The Promise of Multi-Omics and Clinical Data Integration to Identify and Target Personalized Healthcare Approaches in Autism Spectrum Disorders

Abstract

Complex diseases are caused by a combination of genetic and environmental factors, creating a difficult challenge for diagnosis and defining subtypes. This review article describes how distinct disease subtypes can be identified through integration and analysis of clinical and multi-omics data. A broad shift toward molecular subtyping of disease using genetic and omics data has yielded successful results in cancer and other complex diseases. To determine molecular subtypes, patients are first classified by applying clustering methods to different types of omics data, then these results are integrated with clinical data to characterize distinct disease subtypes. An example of this molecular-data-first approach is in research on Autism Spectrum Disorder (ASD), a spectrum of social communication disorders marked by tremendous etiological and phenotypic heterogeneity. In the case of ASD, omics data such as exome sequences and gene and protein expression data are combined with clinical data such as psychometric testing and imaging to enable subtype identification. Novel ASD subtypes have been proposed, such as CHD8, using this molecular subtyping approach. Broader use of molecular subtyping in complex disease research is impeded by data heterogeneity, diversity of standards, and ineffective analysis tools. The future of molecular subtyping for ASD and other complex diseases calls for an integrated resource to identify disease mechanisms, classify new patients, and inform effective treatment options. This in turn will empower and accelerate precision medicine and personalized healthcare.

Introduction

Complex diseases are caused by a combination of genetic, biological, and environmental factors. The determination of disease etiology necessitates the alignment of clinical phenotypes with underlying biomolecular mechanisms. Consequently, researchers have traditionally first identified distinct clinical phenotypes and then identified and compared biomolecular factors that may explain differences in disease manifestation. Biomolecular comparisons across clinical phenotype have been successful in a variety of complex diseases, such as cancer (Kehoe et al., 1999; Ringman et al., 2014; Wong et al., 2015). However, the rapidly expanding availability of genetic and biomolecular expression data from new high-throughput technologies is beginning to shift this traditional phenotype-first approach to a genetics or molecular data-first approach. A molecular data-first approach identifies recurrent genetic variants or expression patterns in order to reduce heterogeneity prior to phenotypic profiling.

Molecular subtyping through gene sequencing, gene expression, and other epigenetic and omics data has been used with great success in cancer to classify subtypes for more effective treatment, understanding prognosis, and identifying disease mechanisms. For example, major molecular subtypes in breast cancer showed marked difference in their clinical features, treatment response, and outcomes (Bertucci et al., 2012; Dvorkin-Gheva and Hassell, 2014; Engstrøm et al., 2013; Schnitt, 2010). Molecular subtypes of lung cancer determined by genetic aberrations were associated with specific tests to assign the subtype and potentially relevant therapies (West et al., 2012; Yauch et al., 2005). Three subtypes of pancreatic cancer identified by transcriptomics analysis showed differences in clinical outcomes and therapeutic response (Collisson et al., 2011). The molecular subtypes of colorectal cancer showed marked differences in survival times (Marisa et al., 2013; Phipps et al., 2015). Molecular subtypes of classical Hodgkin's disease correlated with response to therapy and clinical outcome (Devilard et al., 2002). Three subtypes of gastric cancer established from gene expression data correlated with differences in patients' responses to therapy (Tan et al., 2011). Molecular subtyping of this type is being adapted for other complex diseases such as ASD.

This approach has recently been applied to the study of neurodevelopmental disorders, such as Autism Spectrum Disorders (ASD). ASD is characterized by deficits in social communication and the presence of repetitive, restricted patterns of behavior and interests (American Psychiatric Association, 2013). This neurodevelopmental disorder affects 1 in 68 children, impacting more males than females (Baio, 2012). While individuals with ASD share a core set of features, their genetic etiology and phenotypic presentation are heterogeneous and complex (Betancur, 2011; O'Roak et al., 2012a; Sanders et al., 2012; Stessman et al., 2014). Recently, ASD was redefined in the Diagnostic and Statistical Manual for Mental Disorders, version 5 (DSM-5), to encompass previously distinct classifications of Autistic Disorder, Asperger's Syndrome, Childhood Disintegrative Disorder, and Pervasive Developmental Disorder-Not Otherwise Specified (American Psychiatric Association, 2013). In addition to concerns about reliability and validity of the previous diagnostic criteria (Lord et al., 2012; Sharma et al., 2012) and a lack of behaviorally-defined subtype specific treatments, growing genetic advances failed to find causal differences between these behaviorally-defined subtypes, suggesting instead that a general continuum of autism spectrum disorders with varying levels of severity was more appropriate (King et al., 2014).

