Translational bioinformatics in mental health: open access data sources and computational biomarker discovery.

Mental illness is increasingly recognized as both a significant cost to society and a significant area of opportunity for biological breakthrough. As -omics and imaging technologies enable researchers to probe molecular and physiological underpinnings of multiple diseases, opportunities arise to explore the biological basis for behavioral health and disease. From individual investigators to large international consortia, researchers have generated rich data sets in the area of mental health, including genomic, transcriptomic, metabolomic, proteomic, clinical and imaging resources. General data repositories such as the Gene Expression Omnibus (GEO) and Database of Genotypes and Phenotypes (dbGaP) and mental health (MH)-specific initiatives, such as the Psychiatric Genomics Consortium, MH Research Network and PsychENCODE represent a wealth of information yet to be gleaned. At the same time, novel approaches to integrate and analyze data sets are enabling important discoveries in the area of mental and behavioral health. This review will discuss and catalog into an organizing framework the increasingly diverse set of MH data resources available, using schizophrenia as a focus area, and will describe novel and integrative approaches to molecular biomarker discovery that make use of mental health data.

Introduction

In 2013, mental illness was highly prevalent and estimated as incurring the highest financial burden among medical conditions in the United States, with spending estimated at $201 billion [1]. In light of the considerable cost to individuals and society, mental illness represents a compelling opportunity for discovery and improved patient care. As our ability to untangle biological mechanisms of disease grows, so too does our ability to leverage our richer understanding for better diagnoses, interventions and outcomes. In many areas of medicine, biomarker discovery is causing a shift toward biomarker-based diagnoses that promise better targeted and thus more effective therapies. Precision medicine approaches incorporating molecular and imaging biomarkers into therapeutic decision-making are emerging. Publicly available ‘big data’ resources like TCGA (The Cancer Genome Atlas) are being used and reused in numerous ways, with thousands of downstream citations [2]. Despite the magnitude of opportunity, translation has been slower in mental health (MH) than in other areas of health care [3]. In this article, we first describe the motivation as well as some challenges for the use of biomarkers in MH. We address one major challenge—findability of relevant resources—by providing a catalog of relevant data resources for biomarker discovery in MH, and a framework for their organization. Finally, we give an overview of existing approaches to biomarker discovery using publicly available data.

An exploration of biomarker in MH is especially timely in light of recent announcements from the National Institute of Mental Health (NIMH) exhorting a renewed focus on causal models of disease [4, 5]. Over the past 7 years, NIMH awards have shifted away from clinical research and trials and toward mechanistic biological understanding, coinciding with the NIMH’s launch of Research Domain Criteria (RDoC) in 2011, a framework emphasizing research into mechanisms (rather than clinically observable signs and symptoms) of mental illness [6].

Despite ample motivation, data reuse in MH research remains sluggish even in the presence of available biological resources and an emphasis on data sharing [7]. One important obstacle is the surprising difficulty in identifying available resources, related at least in part to an absence of a systematic approach to cataloging available resources. We seek to propose and use a systematic approach to organizing relevant resources. We then catalog data sets pertinent to MH biomarkers to facilitate secondary use of these data for computational biomarker discovery.

In addition to MH-focused resources and general resources that include MH conditions, a number of rich resources exist for specific MH disorders. However, inclusion of resources for every MH condition would far exceed space limitations for a single review. Therefore, in addition to general MH resources that span multiple disorders, we extend resource cataloging to a single MH disorder, schizophrenia (SCZ) and focus our biomarker discovery literature review on that disorder. SCZ is selected because it is one of the most studied MH disorders, puts heavy burden on the community and co-authors have conducted both large data annotations and various analyses in this area. It is also a prime example of a diagnostic concept well-recognized to be problematic and in need of updating [8, 9]. Biological exploration, biomarker discovery and elucidation of underlying mechanisms are key to addressing this issue. We have also limited our catalog to resources that are publicly available. Private or proprietary data sets or tools that are neither intended nor accessible for secondary research by independent researchers are beyond the scope of this review.

The mind-biology problem: a challenge for another day

The mind-body problem—What is the relationship between the mind (feelings, thoughts, beliefs) and the physical realm (matter, atoms, neurons)?—is commonly recognized in philosophy [10]. We stipulate, for the purposes of this discussion, that psychology becomes neurobiology once a biological mechanism is understood. As we gain insights into the neural basis of normal and abnormal behavior, syndromes historically described in terms of mental constructs can be described in terms of biological constructs. While we do not seek to tackle the philosophical question of whether all mental constructs can be adequately described in biological terms, we do assert that understanding biological mechanisms in MH is valuable, is likely to expand and will benefit from integrative research connecting behavior to biomarkers.

RDoC: ‘Outcomes to Causes and Back’

MH disorders typically include a spectrum of symptoms that affect emotions, thoughts and behaviors [4]. Moreover, two people can be diagnosed with a single disorder such as SCZ, despite having no overlapping symptoms. The NIMH RDoC initiative is an attempt to ‘develop, for research purposes, new ways of classifying mental disorders based on dimensions of observable behavior and neurobiological measures’ [4, 11]. The goal is to generate categories stemming from basic behavioral neuroscience, rather than starting with a highly heterogeneous illness definition and then seeking its neurobiological underpinnings. To this end, RDoC makes no reference to current DSM-based (Diagnostic and Statistical Manual of Mental Disorders) classification but instead proposes an alternative organizing scheme for linking behavior to underlying mechanisms. The RDoC approach is directly relevant to MH biomarkers because it aims to identify specific elements, such as mutations, genes, molecules, cells, pathways, physiological measures or behaviors associated with specific mental constructs across different disorders [3, 11].

