Researchers have split autism into two distinct biological subtypes, one defined by reduced neuronal connectivity and downregulated synaptic genes, the other by heightened immune-gene activity and microglial activation. The finding, drawn from brain imaging of 940 autistic participants and genetic analysis across 20 mouse models, challenges the long-standing practice of treating autism as a single diagnostic category. If the subtypes hold up in clinical trials, they could redirect drug development toward targeted therapies rather than one-size-fits-all behavioral interventions.
Why splitting autism into two biological subtypes changes the clinical calculus
Current autism diagnoses rely on behavioral observation. Two children who score similarly on standardized assessments may carry fundamentally different molecular profiles in their brains. That mismatch has frustrated drug developers for decades: clinical trials that enroll a mixed population tend to wash out treatment effects that might benefit one subgroup but not the other. The new subtyping data offers a way to sort participants before a trial begins.
A cross-species study published in autism connectivity research identified two reproducible functional-connectivity patterns, hypoconnectivity and hyperconnectivity, using fMRI data from 940 autistic individuals and 1,036 controls. Each pattern mapped onto distinct gene-pathway enrichments in 20 genetic mouse models, meaning the subtypes are not statistical artifacts of a single cohort but replicate across species and datasets.
One testable prediction follows from the split. If the immune-upregulated subtype involves active microglial signaling, then compounds that modulate microglial function should reduce repetitive behaviors more quickly in that group than in the synaptic-downregulated group. Pre-and-post fMRI connectivity scores could serve as an objective readout, sidestepping the subjectivity of behavioral rating scales. No published trial has yet tested this hypothesis directly, but the subtyping framework now makes such a trial designable.
Clinically, the prospect of objective, biology-based subtypes is attractive because it could move autism care closer to oncology’s precision-medicine model. Instead of asking whether a drug “works for autism,” researchers could ask whether it normalizes connectivity in the hypoconnected group or dampens immune signatures in the hyperconnected group. Regulatory agencies have already accepted imaging and molecular biomarkers in other neurological disorders; a similar pathway could emerge here if the subtypes prove reliable and predictive.
Mouse models, postmortem tissue, and RNA-seq converge on the same divide
The strength of the subtyping claim rests on convergence across independent methods and labs. The Nature Neuroscience study used unsupervised machine learning on resting-state fMRI to cluster autistic brains without imposing prior assumptions about group boundaries. One cluster showed broadly reduced connectivity across cortical networks; the other showed elevated connectivity, particularly in circuits linked to sensory processing and social attention. Genetic mouse models carrying different autism-risk mutations fell into the same two bins, and each bin was enriched for different molecular pathways: synaptic signaling genes in the hypoconnected group, immune and glial genes in the hyperconnected group.
Postmortem molecular data tells a parallel story. A large-scale cortical transcriptomics study published in cortical gene-expression work documented upregulation of microglial, astrocyte, and neural-immune genes alongside downregulation of synaptic genes across the cerebral cortex in autism. These two expression signatures were not randomly distributed; they formed coordinated modules, suggesting a biological trade-off rather than independent noise.
A separate transcriptomic effort applied unsupervised non-negative matrix factorization to ASD RNA-seq data from 1,711 samples and identified molecular clusters in which immune-system processes were specifically implicated for one cluster while synaptic and neuronal processes dominated others. That independent replication, using a different analytic method on a different dataset, strengthens the case that the two-subtype architecture is real.
Earlier neuropathology work had already flagged neuroinflammation in autism brain tissue and cerebrospinal fluid, documenting neuroglial activation in postmortem samples. And a multi-omics integration study identified a convergent molecular subtype linking histone acetylation marks with transcriptomic differences, pairing neuronal downregulation with immune upregulation at the epigenomic level. Taken together, imaging, gene expression, and epigenetic data all point toward the same fault line: autism biology tends to polarize into a synaptic-weakening profile and an immune-activated profile, even though both may coexist to varying degrees in any given brain.
Gaps in ancestry data, treatment evidence, and longitudinal tracking
The subtyping evidence is consistent across methods, but several gaps limit its immediate clinical use. No published dataset yet links the two connectivity subtypes to longitudinal outcomes or treatment response. Researchers can identify which subtype a person falls into, but they cannot yet say whether that classification predicts which therapies will work, how symptoms will change over time, or whether subtype membership is stable from childhood into adulthood.
Ancestry representation is another open question. The multicenter fMRI cohorts and postmortem tissue collections that anchor these findings have drawn heavily from populations of European descent. Whether the same two-subtype split holds in East Asian, African, or Latin American ancestry groups has not been tested in the primary records. Given that allele frequencies for autism-risk variants differ across populations, replication in diverse cohorts is essential before clinicians can assume that the hypoconnected and hyperconnected patterns are globally generalizable rather than regionally specific.
There are also technical gaps. Resting-state fMRI is sensitive to head motion, scanner differences, and preprocessing choices, all of which can bias connectivity estimates. Although the Nature Neuroscience team harmonized data across sites and validated their clusters in independent samples, routine clinical scanners may not reproduce those conditions. Similarly, RNA-seq and epigenomic assays are currently confined to research labs and postmortem tissue; they cannot yet be deployed as frontline diagnostic tools.
Ethical and practical issues complicate translation as well. Labeling autistic people as “immune subtype” or “synaptic subtype” risks oversimplification and stigma if the labels are misunderstood as fixed categories rather than probabilistic patterns. Many autistic individuals and advocacy groups emphasize that autism is a neurotype, not a disease to be eliminated. Any push toward biologically targeted treatments will need to engage with those perspectives, focusing on alleviating distress and disability rather than erasing neurodivergent traits.
What a path to clinical impact could look like
Turning the two-subtype framework into real-world benefit will require several steps. First, researchers will need prospective studies that assign subtype labels at baseline using standardized fMRI protocols and then track participants over years. Those studies should record developmental trajectories, co-occurring conditions such as epilepsy or anxiety, and responses to behavioral and pharmacological interventions. If subtype membership predicts any of these outcomes, it could justify incorporating connectivity or molecular markers into clinical guidelines.
Second, drug trials will need to stratify or enrich for one subtype. For example, a microglia-modulating compound could be tested primarily in individuals with the immune-activated profile, with pre-registered imaging and molecular endpoints. Conversely, a synapse-strengthening intervention might be targeted to the hypoconnected group. Negative trials that ignore subtype differences may need to be reinterpreted in light of this emerging heterogeneity.
Third, future work must expand beyond European-ancestry cohorts. Building large, harmonized datasets from underrepresented populations, and ensuring that analytic pipelines are robust across sites and scanners, will be crucial for equitable translation. Without that, precision approaches risk deepening existing disparities in autism diagnosis and care.
For now, the two-subtype model does not change how clinicians diagnose autism, but it reshapes how researchers think about the condition’s underlying biology. Rather than a single spectrum with uniform mechanisms, autism increasingly appears to be a set of overlapping neurobiological profiles that share outward behaviors but diverge under the surface. If ongoing work can tie those profiles to prognosis and treatment response, the field may finally move beyond one-size-fits-all interventions toward care that is tailored to the brain in front of the clinician, not just the checklist of behaviors.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