These recent developments are prompting efforts to define the etiology of ASD more clearly by shifting research from the predominantly phenotypic classification of the disorder to a genetics-first, and ultimately a molecular data-first approach. With this broader initiative toward identifying disease subtypes using molecular data, known as molecular subtyping, distinct subtypes of ASD are being explored through genetic testing and omics data (genomics, transcriptomics, proteomics). The use of molecular subtyping and the identification of genes and other omics molecules that affect common functional networks shows promise as a means to reduce heterogeneity and explore similarities and differences between interacting genotypes (Iossifov et al., 2014; Jeste and Geschwind, 2014; O'Roak et al., 2012a; Stessman et al., 2014).

The clinical relevance of molecular subtyping relies on its ability to connect underlying disease mechanisms with clinical and phenotypic data. This is exemplified in ASD research: following the identification of recurrent gene disruptions associated with ASD, neurological and behavioral mapping is taking place through imaging and psychometric testing, identifying distinct phenotypic features that accompany a targeted genotype (Bernier et al., 2014; Frazier et al., 2014; Vandeweyer et al., 2014; van Bon et al., 2015).

The function and pathogenicity of many of the genetic mutations found in individuals with ASD are still unknown. Continued gene discovery requires large sample collections, substantial data infrastructure, multidisciplinary collaboration across research sites, and the ability to work iteratively with families to characterize the clinical presentation adequately (Stessman et al., 2014). This publication is a review of the current state of genetics, omics, imaging, psychometric, and clinical data methods as they relate to ASD subtyping. In addition, a broadly applicable approach to statistical methodology is presented, along with ASD-specific examples of the generation of biologically and clinically significant molecular subtypes. Lastly, this review discusses the need for the integration of genotypic and phenotypic data to inform personalized outcomes and treatment options for patients and their families.

Data in Molecular Subtyping

Molecular data

A broad array of genetics and omics data is being used to carry out molecular subtyping of complex diseases. Molecular subtyping is most often based upon the identification of common genetic mutations and copy number variants, as well as the patterns of gene and protein expression. Regulators of gene and protein expression and activity such as miRNA, DNA methylation, and protein phosphorylation are also increasingly being used for molecular characterization of complex diseases.

Candidate gene discovery and the identification of putative causal copy number variations mark the first steps in a genetics-first approach to subtyping ASD. Exome sequencing projects are well underway and have already identified de novo likely gene disrupting mutations (LGD) associated with ASD (Iossifov et al., 2014; O'Roak et al., 2012a, 2012b; Sanders et al., 2012). Using whole exome sequencing, Iossifov and colleagues (2014) determined that de novo mutations, including copy number variants, account for 30% of simplex autism cases (one affected individual in a family). Of the LGDs identified thus far in individuals with ASD, many belong to shared networks, indicating common biological pathways are at play (Hormozdiari et al., 2015; Iossifov et al., 2014; O'Roak et al., 2012a, 2012b). The use of targeted sequencing, such as molecular-inversion probes (MIP), provides an efficient, cost-effective way to resequence recurrent genetic events in larger populations (O'Roak et al., 2012a; Turner et al., 2009).

