Background

Patient-derived xenografts (PDX) become increasingly important as they provide a new approach for preclinical cancer research. PDX are in vivo models in which human cancer tissues are implanted in animal hosts, typically immunocompromised mice [1]. Exploration of cancer molecular features and drug response in PDX models has yielded a huge amount of information, in some cases leading to changes in cancer patient management [2, 3].

Since the number of researchers working with PDX mice models across Europe is rising quickly, as well as the number of models generated, there was a need to organize information about European PDX mice models. To avoid duplication of efforts, enable extensive collaboration, and to set up shared standards suitable for the PDX mice model facilities in Europe and beyond (to allow for cross-validation studies, among other reasons), the EurOPDX Consortium was established in 2013. Sharing biospecimens, data sets and standardization of laboratory procedures following high quality standards have been the main objectives of the Consortium [2].

To fulfill these goals, EurOPDX has started building a distributed infrastructure for PDX research providing services to other research groups, including PDX mice models biobanking and delivery, as well as drug efficacy testing. The first step towards construction of the EurOPDX Research Infrastructure (RI) has been the joint collection, harmonization and display of PDX-related data from partners, according to existing and newly developed standards. The work we present here is the current status of the EurOPDX Data PortalFootnote 1 which enables integrative search, browsing and exploration of more than a thousand PDX models for preclinical, clinical and molecular information.

Construction and content

Main features of the EurOPDX Data Portal

Through its Data Portal, the EurOPDX Consortium provides access to three main types of PDX-related data: (i) metadata, i.e., all annotations related to PDX model construction, tumor of origin, preclinical and clinical features as specified in the PDX-Minimal Information (PDX-MI) standard [4], (ii) processed molecular data, including cytogenetics and gene expression, mutation and copy number alterations, (iii) additional data, such as model’s drug dosing and patient treatment data.

Figure 1 illustrates the data flow, from partners owning and providing the models and related data, to the end user of the Portal, through the Data Platform.

Fig. 1
figure 1

EurOPDX Data Portal data flows. Four layers are described: (i) Input layer for metadata and processed molecular data provided by partners from the consortium. These data are collected via templates and laboratory information management systems (LIMS); (ii) Data processing layer where the data are collected, semi-automatically standardized and harmonized to be aligned with PDX Finder data services utilized in the Portal as well; (iii) Storage layer consists of the Data Hub where the EurOPDX data are stored; (iv) Data analysis layer is then a graphical user interface enabling (meta)data browsing and further analysis

Full size image

The current data loading process is semi-automatic and occurs through four main steps: (i) EurOPDX data providers fill in a metadata template. (ii) The data is checked for consistency and harmonized (e.g., unifying the gene symbols, diagnoses identifiers etc.) before ingesting into the database. Any additional datasets such as processed molecular data or drug dosing studies are collected in separate files. (iii) Metadata template and datasets are processed and validated by loader components developed within the PDX Finder catalogue [5], (iv) and uploaded into the EurOPDX Data Hub. Details on the whole process are given in the following sections. Technically, the same process and software is used to populate the database of the standalone PDX FinderFootnote 2 (see Fig. 1).

Currently, the Portal displays models and data for a total of 1010 PDXs from 9 PDX providers, across seven European Countries (Table 1).

Table 1 Overview of EurOPDX partners providing the data sets, cancer types and number of PDX mice models represented by these data sets
Full size table

These models represent several cancer systems listed in Table 2.

Table 2 PDX models available at the EurOPDX Data Portal listed according to cancer system
Full size table

Most of the models represent adult cancer models (93%), in 7% of cases age was not specified by the provider.

Figure 2 shows coverage of the models by additional data (molecular, treatment etc.). Gene mutation and copy number alteration are already available for half of the models or more, while other data are being introduced gradually.

