Biomarkers in Autism Spectrum Disorder: Challenges, Advances, and the Need for Biomarkers of Relevance to Public Health

There are numerous hoped-for future insights into the developmental neurobiology of ASD (9), but the condition is still diagnosed behaviorally by the presence of its defining characteristics, through direct behavioral examination and historical information (3). There is, however, vast phenotypic heterogeneity, spanning the entire range of IQ and language function, with variable profiles of strengths and deficits as well as symptom characteristics; change over time; and comorbidities with common psychiatric conditions (e.g., anxiety, mood, and attentional disorders) (10). The most robust clinical features for early diagnosis of autism include reduced interaction with and attention to others, reduced attention to others’ eyes, failure to respond to the calling of one’s own name, and inability to join in imitative games and reciprocal vocalizations (11–13). These indicators are all failures in normative skills that represent milestones in the development of social interaction and social communication skills. Such failures, in turn, become causative factors in subsequent atypical developmental trajectories and in the emergence of more severe symptomatology (14).

Given these multiple layers of complexity, there has been great interest and investment in the identification of biological markers (or biomarkers) for ASD, with the hope of identifying more homogeneous groups for biological study, of aiding in diagnosis (including early detection prior to the emergence or exacerbation of symptoms), and of developing more robust and sensitive markers for individualization of treatment and for measurement of treatment response (15–20). This effort is not distinct from the search for biomarkers in other neuropsychiatric conditions, in which there is an emerging consensus that clinical phenomenology, although still the primary means for classifying individuals into diagnostic categories, does not capture biologically meaningful differentiations (21–23).

Although ASD is a strongly genetic condition, it is important to note that biomarkers need not necessarily be genetically based, familial, or trait dependent (attributes of a biomarker that is also a “developmental endophenotype” [15]). A biomarker candidate can be defined as such if it is a biological variable associated with the “disease” condition and measurable directly in a given patient or in the patient’s biomaterials through sensitive and reliable quantitative procedures (15). As such, candidate biomarkers also need not be measurements resulting from analysis of biosamples (e.g., urine, blood); they can be measurements of behavior, neuropsychological performance, or brain function.

Another important consideration in biomarker research is that the term is used as a tool to capture biological states that may serve a wide range of purposes; indeed, it is likely that there are different biomarkers for different purposes, in regard to ASD as in regard to most neuropsychiatric disorders. The Foundation for the National Institutes of Health has been a major scientific agent in defining evidentiary criteria to support the regulatory acceptance of biomarker use for a wide range of purposes, such as drug-development programs (24). These purposes may relate to indicators of susceptibility or risk, diagnosis or diagnostic subsets, change, prognosis, and prediction of treatment response (including pharmacodynamic response), among others. In ASD, despite major advances in biomarker research in multiple domains—involving, for example, genes and gene expression, metabolism, multimethod neuroimaging, and behavioral experiments using sophisticated technologies and paradigms—no candidate ASD biomarker for any of the referred purposes has as yet been identified that meets the Foundation for the National Institutes of Health’s evidentiary criteria for clinical or preclinical usage or that has gained a measure of clinical utility or viability (18).

The vast genotypic heterogeneity and complexity found among individuals with ASD tempers hopes for a near-term genetic biomarker that might be used as a signature for the condition at the level of the individual patient. Researchers have attempted genotypic subsampling by deploying a genotype-first approach in the study of phenotypic variability in ASD—involving large numbers of individuals sharing recurrent genetic variants known to increase the risk of developing ASD. The results to date have been modest, however, with resultant phenotypes still varying greatly. For example, in a large study of individuals ascertained for the 16p11.2 duplication, previously found to be associated with ASD, behavioral phenotypes ranged from asymptomatic presentation to significant disability, with only a small minority meeting criteria for ASD. Neuropsychiatric symptomatology of affected individuals was predominantly within the realms of intellectual disability, motor delays, and ADHD among children and of anxiety among adults (25). Although these findings are emblematic of the challenges in this line of research, they are not unexpected given the many obstacles resulting from pleiotropy, variable penetrance, and the likely variable additive burden accruing from multiple small-effect risk alleles and interactions thereof, among many others (26).