However, individual sample collections often lack the size to reach statistical significance; large samples of affected and control subjects are needed. This has prompted multi-site collaborations, such as the Autism Sequencing Consortium, which involves researchers who have agreed to share data in order to determine genetic markers more quickly (Buxbaum et al., 2012; Stessman et al., 2014). Patient–clinician–researcher networks, such as that of the Simons VIP Consortium (Simons VIP Consortium, 2012), have been developed for recurrent copy number variations and single gene disrupting mutations associated with developmental disorders (e.g., 16p11.2) in order to determine clinical profiles through comprehensive behavioral phenotyping and neuroimaging. It is the hope that similar research networks can be developed for specific candidate genes as pathogenicity is confirmed.

The utility of identified ASD-associated gene disrupting mutations remains incomplete without further understanding of how these gene disruptions impact transcription, protein expression, and biological pathway modulators. Understanding the molecular mechanisms involved in ASD is the subject of transcriptomic and proteomic research, both of which have indicated abnormal neuronal development and inflammation (Broek et al., 2014). Transcriptomic studies, which use cDNA microarray and RNA sequencing, have identified dysregulated hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptors in individuals with ASD, which have known functions in synaptic plasticity (Isaac et al., 2007; Purcell et al., 2001; van Spronsen and Hoogenraad, 2010). Disruptions to the gamma-aminobutyric acid (GABA) receptor system, which is involved in signal transduction and cellular adhesion, have also been found in ASD samples (Broek et al., 2014; Voineagu, 2012; Voineagu et al., 2011).

Proteomic studies have investigated protein expression in serum, plasma, and postmortem brains of individuals with ASD using techniques such as mass spectrometry and immunoassay (Broek et al., 2014; Schwarz et al., 2011; Taurines et al., 2010). Studies have found increased levels of proteins involved in synapse formation, such as brain-derived neurotrophic factor (BDNF) and glial fibrillary acidic protein (GFAP) in ASD samples (Chauhan et al., 2011; Correia et al., 2010; Fatemi et al., 2002; Riikonen, 2003; Schwarz et al., 2011; Thanseem et al., 2012). Proteomic analyses have also found altered levels of immune system-regulating proteins, such as apolipoprotein in the cerebrospinal fluid and blood of individuals with ASD compared to controls (Corbett et al., 2007; Molloy et al., 2006; Woods et al., 2012; Zimmerman et al., 2005). Proteins show great promise as useful biomarkers for ASD, but large-scale replications are needed. Importantly, despite group differences between individuals with ASD and typically developing counterparts, significant variability is observed within the ASD group underscoring the relevance of using this variance to define subpopulations within ASD.

Clinical data

Molecular subtyping via genetics and omics data gains clinical significance and utility when associated with phenotypic and clinical features. Genetic sequencing provides a “sieve” for the vast heterogeneity of complex diseases such as ASD, sorting individuals at the molecular level prior to phenotypic profiling, thus reducing heterogeneity and simplifying the phenotypic subtyping process (Stessman et al., 2014). While secondary in a genetics-first approach, comprehensive standardized psychometric testing gathers critical information on cognitive, adaptive, behavioral, motor, and neuropsychological levels of functioning. Once recurrent genetic loci have been identified through genetic testing, individuals with ASD and an identified genetic event should be evaluated clinically in order to determine whether phenotypic characteristics suggest a unique ASD subtype (e.g., CHD8) (Bernier et al., 2014). Previous associations have been found between genes associated with ASD and intellectual disability (Kaufman et al., 2010), and already specific gene disrupting mutations identified through molecular subtyping show co-morbidity with significant cognitive deficits (e.g., ADNP, PTEN) (Frazier et al, 2014; Helsmoortel et al, 2014; Vandeweyer et al, 2014).

Clinical phenotyping should also include imaging and physical examination, with an emphasis on head circumference and physical dysmorphology, as these physical measurements are easily, consistently, and reliably collected in clinical settings (Stessman et al., 2014). These latter observations are often part of patient medical records, underscoring the utility of integration of electronic medical records with omics data to parse the heterogeneity of neurodevelopmental disorders such as ASD. Individuals with recurrent disrupted genes belonging to a beta catenin/Wnt signaling-associated protein–protein interaction network show variations in head circumference by subset; macrocephaly is found in individuals with PTEN and CHD8 while microcephaly is predominant in individuals with DYRK1A mutations (Bernier et al., 2014; Frazier et al., 2014; O'Roak et al., 2012a).

