An Autism Case History to Review the Systematic Analysis Of Large-Scale Data To Refine the Diagnosis And Treatment Of Neuropsychiatric Disorders

Abstract

Analysis of large-scale systems of biomedical data provides a perspective on neuropsychiatric disease that may be otherwise elusive. Described here is an analysis of three large-scale systems of data from using Autism Spectrum Disorder (ASD) and ASD research as exemplar of what might be achieved from study of such data. The first is the biomedical literature that highlights that there are two very successful but quite separate research communities and findings pertaining to genetics and the molecular biology of ASD. That is those studies positing ASD etiologies related to immunological dysregulation and those related to disorders of synaptic function and neuronal connectivity. The second is the emerging use of electronic health record systems and other large clinical databases to allow the data acquired during the course of care to be used to identify distinct subpopulations, clinical trajectories and pathophysiological substructure of ASD. These reveal subsets of patients with distinct clinical trajectories some of which are immunologically related and others which follow pathologies conventionally thought of as neurological. The third is genome-wide genomic and transcriptomic analyses which show molecular pathways that overlap neurological and immunological mechanisms. The convergence of these three large-scale data perspectives illustrates the scientific leverage that large-scale data analyses can provide in guiding researchers in an approach to the diagnosis of neuropsychiatric disease that is inclusive and comprehensive.

Introduction

Perhaps the most successful branch of medicine in achieving a precise diagnosis of disease, one directly linked to etiology, has been that of infectious disease. Only a little over one hundred years passed between the identification of microorganisms as the etiological agents for multiple diseases and the consequent development of dozens of therapies, in immunizations and antibiotics, that have had a greater impact on mortality and morbidity than any other medical intervention (1). It is this understanding on the consequences of etiological and precise diagnostic capabilities that were the main drivers of the recent National Academy of Sciences report on Precision Medicine: to use multiple comprehensive measurement modalities to identify which sub group of patients a given patient most resembles and therefore to be able to both assign a diagnostic label and predict clinical course in response to therapeutic intervention. I review here how a systematic approach to large-scale data can make some preliminary and illuminating strides towards a “precision medicine” of neuropsychiatric disease. I use the autism spectrum disorders (ASD) as a prismatic example of the larger opportunity by illustrating how this approach reveals two richly productive but largely separate avenues of research in ASD defined by apparently distinct mechanistic hypotheses. That is, ASD as a disorder of neural connectivity and specifically synaptic connectivity regulation (2, 3) and ASD as a disorder of immunological signaling (4-6).

First some framing is required regarding the task being addressed: diagnosis of the disorder. Here, diagnosis of ASD will be defined in the probabilistic framework used in decision making: the probability of a disease D, given the findings F summarized by the notation p(D∣F). In ASD, we often attempt to diagnose or rule-out a single disease (i.e. autism) even though it is recognized that there are likely to be multiple diseases (i.e. The set of diseases D comprised of {D₁..D_n} that together constitute ASD). A diagnosis will be more useful to the extent that p(D∣F) is high (i.e. close to 1.0) corresponding to the high likelihood of disease or low (i.e. close to 0.0) corresponding to the low likelihood of disease. Further confidence in this likelihood estimate is provided if the error of this estimate is low. The appropriateness of therapy can then be determined by how well it is matched to the disease. This thereby highlights the value of determining which of the diseases that constitute ASD of the set {D₁..D_n} have the highest probability as each therapy will have different efficacy for each of them.

