Charting a Course for Autism Biomarkers

Autism spectrum disorders (ASDs) are defined by qualitative impairments in reciprocal social interaction and the presence of repetitive/restricted behaviors, interests, and/or activities (1). While the precise pathologies are unknown, it is clear that ASD is the behavioral manifestation of early alterations in brain development occurring prior to the emergence of symptoms over the first 2 years of life (2). Aspirations to precision medicine for ASD (3)—getting the right treatment, to the right person, at the right time—have been hampered by a lack of biomarkers. Given the etiological heterogeneity of ASD (4), there is an especially urgent need for stratification biomarkers that can separate subjects into subgroups demarcated by common pathophysiology. Such biomarkers have the potential to maximize the capacity for mechanistically targeted interventions to benefit individuals. Equally needed are (neuro) predictive biomarkers (5) that are sensitive to change in core ASD features. These biomarkers may serve as robust predictors of treatment effects and/or help to prescribe one class of treatment over another. Overall, the lack of both kinds of biomarkers perpetuates the status quo of imprecise treatments, wasted time, and lost resources.

New results from the Infant Brain Imaging Study (IBIS) Network, led by Joseph Piven, signal the end of the status quo. The article—appearing in this issue of Biological Psychiatry (6)—represents the latest returns from the IBIS Network’s extraordinary investment of painstakingly hard and rigorous work directed at collecting longitudinal magnetic resonance imaging data at 6, 12, and 24 months of age from a total of 334 infants, 221 of whom were at high risk for developing autism due to having an older sibling with the condition. In total, 47 of these infants were diagnosed with ASD at 24 months, and their magnetic resonance imaging data were compared with those of other infants who were not diagnosed with ASD. The authors then identified a biomarker of ASD—increased extra-axial cerebrospinal fluid (EA-CSF) at 6 and 12 months—preceding the toddlers’ diagnosis at 2 years. Moreover, levels of EA-CSF at 6 months were positively correlated with the severity of ASD symptoms at 2 years. Critically, in a field where so many findings are never replicated, this study replicates and extends a separate 2013 study from a group led by David Amaral (7). In considering the impact of these new findings, let’s reflect on two Parmenidean characteristics of good biomarkers.

First, biomarkers should be grounded in an understanding of basic developmental mechanisms. The new biomarker fairs well because the prospective longitudinal design uniquely captures the developmental trajectories of the EA-CSF bio-marker. Given the vast range of other information in the IBIS Network’s longitudinal data set (e.g., longitudinal trajectories of key phenotypic characteristics, social functioning, development of neural circuits for social information processing), this new understanding of EA-CSF trajectories can be leveraged to inform the design of more effective procedures for identifying and remediating risk for ASD in certain subgroups of children. It is generally accepted that earlier interventions are more effective for treating a variety of developmental problems. EA-CSF may serve as a robust stratification biomarker that will likely improve early identification, but it might also offer advantages for earlier targeted interventions. That is, because the EA-CSF biomarker relates to the severity of specific ASD deficits, it could help to decipher the characteristic ASD heterogeneity. More targeted treatments could then be developed on a child-by-child basis and implemented earlier in ontogeny, ensuring the most effective course of intervention.

The Shen et al. study (6) was successful in large part because of straightforward and compelling methodological advantages of a longitudinal design. Neuroimaging data are inherently noisy, mostly because individual brains are different from one another. This is widely acknowledged, and it is why nearly all informative neuroimaging studies are conducted within subjects. A longitudinal design is essential to visualize actual developmental changes with reasonable sample sizes. It also makes it more likely to detect relationships between different developmental changes. For example, just knowing that performance on both task A and task B will change between 6 months and 9 months of age tells us almost nothing about those tasks, but individual differences in the response to task A at time 1, or the change in task A performance, are good predictors of individual differences in the magnitude of change in task B and will yield much stronger inferential leverage on which to build lasting theoretical contributions. The multidimensional nature (e.g., brain, genes, behavior) and longitudinal design of the IBIS Network afford exactly this kind of rich contribution to understanding developmental mechanisms (8). These results now set the stage for the opportunity to drill deeper into the mechanism. For instance, this group can next set its sights on determining why youths at risk for ASD have EA-CSF and specifying precisely where (and for whom) increased EA-CSF fits on the causal chain from pathophysiology to ASD manifestation.