While the extent of known genotype–phenotype connections are still limited to small samples sizes, strong associations between identified ASD-associated genes and distinct cognitive and physical phenotypes indicate successful first steps in molecular subtyping. Longitudinal data is needed to better understand the genetic contributions to phenotypic presentation over the course of development. Understanding long-term impact of gene and protein disruption due to disruptive mutations is essential to identifying additional medical and psychiatric risks and informing treatment plans across the lifespan.

Imaging data

Over the past decade, imaging studies of children and adults with ASD continue to investigate whether there is common convergence within the structural and/or functional brain, despite the known heterogeneity of ASD (Anagnostou and Taylor, 2011). A variety of technologies are used, including functional and structural magnetic resonance imaging (fMRI; sMRI), electroencephalography (EEG), and functional near-infrared spectroscopy (fNIRS). The primary goal of ASD imaging research has been to identify biomarkers that are strong diagnostic indicators (Ruggeri et al., 2014). Several studies have implicated possible robust biomarkers, including hypoactive social and language brain areas (Carter et al., 2012; Pelphrey et al., 2011; Williams et al., 2013), and atypical EEG rhythms and components (Bernier et al., 2007; Maxwell et al., 2013; Oberman et al., 2013; Webb et al., 2012). However, many imaging studies report mixed findings, such as conflicting reports of long-range brain connections that suggest under-connectivity (Just et al., 2004) versus reports of over-connectivity (McFadden and Minshew, 2013).

To address these potential inconsistencies, recent work is focusing on how molecular and genetic subtypes may impact patterns of brain activation. In one study using sMRI (Qureshi et al., 2014), a mirror phenotype of brain volume was observed for individuals with ASD and a copy number variation within the 16p11.2 locus, such that compared to controls, deletion carriers exhibited increases and duplication carriers exhibited decreases in brain size. Other work has associated autism risk genes to structural and functional brain connectivity (e.g., CNTNAP2) (Dennis et al., 2011; Rudie et al., 2012). These studies highlight the importance of specifying ASD subtypes when investigating neural biomarkers of ASD, considering the differences that are likely derived from the etiology of subgroups of ASD.

In addition to characterizing atypical neural patterns of individuals with ASD, other recent imaging initiatives seek to detect neuroendophenotypes in individuals with a typical clinical phenotype but also a genetic vulnerability for developing ASD (e.g., unaffected relatives). Neuroendophenotypes are heritable indicators that persist in unaffected individuals, regardless of whether the pathology developed (Gottesman and Gould, 2003). In other words, this approach provides an opportunity to observe the neurobiological mechanisms by which high-risk individuals (e.g., siblings of children with autism) overcome genetic susceptibility (Constantino et al., 2010). For instance, unaffected children who have a sibling with ASD exhibit increased activation within key social perception brain regions, such as the superior temporal sulcus and ventromedial prefrontal cortex, above and beyond typically developing children (Kaiser et al., 2010). Similarly, other studies have targeted other social brain regions, such as the amygdala, as being functionally and structurally distinct for unaffected siblings (Dalton et al., 2007; Segovia et al., 2014; Spencer et al., 2011). These compensatory mechanisms may highlight areas of strength for unaffected relatives that can be used for targeted treatment of children with ASD and may better elucidate molecular etiologies in tandem with state-like biomarkers of ASD.

Data resources

An increasing array of publically available genetics and omics data is helping to greatly expand the use of molecular subtyping in complex diseases. This is most readily apparent in cancer where the Cancer Genome Atlas (TCGA) (McLendon et al., 2008; TCGA Network, 2011, 2012) maintains a repository of omics data including sequencing, gene and protein expression, SNPs, miRNA, and methylation for thousands of tumors across dozens of types of cancer. GEO and Array Express maintain huge repositories of gene expression data (Barrett et al., 2010; Kolesnikov et al., 2015) and other repositories are now storing raw proteomics data (Farrah et al., 2014; Vizcaíno et al., 2013). Other resources such as the Multi-Omics Profiling Expression Database (MOPED) are also processing these data to present more accessible and standardized views of gene and protein expression data (Higdon et al., 2014; Kolker et al., 2012; Montague et al., 2015, 2014).