Fig. 2
figure 2

Visualization of data types available in the Data Portal, by percentage. Out of 1010 deposited models, 653 include gene mutation data (65%), 505 include copy number alteration data (50%), for 283 models there are available drug dosing studies (28%), 270 models include gene expression data (27%), 103 models cytogenetics (10%), and 69 models include patient treatment data (7%)

Full size image

EurOPDX Data can be browsed and retrieved by the end user through two main graphical interfaces: the Data Portal Search InterfaceFootnote 3 and the EurOPDX cBioPortal Interface.Footnote 4 While the Search interface allows mainly searching models by metadata, the cBioPortal interface allows in-depth search and analysis of gene-level molecular data [6, 7], see examples at Fig. 5.

Input layer

Metadata is collected via Excel templates as this format has been the most preferred based on user feedback. The European Molecular Biology Laboratory - European Bioinformatics Institute (EMBL-EBI) and University of Turin teams developed and refined 9 collection templates to ensure a global community coverage and compliance with the PDX-MI standard [4]. This standard defines the minimal information for describing the clinical attributes of a patient’s tumor, the details of model creation, quality assurance methods, and metadata associated with the model’s availability for use in cancer research. These templates have been re-used by the global PDX community and new templates are developed or updated on an ad hoc basis to cover new data types or attributes. Current templates include:

PDX mice models and sample Metadata templates

  • Basic Metadata template collecting patients’ clinical and PDX mice models metadata

  • Sample Metadata template collecting metadata for molecular data

Individual templates for data (molecular and treatment)

  • Molecular datasets templates

    • ° Cytogenetics

    • ° Mutation

    • ° Copy number alterations

    • ° Fusion

    • ° Transcriptomic

  • Treatment template (patient and PDX mice models)

    • ° Patient treatment

    • ° Drug Dosing of the PDX mice models

Processing layer

Metadata provided by partners are validated for adherence to the PDX-MI standard [4]. Metadata harmonization is achieved by mapping biologically identical histological concepts provided by different sources. Specifically, to support consistent searching across resources, we use different attributes such as original histological term and the primary tissue provided by the resource. For example, histological concepts “Adenosquamous”, “adenosquamous carcinoma”, “Ad and SC carcinoma” share the same primary tissue “lung” and are mapped to the National Cancer Institute thesaurus (NCIt) [8] ontological label “Adenosquamous Lung Carcinoma”. Moreover, concepts are aggregated based on meaningful groupings like cancer by anatomical system or cell morphology. This approach allows a search for “lung cancer” models to display hits across all subclasses of lung cancer models in a single query.

Processed Molecular Data Sets are retrieved by the EMBL-EBI team from existing repositories/locations as provided by the data owners. Data is then validated and uploaded to a Neo4J database (details in the following section). To populate the database a bespoke Java Extraction-Transformation-Loading (ETL) pipeline was written to extract relevant attributes corresponding to the PDX-MI standard from the data provided by the PDX mice models providers. Software is freely available on GitHubFootnote 5 under an Apache 2.0 license.Footnote 6

A major bottleneck to molecular data integration is the heterogeneity of the data produced from multiple sources that involves a variety of sequencing platforms, laboratory protocols and analyses. This heterogeneity introduces strong technological biases and causes data inconsistencies, hindering data integration efforts. To control this technical variability, we have developed a service to remap all PDX mutation data to the same genome assembly (GRCh38). We then re-annotated variants on a single annotator (Variant Effect Predictor – VEPFootnote 7). This harmonization provides accurate and standardized annotation of variants, following Human Genome Variation Society (HGVS) standards,Footnote 8 and it ensures consistent searching and linking to cancer annotation databases like CivicDB [9], OpenCravat [10], and COSMIC [11].

In addition to the PDX-MI standard, data from the models in the EurOPDX Data Portal comply with other nomenclature and metadata standards accepted by the community: NCI thesaurus [8] for cancer type, diagnosis and other cancer attributes, Human Genome Organization Gene Nomenclature Committee [12] standards for the names and symbols of the human genes, and International Committee on Standardized Genetic Nomenclature for Mice [13] for host mouse strain nomenclature. NCIt [8], ChEBI [14], CHEMBL [15] and PubChem [16] are used to standardize drugs and compound names.

Storage layer

The EurOPDX Data Hub constitutes the repository layer of the RI and provides an application programming interface (API) for exporting selected data sets to chaining tools like cBioPortal, see Fig. 3.