The Published Literature for Large Scale Characterization of Research

In the recently published Diagnostic Standards Manual 5, Autism Spectrum Disorder is defined as including persistent deficits in social communication and social interaction and restricted, repetitive patterns of behavior, interests, or activities. This new single disorder replaces several previously defined disorders including Autistic Disorder, Asperger’s Disorder, and Pervasive Developmental Disorder Not Otherwise Specified. This redefinition will surely lead to a change in diagnosis for many individuals and possibly a change in funding of support services. The controversy that emerged prior to and after this publication illustrates the challenge posed by the diagnostic and prognostic task when applied to a disease complex that many recognize to be a constellation of heterogeneous pathophysiologies (5, 7-12), some of which have genetic etiologies, some environmental or a combination thereof. A multi-dimensional characterization of the patient population of interest, that measures the multiple genetic, molecular, clinical and environmental-exposure features of each patient to derive the overall landscape of the constellation of heterogeneous diseases that distinguish that population, provides the most comprehensive and systematic viewpoint (13). Of course, such integrative data sets are currently far and few between, with the Simon’s Simplex Collection (14) constituting an notable example of what such integration can yield (and the effort and investments required to bring it together). With the steady accretion of clinical and research data sets, we can anticipate such multidimensional assessment to grow. Therefore it will become essential to determine which of the set of diseases comprising ASD in {D₁..D_n} are being diagnostically evaluated. Merely making this determination of which diseases are being considered as part of the ASD set is challenging. This challenge is best illustrated by a large-scale database available to all ASD researchers: that of the published literature. If we focus on those recent publications that were NIH supported and therefore deposited in the Open Access Pubmed Central NIH repository (15), then as illustrated in Figure 1, not only is the primary literature balkanized, but even the citations made by the authors of this literature largely address disparate domains of biology. If we label the autism and genetics literature as pertaining non-exclusively to four sets: neuronal synaptic function (N), immunological function/disorders (I), with N cit. and I cit. denoting the literature cited by these first two sets, then as shown in Figure 1, the overlap is remarkably slight. For example, of 290 publications in N only 18 are also in I and of the 12391 cited by the publications in N, only 1551 are cited by I. At best, this suggests that either the set of findings or the set of diseases considered in developing a precision diagnosis of the ASDs is incomplete, depending on which research community is addressed. This raises the question of what population studies can reveal regarding this apparent dichotomy? By way of example, large-scale population genomics have revealed previously poorly defined or unsuspected subtypes of disease within breast cancer (16), non small cell lung carcinomas (17) and leukemia (18). However preceding the advent of genomics by more than a century, physician-scientists have used observational studies to define disease subtypes. Jean Martin Charcot, for example, systematically and comprehensively studied the patients in a large neurological hospital in Paris and was thereby able to define new and lasting disease entities out of a pool of previously monolithic and broad neurological diagnoses (19). A century and a half after Charcot, can we undertake large scale observational studies of patients enabled by the recent acceleration in electronic health record systems deployment to augment our ability to generate an integrated view of p(D∣F) for ASD?

Figure 1.

Illustration of the incomplete overlap in research of ASD genetics based on investigations of synapses and research in ASD genetics based on investigations of the immune system. Four ellipses are shown corresponding to four corpora all selected from Pubmed Central. N denotes those publications focused on genetics and synapses, I denotes those publications focused on genetics and immune system. I cit. denotes those publications that were cited by those in I. N cit. denotes those publications that were cited by those in N. The intersection between N and I accounts for only 4 percent of the combined publications and the intersection between N cit. and I cit. accounts for only 8 percent of the combined citations.

Electronic Health Records for Large Scale Characterizations

The acceleration of the adoption of electronic health records (EHR’s) in clinical care through the HITECH Act of 2009 (20) may or may not increase the productivity or safety of healthcare delivery but it certainly has provided a large source of detailed clinical documentation of patients. This enables researchers adept in the “secondary use” of EHR data to identify patients with the clinical phenotype of interest and then use the samples acquired in subsequent visits for clinical diagnostics for the purposes of genotyping, resequencing and even epigenetic characterization, as reviewed in (21, 22). In addition to structured or codified data (e.g. laboratory test, medications, diagnostic and procedure billing codes), the development of “natural language processing” (NLP) techniques (23-27) enables the narrative text of clinical notes to be mined to obtain a far more accurate phenotypic assessment of the patients than from the codified data. Given that the codified billing data is well known to be biased for reimbursement and insufficiently fine grained, this is not surprising. However, when the codified data is combined with the NLP-derived data the phenotyping accuracy is higher than with either clinical source alone (22). Furthermore, this automated phenotyping has been shown to be generalizable, portable and reproducible across healthcare systems (28, 29). These very encouraging early studies should not obscure the methodological challenges that these observational data sets entail. The time span covered by most EHRs is of short duration in most systems because of their recent adoption (30). NLP techniques currently require effortful fine-tuning based on iterative comparison of their performance selecting the “right” patients relative to that of experts manually reviewing a subset of the same records. Moreover, whereas the claims data may be biased for reimbursement, they do cover populations through the entirety of their paid health encounters whereas electronic healthcare data may have greater detail but often only pertain to a fraction of these encounters (31). For example, an academic center’s EHR may include documentation of the initial ASD diagnosis and subsequent episodes of acute morbidity. However, they often lack the documentation of the growth and development of these children noted in the community pediatric practices. All these sources of bias and complexity suggest that the use of these data requires at least as much care and multidisciplinary expertise (31) as genomic data analysis early in the adoption of a new sequencing platform.