In the field of autism research, large databases such as the National Database for Autism Research (NDAR, ndar.nih.gov) have been developed to house multidisciplinary biomolecular data, including exome sequencing, brain imaging, and clinical diagnostic data. A large-scale genome mapping study, AUT10K (funded by Autism Speaks) is using Google Cloud to manage, analyze, and disseminate its data, which provides new opportunities for broader access. However, the utility of these resources is significantly limited due to the lack of detailed clinical and phenotypic data being explicitly linked to the molecular data.

Summary of data in molecular subtyping

With rapid advances in genetics, omics, imaging, and clinical research on ASD, the need for collaborative data sharing and integration becomes imperative for building a comprehensive understanding of ASD etiology. In order to initiate effective avenues for data integration, a clear understanding of the commonly used methods and resources for data sharing and analysis in molecular subtyping is needed, as they relate to subtype identification and patient classification (Dumbill and Kolker, 2013; Field et al., 2009; Higdon et al., 2013; Kolker and Stewart, 2014). Figure 1 shows a sample database schema for integrating ASD data that can be used for the generation of molecular subtypes.

FIG. 1.

A model schema showing data used for molecular and disease subtyping in ASD.

Data Analysis and Integration for Molecular Subtyping

Figure 2 outlines an analysis approach to generating molecular and disease subtypes. This subtype and analysis approach can be used to characterize disease mechanisms, relate molecular data to disease phenotypes through clinical and environmental data, classify new patients, and identify optimal and potential new treatments for different subtypes.

FIG. 2.

Approach to molecular and disease subtyping in ASD.

Classifying patients into subtypes

Following molecular data collection, the methods used to classify patient genetic data into molecular subtypes vary greatly depending on type and amount of data. When subtyping by identified genetic mutations or CNVs, which are often sparsely distributed across patient populations and rarely occur with high multiplicity in any individual patients, the mere presence of these mutations is enough to generate molecular subtypes. However, differential gene expression and common biological pathways can make phenotypic manifestations difficult to differentiate at a clinical subtyping level (Stessman et al., 2014). In the case of ASD, it is anticipated that while some gene mutations may be highly penetrant and possess distinct clinical features indicative of syndromic subsets of ASD, other mutations are linked to common pathways such that phenotypic profiles are highly interconnected and more difficult to tease apart (Stessman et al., 2014).

Due to this variability in genetic background, gene expression data obtained from microarrays or RNA-Seq provide a more quantitative basis for generating molecular subtypes. The use of clustering methods with gene expression data is well established and widespread (de Souto et al., 2008; Eisen et al., 1998). The simplest and most commonly used approach is hierarchical clustering where patients are iteratively grouped by using a distance metric based upon expression values. This approach has been used in many previous molecular subtyping studies (Prat et al., 2010; Rouzier et al., 2005; Sørlie et al., 2001). Iossifov and colleagues (2014) clustered functional classes to determine enrichment of LGDs in individuals with ASD and their siblings in the following functional domains: Fragile-X mental retardation protein (FMRP) target genes, chromatin modifying genes, genes encoding postsynaptic density proteins, and those expressed in embryonic development.

There are many other clustering and unsupervised learning approaches, two of the most frequently used for molecular subtyping are k-means and self-organizing maps (SOM) (Borkowska et al., 2014; Hartigan and Wong, 1979; Kohonen, 1989). These approaches can often create more definitive and interpretable clusters than hierarchical methods. An approach that has become popular in molecular subtyping for cancer is consensus clustering. In this approach, a consensus is created from repeated clustering using multiple subsets of data and different numbers of clusters (Monti et al., 2003).