The vast phenotypic heterogeneity of ASD also accounts for why the identification of biomarkers for the condition has proved to be so elusive. Researchers have primarily focused on mapping biomarkers onto clinically defined categories, but the categories do not capture the multidimensional and complex clinical, cognitive, and behavioral phenotype associated with ASD and its overlap with other disorders (18). Additionally, developmentally invariant biomarkers for ASD are particularly challenging, because the phenotypic manifestations unfold over time, especially in early childhood, reflecting dynamic developmental interactions among multiple risk factors (27). Phenotypic heterogeneity is also a significant hurdle in regard to replicability and generalizability of findings emerging from studies involving small sample sizes or samples of convenience. Population-based studies involving biological or neuropsychological variables, which could achieve greater coverage of the phenotypic spectrum as well as statistical power, are still quite scarce. These examples typify a wide array of challenges slowing down advances toward reproducible ASD biomarkers, which can be briefly summarized in the following general categories.

Until fairly recently, most biomarker research in ASD focused almost exclusively on the goal of segregating individuals with ASD from other populations—the diagnostic purpose. Nevertheless, there is now a concerted effort to generate ASD biomarkers specifically for clinical trials (38) and for mechanistic studies of biological states (39). In addition, most research has focused on cross-diagnostic group results, failing to involve large enough samples to meaningfully probe performance accuracy at the individual patient level. This situation has been changing markedly in the past decade, with some studies beginning to yield encouraging results (15, 19).

Given that ASD is one of the most strongly genetic of all neuropsychiatric conditions (40), it is little surprise that ASD research has generated a large number of candidate genetic biomarkers (41). Genomewide association studies have identified, with replication, de novo variations that are strongly implicated in ASD (42, 43). Several commercial operations are already marketing genetic testing for ASD on the basis of clusters of genes associated with ASD risk and gene-expression profiles. To be sure, genetic testing is fully established as an important element of “standard of care,” with tangible benefits to patients (44, 45) and results that may be helpful in clinical management of ASD. No genetic biomarker, however, has yet been shown to be ready for screening, diagnosis, or treatment of ASD (46) beyond the discrete examples of well-known, highly penetrant genetic disorders (e.g., Rett syndrome, fragile X syndrome, tuberous sclerosis, and neurofibromatosis) (18).

The most established search for ASD biomarkers, however, has been in biochemical measurements (19). As noted, hyperserotonemia has consistently been recorded in 25%−41% of individuals with ASD (47). The pursuit of a biochemical marker such as this has engendered a great deal of clinical excitement, because such findings could lead to targeted treatments addressing the metabolic imbalance directly. This hope, however, has not yet yielded results robust enough to affect clinical practice, and psychopharmacological treatment of ASD is still symptom based, not biomarker based (48). As a sign of what the future holds, however, metabolomics as a tool for discovery of ASD biomarkers, which uses a broad range of metabolites, has already yielded high diagnostic accuracy in some studies (49). It remains to be seen whether such panels retain their sensitivity and specificity characteristics when used with larger population-based samples and in studies involving more complex samples that include individuals with a wide range of psychiatric and medical conditions.

Another class of candidate biomarkers is emerging from electroencephalography (32). Some studies have detected abnormalities as early as 6 months of age, thus raising the possibility that such a procedure could also facilitate early detection of risk, before symptoms emerge in the second year of life (50, 51). Magnetic resonance imaging has also been a fertile ground for ASD biomarker research, with measures of brain structure, function, maturation, connectivity, and metabolism beginning to yield fruit (16). It is still difficult to judge whether findings are contributing factors to the development of ASD or are a result of another underlying abnormality. Although there have been claims of ASD neural signatures for several years (52), only recently were neuroimaging findings tested specifically as ASD biomarkers (39).