Data, like life, is not always FAIR (Findable, Accessible, Interoperable, Reusable)

One of the expected and desirable results of the NIMH funding shift has been the generation of large biological data sets relevant to MH that are expected to be shared and reused. Recognizing the urgent need to improve infrastructure around data discoverability and reuse in the big data era, a group of stakeholders from academia, industry, funding organizations and publishers came together to design and endorse a set of measurable principles to act as guidelines for best practices in data sharing [12]. The resulting framework is known as FAIR principles—Findable, Accessible, Interoperable, Reusable. FAIR principles put particular emphasis on enhancing the ability for computers to find and use existing data. Findable refers to whether a researcher who would want to use the data set in question is able to discover that the data exist. This requires clear, persistent and searchable metadata. Accessible refers to whether the data are available to be downloaded. Are they retrievable through a standard communications protocol that enables authentication and authorization? Interoperable considers whether appropriate data and metadata standards are used for knowledge representation. Reusable addresses whether the data and provenance are represented in sufficient detail, with clear guidelines for usage.

While some researchers express concerns about reuse of clinical data in particular [13], many in the scientific community see significant benefit to be gained by data sharing and reuse [14]. The National Institutes of Health (NIH) has launched a Data Commons initiative to establish a virtual environment to facilitate the use, interoperability and discoverability of shared digital objects used for research [15]. This review focuses on those resources, data sets and publications that adhere to the spirit of the FAIR principles [12].

Biomarkers in mental health: What, why and how?

What is a ‘biomarker’ anyway?

A biomarker traditionally is defined as ‘a characteristic that is objectively measured and evaluated as an indicator of normal biological processes, pathogenic processes, or pharmacological response’ [16]. Biomarkers can be generally classified as (1) diagnostic or trait markers that indicate the presence of a disease, (2) prognostic markers that indicate the likely course of a disease or (3) theranostic markers that predict how an individual is likely to respond to a certain treatment [17, 18]. As yet, no clinically actionable biomarkers have been approved for use in MH [11]. However, there is increasing recognition of the biological underpinnings of MH, the importance of biomarker discovery and the significant opportunity that MH poses in this regard. To this end, substantial research efforts have been devoted to biomarker discovery, and a number of publications describe promising leads [19-24]. Importantly, many of these studies have made their data publicly available to varying degrees, enabling secondary research and innovative approaches to analysis, in some cases through novel, integrative methods that could not have been done with the original data alone.

Biomarker types

Physiological biomarkers span a wide range of modalities and data types and may be categorized as either microscopic or macroscopic in scale (Figure 1).

Figure 1

Overview of micro- and macro-level biomarkers. indels, small insertions/deletions; SV, structural variants.

Micro-scale biomarkers: all things omics

Micro-scale biomarkers refer to biomarkers at the molecular level. The various and ever-increasing number of ‘-omic’ data-based biomarkers has been documented elsewhere [25, 26]. Genomic biomarkers generally refer to DNA sequence, including single-nucleotide variations (SNVs), copy number variations (CNVs), insertions, deletions, structural variants, etc. Transcriptomics refers to RNA expression, including both coding and noncoding RNA. Epigenomics refers to features of DNA other than the sequence itself, e.g. methylation, histone modification, etc. Proteomics refers to the presence, quantity and posttranslational modification state of proteins and peptides. Metabolomics refers to identification, quantification and ratios of various metabolites generated through the organism’s metabolism. Genomic and transcriptomic biomarkers have arguably received the most attention in the past decade, in part because they have become relatively low hanging fruit: microarrays and sequencing technologies make it fairly straightforward and increasingly inexpensive to make observations across the entire genome and transcriptome.

Macro-scale biomarkers: tissue and system level

Macro-scale biomarkers are observed at the tissue or system-level, generally through imaging technologies. Advances in brain imaging technology over past 20–30 years have enabled application to MH and illness. Commonly used imaging modalities include magnetic resonance imaging (MRI), magnetic resonance spectroscopy, positron-emission tomography (PET), single-photon emission computed tomography and diffusion tensor imaging (DTI). Other methods of neuroimaging involve recording of electrical currents or magnetic fields, for example electroencephalography (EEG) and magnetoencephalography (MEG). The additional biomarker types listed below are all macro-scale biomarkers.

Structural biomarkers of the brain

Structural imaging provides qualitative and quantitative information about the brain that describes the shape, size and integrity of gray and white matter structures in the brain. Typically, morphometric techniques measure the volume or shape of gray matter structures and white matter tracts. Structural MRI is used for identifying density or volume of brain matter, and DTI provides images of anatomical pathways and circuits especially of white matter [27].

Functional biomarkers of the brain

Whereas structural imaging provides static anatomical information, functional imaging provides dynamic physiological information [28]. Functional MRI (fMRI) and PET measure localized changes in cerebral blood flow related to neural activity, while EEG and MEG measure electrical currents and magnetic fields that vary with function.

