that link distinct data modalities. In order to rapidly find and assemble user-specific research questions through the SUBA portal, the search interface was upgraded to a tab-category system (Supplementary Figure S1). The tab-category system separates the query building blocks in intuitive sections for subcellular, protein properties, PPI, affiliations and BLAST questions (Figure 3). The subcellular tab shows filter options for experimental MS and FP data or for predictive subcellular locations by the 22 predictors and the SUBAcon classification. The newly integrated suborganellar categories for experimental data can be accessed by clicking on the arrow ‘»’ for each organelle. Any of these can be chosen as a search filter. For ease of use, the cell schematic for the 11 major locations is provided for users to click on the desired organelle. In SUBA4, more than one organelle can be selected at the same time, before adding the filter to the main query. The locations will be connected by an ‘OR’.

The protein properties tab contains all filters regarding physical protein properties, genetic locations as well as all AGIs, TAIR descriptions, and all other searchable annotations (Supplementary Figure S2). The latter allows users to retrieve whole protein families or groups associated with specific GO, PFAM terms and other annotations and assess their subcellular locations.

In order to manage complex queries specifically associated with the PPI data, a PPI tab was generated (Figure 3). Beside the conventional queries identifying existing PPI partners, users can now also filter for methodology and observed PPI subcellular location. The subcellular location search layout for PPI data is the same as for MS and FP. PPI localizations are derived from manually curated bifocal completion and fluorescence resonance energy transport assays in studies and constitute an additional localization data set to existing FP and MS data.

Newly available bibliography metadata from research studies can now be searched through the affiliations tab (Figure 3). It contains a large number of filter options that can be combined to a specific query about the age of data and the research groups responsible. The world map can be used to quickly assess which countries have contributed data sets to SUBA4 (Supplementary Figure S3) and the full detail of the global distribution of the studies can be seen under the Locations tab in the top menu.

The last tab contains a BLAST sequence alignment tool new to SUBA4 (Figure 3). The BLAST tool is linked to the SUBA search function to enable users with sequence fragments from all species to find the closest match in TAIR10 and directly retrieve data from SUBA. The alignment and BLAST score of the submitted sequence to a matching AGI in SUBA are shown in the results view for each hit. This extends the SUBA data user base and aligns with the general global efforts of connecting species data more easily (33). A quick link window to the BLAST tool can also be found directly on the SUBA4 homepage.

The query-building window is connected to all five tabs (Figure 3), allowing the user to choose filters from any tabs to build one query. The new SUBA4 tab system allows for scalability to future parameters and added functionality without the need of interface changes.

A toolbox to view and interrogate large-scale subcellular location data

The toolbox in SUBA4 contains three distinct tools, namely the Multiple Marker Abundance Profiling (MMAP) tool, Coexpression Adjacency Tool (CAT) and PPI Adjacency Tool (PAT). Each tool provides a unique link between the subcellular location consensus (SUBAcon) and protein abundance (MMAP), coexpression (CAT) and PPI (PAT) data, respectively, giving the SUBA4 user straightforward data analysis options. The detailed descriptions of the development of the MMAP, CAT and PAT and an assessment of their uses will be described in detail elsewhere (personal communications, H.T. Parsons et al.). However, a brief introduction to each tool is given here.

The MMAP tool provides an estimate of the proportion of different subcellular protein structures in a user provided list of AGIs by combining information from SUBA and MS experimental observations of proteins in the MASCP gator database. All spectral counts observed in MASCP gator were scaled to estimate a relative abundance score of each protein as a measure of observed spectral counts across all experiments. A marker list of single targeted proteins with high confidence was generated that in combination with the observation abundance score was used to estimate the relative abundance of organellar protein in a given AGI list. Therefore, users can submit custom AGI lists to the MMAP tool and receive the number of distinct proteins per each organelle as well as an estimate of relative protein abundance composition compared to expected subcellular abundances in green (high-light) or other (low-light) tissues (Figure 4).

The CAT tool allows the coexpression of genes encoding specific proteins in a list to be assessed in context of the final subcellular location of their proteins. It makes use of the fact that when two transcripts are coexpressed there is a potential relationship between the proteins they encode. The user can assess a list of AGIs for existing coexpression partners that are both within the list or with only one AGI within the list. The search can be filtered by coexpression strength parameters - rank and correlation coefficient (Figure 5). The tool offers filter options for subcellular locations of ProteinA and coexpressed ProteinB or allows the retrieval of protein partners with the proximity relationship categories: match, adjacent, distant or secretory. This proximity relationship classification was developed using the paired call of the SUBAcon classification of both proteins (Supplementary Table S2). After submitting the list of AGIs, the tool uses the SUBAcon calls, collates them and compares them to the key, the proportions of the proximity relationships in the queried data set is then displayed in the results offering an instant analysis of the coexpression data in a subcellular context (Figure 5).