In one study (53), increased extra-axial cerebrospinal fluid successfully predicted which infants at high familial risk for autism actually developed autism at the later point of diagnostic ascertainment. In another study of at-risk infants, hyperexpansion of the cortical surface area between 6 and 12 months of age preceded brain-volume overgrowth observed between 12 and 24 months, and the latter was linked to the emergence and severity of autism-related symptomatology (54). The authors then deployed a deep-learning algorithm using brain surface-area information at 6–12 months to predict ASD diagnosis at 24 months, achieving a positive predictive value of 81% and sensitivity of 88%. These impressive findings, of course, await replication in other samples, but they clearly suggest a future in which brain biomarkers could be a tool for the early identification of children with ASD.

The examples provided here are but a small sample of the large number of brain structure and function studies of candidate neuroimaging biomarkers in ASD (16, 17). Such studies have included evoked potentials, single-wave and complex EEG patterns, structural imaging, diffusion tensor imaging, resting-state functional MRI, and magnetic resonance spectroscopy. Although the emerging body of findings is indeed very promising, it is still the case that, in clinical practice, the etiologic yield of neuroimaging methods is very low (55, 56). These procedures are not clinically indicated in diagnosis or treatment of ASD without accompanying symptoms that necessitate them (e.g., suspicion of seizures).

Across the field, despite a massive number of studies (16), consensus on replicable and individual-patient neural measures mapping onto behavioral phenotypes (clinical or experimental) is still quite low insofar as their clinical utility potential is concerned. Research has yielded very consistent brain findings at the group level, however, primarily in regard to activation of circuitries associated with social function (57). This state of affairs may change as studies become more robust, with larger sample sizes, pooled data across research sites, and more standardized analytical approaches.

A major issue in ASD biomarker research, represented by the sampling of candidate biomarkers briefly outline above, is that the more cost-effective and clinically viable methods are yet to achieve performance levels that would make them clinically relevant. In addition, a small number of biomarkers that have yielded exciting preliminary results, with high levels of accuracy and predictive power, are probably too expensive or technically complex to be considered in clinical practice anytime soon. Yet it is exactly this combination of attributes—high accuracy and viable cost-effectiveness—that the kinds of biomarkers most likely to have an impact on clinical management of individuals with ASD need to achieve.

In medicine, the most widespread biomarkers result from analyses of relatively easily obtained biosamples (e.g., blood, urine). To date, metabolomics appears to hold the greatest translational promise in ASD diagnostic biomarkers (49, 58), but studies thus far have involved very small samples and await replication. Standard radiological and neuroimaging methods have, of course, also had a profound impact on biomarker discovery research in an array of medical conditions and, as described, have also shown great promise in ASD biomarker research. However, the most robust findings have resulted from procedures that would be too costly or difficult to execute in the nonspecialized clinical centers. Examples of relevant considerations include high cost of neuroimaging and viable financial coverage (in the absence of a specific clinical justification for the procedure); standardization of data collection parameters as well as of data-analytic approaches; and viability of the procedure in the case of young patients and patients with intellectual disability (in most cases, sedation might not be an ethical alternative in the absence of a well-justified clinical indication for the procedure).

None of these challenges, however, is insurmountable if robust, reliable, sensitive, and specific biomarkers are found. From a clinical standpoint, though, such biomarkers would still have to match well to current, expert-clinician-based standards of cost, access, viability, and utility in diagnostic and clinic evaluations. Therefore, for ASD biomarkers to become clinically translatable, it is very likely that they will also have to be cost-effective and viable in more general clinical settings.

Some progress along these lines has already been achieved in biomarkers for early identification and diagnosis of very young children (17). From a public heath standpoint, few—if any—topics in ASD research have been prioritized more strongly than the need for better tools and procedures for early identification and diagnosis. There is consensus about the criticality of early diagnosis (59, 60) and intervention (61, 62) to promote optimized outcomes, as well as about the need for much greater understanding of early ASD pathogenesis in social brain and social behavior (13, 63).

The American Academy of Pediatrics (59, 64) strongly recommends universal early screening for autism in the second year of life because early intervention significantly improves outcomes (65–68) and may even normalize aspects of brain function (68, 69) for children with ASD. However, with very few exceptions, early intervention typically requires the diagnosis of ASD, which, in, turn awaits the emergence of symptoms that can only be identified as early as in the second or third year of life (70, 71). Nationally, the median age of ASD diagnosis still hovers around 4.5 years (72).