The connectome: a ‘wiring diagram’ for the brain

The brain connectome defines the connectivity architecture and network organization of the neural components of the brain in terms of both structure and function. The connectome is represented as a large graph with nodes (brain regions) and edges (pathways) and has been enabled by advances in neuroimaging including structural MRI, fMRI and diffusion MRI [29]. Connectivity analysis based on graph theory is used to explore variations in the type and strength of connectivity between brain regions. Current evidence demonstrates alterations in both large-scale network and local network connectivity in mental health, and these alterations define distinct clinical and cognitive phenotypes [30].

Biomarker data resources

A framework for resource classification

A surprising challenge awaits a novice attempting integrative analyses: simply identifying what resources are available, how they relate to each other and what each one can and cannot provide is surprisingly difficult. In writing this review, we initially set out to catalog a list of publicly available data resources relevant to mental health. In the course of due diligence to identify these resources, certain categories and attributes emerged. Thus, our effort to catalog available resources also informed the creation of a candidate framework for classifying and organizing the different resource types.

Data resources can be classified as one (or sometimes more than one) of four high-level categories: (1) Organizational entity; (2) Initiative; (3) Platform; or (4) Data set (Figure 2). Examples of organizational entities include federal agencies, such as the NIMH, and nonprofit organizations, such as the Allen Institute for Brain Science [31]. Initiatives are activities or groups organized around activities aimed at creating, collecting or cataloging data for research. Examples include PsychENCODE, BioCADDIE and the Psychiatric Genomics Consortium (PGC) [32-34]. Data sharing platforms are Web-based applications that enable a researcher to search for data sets using metadata and to download the data. Examples include Sage Bionetworks’ Synapse platform or the Gene Expression Omnibus (GEO) [35, 36]. Finally, specific data sets may include data resulting from experimental assays, e.g. various data sets available in GEO, or curated knowledge bases like SZGR [28]. As shown in Figure 2, the relationships between different categories do not form a simple hierarchy but are instead are many-to-many. An initiative may be associated with one or more organizational entities, whether through funding or logistical or administrative support, while an organizational entity may be associated with one or more initiatives. A given organizational entity or initiative may rely on one or more platforms. A platform may contain (or point to) one or more data sets from one or more initiatives or organizational entities. A given data set is generally stored in one platform, but may also be accessed through other platforms, whether because it is replicated there or because some platforms serve as portals to federated data sets. These categories are not strictly mutually exclusive, and some resources blur the boundaries between them. For example, it can be hard to differentiate between an organization and an initiative. As a general rule, if an organization was created primarily for the purpose of creating or collecting data, we consider it an initiative. In addition, a curated knowledge base may import and redistribute some data sets on which it is based making it both a platform and a data set.

Figure 2

A framework for classification of data-related resources. Nodes denote resource types (Entities, Initiatives, Platforms and Data sets), and edges show the many-to-many relationships among them.

With respect to platforms, several attributes are especially salient. Some platforms such as Open fMRI focus on a single data type. Other platforms such as Synapse are meant to be general-purpose. In addition to storing different types of data, Synapse is disease-agnostic, storing data from many different diseases and medical domains. Other platforms, for example the Stanley Neuropathology Integrative Database focuses on a specific set of mental health conditions. Figure 3 shows where major data-sharing platforms relevant to mental health fall along the spectra of data-type specificity and disease focus. Some platforms, e.g. DataMed developed by the bioCADDIE project team for the NIH BD2K Data Discovery Index (DDI), are essentially portals to a federated collection of data sets that reside in still other platforms. Finally, some resources solely house data or information, while others are associated with biospecimens that may be available for generating additional data.

Figure 3

Visual representation of data platform attributes. See Table 1 for abbreviations.

Table 1

Open data resources for biomarker discovery in mental health, particularly in schizophrenia