Typically, gene expression datasets contain data on thousands of genes, which create a level of noise and complexity that can make clustering algorithms inefficient and results hard to interpret. This has led to the use of data reduction methods such as principal components analysis (PCA) (Yeung and Ruzzo, 2001) and partial least squares (PLS) (Nguyen and Rocke, 2002) to reduce the dimension of data. Other methods limit the number of feature or genes used for clustering, creating both tighter and easier to interpret clusters; among these are sparse clustering, gene shaving, and dendogram sharpening (Hastie et al., 2000; Stanberry et al., 2013, 2003; Witten and Tibshirani, 2010).

The methods used to analyze gene expression have been adapted for other types of data platforms as well, such as miRNA, methylation, SNP array, proteomics, and metabolomics. Difficulties arise when the generation of subtypes needs to be based upon multiple platforms. Clustering can be applied individually to different sets of data and then the clusters can be merged using various ad hoc approaches. Alternatively, clustering can be applied jointly to different sets; indeed, several approaches have been developed to do that (Shen et al., 2010). All of these approaches will benefit for consistently processed, normalized, and analyzed expression data (Holzman and Kolker, 2004; Kolker et al., 2011).

Characterizing shared pathways and mechanisms in molecular subtyping

The use of clustering algorithms applied to molecular data can easily generate subgroups, but these are of little value if they do not help characterize the underlying disease mechanisms. Once molecular subtypes are established, they are most often characterized by describing their common genetic mutations and molecular features or by the over or under expression of specific molecules (genes, proteins, or other omics).

As mentioned previously, in the case of ASD, further disease mechanism characterization is performed by linking different genes according to their involvement in biological pathways or interaction networks (Hormozdiari et al., 2015; Iossifov et al., 2014). Many public (e.g., Reactome, Panther, BioCyc, KEGG) (Caspi et al., 2014; Croft et al., 2014; Kanehisa et al., 2014; Mi et al., 2013) and commercial pathway resources (e.g. Ingenuity) are available, as well as interaction databases (e.g., String, Intact) (Franceschini et al., 2013; Kerrien et al., 2011). Some methods take advantage of these resources to jointly combine pathway information with clustering in order to generate subtypes that are more directly interpretable and connected with existing pathway knowledge (Milone et al., 2014). Joint clustering of genes and patients may be able to identify sets of genes operating in concert outside of known pathways (Shen et al., 2010).

When subtypes are based on expression data, conventional analysis tools can be applied to molecular subtypes in order to characterize them. This includes identifying genes (or other molecules) that are highly differentially expressed across subtypes. Popular models for differential expression analysis include linear models for microarray analysis (LIMMA) and significance analysis of microarrays (SAM) (Smyth, 2004; Tusher et al., 2001). For example, a study by Zeidan-Chulia and colleagues (2014) used LIMMA modeling to identify altered expression of Alzheimer's-related genes in the NOTCH and Wnt signaling cascades in a sample of individuals with ASD, particularly the downregulation of mitochondria-regulating genes.

In addition, gene set approaches (Subramanian et al., 2005; Wu et al., 2010; Wu and Smyth, 2012) can be used to identify important pathways or networks that differ across subtypes in their expression pattern. Improvements to these approaches incorporate pathway structure to help identify important sub-pathways as is done in the Differential Expression of Pathways (DEAP) method (Haynes et al., 2013). It is important to note that if the same data are used for characterizing subtypes as was used for generating subtypes, then measures of statistical significance (p-values, etc.) will be highly biased.

Statistical models for connecting molecular subtypes with clinical phenotype

In order to understand and treat a complex disease such as ASD, molecular characteristics must be linked with the clinical manifestation of the disease. Standard models can be used to compare clinical data across established molecular subtypes, including linear (for continuous data), generalized linear (categorical, dichotomous, ordinal or count data), or Cox (for survival or time to event data) models (Cox and Oakes, 1984; McCullagh and Nelder, 1989). These models can be adjusted for the interaction of molecular data with environmental data (race, gender, parental factors, exposure during pregnancy) by adding terms to the model.

