The Clinical Problem: High Heritability, Low Predictability
Every SLP, OT, or school psychologist who has sat with the family of a newly-diagnosed toddler's younger sibling has faced the same question: what should we be watching for? It is tempting to want a clean answer—a gene, a behavior, a single confirmatory sign that tells a team when to act. The genetics say this instinct is understandable but the wrong ask. A 2013 review, Exploring Links between Genotypes, Phenotypes, and Clinical Predictors of Response to Early Intensive Behavioral Intervention in Autism Spectrum Disorder, lays out just how strong the genetic signal is: monozygotic twin concordance runs 70–90%, dizygotic twin concordance is around 10%, and first-degree relatives carry more than a 20-fold increase in risk compared to the general population. As the review's own authors put it, autism spectrum disorder is amongst the most familial of psychiatric disorders.
And yet, as the same review states plainly, the relationship between different aspects of the behavioral and cognitive phenotype and their underlying genetic liability is still unclear. High heritability tells a family that risk is real and elevated. It does not tell a clinician which child in front of them will be affected, when signs will appear, or what those signs will look like. That gap between heritability and predictability is the whole clinical problem, and it is caused by heterogeneity operating at two levels at once—genetic and phenotypic.
Why One-Gene, One-Pathway Models Fail
The 2013 review proposes one influential model to explain how genetic vulnerability becomes a clinical phenotype: building on genetic risk, a deficit in attention to social stimuli leads to impaired interactions with caregivers, which in turn disrupts the developing neurocircuitry for social cognition, which then affects downstream domains like language. It is a plausible cascade, and it is consistent with better outcomes from earlier treatment. But notice what the model itself predicts: because the cascade can be entered or interrupted at different points, and because the underlying genetic liability is not one variant but many, the clinical phenotype that results is necessarily heterogeneous. There is no reason to expect one gene, one brain pathway, or one behavioral marker to capture it.
This is exactly why researchers have turned to endophenotypes—intermediate, more measurable traits (a neurocognitive score, a symptom subscore, a brain response) that sit between genotype and the full clinical diagnosis, in the hope that they carve autism into more genetically homogeneous pieces than the diagnostic category itself does.
The Endophenotype Strategy in Practice
A 2022 study, Quantitative Trait Locus Analysis for Endophenotypes Reveals Genetic Substrates of Core Symptom Domains and Neurocognitive Function in Autism Spectrum Disorder, shows what this strategy currently delivers. Drawing on the Autism Genetic Resource Exchange—demographic and phenotype data on 11,961 individuals, with whole-genome sequencing available for 3,833 of them through the MSSNG and iHART datasets—the researchers ran genome-wide association analyses across 29 endophenotype scores and roughly 0.58 million common genetic variants. They found nine variants significantly associated with six endophenotype scores covering neurocognitive development and core symptom severity. The strongest and most specific result: multiple variants in the VPS13B gene were significantly associated with the Stereotyped Behaviors and Restricted Interests score on the Autism Diagnostic Observation Schedule, Module 3. VPS13B is already known as the causal gene for Cohen syndrome and a candidate gene for syndromic autism, so this is a biologically plausible hit, not a fluke.
But the authors are explicit about the limits: the effect sizes are small, the sample is moderate, and the findings require validation in another cohort before anyone should treat them as established. Their own conclusion is worth repeating to a team that wants a genetic shortcut: these candidate genes may be responsible for specific traits that constitute core symptoms and neurocognitive function of ASD rather than the disorder itself. In other words, the best current genotype-endophenotype work explains a sliver of variance in one symptom domain in one gene—not autism as a diagnosable category, and not any given infant's trajectory.
What Infant-Sibling Studies Show: More Than One Pathway
If genetics does not resolve into one pathway, does behavior in infancy? Prospective studies of infants with an older diagnosed sibling—who carry elevated familial risk and can be followed from birth before diagnosis is possible—are the main tool for answering this. A 2014 review, Developmental Pathways to Autism: A Review of Prospective Studies of Infants at Risk, frames the goal of this literature as characterizing how ASD symptoms emerge from a complex interaction between pre-existing neurodevelopmental vulnerabilities and a child's environment, modified by compensatory skills and protective factors—not from a single fixed unfolding. Within that frame, it reports that children later diagnosed with ASD show social and communication difficulties emerging in the second year of life, and that early neurocognitive markers include atypical neural response to gaze and slowed disengagement of visual attention. The review is explicit that this is early-stage work: it identifies methodological challenges still unresolved in the field and calls mapping how ASD unfolds from birth central to improving both identification and the interventions available once a child is identified.
A separate 2014 review of the same infant-sibling literature, From Early Markers to Neuro-Developmental Mechanisms of Autism, complicates the tidy social-cascade story directly. It reports little evidence for decreased social orienting or social motivation in the infant cohorts it reviewed—a finding that sits uneasily next to a model built around early social-attention deficits. Instead, it finds evidence for atypical development of sensory and attentional systems, and argues the field should move away from models that localize autism to a single "social brain" region toward models allowing for brain-wide involvement. Its own stated conclusion is careful and non-specific: there is some evidence for multiple developmental pathways to autism. It does not say what those pathways are, and it does not sort individual children into named subgroups.
It is tempting to read these two reviews together as evidence for two clean subgroups—one social-attention-led, one sensory/attentional-led. That is an inference on our part, not a finding either paper reports, and it deserves to be labeled as such. What the two reviews actually license is a narrower claim: markers are not confined to the social domain that the 2013 cascade model predicts, and when a study fails to find a social-attention deficit in a cohort, that does not mean the cohort is developing typically—divergence may be showing up in a different domain, such as sensory or attentional processing, instead. That is a real and clinically useful point. It is not the same as saying the field has identified two distinct presentations.