Resource	Type	URL	Notes
Enhancing Neuro Imaging Genetics Through Meta Analysis (ENIGMA)	O	http://enigma.ini.usc.edu/ongoing/enigma-schizophrenia-working-group/	The ENIGMA Network brings together researchers in imaging genomics to understand brain structure, function and disease, based on brain imaging and genetic data. Includes Schizophrenia Working Group (ENIGMA-SCZ)
NIMH	O	https://www.nimh.nih.gov/index.shtml	The institute within the NIH that focuses on mental health and disease. The NIMH is one of 27 institutes and centers within NIH, which is part of the US Department of Health and Human Services
Open Translational Science In Schizophrenia (OPTICS)	O	https://sites.google.com/site/opticsschizophrenia/home	A time-limited proof of concept pilot project designed to provide a forum for translational science based on Janssen clinical trial data made available to qualified investigators
Stanley Medical Research Institute (SMRI)	O	http://www.stanleyresearch.org/	A nonprofit organization supporting research on the causes of, and treatments for, SCZ and bipolar disorder
Mental Health Research Network (MHRN)	O	http://hcsrn.org/mhrn	Consortium of 13 health system research centers dedicated to improving patient mental health through research, practice and policy. Supported by a cooperative agreement from the NIMH. The MHRN conducts pragmatic research in health systems serving over 12 million patients
Common Mind Consortium	I	http://commonmind.org	Public–private partnership to generate and analyze large-scale genomic data across several brain regions from human subjects with neuropsychiatric disease and to make these data and the associated analytical results broadly available to qualified investigators
Human Connectome Project (HCP)	I	http://www.humanconnectome.org/	Large NIH-funded project for integrating genomics, behavior and brain imaging. Currently, high-resolution imaging data are available on 1200 individuals. Primary modalities measure brain activity (resting state fMRI and task-evoked fMRI), white matter integrity (diffusion imaging and T2 FLAIR) and oscillatory brain activity (EEG and)
NIMH Human Genetics Initiative	I	https://www.nimhgenetics.org/nimh_human_genetics_initiative/	Intended to establish a national resource of clinical and diagnostic information and immortalized cell lines from individuals with SCZ, bipolar disorder or Alzheimer's disease and their relatives, available to qualified investigators for research on the genetic basis of these disorders
PsychENCODE	I	https://www.synapse.org//#! Synapse: syn4921369/wiki/235539	Funded by the NIMH with the goal of accelerating discovery of noncoding functional genomic elements in the human brain and elucidating their role in the molecular pathophysiology of psychiatric disorders
Stanley Neuropathology Consortium (SNC)	I	http://www.stanleyresearch.org/brain-research/neuropathology-consortium/	A collection of 60 brains, consisting of 15 each diagnosed with SCZ, bipolar disorder or major depression, and unaffected controls. Samples may be requested for research purposed. Associated data are available in the SNC Integrative Database (SNCID)—see below
Psychiatrics Genomics Consortium (PGC)	I	http://www.med.unc.edu/pgc	Founded in 2007, the PGC includes over 800 investigators from 38 countries with the goal of conducting meta- and mega-analyses of genomic data for psychiatric disorders. The initial focus was on autism, attention-deficit hyperactivity disorder, bipolar disorder, major depressive disorder and SCZ. More recently, the scope has expanded to other conditions and other types of genetic variation beyond SNVs
Neuroscience Information Framework (NIF)	I/P	https://neuinfo.org/	An NIH-funded framework for identifying, locating, relating, accessing, integrating and analyzing information from the neuroscience research enterprise. NIF has come to refer to both this initiative and the set of tools and platforms that make up that framework including the registry of electronic resources and the discovery portal for searching those resources. NIF includes >4500 curated resources and access to > 100 databases
Allen Brain Atlas/Data Portal	I/P	http://human.brain-map.org/	The Allen Institute for Brain Science is dedicated to understanding how the human brain works in health and disease. The Allen Human Brain Atlas integrates anatomic and genomic information across the brain. Data modalities include MRI, DTI, histology and gene expression data derived from both microarray and in situ hybridization (ISH) approaches. Microarray data are spatially mapped to the MRI. Complete microarray and RNA-seq data are available for six human brains. ISH data are available for ∼50 SCZ brains
NIMH Repository and Genomics Resource (RGR)	P	https://www.nimhgenetics.org/available_data/schizophrenia/	Includes 100+ studies, including CommonMind, PsychENCODE. Formerly the Center for Collaborative Genomic Studies on Mental Disorders, the RGR was established in 1998 through the NIMH Human Genetics Initiative to leverage and increase the value of human genetic samples and data produced through NIMH-funded research. It contains a collection of > 150 000 well-characterized, high-quality patient and control samples from patients with a range of mental disorders. The RGR’s Biologic Core and a Data Management Core are external to NIH
Function Biomedical Informatics Research Network Data Repository (FBIRN DR)	P	fbirnbdr.nbirn.net: 8080 (BROKEN)	FBIRN was initially focused on assessing major sources of variation of fMRI data generated across different scanners. The FBIRN Phase 1 data set consists of a traveling subject study of five healthy subjects, each scanned on 10 different 1.5 to 4 T scanners. The FBIRN Phase 2 and Phase 3 data sets consist of subjects with SCZ or schizoaffective disorder along with healthy comparison subjects scanned at multiple sites. The BIRN Data Repository (BDR) includes imaging, clinical, cognitive and physiological data
OpenNeuro (previously OpenfMRI)	P	https://openneuro.org/(https://openfmri.org/)	A neuroimaging repository to enable reproducible analysis and data sharing. Started in 2010, it initially focused only on task-based MRI, but is now open to all forms of neuroimaging data, reflected in the name transition from OpenfMRI to OpenNeuro. Data are anonymized before distribution to protect the confidentiality of participants and distributed using a Public Domain license
Research Domain Criteria Database (RDoC DB)	P	https://data-archive.nimh.nih.gov/rdocdb/	A data repository for the harmonization and sharing of research data related to the RDoC initiative and mental health research more generally. The actual platform uses software designed to host the NIH’s National Database for Autism Research (NDAR)
SchizConnect	P	http://schizconnect.org/	Federated access to several neuroimaging databases with images acquired on SCZ subjects. Data sources include FBIRN, NUSDAST, COINS and MCIC (maintained by the Mental Illness and Neuroscience Discovery Institute, now the Mind Research Network). More than 1100 subjects with >1000 have imaging data, including resting state fMRI, task-related fMRI, structural and diffusion imaging
SNCID	P	http://sncid.stanleyresearch.org/	Web-based tool for exploring neuropathological traits, gene expression and associated biological processes in psychiatric disorders generated by the SNC within the SMRI
Australian Schizophrenia Research Bank	P	http://www.schizophreniaresearch.org.au/bank/	A research database and storage facility that links clinical and neuropsychological information, blood samples and structural and fMRI brain scans from people with SCZ and healthy nonpsychiatric controls, and currently has data on ∼900 cases and 900 controls
Internet Brain Volume Database (IBVD)	P	http://ibvd.virtualbrain.org/	Centered around publications as the central data structure, IBVD is a Web-based searchable database of brain neuroanatomic volumetric observations that enables electronic access to the results in the published literature
dbGap	P	https://www.ncbi.nlm.nih.gov/gap	Developed by the NIH’s NCBI to archive and distribute the data and results from studies that have investigated the interaction of genotype and phenotype. While the focus is on genomic data, other data types are included as well, for example metabolomic data and laboratory values
Metabolights	P	http://www.ebi.ac.uk/metabolights/	A database for Metabolomics experiments and derived information. Metabolights is the slightly more established European counterpart to the NIH’s MW and the recommended metabolomics repository for a number of top journals
DataMed	P	http://datamed.org/	Data search engine portal to enable users to search for data across different repositories developed for the NIH BD2K DDI by the bioCADDIE project team. The initial prototype release (v2.0) features a set of data repositories selected by the bioCADDIE team, with a form to suggest additional repositories for inclusion
Metabolomics Workbench (MW)	P	http://www.metabolomicsworkbench.org/	A repository for metabolomics data and metadata, MW provides analysis tools and access to metabolite standards, protocols, tutorials and training
PRIDE	P	https://www.ebi.ac.uk/pride/archive/	A centralized, standards compliant, public data repository for proteomics data, including protein and peptide identifications, posttranslational modifications and supporting spectral evidence. Most of the data sets related to mental health disorders in PRIDE are derived from animal models
Synapse	P	https://www.synapse.org/	Sage Bionetworks’ software platform for data sharing and provenance tracking. Synapse enables researchers to carry out, track and communicate research in real time and enables co-location of scientific content (data, code, results) and narrative descriptions of that work. The platform is agnostic regarding biomedical domain or data type and hosts a number of different file types and projects funded by a number of different sources
GEO	P	https://www.ncbi.nlm.nih.gov/geo/	An international public repository developed by the NIH NCBI that archives and freely distributes microarray, next-generation sequencing and other high-throughput functional genomics data submitted by the research community
AE	P	https://www.ebi.ac.uk/arrayexpress/	The European counterpart to GEO. AE is an archive of functional genomics data from high-throughput functional genomics experiments. A subset of experiments is imported from GEO, while others are submitted directly
GEMMA	P	http://www.chibi.ubc.ca/Gemma/	Gemma is a website, database and a set of tools for the meta-analysis, re-use and sharing of genomics data, currently primarily targeted at the analysis of gene expression profiles
OmicsDI	P	http://www.omicsdi.org	Enables data set discovery across omics data resources spanning eight international repositories, including both open and controlled access data resources. The resource provides key metadata for each data set and uses this metadata to enable search capabilities and identification of related data sets. OmicsDI helps researchers to idenitfy groups of related, multi-omics data sets across repositories