ASD is also a lifetime condition associated with economic burdens of more than $2.4 million per child or family (73) and more than $120 billion per year in the United States alone, most of which is related to supports in adult life (74–76). These facts have prompted the U.S. Department of Health and Human Services’ Interagency Autism Coordinating Committee to repeatedly prioritize this domain of research, to accelerate progress (77). This consensus comes from several advances in developmental neuroscience, which, in turn, translate into research imperatives.

First, there is a need to capitalize on maximal neuroplasticity. The first two years of human life represent the period of greatest brain transformation: A newborn’s brain doubles in size in the first year of life and increases again by another 35% by year 3 (78, 79). Synaptic density, a marker of experience-dependent brain specialization, quadruples within year 1 alone, and it reaches levels 200%−300% greater than the synaptic density of an adult by the end of the third year (with concurrent and subsequent pruning and strengthening) (80, 81). Of importance, longitudinal gene expression associated with synaptogenesis over the first two years of life, over a wide range of brain structures, is characterized by maximal values across the 6- to 12-month window, then decreases drastically after 15 months, or much before the time at which ASD symptoms emerge and the condition can be reliably diagnosed (82). Therefore, treatment that is conditional on ASD diagnoses misses important windows of opportunity, and identified prodromal risks need to be addressed (13, 83–85).

Second, there is a need to better understand the first two years of life of children with autism to identify early opportunities for intervention. This consensus comes from findings suggesting that developmental disruption of early-emerging mechanisms of socialization drives pathogenesis and results in autism symptoms (84–87). For example, mechanisms such as preferential eye fixation move social and communication development forward, because the eyes and gaze are critical for extraction of socially adaptive information (88) and are potent enhancers of social-information neural processing throughout the lifespan (89, 90).

Developmental social neuroscience work in this area has already yielded some reliable and replicated quantitative markers of prodromal ASD. These span the spectrum of social-communication symptoms (84), represent variable instantiation of genetic vulnerability that likely represents dosage and timing of disruption (63), focus on skills that are present from the first days and weeks of life (89, 91–93), and are population-wide quantitative traits under stringent genetic control (85). If these findings are fully replicated in larger samples, they will strengthen the hypothesis that both symptomatology and outcome levels result from these early disruptions and, as such, might be significantly attenuated and optimized (94, 95).

There is potentially great significance to this hypothesis if fully confirmed: As noted, most cases of ASD are tied to highly complex polygenic profiles of genetic burden (7, 96). In only a small minority of cases can single gene mutations be thought of as causal in ASD (97). Thus, the notion that the vastly heterogeneous nature of ASD is well captured by a syndromewide entity (as in its current DSM formal definition [1]) and is reliably diagnosed with standardized instruments (e.g., the Autism Diagnostic Observation Schedule and Autism Diagnostic Interview [98, 99]) might be possible not because of commonalities across the hundreds of initial causes (the so called “autisms” [100]) but because of commonalities in what these causes disrupt: infant-caregiver reciprocal social engagement. This connection between infant and caregiver can be considered not only as the platform for infant survival but also as the co-opted platform for social and communication brain-behavior development (63).

In fact, disruptions in patterns of reciprocal social engagement may occur among children with unrelated medical conditions, such as prenatal and perinatal suboptimalities (e.g., extremely preterm babies [101, 102]) and diseases (e.g., congenital heart disease [103]). These children then develop ASD-related outcomes at higher than expected levels. In summary, the phenotypic heterogeneity of ASD should not surprise us given the equally baffling genotypic heterogeneity. What might be surprising is, in fact, the homogeneity found in ASD despite such etiologic complexity (63)—sufficient homogeneity to constitute one of the most robustly validated and reliable DSM diagnoses of a neuropsychiatric disorder (104).

Several developmental neuroscience research groups using cheaper and more scalable tools than electroencephalography and MRI are capitalizing on these concepts to test experimental procedures that might serve the purpose of biomarkers for early detection and diagnosis in ASD, maybe even at the population level. In one example, Pierce and colleagues (105) used eye-tracking-based measurements to show that abnormal visual preference for geometric images (relative to social images) among toddlers had both diagnostic and predictive value, in that these measurements were associated with increased symptom severity. The performance of this test, however, had very low sensitivity (21%) and displayed exceptional specificity (98%) for ASD. The authors proposed, therefore, that this low-cost procedure might be a potential biomarker for an ASD subtype with more severe symptoms.