Where the Evidence Disagrees
It is worth being explicit about the disagreement rather than smoothing it over, because this is where a team's assumptions are most likely to go wrong. The 2013 genotype-phenotype review's cascade model treats an early deficit in social attention as the engine that drives everything downstream. The 2014 infant-sibling mechanisms review finds little evidence for that specific deficit in the samples it examined, and instead points toward sensory and attentional atypicality as at least as important. These are not the same claim, and they should not be merged into a single "autism starts with social attention problems" narrative.
There is a second tension worth naming directly. The 2014 mechanisms review's own highlights state that behavioral and brain markers differentiate infants who go on to develop autism from controls during the first year of life—markers exist that early, at the group level. A separate review of the broader infant-sibling literature—not one of the four studies formally reviewed here—cautions that few markers have been identified before 12 months and that the markers identified at 12 months are not good predictors at an individual level. These two statements are not contradictory: a marker can separate a group of at-risk infants from controls on average while still being too noisy to predict what one specific infant will do. Conflating group-level differentiation with individual-level prediction is exactly the mistake that leads teams to over-trust an early sign.
A third point of caution concerns the two 2014 infant-sibling reviews specifically: both are literature reviews drawing on the same broad infant-sibling research literature rather than on two independent samples. We have not confirmed they cite identical cohorts, and the BASIS-specific claim in an earlier draft of this point was not supported by either abstract, so this is offered as a reasonable inference, not a documented fact. When the two reviews agree that the picture is heterogeneous, that is not full independent replication—it is closer to two readings of overlapping evidence arriving at a similar interpretation. That still counts for something, but it should not be oversold as four separate datasets converging on one answer.
Finally, the 2022 genetic study's VPS13B finding is a real, biologically grounded association—but it is a small-effect-size result from one dataset, explicitly flagged by its own authors as needing replication before it should inform any clinical judgment about an individual child.
What Actually Holds Up Across the Evidence
Read carefully rather than smoothed over, twin and family genetics, genotype-endophenotype association, and infant-sibling behavioral research each document heterogeneity within their own domain—though not all of them independently discover it: the 2022 endophenotype study treats phenotypic and genetic heterogeneity as already established at the outset and searches for genetic substrates within it, rather than arriving at heterogeneity as a new finding. High heritability (2013 review) does not resolve into one genetic mechanism (2022 study). Elevated familial risk in infancy, tracked prospectively from birth (2014 developmental pathways review), does not resolve into one behavioral trajectory (2014 mechanisms review)—keeping in mind that these last two draw on overlapping literature rather than independent samples, so their agreement is suggestive rather than proof. Taken together, these separate literatures are each consistent with heterogeneity rather than a single mechanism within their own domain—though Study 1's social-attention cascade and Study 3's infant-sibling findings actively conflict, so this is not a case of independent lines of evidence converging on identical detail. Heterogeneity is not a gap waiting to be closed by the next study. Treating it as a temporary research failure, rather than as the actual shape of the evidence, leads teams to keep searching for the one sign that will let them stop watching. The research says that sign does not exist yet—whether it ever will is a question the current evidence base cannot answer.
Clinical Takeaways for SLPs, OTs, and School Psychologists
None of this means early identification is hopeless—it means the identification strategy needs to match what the evidence actually shows.
- Do not wait for genetic results to justify starting surveillance or intervention. Even the strongest current genotype-endophenotype finding (VPS13B and repetitive behavior severity) explains a small slice of variance in one symptom domain and is unreplicated.
- Keep screening multi-domain, not just social-communication-focused. One infant-sibling review found limited evidence of early social-orienting deficits but did find sensory and attentional atypicality; a protocol built only around gaze and joint attention risks missing infants whose earliest divergence shows up in a different domain. No study in this set has directly tested screening yield against sensory or attentional markers, so treat this as a reasoned inference to guard against, not a proven failure mode.
- Do not treat a single early marker, or the absence of one specific expected marker, as a clean answer either way. A negative finding for a gaze or joint-attention marker at 12 months does not rule out an emerging diagnosis, and a positive genetic finding for a variant like VPS13B does not, on its own, establish one.
- When talking with families about elevated genetic risk from an older diagnosed sibling, be precise about what the heritability numbers do and do not say: risk is real and substantially elevated, but it does not specify which domain will be affected, when signs will appear, or how severe the eventual presentation will be.
- Plan surveillance as repeated and broad-based rather than a single checkpoint. This is a reasoned extrapolation rather than a direct finding—no study reviewed here tracked within-sample timing of divergence across domains for the same infants—but it is supported by the age-dependence of screening accuracy and by how many later-diagnosed toddlers look typical at early checkpoints.
- Follow the AAP/CDC schedule rather than improvising one: general developmental screening at 9, 18, and 30 months, autism-specific screening at 18 and 24 months, and screening at any other visit where a concern is raised. Screening accuracy is age-dependent—one study found M-CHAT/F positive predictive value was 0.69 at 20+ months versus 0.36 under 20 months—and rescreening matters because roughly a third of toddlers later diagnosed look typical at 18 months, only about 18% of children diagnosed at 36 months had already been diagnosed at 18 months, and Q-CHAT follow-up work at age 4 has identified new cases missed at the toddler screen.