Note: Type: O, organizational entity; I, initiative; P, platform.

Resource identification

Because the topic of data for biomarker discovery in mental health is so broad, a simple PubMed query was not feasible. (For example, a query for ‘[mental health OR behavioral health OR psychiatric] AND [biomarker OR genomic OR imaging] AND data’ yields >25 000 hits.) A preliminary list of resources was established based on co-authors’ prior knowledge in the domains of open data, FAIR principles and MH. The list was then augmented through a series of searches in PubMed and Google Scholar using combinations and variations of the following terms: mental health, SCZ, open data, database, imaging, genomic, proteomic, metabolomic and biomarker. In addition to direct hits yielded by these terms, PubMed’s ‘similar articles’ provided valuable additional results. Finally, a search in BioCADDIE’s DataMed data search engine yielded additional sources. Inclusion criteria for resources were: (1) Scope includes one or more types of biomarker data (beyond clinical phenotype data); (2) Data accessibility, or at minimum some indication of how to request the data; and (3) Coverage of MH phenotypes, or in the case of disease-specific resources, SCZ. The resulting list of data resources and their metadata is provided in Table 1. Figure 4 gives a high-level landscape overview for the MH-specific organizational entities, initiatives and platforms and how they relate to each other.

Figure 4

Overview of landscape of organizational entities, initiatives and data sharing platforms.

A number of potentially useful resources were deemed out of scope for this review because they lacked either -omics data (e.g. National Database for Clinical Trials Related to Mental Illness, NDCT, Yale Open Data Access Project YODA [37]), or psychiatric phenotype data (Exome Aggregation Consortium, ExAC [38], Genotype-Tissue Expression (GTEx) project [39]).

Data sets

Genomic data

The two main data repositories for gene expression or transcriptomic data are GEO and ArrayExpress (AE). GEO is an international public repository developed by the US NIH’s National Center for Biotechnology Informatics (NCBI) that archives and freely distributes microarray, next-generation sequencing and other high-throughput functional genomics data submitted by the research community [35]. AE is the Europe-based repository, hosted by the European Bioinformatics Institute within the European Molecular Biology Laboratory (EMBL-EBI). Data are imported from GEO into AE on a weekly basis making GEO a subset of AE. To be uploaded to these data repositories, data sets need to be in a specific format, such as GEOarchive, SOFT or MINiM. They must also include appropriate metadata about the clinical and experimental data. Both AE and GEO enable programmatic access to data via tools like R/Bioconductor. Data sets in GEO therefore are able to satisfy the F (findable) and R (reusable) FAIR criteria. Gene expression data in GEO are generally considered to be de-identified, and are thus freely available for public use.