The PAT tool is built on similar statistical principles and relationship categories to the CAT tool but using experimental PPI data. SUBA4 contains 26 327 unique experimental PPI observations, of which only 799 have been experimentally localized during the PPI experiment. Using SUBAcon, all proteins from the PPI data set can be assigned to a subcellular location, that when paired indicate a proximity relationship of the PPI couple. The PAT tool was developed to enable interpretation of protein-protein relationships in the context of proteins proximity in the cell. The user can identify a number of experimentally verified PPI partners within a list or associated with a protein in a submitted AGI list through the PAT tool. The user can also take advantage of filter options regarding the classified subcellular location of proteinA and proteinB or filter by the proximity relationship of the protein pairs (Supplementary Figure S4).

The CAT and PAT tools will aid the interpretation of typical subsets of AGIs derived from clustering, filtering of transcript data including sets of proteins changing in abundance in particular conditions or genotypes, protein lists from immunoprecipitations, organellar or membrane preparations, lists of proteins in pathways or gene families and many other custom questions posed by the user.

A simplified results view with more customizable options

When submitting a query, the results tab automatically becomes activated and shows the hits in tabular format. Since SUBA3, the default content presentation of the retrieved SUBA data (AGI, protein description, SUBAcon location calls, predictions and experimental data) has been simplified to avoid overloading the viewing field. The user can choose from a number of additional display options including locations predicted, external curation, experimental data, PPI, and bibliography metadata. The experimental data in the SUBA4 results table is summarized per organelle indicating the amount of experimental data available from FP and MS. The last column displays the PPI data including the AGI and now the SUBAcon call for any interacting proteins. The simplification and added flexibility for data display in the results table caters for users with distinct questions. A user can now revisit the query that led to the results by clicking on the ‘What's this Query?’ button. The ‘Download’ button next to it will initiate the download of the query result data table for use of SUBA data for downstream applications. For detailed investigation of individual proteins, the user can click on the AGI in the results view and open the individual AGI factsheet.

A factsheet that visually collates subcellular and annotation data

The additional suborganellar data, PPI parameters, annotations and data set linkage substantially increased the amount of data for each AGI available. The SUBA4 factsheet has been redesigned to separate prediction from experimental localizations as well as include protein-protein proximity relationships (Supplementary Figure S5). The factsheet, similar to SUBA3 provides a data summary (SUBAcon, ASURE and schematic), predictions and external curations at the top. The increased depth of subcellular location is shown for FP, MSMS and PPI studies as a column, while the PMID, year, AGI of PPI partner PPI method and observed location (if available) is visible for PPI localizations associated with each AGI. A further category of protein-protein relationships was added that links SUBAcon calls directly to coexpression data as well as all PPI data. Thereafter, associated AGIs by coexpression (rank 1–10) and PPI are collected and linked to individual SUBAcon calls. The output gives the viewer an immediate overview of affiliated gene networks and PPI links and their subcellular locations. If the overview on the factsheet reveals a lead, more detailed investigations can be performed using the CAT and PAT in the toolbox. Finally, while retaining the conventional parameters of the SUBA3 factsheet, SUBA4 also contains all newly incorporated annotations that link directly to the respective data source for further exploration.

Current implementation status and future plans

SUBA provides a landing portal for accessing and linking existing subcellular data sets fast, which increases visibility and reuse of existing valuable experimental data. SUBA has become a central go-to for subcellular information in plant biology over the past decade. In context of ongoing global data maintenance efforts, SUBA has evolved alongside guidelines of data structuring, provenance and standardization consensus agreements around TAIR and Ensembl Plants.

SUBA4 now highlights the value of data linkage between subcellular data and other data entities from transcriptomics (AttedII) and proteomics (MASCP gator) in the SUBA4 toolbox. In reverse, many omics analysis pipelines benefit from SUBA-derived subcellular annotations. Currently, SUBA is integrated in the Bio-Analytic Resource platform (http://bar.utoronto.ca/) aiding transcriptome data interpretation as well as in Araport (https://www.araport.org/) as an intergrated API for downstream analysis pipelines. Further efforts aim to link SUBA to global genome and proteome browsers such as Gramene and Uniprot, respectively.

Future plans see SUBA utilize subcellular-refined pathway mapping to predict and define subcellular reaction rooms in plant cells. More than 35 000 Arabidopsis proteins are predicted to exist of which only 10% have been used in genome-wide metabolic models for calculating the energy household of plant cells (34–37). In fact, all of these models have heavily relied on SUBA to generate the compartmentalization of metabolism, which in turn achieved a better representation for estimating the energy required for cell homeostasis (14). With the steadily growing subcellular resolution, SUBA will remain an important part of building increasingly compartmentalized and complete genome-wide metabolic models. SUBA4 has now directly linked several pathway-mapping services including KEGG and Biocyc to the subcellular location data. Together with pathway prediction tools and subcellular location, the SUBA toolbox plans to provide useful cataloguing pipelines that help predict pathways, pathway linkage across compartments as well as necessary transporters into balanceable reaction rooms for modelling bioprocesses more accurately. Similarly, new quick tools for linking SUBA with other data entities such as phenomics data and epigenetic data will be an important step toward linking molecular data to plant traits. The latter data sets, while currently challenging to link, are under development for implementation into the SUBA toolbox for future SUBA releases.