Blood Metabolites and AI Reach Over 80% Accuracy in Supporting Autism Diagnosis

A blood test that could help clinicians diagnose autism spectrum disorder in toddlers has moved a significant step closer to real-world use, thanks to a new translational study published in the Annals of Biomedical Engineering. Researchers led by Juergen Hahn of Rensselaer Polytechnic Institute, working with clinical partners in Arizona, Tennessee, and the biotechnology company BioROSA, report that measurements of metabolites from two interconnected biochemical pathways, combined with machine learning classification, predicted whether children on a diagnostic waitlist had autism with greater than 80 percent accuracy. The work, registered as the Metabolic Autism Prediction (MAP) Study, is notable not because it discovered a new biomarker, but because it attempted something far rarer: testing an existing biomarker concept prospectively, in a double-blind design, in the exact clinical setting where it would eventually be deployed.

The scientific foundation of the study rests on the folate-dependent one-carbon metabolism (FOCM) pathway and the transsulfuration (TS) pathway, two tightly linked biochemical networks that together govern methylation reactions and the body’s antioxidant defenses. The FOCM pathway transfers single-carbon units derived from folate to processes such as DNA methylation and nucleotide synthesis, while the transsulfuration pathway channels the sulfur of the amino acid homocysteine into the production of cysteine and, ultimately, glutathione, the cell’s principal antioxidant molecule. Disruptions in these pathways have been repeatedly documented in autism research. Earlier work by several of the same investigators, including studies of oxidative stress and methylation capacity in children with autism, found consistent metabolic imbalances: altered levels of methionine, S-adenosylmethionine, S-adenosylhomocysteine, cysteine, glutathione, and related redox species. What had been missing was evidence that these differences could be measured reliably in a general clinical population, before diagnosis, and translated into a decision-support tool.

To address that gap, the team designed a double-blind case-control trial in which blood samples were collected from children who had been referred to developmental pediatricians because of concerns about their development. Crucially, none of the children had a confirmed diagnosis at the time of enrollment. The cohort comprised 140 children between 18 and 60 months of age, recruited at two developmental pediatric clinics. Alongside the blood draw, each child underwent comprehensive gold-standard clinical evaluations, including the Autism Diagnostic Observation Schedule (ADOS), the Mullen Scales of Early Learning (MSEL), and the Vineland Adaptive Behavior Scale (VABS). Combined with a complete medical history and physical examination, these assessments allowed clinicians to confirm or rule out suspected autism using DSM-5 criteria. The diagnostic outcome was withheld from the analytical team, and the metabolite measurements were withheld from the clinicians, preserving the double-blind character of the trial.

The diagnostic outcomes divided the cohort into two groups: 114 children received an autism spectrum disorder diagnosis, while 26 were found to have non-autism-related developmental delays. This composition reflects the reality of developmental clinics, where a majority of referred children do indeed receive an autism diagnosis, but a meaningful minority present with other conditions such as language delay, global developmental delay, or other neurodevelopmental concerns. The inclusion of the latter group is methodologically important. Many earlier biomarker studies compared children with autism to typically developing peers, a comparison that is scientifically informative but clinically less relevant, since a physician’s practical question is not whether a child differs from typical development but whether the child’s profile indicates autism rather than another developmental condition.

On the analytical side, the researchers measured concentrations of metabolites spanning the FOCM and TS pathways, quantified in nanograms per milliliter, and then applied artificial intelligence-based classification algorithms to distinguish the two diagnostic groups. Rather than feeding raw metabolite concentrations directly into a classifier, the team engineered additional features from the measurements, including metabolite ratios and other statistically derived combinations. This feature engineering reflects a key insight from metabolomics: the information content of a metabolic profile often lies in the relationships between metabolites rather than their absolute levels. Ratios between methylation-cycle intermediates and transsulfuration products, for example, can capture the balance between methylation capacity and antioxidant synthesis more sensitively than any single concentration. The study also employed rigorous statistical practices, including normality and skewness testing of the metabolite distributions, cross-validation for model assessment, and Monte Carlo repetitions to characterize the uncertainty of reported performance metrics, with 95 percent empirical percentile intervals reported across repetitions.

The headline result is that classification algorithms achieved over 80 percent accuracy in predicting whether a blood sample came from a child diagnosed with autism. In a prospective, double-blind, clinic-based cohort, that level of performance is meaningful. It is not presented as a replacement for clinical judgment. The authors are explicit that the results need to be replicated in larger studies, particularly ones that include more children with non-autism-related developmental delays, since the control group of 26 children limits how precisely the model’s specificity can be estimated. Still, the study demonstrates that a physiological measurement, coupled with machine learning, can support autism diagnosis in a clinically relevant setting rather than only in retrospective case-control comparisons.

