Why Autism Epigenetics Is Not Yet a Clinical Test

What the review found

Before a biological measurement can be used to identify a condition in a child, it has to clear two hurdles. Independent laboratories must find the same marker altered in the same direction, and the marker must sort affected children from unaffected ones accurately enough that the result adds something a clinician did not already have. Most candidate markers stop at the first.

Eighteen studies are all that exist on this question in young children. They were assembled in a systematic review from the University of Pisa and the IRCCS Stella Maris Foundation, published in the International Journal of Molecular Sciences in 2023, with Andrea Stoccoro and Eugenia Conti as lead authors. Searches of PubMed, Scopus and Web of Science returned 384 unique papers, of which 18 met the criteria: children with idiopathic autism, meaning autism with no identified genetic cause, at a mean age of eight or under.

DNA methylation is a chemical tag that sits on DNA and changes how readily a gene is read, leaving the genetic sequence underneath unaltered. The tags shift with development and with environmental exposure, which is the source of the interest. They might record something about a child that the genome alone does not.

The review confined itself to tissue a clinic could realistically collect. Fourteen studies used peripheral blood, two used cells swabbed from the inside of the cheek, and two used dried blood spots taken at birth from newborns later diagnosed. Thirteen were case-control comparisons; three followed children over time. Sample sizes ran from 14 autistic children against 10 controls up to 629 against 634.

Nearly every study reported differences, with genes surfacing across both candidate-gene work and genome-wide scans, among them OXTR, RELN, HTR4, ESR2, ST8SIA2 and TGFB1. No gene was found altered in more than one of the eighteen.

The two largest studies returned nothing at their own thresholds. Andrews and colleagues, comparing 453 autistic children with 515 typically developing children, found no site meeting the genome-wide correction of p below 1.12 × 10⁻⁷, and reported seven sites at looser thresholds instead. Hannon and colleagues, working with dried blood spots from 1,263 newborns of whom 629 were later diagnosed, also found nothing at the epigenome-wide threshold and reported twenty sites at p below 5 × 10⁻⁵.

One result did converge. Three separate groups, using three different laboratory techniques, each found the genome as a whole slightly less methylated in autistic children than in controls. That is the single observation in the review with more than one laboratory behind it.

What a test would still require

Discovery and validation are distinct stages of biomarker work. A discovery study finds a difference between groups. A validation study carries that specific difference to a fresh sample collected by other people and asks whether it survives. The reviewers state plainly that the most interesting results have not been replicated by independent research groups, and they attribute the failure to small samples, differing designs, differing genetic backgrounds and the heterogeneity of autism itself.

Methods varied so widely that no statistical pooling of the eighteen studies was possible. Blood, cheek cells and neonatal blood spots were read by array scans, pyrosequencing, liquid chromatography, antibody assays and several forms of methylation-specific PCR. A marker measured one way in one tissue is not straightforwardly the same marker measured another way in another.

Only one of the eighteen studies included children with a different diagnosis as a comparison group, setting autistic children alongside children with Fragile X syndrome. The rest compared autistic children with typically developing children, or examined autistic children alone. A test in clinical use faces the harder version of the question, since the children who arrive for assessment are rarely typically developing; they are children with language delay, with intellectual disability, with attention difficulties, or with several of these at once.

The review reports no accuracy statistics for any marker. Nothing on how often a marker correctly identifies an autistic child, nothing on how often it points at a child who is not. Those figures separate a group difference from a test, and the primary studies were not designed to produce them.

The route from here is well mapped elsewhere in medicine. A specific marker is named in advance, measured by a standardised method in a large sample, then tested again in an independent sample gathered by a different team, against the children a clinic actually sees. None of the eighteen studies meets that description, which the reviewers acknowledge in their own conclusion.

Commercial epigenetic tests already exist for other conditions, and the review cites them as evidence of what the technology can eventually support. That infrastructure runs ahead of the autism evidence. A laboratory offering an epigenetic autism screen can be asked which published marker it measures and in which independent sample that marker was confirmed. If the answer names only the study that discovered the marker, you are being offered a hypothesis.