Fig. 3
figure 3

EurOPDX Data Hub schema. The Hub includes the Neo4J database containing data accessible via the Data Portal Search page, and the MySQL database providing temporary operational data. There is a REST API available within the Hub accessible to the chaining tools

Full size image

The Data Hub is designed to handle multiple types of datasets and databases. Three types of data are stored:

  1. 1.

    Clinical data that contain pseudonymized information about patients (e.g., age at the collection, sex, diagnosis) and about PDX mice models (stage of cancer, site of the tumor, primary/metastasis status).

  2. 2.

    Metadata which describe how and where the PDX mice models were created/prepared (e.g., implant site of engraftment, an identifier of the mouse).

  3. 3.

    Genomic data – this comprises an enormous set of different data types. Currently expression, copy number alteration, and mutation.

The storage is based on a Neo4J database (DB) that follows the structure from PDX Finder’s [5] Neo4J DB records of EurOPDX partners/data providers (the N4J DB schema is shown in Fig. 4) models, patients, clinical data, and molecular data. Besides Neo4J there is a MySQL DB for storing temporary data in Data Hub, and eventual extensions not covered by the PDX Finder data schema (it is not required currently but it was the case of the previous version before gene expression data became supported by PDX Finder).

Fig. 4
figure 4

Schema of the Neo4J Database. Each node of the graph database schema is an entity or event captured in the data. Arrows, so called edges, show the relationships between the nodes

Full size image

Data loading tools of PDX Finder – including automatized standardization and harmonization – are used for loading data to the shared database of the Hub. The Hub design is not restricted to using the Neo4J database only. On the contrary, the previous versions complemented it with a relational database to store additional data, being populated with other tools. The current setup is the result of unification developments.

The RESTful APIFootnote 9 of the Hub provides a unified way for export of clinical data, metadata and molecular data for a specified set of PDX models. This API is designed to be complete in the sense that any chained visualization on data processing tool (cBioPortal currently, GenomeCruzerFootnote 10 foreseen) can retrieve all data it needs via this API. In this way, strict consistency of the data among the tools is ensured. Technically, we manage the tool integration by wrapper scripts, which retrieve the data using the API, format them appropriately, and feed to the wrapped tools.

The Data Hub API is based on Python’s Flask framework. To ensure automated deployment and effective maintenance, the Data Hub runs with its components (Neo4J, MySQL, API and Proxy) within Docker containers managed by docker-compose [17].

Output layer

The user interface of the above-described database is integrated with EurOPDX Data Portal web pages in Umbraco Content Management System [18]. The principal entry pointFootnote 11 is based on branded PDX Finder software with minor extensions (classifying models by their availability for Trans-national access via the EurOPDX RI, search by specific model identifiers, etc.). This interface allows search and selection of the models by metadata (origin, diagnosis, treatment, etc.). An example of the search screen is shown in Fig. 6.

The same data are available by the EurOPDX cBioPortal interface, populated by a setup procedure which calls the Data Hub API as described above. This interface is focused on more complex analyses and visualizations of the molecular data (gene expression charts, mutations, and their correlations), see Fig. 5 for several examples of the visualization tools available within the cBioPortal platform.

Fig. 5
figure 5

EurOPDX data visualization in the EurOPDX cBioPortal instance. A Example of plots which can be designed according to the user’s needs. This is a plot showing EGFR gene mutations compared with age of the patient within a selected study; B OncoPrint of queried genes. Provides information about number of samples per patient within a selected study, data types availability per each sample and gene mutations or alterations; C Example of mutation analysis of a chosen dataset – MutationMapper tool. It is possible to browse mutations within the selected gene - mutations mapping on a linear protein and its domains, type of the mutation – and get detailed sample information in which mutations are present. 3D visualizations of the protein interactions and other mutation information is available by linking external databases

Full size image

User registration and authentication

The Portal is freely accessible to any researcher from the internet after a lightweight, self-service registration, required for maintaining minimalistic records of the service usage enforced by policies of the underlying cloud service provider, as well as to keep track of users’ agreement to the terms of Acceptable Use Policy (which is, besides forbidding unlawful use, mostly informative only, not restricting the use of data for non-commercial purposes), and to offer personalized services eventually. We adopted the technical solution developed in the ELIXIR research infrastructure [19], which merges the possibility to use the user identity of any institution supporting the eduGAINFootnote 12 federation, as well as social identities (Google, LinkedIn, ORCID); migration to the emerging European Life-Science authentication infrastructure is foreseen.