The clinical motivation for this line of research is considerable. Autism spectrum disorder affects a substantial and growing number of children, with surveillance data from the United States indicating continued increases in prevalence and in the identification of children at ages 4 and 8. The economic burden is correspondingly large, with lifetime social costs estimated in the trillions of dollars nationally, and diagnosis is frequently delayed. Families often wait many months, sometimes years, between first concerns and a definitive diagnosis, because the gold-standard evaluation depends on specialized clinicians who are in short supply. This delay matters: a substantial body of evidence shows that early interventional approaches, including intensive behavioral treatment beginning in toddlerhood, can improve long-term outcomes in language, cognition, and adaptive behavior. A blood-based test that could be administered at the point of referral, while a child waits for a full diagnostic evaluation, could help triage referrals, shorten effective wait times, and give families earlier access to services.

The study also situates itself within a broader effort to develop multivariate biomarker-based diagnostics for neurodevelopmental disorders. Previous work from the same group demonstrated that multivariate analysis of oxidative stress and DNA methylation markers could classify children with autism against typically developing peers with high accuracy, and subsequent validation studies extended this approach. Other teams have explored metabolomic screening in the Children’s Autism Metabolome Project, salivary microRNA signatures, and microbially derived metabolites as candidate diagnostic or screening tools. What distinguishes the current work is its translational framing: the enrollment of children on diagnostic waitlists, the use of DSM-5-anchored clinical outcomes, the double-blind design, and the registration of the trial on clinicaltrials.gov under identifier NCT04672967. The study received institutional review board approval from the Biomedical Research Alliance of New York on August 12, 2021, and informed consent was obtained from all participants, with procedures adhering to the Declaration of Helsinki.

There are, of course, important caveats and open questions. The imbalance between the autism and non-autism groups means that future studies should deliberately enrich the comparison group with children who have other developmental conditions, to better establish how the test performs in the full differential-diagnostic context. The metabolite panel is drawn from a specific pair of pathways, and it remains to be seen whether combining FOCM/TS markers with other biomarker classes, such as mitochondrial, immune, or microbiome-derived measures, would improve performance further. Regulatory pathways, assay standardization across laboratories, and cost-effectiveness will all need attention before such a test could become routine. The authors also disclose relevant commercial interests: several authors are employees of BioROSA Technologies, which has licensed intellectual property from Rensselaer Polytechnic Institute related to diagnosing autism, and two academic authors hold related intellectual property, considerations that are common in biomarker translation but worth noting.

Even with those caveats, the study represents a template for how physiological biomarkers might realistically enter autism care. Rather than promising a standalone diagnostic, the researchers frame the metabolite-based classifier as a support tool, one that could complement, rather than replace, the observational expertise of developmental pediatricians and psychologists. If the findings replicate at scale, a simple blood draw at the moment of referral could provide clinicians with an additional, biologically grounded data point during a period when families are often left waiting in uncertainty. In a field where diagnosis still rests entirely on behavior and development, the prospect of a validated biochemical adjunct, tested prospectively in the clinics where it would actually be used, marks a tangible advance toward faster, more informed answers for children and their families.

Subject of Research: A double-blind translational trial using folate-dependent one-carbon metabolism and transsulfuration pathway metabolites with machine learning to support autism spectrum disorder diagnosis in children on diagnostic waitlists.

Article Title: Translational Study of Using FOCM/TS Metabolites for Supporting Autism Spectrum Disorder Diagnosis

Article References: Arici, H., Causey, M., Patra, S., Kruger, U., Villegas Uribe, C. A., Melmed, R., Ciuk, C., Crisler, S., Marler, S., Witters-Cundiff, A., Bhadresa, S., Slattery, J., & Hahn, J. (2026). Translational Study of Using FOCM/TS Metabolites for Supporting Autism Spectrum Disorder Diagnosis. Annals of Biomedical Engineering. https://doi.org/10.1007/s10439-026-04365-6

Image Credits: AI Generated

DOI: 10.1007/s10439-026-04365-6

Keywords: autism spectrum disorder, FOCM, transsulfuration pathway, metabolomics, biomarkers, machine learning, clinical trial, early diagnosis, oxidative stress, methylation, developmental pediatrics, translational research