There has been an increase in genomic profiling of data related to MH in the last few years. Taking SCZ as an example, a search in AE for published SCZ data sets shows 92 data sets in humans (Supplementary Table S1). Only two data sets were published in 2007 as compared with 11 data sets published in 2016. Until 2010, the majority of published data sets concerned transcriptome profiling. In 2012 and 2013, other genomic methods had gained popularity including methylation and next-generation sequencing technologies, exploration of noncoding regions and gene expression and splicing. Since 2014, many studies have been published using newer genomic platforms including chromatin immunoprecipitation sequencing, RNA sequencing (RNA-seq) and microRNA-seq, amounting to an approximate 15 published studies in 2014, 13 studies in 2015 and 11 studies in 2016. As of July 2017, we found that of the 92 data sets, 80 had been cited in one or more subsequent publications. Using Google Scholar queries on data set accession identifiers, it was determined that these 80 publications have been cited 6710 times (Supplementary Table S1). Note that citation does not necessarily imply analysis: many publications called attention to the existence of a data set without performing any additional analysis.

NCBI’s database of Genotypes and Phenotypes (dbGaP) contains archived data and results from studies that have investigated the association between genotype and phenotype in Humans [40]. The European equivalent is the EGA (European Genome-phenome Archive) [41]. As with the gene expression repositories, data sets need to be in a specific format along with minimal metadata to be submitted into dbGaP or EGA. Note that these repositories contain sequencing data that are unique to the individuals from whom they were derived, and thus cannot be considered completely de-identified. Users must therefore submit a data request form detailing the goals of their project and how they intend to use the data and observe data use policy for approval by a data request committee. This approach has implications for meeting the accessibility aspect of FAIR criteria but represents a balance between data accessibility and data privacy for research participants. dbGaP contains a number of MH-related data sets, including SCZ. Of the 154 studies returned based on a query for the term ‘schizophrenia’, only a small subset was targeted at SCZ as determined by manual inspection. In this case, findability is hampered by the number of false positives (Table 2). The vast majority of the 24 studies returned in a search for ‘schizophrenia’ in EGA are either focused on SCZ or have some number of samples included with a SCZ diagnosis.

Table 2

SCZ data sets in dbGaP

Data set ID	Name	# Participants	Platform	Publication (PMIDs)	Citations	Data type
phs000979.v1.p1 (PRJNA293910)	Gene Expression in Postmortem DLPFC and Hippocampus from Schizophrenia and Mood Disorders	914	HumanHap650Yv3.0, Human1M-Duov3_B, Human HT-12 Expression Bead Ch	28070120	[4]	SNP array, mRNA expression
phs000473.v2.p2 (PRJNA157243, PRJNA94281)	Sweden-Schizophrenia Population-Based Case-Control Exome Sequencing	12 380	SureSelect Human All Exon v.1 Kit, SureSelect Human All Exon v.	22641211	[15]	WES
phs000738.v1.p1	Exome Sequencing in Schizophrenia Families	216	SeqCap EZ Human Exome Library v2.0	23911319, 24317315	[1]	WES
phs000687.v1.p1	Bulgarian Schizophrenia Trio Sequencing Study	1826	SureSelect Human All Exon v.2 Kit, SureSelect Human All Exon v3-50Mb, SeqCap EZ Human Exome Library v2.0	23040492, 22083728, 24463507	[1]	WES, SNP Genotype
phs000608.v1.p1	Whole-Genome Profiling to Detect Schizophrenia Methylation Markers	1459	MBD-seq	23244307	[42]	Methylation
phs000448.v1.p1	Genetics of Schizophrenia in an Ashkenazi Jewish Case-Control Cohort	3096	HumanOmni1-Quad_v1-0_B		[4]	SNP array
phs000021.v3.p2	Genome-Wide Association Study of Schizophrenia	5064	AFFY_6.0	16400611	[43]	SNP array
phs000167.v1.p1	Molecular Genetics of Schizophrenia-nonGAIN Sample (MGS nonGAIN)	3029	AFFY_6.0	16400611	[44]	SNP array

Proteomic and metabolomic data sets

EBI’s Metabolights and NCBI’s Metabolomics Workbench (MW) are two major metabolomics data repositories. Metabolights has no SCZ data sets; MW has one but data are not downloadable. A limited number of data sets appear to be available for other mental health phenotypes such as Alzheimer’s disease and autism spectrum disorder. PRIDE (PRoteomics IDEntifications), the leading proteomics data repository, has only three SCZ-related data sets: two in rat models and one in a mouse model. Data sets related to other MH disorders are similarly limited and largely generated from animal models.

Imaging repositories

SchizConnect is a federated portal that integrates data from three neuroimaging consortia on SCZ: FBIRN's Human Imaging Database (HID), MRN's Collaborative Imaging and Neuroinformatics System (COINS) and the Northwestern University Schizophrenia Data and Software Tool (NUSDAST) project [45].

A number of general purpose brain imaging repositories exist, such as OpenNeuro (formerly OpenfMRI) [46], and the Neuroimaging Informatics Tools and Resources Clearinghouse Image Repository (NITRC IR) [47]. However, to date, the publicly available data sets in those resources appear to be more cognition-oriented (e.g. classification learning, visual and auditory functions, attention) than psychiatric. Notable exceptions in SCZ include [42, 48].