In another example, Jones and Klin showed that eye-tracking-based measurements of social-visual engagement (the way infants visually explore, engage, and, ultimately, learn from and adapt to their surrounding world) could segregate babies who were later diagnosed with ASD from babies who were not, from as early as 2–6 months of age. These data, moreover, were predictive of both diagnostic classification and level of disability at the time of diagnostic ascertainment, at 24 and 36 months (84). Although the densely sampled design used in that study is not community viable, similar measures used cross-sectionally at the age of 21 months or so yielded similarly accurate performance (85).

Eye-tracking technologies offer several initial advantages over neuroimaging methods. Relative to EEG- and MRI-based procedures, they are cheaper to acquire, simpler to set up, and more convenient to the patient, and procedures are briefer. This allows for repeated sessions, as one might require in a clinical trial. Procedures are also more robust relative to movement or other artifacts, which makes it possible to achieve high levels of success with all age groups and levels of function, including toddlers and intellectually disabled individuals (provided there are no major impairments in sensory or motor function).

Data acquisition can be brought to acceptable standards with multiple kinds of data-acquisition technologies, with many fewer parameters requiring standardization, and validation of data can more easily be established through standard analytic pipelines (e.g., ensuring integrity of the visual system and adequacy of equipment calibration). Data-analytic approaches, however sophisticated and computation heavy, can be executed agnostic to the eye-tracking technology used (because, ultimately, eye-tracking data refer to simple space-time coordinates relative to standardized stimuli, e.g., preprogrammed static or dynamic images or naturalistic video). Although data analysis may require sophisticated approaches, including algorithms resulting from machine-learning methods, these can be automatized in pipeline fashion and thus hidden from a clinical end user who would only need access to the clinically relevant information from a patient’s data analysis relative to normative or clinical benchmarks.

In summary, although the technology, methods, and data-analytic strategies used in eye-tracking research are advanced, this procedure, at the point of “service delivery,” may amount to a noninvasive behavioral protocol administered by a technician and lasting, depending on paradigm, maybe 10–12 minutes. Current technology allows for transmission of eye-tracking data to a central location for analyses through secure web portals, where pipelines for automated data processing and analysis can be implemented. This blueprint could result in a low-cost, high-throughput biomarker.

To date, however, none of this infrastructure has been adequately validated by any of the various research groups working on this research line. Yet progress toward this goal is being made. Early intervention offers enormous potential for optimizing outcomes for children with ASD (60–62, 106, 107), although the challenges in achieving universal screening that is robust, sensitive, and specific (62) are still pervasive. Moreover, early identification holds great promise for translating screen positives into access to diagnosis and for establishing effective access to treatment and intervention (72, 108, 109). Given these possibilities, the attainment of community-viable biomarkers for early identification of ASD might be one of the most critical “game changers” in clinical management of the next generations of individuals with ASD.

Although ASD biomarker research has yielded substantial advancements to date, with candidate biomarkers spanning multiple domains of inquiry and multiple tools of measurement—from genetics and genomics to metabolomics, brain structure and function, and neuropsychological performance—none has as yet achieved clinical acceptance. Clinical management of individuals with ASD remains the domain of expert clinicians using symptom-based approaches. Genotypic heterogeneity and complexity continue to challenge progress in the discovery of genetic and genomic signatures of the condition; similarly, phenotypic heterogeneity continues to be a formidable challenge to mechanistic biological studies dependent on more homogeneous samples. However, near-term biomarker impact on efforts to optimize clinical outcomes is a reasonable expectation given more recent, cost-effective, and community-viable tools to aid in early screening, identification, and diagnosis of ASD. These, in turn, might strongly enable policy promoting increased access to early intervention, which is known to have a favorable lifetime impact. If so, this effort provides us with a promising opportunity to optimize outcomes for the next generations of individuals with ASD.