Importantly, at a time when genomic studies of neuropsychiatric disease require tens of thousands of subjects, EHR-driven phenotyping coupled to the genomic characterization of discarded samples is one to two orders of magnitude faster and less costly in identifying patients of interest than conventional study cohort techniques (21). This EHR-driven phenotyping has been performed successfully for several neuropsychiatric phenotypes including major depressive disorder (32, 33) and bipolar disorder (34) and several groups are currently working on similar approaches to ASD. It remains however, that even for diseases that are as common as 1 in 100, any single healthcare system may not have sufficient numbers of patients to enable a statistically robust characterization of these diseases. This is even more problematic when the disease is not monolithic but rather a constellation of the many rarer diseases {D₁..D_n} of which they are composed. This can be addressed by enabling queries that cross multiple healthcare systems. For example, we have developed a system—the Shared Health Information Network (SHRINE) (35, 36)—which has been used, with appropriate governance, oversight and privacy protection measures, to issue real-time queries across multiple healthcare sites comprising records of millions of patients to both identify rare events (14) and enable phenotyping for neuropsychiatric genomic studies to occur at the scale of hundreds of thousands of individuals. SHRINE has been adopted for the Harvard affiliated hospitals (comprising 6 million patients and 10 billion facts), and the University of California for its UC REX system (37) covering over 11 million patients.

In this context, a recent use of SHRINE to study the co morbidity landscape of ASD presages future large EHR system studies of the neuropsychiatric diseases. This study, one of the largest to date, covered over 14,000 patients with ASD over a 15 year period (38) representing at least 0.5% of the hospital populations. Many of the co morbidities fall squarely into categories that are commonly thought as related to neuronal and synaptic function including increased seizure frequency (19.44%) and increased sleep disorders (1.12%), bowel disorders—excluding inflammatory bowel disease (11.74%) and schizophrenia (2.4% increasing to 8.76% after age 18). All these prevalences were highly significantly (often an order of magnitude) higher than either the general population or even utilization-matched hospital populations. Conversely, several diseases that were anecdotally reported to include increased frequency of ASD were confirmed as such, including muscular dystrophy with 5% ASD prevalence. With regard to the aforementioned divide between the immunological and synaptic studies, several diseases with an autoimmune component were identified with much higher prevalence than both the general population and the matched hospital populations: type 1 diabetes (0.67% rising to 2.08% after age 18) and inflammatory bowel disease (0.68% rising to 1.99% after age 18). There have been previously many case reports about these co-morbidities but the absence of a systematic population view has made it understandably easy to treat their biological import with some hesitation. Moreover, because a single developmental medicine specialist seeing 1000’s patients with ASD is unlikely to see more than 10 patients with IBD or type 1 DM, these claims might not be consistent with their impression of their own population. Understandably, this has lead some to question the validity these electronic health record-based diagnoses. Detailed review suggests that they are indeed valid. For example, in a comparison of “gold standard” IBD diagnoses by expert gastroenterologists, the combination of natural language processing and codified data from the electronic health record attained specificities in the 95-97% range (39).

The insight provided by a systematic population perspective is enhanced further, by having longitudinal, if retrospective, follow-up of these patients over 15 or more years (40). Just as in the early expression microarray experiments (41), the patients are hierarchically clustered together based on their similar trajectories but instead of characterized by gene expression, they are characterized by the co-morbidities noted at each six-month interval. As summarized by Figure 2 below there are at least 3 distinct clusters that are currently identifiable. One cluster is highly enriched for seizures with a prevalence as high as 80%. This is in contrast to the alternative hypothesis which would be a homogenous random distribution of epilepsy across the population with autism if the epileptogenesis was due to a common etiology across ASD. Another cluster includes individuals with increased prevalence of ear infections, sinusitis as well as multiple upper respiratory infections and (not shown) inflammatory bowel disease. A third cluster is characterized by multiple neuro-behavioral disorders such as ADHD and anxiety and at a lower frequency (not shown) schizophrenia, the latter becoming much more prevalent in this population after age 18 (42).

Figure 2.