The Functional Connectomes Project is also available through NITRC [49]. It comprises data from >1400 healthy subjects who underwent fMRI scans that assessed their brain activity when their minds were at rest. Included in the 1400 is a subset known as the COBRE (Center for Biomedical Research Excellence) data set, which includes anatomical and functional MR data from 72 patients with SCZ and 75 healthy controls. These data have been analyzed in a number of different ways by different groups [50-53].

Curated knowledge bases

SZGR [28], SZGene [44] and SZDB [54] are three distinct but significantly overlapping knowledge repositories that include curated information regarding SCZ-related genes and data sets. The SZGR, available since 2009, is a ‘one-stop shop’ for genes and variants in SCZ, along with their function, regulation and drug information. It was created through systematic review and curation of multiple lines of evidence and includes ∼4200 common mutations and ∼1000 de novo mutations [28, 55]. SZGene is affiliated with the Schizophrenia Research Forum and contains data from 1700 studies. It enables the user to search by gene, protein, polymorphism, study or keyword to return the specific publications addressing those features [44]. However, the resource only contains data from studies before 2012 and is no longer supported. (Unfortunately, this is not uncommon for resources that require maintenance over time.) Finally, SZDB includes genomic, transcriptomic, molecular network data and functional annotations [54].

The DisGeNET database (http://www.disgenet.org) integrates human gene–disease associations from various expert curated databases and text-mining-derived associations including Mendelian, complex and environmental diseases. A search in July 2017 for genes and single-nucleotide polymorphisms (SNPs) associated with SCZ yielded 1871 and 1635, respectively.

Genome-wide association studies and beyond: innovative, integrative approaches to biomarker discovery in mental health

At the most basic level, data sharing can enable integrative analysis for biomarker discovery by allowing researchers to combine two or more comparable data sets to increase statistical power through increased sample size. Researchers have developed many creative methods to combine data, enabling the discovery of patterns not apparent when analyzing just a single data type. Genotype data, including those identified through both microarrays and next-generation sequencing, can be combined with other data types such as protein–protein interaction networks, biological pathways, gene expression and co-expression, methylation and microRNA regulation data [43, 56–60]. In some cases, researchers have been able to combine three or more data types in creative ways for biomarker discovery [61, 62].

Most studies start with the purpose of discovering novel variants and then make use of public resources to validate and support their initial discovery. Methods may be categorized at the gene level [63-66], pathway level [67] and network level [68-70]. Another way to categorize the methods is based on the multi-omics data. Some integrative studies involve only genetics and eQTL (expression quantitative trait loci) data [71], while others are more comprehensive, involving multiple genome-wide association studies (GWAS) and/or other dimensional data [72, 73]. Recently, with the dramatic increase of GWAS data, especially by the PGC, an increasing number of studies have been published for integrative analyses using multiple-disorders or multiple-omics data, aiming to identify shared or unique genetic variants among different MH disorders [74]. To generate an overview of these integrative studies, we used PubMed to systematically search for integrative studies using keywords listed in Table 3. In total, we obtained 595 publications for integrative studies of SCZ. A majority of them (497) were published after 2010, likely due in part to the curation of omics data in recent years. As shown in Figure 5, the publication of integrative studies in SCZ has been increasing sharply in recent years, mostly in the category of eQTL. Recurring themes among these integrative methods include overlap between variants from different omics modalities and randomization and permutation tests for statistical significance.

Table 3

Keywords and counts for integrative biomarker studies in schizophrenia published before May 2017

Keywords	N
schizophrenia [TIAB] AND GWAS AND expression	285
schizophrenia [TIAB] AND SNP AND expression	242
schizophrenia [TIAB] AND GWAS AND network	140
schizophrenia [TIAB] AND SNP AND network	75
schizophrenia [TIAB] AND GWAS AND methylation	36
schizophrenia [TIAB] AND GWAS AND eQTL	35
schizophrenia [TIAB] AND SNP AND integrative	32
schizophrenia [TIAB] AND GWAS AND quantitative traits	26
schizophrenia [TIAB] AND GWAS AND transcriptome	26
schizophrenia [TIAB] AND SNP AND methylation	20
schizophrenia [TIAB] AND SNP AND eQTL	19
schizophrenia [TIAB] AND SNP AND quantitative traits	11
schizophrenia [TIAB] AND SNP AND transcriptome	10
schizophrenia [TIAB] AND GWAS AND integrative	5
schizophrenia [TIAB] AND SNP AND transcriptome	4
schizophrenia [TIAB] AND genotyping AND transcriptome	3
schizophrenia [TIAB] AND SNP AND ATAC-seq	1

Figure 5

Publication summary of SCZ integrative studies.

Note: Publications in 2017 were estimated based on the data between January and May in 2017.

Big big-data: large-scaleGWASs

Since 2008, GWAS have reported a number of genetic variants associated with SCZ [63-66]. The Schizophrenia Working Group of the PGC conducted the largest GWAS in SCZ to date (36 989 SCZ and 113 075 controls) and identified 108 loci [75]. The largest ancestrally and phenotypically homogeneous GWAS study of SCZ (11 260 cases and 24 542 controls) reported 50 novel SCZ risk loci [76].

Despite these large-scale studies, the genes or functional DNA elements through which these variants exert their effects remain unknown. An emerging trend is to integrate multidimensional data from genetics, epigenetics and transcriptomics to prioritize biomarkers and also to understand the underlying mechanisms by which they act.