An important aspect of complex diseases such as ASD is progression of the disorder over time. Models that can incorporate longitudinal data, such as linear or generalized linear mixed models, are important for identifying mechanisms related to progression of the disorder or response to treatment (Breslow and Clayton, 1993; Laird and Ware, 1982). Longitudinal models incorporating molecular data have begun in studies on the progression of ASD in early development, such as the work of Glatt and colleagues (2012), who associated mRNA biomarkers in infants who showed early signs of ASD between 12 and 36 months of age. This ability to detect ASD subtypes during early infancy will accelerate opportunities for immediate and personalized treatment decisions. Other methods have been developed for complex diseases that combine the generation of subtypes with analysis of clinical data, such as survival, allowing for analysis when subtypes are not pre-defined (Bair and Tibshirani, 2004).

New patient classification and treatment identification

Molecular subtyping can greatly enhance the classification of new patients to disease subtypes, a crucial step to understanding a prognosis and appropriately targeting treatments. Since new patients will often have more limited data than the set patients used to generate the molecular subtypes, new classifications need to be based on simpler, more easily obtained data. In some cancer subtyping, small panels of biomarkers are used to classify patients to subtypes (Bastien et al., 2012; Choudhury et al., 2015; Taylor et al., 2012). Previous analyses can be used to help identify sets of potential predictor variables. The data can be randomly divided into training and test sets to help build and validate the model. If the amount of available data is small, cross-validation approaches can be used to evaluate the model. A wide range of supervised learning models are available to build a classifier such as logistic discriminant analysis, support vector machines, nearest neighbor and Bayes classifiers (Hastie et al., 2009). Software such as WEKA (Hall et al., 2009) can implement a wide range of models, so both the predictors and the type model can be easily evaluated. Additionally, electronic medical records can be a simple, more easily obtained resource that can be mined for the validation of subtype classification.

Two of the key objectives for molecular subtyping are early identification of patients (i.e., early infancy) to target the most effective treatment for individuals and to identify new treatment targets. Genes and pathways identified in the characterization of subtypes can be tied to different pharmacological databases such as PharmGkb and DrugBank (Law et al., 2014; Whirl-Carrillo et al., 2012). These resources can be used to identify compounds that inhibit or induce specific gene or protein expression or that block or stimulate specific pathways. Clinical data regarding response to treatment can be related back to molecular subtypes to identify those with the best responses to particular treatments. Linking longitudinal data to etiologically-derived subtypes can reduce the variability in phenotypic presentation, unmasking common behaviors or biomarkers that aid in diagnosis and predict prognosis.

Genetics has already informed pharmacological treatment exploration for ASD (Jeste and Geschwind, 2014). For example, CNTNAP2 variants have been associated with ASD and other neurodevelopmental disorders (Alarcon et al., 2008; Arking et al., 2008); this variant has been shown to have increased expression in frontostriatal circuits of the brain (Abrahams et al., 2007). CNTNAP2-mutant mouse models, which present ASD-like symptoms, have shown alleviated repetitive behaviors, but no change in social deficits when treated with risperidone, a dopamine antagonist (Penagarikano et al., 2011; Penagarikano and Geschwind, 2012). As molecular subtyping continues in the field of ASD research, it is expected that personalized treatment will emerge for distinct molecular subtypes, better equipping medical professionals to address symptoms of ASD, as well as co-morbid conditions.

Emerging Molecular Subtypes in Autism

Even in its early stages, molecular subtyping with multiple integrated data methods has been shown to be successful in ASD research (Bernier et al., 2014; Frazier et al., 2014; Vandeweyer et al., 2014). Using exome and targeted sequencing technology, recurrent de novo disruptive mutations, such as CHD8, ADNP, DYRK1A, and PTEN have been found in individuals with ASD (Iossifov et al., 2014; O'Roak et al., 2012a, 2012b). Comprehensive phenotyping of a growing number of individuals with these disruptive mutations indicates a high likelihood of autism and unique medical, psychiatric, and morphological characteristics that suggest specific genetic subtypes for ASD (Bernier et al., 2014; Frazier et al., 2014; van Bon et al., 2015). Autism, intellectual disability, and dysmorphic features have been found in individuals with a disruptive mutation to ADNP, along with multiple reports of visual and cardiac defects (Vandeweyer et al, 2014; Helsmoortel et al, 2014).