Trajectories of Comorbidities Characterizing Three Distinct Sub-clusters of ASD Defined by Electronic Health Record Data. Shown in red are the top three co-morbidities from cluster 1 where seizures rise to a prevalence of 80%. Shown in blue are the top three co-morbidities from cluster 2 with an early childhood peak of infections. Not shown (because lower ranked) is a rise in inflammatory bowel disease that continues to rise through adolescence. Shown in green are the top three co-morbidities in cluster 3 which are characterized by anxiety and hyperkinetic activity. Not shown (because lower ranked) is a rise in schizophrenia that accelerates with onset of adolescence.

The significance of these clusters here is that they represent two important consequences for diagnosis and prognosis. First of all, they are instances, albeit preliminary, of the distinct pathophysiologies of children who all have the label of autism but in fact appear to have very different diseases. That is the patients who are members of these clusters have clinical manifestations that appear to belong to different underlying diseases in the set {D₁..D_n} currently comprising ASD. For example, cluster 3 appears much more as neuropsychiatric clinical manifestation whereas cluster 2 appears more immunological or infectious-related but all of them share in common the manifestations of autism. These immune or infection related etiologies are also supported by large epidemiological studies such as those documenting increased ASD prevalence in children whose parents have rheumatoid arthritis or type 1 DM (43) and increased ASD in pregnancies characterized by high C-reactive protein (44). Of course, these early studies at the population level are encouraging but follow-up studies are required to determine if these distinct clusters correspond to the aforementioned mechanisms previously described in the literature. The trajectories shown are also relevant in that they provide a chronological signature. So for example whereas some of the neuropsychiatric disorders appear to increase with time, some of the immunological disorders such as sinusitis and otitis media peak early in childhood. Others, such as inflammatory bowel disease, type 1 diabetes and schizophrenia increase in prevalence with age. Another contribution to diagnostic precision may be enabled by the identification of these phenotypic subclusters. Genetic studies that are focused on these subgroups, rather than the undifferentiated group of patients that fall under the ASD rubric, may provide greater biological homogeneity and therefore have higher power to find genetic contributions to risk.

EHR data sets are perhaps the fastest growing source of observational clinical data and therefore they will likely overlap and complement the membership of other cohort studies such as the Avon Longitudinal Study of Parents and Children (45). This presents at least two opportunities: the validation and calibration of findings in the EHR-derived populations against the more systematically acquired study cohorts (46) and testing the generality of those cohort studies by comparing them to geographically distant clinical populations from EHR-equipped health systems.

High Throughput Large-Scale Data for Integrative Characterization

Genome-wide assessment of genetic variation (e.g. in exome studies) and function (e.g. transcriptomic or epigenomic measures) promise an unbiased perspective on disease processes. The former captures the heritable component whereas the latter integrates environmental and genetic influences. If these are unbiased then why does the literature derived from them, as described above, appear to have such a dichotomous nature? One argument is that the underlying disorders discussed here, immunological vs. synaptic/neural-connectivity dysregulation are inhomogenously split across the environmental component and inherited component. For example it has been argued that the immunological signature is environmentally mediated rather than inherited (47). However, close analysis of the results of these high-throughput data reveals other potential reasons.

First, is a cognitive bias that results from the history and context within which a gene’s function was discovered. For example, many chemokines and other inflammatory mediators thought to be characteristic of the immune system have been now shown to be powerful and essential morphogens in the normal development of the mammalian brain (48, 49). So much so, that it is likely that if they had been first discovered by neuroscientists, they would be universally called neurokines (50). Similar arguments apply to the regulation of mTOR mediated autophagy processes which might have been labeled as synaptic pruning functions if first discovered in the CNS (51). From this perspective, many of the genes implicated in ASD have both a synaptic or neuronal connectivity function and an immunological function. For example, of the genes implicated in autism by the Simons Foundation (see Table 1), 10% of them overlap with the GO categories covering immunological function. Similarly the genes in T receptor signaling pathway overlap with 21% of the genes in the long term potentiation pathway (one of the mechanisms underlying synaptic plasticity) as do the genes in the GO immune genes. From this perspective, the ASD immunological and synaptic genetics research communities might be much closer in their focus than is apparent from their literature. It also implies that some of the inherited variation could be as easily labeled as immunological as it is labeled synaptic/neuronal connectivity.

Table 1.

Overlap between sets of genes found to be implicated in autism and those in immunological regulation and long-term potentiation.

