Causes and genetics

Two claims about what causes autism are settled in the research, even though both are still argued over in public. Vaccines do not cause autism: national cohorts in Denmark and the United States and pooled analyses agree, across the measles-mumps-rubella vaccine, thimerosal-containing vaccines and aluminum-adjuvanted vaccines.^1,2,3 Prenatal acetaminophen is not shown to raise the likelihood of autism: studies that compare siblings in the same family, one exposed before birth and one not, find no association, while ordinary cohort studies that cannot control for family factors report a small one.^4,5,6 A third claim, that leucovorin treats autism, is not settled, and the trial most often cited for it was retracted on 29 January 2026.⁷ The entry also records, as dated documents rather than evidence, what US health bodies said and did about these three subjects in 2025 and 2026.

Heritability is the share of the differences between people in a trait, in one population at one time, that is linked to genetic differences, and it is not a person's chance of being autistic.⁸ Pooling the twin studies that met a systematic recruitment standard gave about 64% to 91%, and a five-country register analysis of about 2 million children put the median at 80.8%, with country estimates from about 51% to 87%.^9,10 Most of the genetic contribution lies in common variants, each with a small effect; rare de novo mutations matter greatly to the person who carries one but account for only a small share of the variation in the population.¹¹ Sub-diagnostic autistic traits are more common in the relatives of autistic people, a pattern called the broader autism phenotype.¹² A sibling's risk is raised: in a prospective high-risk cohort about 1 in 5 younger siblings of an autistic child were autistic, a higher figure than whole-population registers give.^13,14 No study found as of October 2026 has measured directly how often the children of autistic parents are autistic.

A large share of autism's genetic contribution comes from common variants, each with a small effect: a genome-wide association study pooling 18,381 autistic people found five associated loci.¹⁵ Rare and de novo variants matter in a minority of cases, with copy-number variants carrying the greatest relative risk, and the genes they disrupt cluster in early brain development and in how neurons communicate.¹⁶ Curated gene lists grow with the sample and the threshold: 102 genes in one exome study, 72 at a stricter threshold in a larger analysis, and 253 in a 2026 preprint.^17,16,18 A minority of autistic people have a single-gene condition, such as fragile X syndrome or tuberous sclerosis complex, where about 30% of children are autistic, or a recurrent copy-number variant.^19,20 Most genomic samples are of European ancestry, and polygenic scores built from them transfer poorly to other ancestries.²¹ A polygenic score for autism is a research tool, with a correlation of about 0.16 with a diagnosis and no clinical use.²² Clinical testing finds a pathogenic result in a minority of children, and the most common result is that no single-gene cause is found.^23,24

Almost every finding about pregnancy exposures is an association, and the strongest designs compare siblings who differ in whether they were exposed.²⁵ Gestational valproate is the one exposure here with evidence pointing toward a causal effect: a US cohort found a hazard ratio of 2.67 while the drug it served as a positive control for was null, a 2025 meta-analysis of eight cohorts pooling more than six million people found 3.10, rising above about one gram a day, and the association survived a sibling comparison.^26,27 The US label treats it as a known risk, and as of October 2026 it says exposure in pregnancy "may be associated with an increased risk" of autism, with adjusted estimates across studies from 1.6 to 6.4.²⁸ Parental age is associated with autism, at about 1.4 to 1.6 times the likelihood in the highest age categories, a modest association whose proposed mechanism is untested.²⁹ Antidepressant use, maternal diabetes, preterm birth and infection are associated in cohorts, and the family and negative-control designs that have been run attenuate most of them: the sibling comparison for infection fell to no association, and a father's antidepressant use was associated too.^30,31,32 Air pollution and phthalates show small and inconsistent associations, and the folate finding is a protective association that is probably confounded.^33,34 The maternal-antibody hypothesis holds that some autism begins before birth when a mother's antibodies bind proteins in the developing brain; the antibodies are found more often in mothers of autistic children, and no human study shows that they cause autism.³⁵

Preimplantation genetic testing for polygenic disorders genotypes embryos created during IVF and returns a polygenic risk score for each one. As of October 2026 at least one company advertises an autism risk estimate to people having IVF.^36,37 The models agree that the absolute gain is small: about 3 centimeters of height and 3 IQ points from choosing the highest-scoring of about ten embryos, modeled relative risk reductions of 15 to 35 percent for a single disease in a second analysis, and about 1 to 3 percent in a 2026 preprint that modeled real IVF cycles.^38,39,40 A polygenic score for autism has a meta-analytic correlation of about 0.16 with a diagnosis and no clinical utility, and no study found as of October 2026 reports a validated polygenic embryo score for autism.^22,37 Most professional bodies that have issued statements, including the European Society of Human Genetics in 2022 and the American Society for Reproductive Medicine in 2026, say the practice is not ready for clinical use; the companies dispute that.^41,42 Surveys of autistic adults find most raise concerns about eugenics, and a nationally representative survey of US adults found disapproval of screening for traits was more strongly tied to concern about eugenics than disapproval of screening for conditions.^43,44