Combining expression and sequence data: eQTL

Several large-scale studies have markedly expanded the scope of known eQTLs [77-79]. One recent example integrates SCZ postmortem brain gene expression with GWAS signals at the pathway levels [80]. Building on these eQTL resources as well as gene expression profiles, candidate genes have been identified by integrative studies [81-85]. Using expression data from 647 postmortem human brain samples collected by Stanley Medical Research Institute and the GTEx project, the gene complement component 4 (C4) in major histocompatibility complex region was identified as contributing to SCZ risk [86]. In addition, the CommonMind Consortium (CMC) generated RNA-seq data from postmortem dorsolateral prefrontal cortices from 258 subjects with SCZ and 279 controls, and they identified a list of genes whose expression was significantly affected by SCZ risk variations [87]. Using the CMC data, a list of brain splicing quantitative trait loci was identified that are causally associated with SCZ [88]. In another study, SCZ risk genes were identified using summary data from GWAS and eQTL in which gene expression data were generated from 5311 peripheral blood samples [89]. More examples include valuable eQTLs located at NMDAR [90], CTCF and CACNB2 [91], and 17q25 locus [92], most of which used public eQTL data such as those from GTEx. These eQTLs are mostly common variants. Rare noncoding SCZ risk variants were also identified [93].

Combining genomic data with multiple phenotypes: pleiotropy

Genomic data combined with multiple phenotypes can enable the discovery of pleiotropy-associated genes, i.e. alleles that impact two or more apparently unrelated effects. The PGC provides a good example of this. They initially, and intentionally, focused studies on five major psychiatric disorders: autism, attention-deficit hyperactivity disorder, bipolar disorder, major depressive disorder and SCZ [74, 94]. A number of clinical features transcend these disease classifications, and previous research had suggested overlap in familial and genetic liability for different combinations of these disorders [74].

Combining genomic and imaging data

Another common integrative approach in the study of MH is the combination of genomic data with imaging data. The Enhancing NeuroImaging Genetics through Meta-Analysis (ENIGMA) consortium provides tools and protocols to meta-analyze genome-wide and neuroimaging data from research teams worldwide [95]. The consortium does not require participating investigators to contribute raw data nor provide access to such data for research or public use. Instead, ENIGMA provides standardized protocols with predefined covariates to allow sites to conduct GWAS and imaging studies locally and report meta-analyzed data, which are made publicly available through an interactive visualization tool, ENIGMA-Vis [96]. For example, the ENIGMA SCZ working group conducted a collaborative, prospective meta-analysis of neuroimaging data from >4500 study participants (2028 were subjects with SCZ and 2540 were healthy control) across 15 sites [97].

Combining genomic and epigenetic data

DNA methylation is important for epigenetic regulation of gene expression. The 108 SCZ-GWAS risk loci identified by the PGC [75] were evaluated systematically as methylation quantitative trait loci in postmortem prefrontal cortex from 191 SCZ and 335 controls [98], 689 SCZ and 645 controls [99] and 1163 postmortem brains of European ancestry [100]. The loci were also systematically analyzed using 166 human fetal brains [101]. The evidence showed a much stronger differential DNA methylation enrichment in genes associated with SCZ, even using a medium sample size (<100) of postmortem brains [102].

Transposase Accessible Chromatin followed by sequencing (ATAC-seq) is another promising technique that can be used to map chromatin accessibility. Recently, ATAC-seq has been used to study spatiotemporal regulation of gene expression of neuronal and nonneuronal nuclei isolated from frozen postmortem human brain to map chromatin accessibility for SCZ risk loci [103].

Concluding remarks

Our review of data resources and integrative biomarker discovery in MH with a focus on SCZ suggests a recent increase in the number and quality of resources and an even more recent growth in their use. Indeed, we are starting to see high-profile papers that leverage some of these existing data sets. For example, a recent paper in Nature Genetics described a GWAS study in 36 000 individuals of Chinese ancestry that was combined with data from the PGC [75] to perform trans-ancestry meta-analyses yielding 30 novel risk loci for SCZ [104]. However, there remains significant untapped potential: many resources that could be made available under FAIR principles are not, and those resources that are available remainunderused. While we leave other MH disorders for a future paper, we do not believe the results will be significantly different for anxiety or depression, or for most other disorders in mental health with the possible exception of autism spectrum disorder, where a concerted national effort by both public and private entities has created a large concentration of findable and accessible shared data [105, 106].

The challenges facing those seeking to reuse resources for integrative research are numerous and formidable, but we believe that many seeking to enter the fray are tripping over the threshold of findability. We hope that both the suggested organizing framework and the catalog of resources presented here will help new entrants into the space cross the threshold successfully and focus on more substantive challenges of data integration and analysis. We also believe the proposed framework is useful more broadly in other biomedical domains to facilitate categorization and dissemination of information about data resources to support emerging precision medicine initiatives.

As noted recently in a memo from Dr Joshua Gordon, director of NIMH, understanding the underlying biology of MH is more important than ever, and increasingly within reach given recent technological developments [4]. A consensus on FAIR principles, the NIH push toward data sharing, and NLM support for best practices, mean those who continue to develop innovative approaches to the vast and ever-increasing amount of publicly available data will help the rest of us gather valuable insights about mental health diagnosis and treatment.

Click here for additional data file.