Two sets intersected	Genes in the intersection
Gene Ontology Immune Genes (http://bit.ly/KIcYOZ) and the Simons Foundation autism gene database (https://gene.sfari.org)	ADA CD44 HLA-A C4B ITGA4 NOS2A PTGS2 IL1RAPL1 APC ITGB3 ITGB7 ADORA2A ALOX5AP HRAS ADRB2 NRP2 RPS6KA2 LAMB1 (~10% of SFARI genes)
Gene Ontology Immune Genes (http://bit.ly/KIcYOZ) and KEGG Long Term Potentiation (hsa:04720)	RAF1 PRKCA CREBBP MAP2K2 BRAF MAP2K1 RPS6KA2 MAPK3 MAPK1 HRAS EP300 CAMK2A CAMK2G PRKACA ATF4 (~21% of KEGG LTP genes)
KEGG T cell receptor signaling (hsa:04660) and KEGG Long Term Potentiation (hsa:04720)	RAF1 PPP3R2 PPP3CC PPP3R1 MAP2K2 MAP2K1 NRAS CHP2 MAPK3 MAPK1 HRAS KRAS PPP3CB PPP3CA CHP (~21% of KEGG LTP genes)

It should be acknowledged that in contrast to classical Mendelian disorders, complex diseases such as ASD are fertile ground for the cognitive biases outlined above. With so many genes in common, with the phenotypic pleiomorphy of ASD, and with multiple non-CNS immunological co-morbidities (e.g. type 1 DM, inflammatory bowel disease, rheumatoid arthritis) there are plenty of opportunities for investigators focused on a single system or single organ to observe reflections of the same genetic dysregulation, but in their tissue of interest. Likewise, the overlap is one possible explanation of why peripheral blood RNA or protein expression levels differ in ASD and non-ASD subjects (52, 53) and that the difference can be used to classify these patients characterized by many of the same pathways identified in genomic studies (7, 54).

The consequence of cognitive bias results in another kind of bias: that of narrative bias. For example, in a study summarizing thousands of findings in a whole genome study, there is inevitably a process by which the investigators will choose which mechanisms/genes are highlighted in the limited space available in their publication. In an important study of CNV’s in ASD (55), for example, enrichment was also found in major histocompatibility complex MHC-I related gene-sets as noted in the Supplementary Materials. However, the investigators understandably chose to omit the finding from their main text because it did not relate to the other molecular themes they had chosen to focus on. In the literature that then cites that article, this immunological component rarely appears if at all. The narrative bias thereby leads to another well-known phenomenon: citation bias (56) . Citation bias leads to the insular interpretation of findings that focus on mechanisms that do not fit into that bias. So for example previously early evidence of the familial clustering of autoimmune disorders in families with ASD (57) and HLA-DR4 association with ASD (58, 59) is only cited by the immunologically oriented literature in ASD.

The aforementioned balkanization of neuropsychiatric investigations may be increasingly a phenomenon of the past. Data sources such as the NDAR repository at NIMH (60) and the Psychiatric Genomics Consortium (61) provide investigators with the a set of integrative measurements previously unavailable. These more comprehensive data resources enable analyses across disorders (62-65) which allows the common and distinguishing aspect of the spectrum of these disorders to be studied phenotypically and etiologically. This broader perspective is also reflected in recent reviews (66-68) which bridge the gap illustrated in Figure 1.

Conclusion

As in many other domains of human disease, neuropsychiatric disorders are prone to the natural tendency to focus on specific aspects that do not reflect the entirety of the manifestation or mechanisms of these disorders. Here I have illustrated how three large-scale data sources: the literature, electronic health records and high throughput genome-scale measurements illustrate the extent and balkanization of the study of neuropsychiatric disease, specifically in the case of ASD. At the same time, these large-scale data sources provide the means to attain a comprehensive perspective. That is, by systematically analyzing large-scale data sources, we can identify the molecular and clinical characteristics of the disparate disorders {D₁..D_n} of which ASD serves as a unifying, if temporary, label. In doing so, we enable selectivity in our therapeutic trials and ultimately therapeutic decision process.

The three large-scale data sources discussed are only the most currently accessible of those relevant to ASD. There are several others that are highly likely to be informative. Chief among these are unbiased approaches to measuring human environmental exposures (69-71) at the population level as well as the broader instrumentation of behavioral/cognitive performance (72) which is only glimpsed during formal clinical evaluation. Such comprehensive environmental and behavioral assessments are essential if we are to understand the large proportion of the variance in the disorders that lies outside their inherited predispositions, which in the case of ASD is at least 30-40%.