Second, biomarkers should be sensitive reliable measures of targeted atypical processes that are informative at the individual level. Here too, the EA-CSF biomarker receives high marks. Shen et al. developed an automated segmentation algorithm that was found to be valid and reliable relative to gold-standard manual methods (6). This accomplishment offers the potential for developing a highly scalable approach that could be widely deployed in common medical settings. Furthermore, in addition to demonstrating group-level differences, the authors used cross-validation to model the sensitivity and specificity of their method at the level of the individual. This is absolutely essential if the biomarker is to be useful for predicting treatment response and measuring treatment-induced changes.

The EA-CSF measurements, early in development, were also predictive of individual differences in ASD symptom severity. It will be interesting to see how this biomarker fares as researchers use it outside of the high-risk/infant sibling context. Shen et al. discuss the importance of this essential next step (6). In so doing, they are recognizing the fact that the spectrum from mental health to mental illness is continuous and not categorical—Mother Nature has not yet read the DSM-5. Thus, a dimensional individual differences approach is crucial to understanding social functioning and its development. With their current study, the authors have built the methodological foundation to predict individual differences in social functioning, the core of ASD, from growth curves of developmental change in EA-CSF and associated changes in brain structure and connectivity across the first 2 years of postnatal life. This is a landmark achievement.

It will be crucial to test Shen et al.’s approach (6) in samples where infants and children are not excluded or included by virtue of possessing any particular categorical psychiatric diagnosis or by their family history of the same. This will allow for the development of much more generalizable biomarkers. Moreover, it will be important to test this approach in samples that include equal numbers of male and female subjects across levels of risk. The authors statistically covaried participant sex in their analytic models, but an inspection of the male-to-female ratios across the subject groups reveals a familiar pattern of sex differences. Among those at high risk for ASD, the male-to-female ratio for those receiving an ASD diagnosis at 36 months age was ~8.5:1. Among those who did not receive an ASD diagnosis, but who were at increased familial risk, this ratio was ~1.1:1. Among those at low (regular) familial risk, and who did not receive an ASD diagnosis, the ratio was ~1.6:1. This is striking in light of recent discoveries concerning sex differences in the neural and genetic mechanisms of ASD [e.g., (9,10)]. Furthermore, while the authors accurately report the absence of a statistically significant group × sex interaction in levels of EA-CSF, they did see an intriguing trend toward such an interaction (F_2,398 = 2.46, p = .09). Our own recent experience in biomarker development left us very cautious about presuming that biomarkers for ASD will work equally well in male and female individuals (10).

In sum, the IBIS Network has given us a national treasure, a large, carefully cultivated, well-characterized prospective longitudinal data set of infants at increased risk for developing ASD. This research marks the start of a new era in which advanced, automated, scalable neuroimaging techniques will evolve into an integral part of a translational research chain. Novel behavioral treatments and pharmacotherapies for ASD may be further developed with the tremendous benefit of directly and more precisely assessing impairment and change in neural circuits and behavior. Objective, neuroimaging-derived biological markers could someday soon be used, at the outset, to make treatment decisions related to dose, duration, intensity, and approach, as well as to the use of concurrent pharmacological intervention, specific to individuals’ profiles. Increased understanding of neural mechanisms and individual differences in treatment response, leading to the development of biomarkers for treatment stratification, will inform treatment approaches for core ASD symptoms and improve the ability to select patients who are likely to benefit from particular treatments toward the goal of personalized care. This is a transformative new era of discovery and translation.