Beyond the Landmarks: Where to Next With Biomarkers of Autism?

In this issue, the Autism Biomarkers Consortium for Clinical Trials (ABC-CT) offers a landmark contribution to the field of autism science (1). Despite decades of conjecture (2) regarding the empirical status of specific neural markers associated with this clearly neurodevelopmental disorder, no study to date has employed the level of rigor, consistency, and control to directly test core psychometric requirements associated with qualification as a biomarker. The methods utilized in this consortium also offer a primer in scientific humility, as evidenced by the authors’ consistency and registered clarity regarding the criteria required to declare any assay “diagnostic,” still cautiously endorsing specific responses (particularly latency of face-specific event-related potentials [ERPs]) as promising biomarkers for incremental change that may prove useful for clinical trials. In particular, the N170 ERP, a component that is larger and faster in response to faces, the P100 ERP, a marker of exogenous, automatic, early attention, and resting power spectrum slope and gamma activity demonstrated moderate stability over 6 weeks. While group differences did not emerge for every EEG/ERP marker, face memory was associated with faster N170 latency, larger P100 amplitude, lower gamma power, and more negative resting power spectrum slope. In their report on the ABC-CT data, Webb et al. note that the stability of these ERP and EEG markers, lack of discriminant validity between groups, and correlations with phenotypic variables suggest that these potential biomarkers are not suitable for diagnosis, but may be viable inclusion or stratification biomarkers in clinical trials. With this in mind, we hope to augment this discussion of where the field may go—and what pitfalls may arise—as we move forward from this landmark.

Strengths as Limitations, Directions, and Opportunities

The rigor of the ABC-CT offers a case study in the double-edged sword of tight controls. On the one hand, confidence in signals obtained across sites can undoubtedly be stronger than even those sought through careful replication. On the other hand, questions of generalization bear consideration after this rigorous investigation. First, do these findings generalize to EEG setups other than the system used by the ABC-CT consortium, including those with sparser arrays, different administration procedures, varying signal processing pipelines, and differing behavioral protocols (3)? These methodological differences may not be trivial: prior work has shown that the error-related negativity (ERN) ERP, a neural index of self-monitoring, such as detecting errors when one commits them, which has been associated with anxiety symptoms, varies in stability by reference, baseline correction, amplitude measurement, and site recordings (4). Could an exploratory approach, such as the one used to assess the ERN, be applied to ERPs of interest in the ABC-CT to enhance the reliability of obtained signals across acquisition methods? A second issue pertains to diversity—both of stimuli and of participants. For instance, it appears that, for the face paradigm, a very small number of faces were used, to reduce variability; however, this introduces questions regarding representation across racial, ethnic, and cultural groups. Relatedly, the stimuli presumably included only neurotypical faces and facial expressions; how might these effects replicate (or not) when the facial stimuli are produced by autistic individuals, given neurotypical difficulties in reading emotions in such faces (5)? Likewise, it is well established that individuals with coarse/curly hair, as is more common in Black/African American populations, are often systematically excluded from EEG studies (6), particularly when standardized dense arrays are used. Can these paradigms be applied to sparser array setups to increase diversity?

The ABC-CT shows that it is possible to obtain reliable signals across sites using tight controls; however, evidence of the robustness of these signals will depend on future work increasing the diversity of equipment, stimuli, and participants. These questions grow larger when we begin to consider the future of ERP research (2) and accessibility of clinical trials to people from all backgrounds, where it is likely that increasing use of portable EEG setups will give way to in vivo measurement of the components identified by Webb et al. Will these components retain their psychometric strength in this setting? More generally, to what in vivo social behaviors do the obtained components map on—and what does this tell us about the ecological validity of these components?

Potential of Potentials for Clinical Trials

Of course, the stated aim of the ABC-CT is to identify biomarkers of autism that can be useful for clinical trials. Latency of P1 and N170 ERPs to faces were identified as the most promising candidates. This, though, quickly begs practical follow-up questions regarding the parameters of those clinical trials. First, especially given the fact that these biomarkers were not found to yield diagnostic differentiation, it is important to determine: clinical trials for whom? Are these biomarkers meant to be useful for those who meet diagnostic criteria for autism spectrum disorder (ASD), or a specific clinical or demographic subset of the population (e.g., older youths, or those with face memory challenges, given the phenotype correlations obtained)? Might the answer to this question vary as a function of race, ethnicity, language, or other intersectional factors that otherwise led to the historical exclusion of subgroups of the population (7)? Are these biomarkers applicable to individuals who (at specific times or in discrete contexts) pass as nonautistic (8)?

Second, what kind of intervention might be especially amenable to impacting these ERPs? While there is some evidence that the N170, at least, may show change during randomized controlled trials of performance-based social interventions for adolescents (9), these data bear replication and extension. How might these effects vary, however, as a function of age (e.g., early intervention or adult supports), and intensity (i.e., what might be the dosage threshold necessary to elicit change)?

Third, the results reported by Webb et al. offer intriguing opportunities for exploration regarding which biomarkers may translate best to which type of intervention. For instance, while face-responsive ERPs were related to several social phenotypic variables (e.g., face memory, adaptive functioning scores), suggesting that they may be promising as biomarkers of these outcomes, data loss was higher (77%) for this task, and was related to several indices of functioning. Since autistic youths with greater social functioning difficulties are most likely to need—and enroll in—social interventions, how do we resolve this confounder? For individuals with more restricted and repetitive behaviors, resting-state gamma power or power spectrum slopes may hold more promise given the decrease in data loss (86%), which was unassociated with restricted and repetitive behaviors, and therefore could potentially be used in interventions designed to enhance flexibility (10). Findings from the ABC-CT highlight that biomarkers are likely not “one size fits all,” and instead provide a first step in identifying which biomarker may be appropriate for a given intervention and a specific child.

Finally, and perhaps most important, how can we leverage biomarkers to learn about mechanism of change in interventions (11)? In the imagined scientific ideal, we would wish to say that specific ERPs tell us about specific processes in interventions (e.g., an intervention that affects social functioning by engaging and improving attention to faces may be hypothesized to yield change in a specific facial attention-linked ERP that is detectable during the course of the intervention, and in turn mediates the relationship between condition assignment and social functioning outcomes). However, the actual on-line function of any given ERP (i.e., its manifestation in vivo if obtained during real-world activities) may be far more dynamic—particularly in the course of a clinical trial—than these static constructs suggest. It will thus be vital to use both clinical trial methodologies (e.g., dismantling designs) and statistical approaches (e.g., multilevel growth models that allow for modeling within- and between-person change in tandem) that allow for testing such dynamics.

The ABC-CT supported further assessment of three paradigms (faces, VEP, and resting-state) to determine whether markers from these paradigms may be useful in clinical trials of autism. Although EEG and ERP markers from these experiments were not determined to be promising diagnostic biomarkers, many of these markers showed reliability and stability and correlated with phenotypic outcomes (e.g., face memory and parent-report measures) indicating that they may provide utility as inclusion or stratification markers; this was particularly the case for those obtained during the faces paradigm. Overall, these landmark findings from the ABC-CT both provide some foundational answers to key psychometric questions about potential biomarkers of autism, and invite many more new questions—opening new vistas of the landscape. Perhaps the search for these elusive psychiatric biomarkers can tell us more than we had thought—but also less than we had hoped. These findings fundamentally highlight the importance of further bridging the study of key complex social outcomes across levels of analysis (12) to ultimately reveal their practical potential.