Individuals with mutations to CHD8 have enriched instances of chronic gastrointestinal complications, distinct facial dysmorphology, and macrocephalic head size (Bernier et al, 2014). In contrast, those with DYRK1A mutations have greater likelihood of microcephalic head size and early growth difficulties (van Bon et al, in press). In a recent study by Frazier and colleagues (2014), individuals with PTEN mutations showed abnormal white matter brain volume in addition to autism symptoms. These data, while still preliminary due to small sample sizes, suggest that the genetic heterogeneity of autism can be successfully reduced to smaller subsets of recurrent disruptive mutations in biologically related networks (Stessman et al., 2014). Figure 3 shows an example comparative analysis of gene expression across different clinical features and phenotypes for molecular subtypes of ASD.

FIG. 3.

Comparison of expression levels across clinical features and phenotypes of mutated genes characterizing different molecular subtypes of ASD.

Future Directions

The potential of molecular subtyping for advancing research, guiding personalized treatments, and providing a clear understanding of disease prognosis and progression has clearly been shown in diseases such as cancer and is emerging in ASD. Further, the emergence of family groups and communities centered on etiologically driven subgroups of neurodevelopmental disorders, such as Simons VIP Connect (www.simonsvipconnect.org) (Simons VIP Consortium, 2012), allows for improved quality of life for families. However, a broader use of molecular subtyping is impeded by data heterogeneity, diversity of standards, ineffective analysis tools, and a lack of rich clinical phenotypic and clinical data linked to molecular data. This limits the reproducibility and usage of subtypes across patients, experiments, and diseases.

Challenges arise when trying to generate robust and reproducible molecular subtypes from different experiments and datasets with widely varying data types and experimental protocols. This issue has impacted the generation of molecular subtypes for diseases such as ASD (Stessman et al., 2014). In addition, as new data and methods become available, the need to update subtypes must be addressed.

Achieving reproducible and robust molecular subtyping will require resources and technology that provide data integration using community standards, proper normalization, and sufficient meta-data across different omics (Chain et al., 2009; Dumbill and Kolker, 2013; Field et al., 2009; Galperin and Kolker, 2006; Garrity et al., 2008; Higdon et al., 2013, 2008; Hogan et al., 2006; Holzman and Kolker, 2004; Kolker and Stewart, 2014). These resources should enable molecular subtyping based upon only the highest quality data, experiments, and standards, use only the most robust models and analysis methods, and facilitate validation on sets of very well characterized subjects (Hather et al., 2010; Higdon et al., 2004; 2007; 2011; Higdon and Kolker, 2006; Kolker et al., 2011). Robust models need to be available to accurately predict the subtypes of new patients. Finally, molecular subtypes need to be connected to rich, but easily attainable, clinical and phenotypic data so that molecular subtyping can be used to create personalized treatments for patients. The development of such resources to assist the molecular subtyping of ASD has the potential to accelerate both the classification of new patients and the development of treatment regimens tailored to the specific presentation of a given subtype. As a result, these resources will empower and accelerate precision medicine and personalized healthcare and will ultimately serve families in more proactive and comprehensive ways.

Acknowledgments

We thank Maggie Lackey for her critical reading. This article was supported by National Science Foundation under the Division of Biological Infrastructure [0969929]; The National Institutes of Health R01-HL-060666-10A1; The Robert B. McMillen Foundation; Seattle Children's Research Institute-Northeastern University [to EK]. Additional support was provided by the National Institutes of Health R01 MH101221, R01 NIMH MH 092367, R01 MH100028, and R01 MH100047 [to RB].

Author Disclosure Statement

The authors declare that there are no conflicting financial interests.