Utility and discussion

User interface

Users can reach the EurOPDX Data Portal homepage at https://dataportal.europdx.eu/. This page displays information about competitive calls by the EurOPDX RI for access to PDX models, model data availability and a link to the Search page at https://dataportal.europdx.eu/search.

User registration/login is required to access the Search page.

As Shown in Fig. 6, the Search page is subdivided into two sections. On the left, multiple filters allow selection of models by basic features, molecular data, treatment data, and patient/tumor data. Each filter category is further subdivided into sub-categories, and the filters can be mixed for accurate queries based on the user’s needs. For users interested in accessing models through the EurOPDX RI, a filter enables selection of models available for Trans-national access.Footnote 13 Models are also accessible on a collaborative basis, by contacting the PDX owners. The right part of the Search page displays the results of the query in tabular format, with each model in a separate row. Key features of the models are presented in columns and include model identifier (ID) and its original provider, tumor histological classification, site of primary tumor, collection site, and links to available datasets (Table 3).

Fig. 6
figure 6

Screenshot of the Search page from the EurOPDX Data Portal

Full size image
Table 3 Links to publicly available datasets
Full size table

Users can navigate to a model/patient/tumor page or to the data of interest by clicking on the unique PDX Model ID or data links in each row. At the lower-right corner of the page, users find options to download search results in a tabular format or to export them to a dedicated instance of cBioPortal for Cancer Genomics [6, 7] for deeper investigation of available molecular data, as described in the next section. A pre-compiled instance of the EurOPDX cBioPortal, including all main datasets currently stored in the Hub is publicly available at https://cbioportal.europdx.eu/.

A video tutorial containing the Portal functionalities can be found on EurOPDX RI YouTube channel next to other informative videos, please see https://youtu.be/l2AfjjcfT2Y.

Tool chaining and consistent data availability

While the Data Portal allows to search and select models and samples according to certain criteria, the cBioPortal provides additional visualization and analysis capabilities. To link these two services, we leveraged the thorough design of the Data Hub API to extract a model selection in Data Portal and create a dedicated cBioPortal populated only with the data associated to the specific models selected. In order to speed up the startup time of such cBioPortal personalized to a user’s interest, snapshots of partially prepared (namely populated with all static data) cBioPortal Docker containers are reused.

Conclusions

The EurOPDX Data Portal provides essential search functionality as well as complex molecular data analysis and visualization for cancer researchers interested in accessing PDX mice models. In particular, the Portal is the primary entry point for users interested in the free-of-charge research services offered by the EurOPDX RI. Indeed, from the perspective of the user looking for PDX models that are relevant to his/her research, retrieval of this information in the Portal serves as a background to request specific trans-national access services provided by the EurOPDX RI -- frozen PDX tumour samples shipment, in vivo studies on the selected PDX model(s), deposition of PDX model(s) (unlike the open access to data in the Portal, these services may be subject to further restrictions). The Portal is populated with curated data, special attention is paid to their thorough harmonization so that queries across multiple studies are possible. The procedure of data curation is well-defined and automated as much as possible to both speed up the process and avoid unintentional human errors. The data are shared among the Portal components, making sure by technical means that they are always mutually consistent.

At the time of writing this article, the Portal offers data on 1010 PDX models referring to 7 distinct cancer types, and the number models is expected to grow steadily. As such, the Portal provides the technical prerequisites to achieve the principal missions of the EurOPDX Consortium, which is to enable data sharing among the partners as well as to the broader research community, thereby improving the efficiency of PDX research and reducing duplication of efforts and unnecessary animal experiments.