Citation
Litman, A., Sauerwald, N., Green Snyder, L., Foss-Feig, J., Park, C.Y., Hao, Y., Dinstein, I., Theesfeld, C.L., & Troyanskaya, O.G. (2025). “Decomposition of phenotypic heterogeneity in autism reveals underlying genetic programs.” Nature Genetics, 57, 1611–1619. doi:10.1038/s41588-025-02224-z. Princeton University / Flatiron Institute / Simons Foundation.
What this paper is
A top-tier genetics paper (Nature Genetics) that takes a person-centred approach to autism heterogeneity. Rather than looking at single traits and asking which genes are associated, it models whole individuals across 239 phenotypic features and asks: do autistic people fall into distinguishable groups, and if so, do those groups have distinct genetic architectures? The answer to both is yes.
Key findings
- Cohort: 5,392 autistic individuals from the SPARK cohort (a large US research effort), with matched genetic data and 239 item-level phenotype features drawn from the SCQ, RBS-R, CBCL, and developmental milestone questionnaires.
- Method: Generative Finite Mixture Model (GFMM), a person-centred statistical approach that clusters individuals by their full phenotypic profiles rather than fragmenting each person into separate trait scores. The model was tested with 2–10 latent classes; four emerged as optimal.
The four phenotypic classes
| Class | n | Profile |
|---|---|---|
| Social/behavioral | 1,976 | High social communication and repetitive behaviour difficulties, plus attention deficit, disruptive behaviour, anxiety — but no developmental delays. Enriched in ADHD, anxiety, major depression. |
| Mixed ASD with DD | 1,002 | Strong developmental delays, language delay, intellectual disability, motor disorders. Lower rates of ADHD, anxiety, depression. Earlier age at diagnosis. This is the class most relevant to autistic people with intellectual disability. |
| Moderate challenges | 1,860 | Lower scores across all seven measured categories compared to other autistic children. Still significantly above non-autistic siblings on diagnostic measures. |
| Broadly affected | 554 | High scores across everything — social, behavioural, developmental, psychiatric. Highest rates of co-occurring conditions. Highest numbers of interventions. |
- Classes map to distinct genetic programs. Patterns in common genetic variation (polygenic scores), de novo mutations, and rare inherited variation differ between classes. This is not just phenotypic clustering — the biological architecture is genuinely different.
- Developmental timing matters. The sets of genes disrupted in each class are active at different points in brain development, and those timing differences align with the clinical differences between classes. The “Mixed ASD with DD” class, for example, shows disruption in genes active during earlier developmental windows.
- Validated and replicated in an independent cohort. The four-class structure is stable under perturbation and robust to various statistical tests.
Method in brief
SPARK cohort, 239 phenotype features (item-level responses on SCQ, RBS-R, CBCL, plus developmental milestones). GFMM trained with latent classes 2–10; four selected by BIC, validation log likelihood, and clinical interpretability. Genetic analyses: polygenic scores for autism, educational attainment, and related traits; de novo variant analysis; rare inherited variant analysis; pathway enrichment; developmental transcriptomics to link disrupted genes to brain development timing.
Relevance to this wiki
This paper matters in three ways:
-
The “Mixed ASD with DD” class is the core population this wiki focuses on. The paper gives that population a genetic identity — not just a clinical description but a distinct biological profile with different genetic programs, different developmental timing, and different pathways disrupted. This strengthens the case that autism with intellectual disability is not simply “more severe autism” but a qualitatively different thing.
-
The open question gets partially answered. This wiki’s Sensory processing in autism and intellectual disability page asks whether ASD + ID should be treated as a distinct condition. This paper stops short of saying yes outright, but its architecture — four classes with different genetics, different developmental timing, different clinical trajectories — is consistent with multiple “autisms” rather than one spectrum.
-
The person-centred methodology is aligned. The insistence on individual prikkelprofielen rather than generic profiles is the clinical equivalent of what this paper does computationally: model the whole person, not isolated traits. If sensory processing data were added to the SPARK feature set, one might expect the four classes to show distinct sensory processing signatures as well — a testable prediction the wiki should track.
Limitations
- No sensory processing data in the phenotypic features. The 239 features come from the SCQ (social communication), RBS-R (repetitive behaviour), CBCL (child behaviour), and developmental milestones. Sensory processing is not directly measured. This is a significant gap given that hypo/hyperresponsivity is a DSM-5 criterion.
- US-only sample. The SPARK cohort is a North American effort. Cultural, diagnostic, and healthcare system differences may limit generalisability to Dutch and European populations.
- Children only. Participants are children; how the four classes map onto adult presentations is unknown.
- The model discovers structure, not truth. Four classes is a statistical optimum, not a biological claim about exactly four kinds of autism. The real landscape may be more continuous.
Related pages
- Genetic heterogeneity in autism — topic overview seeded from this paper
- Sensory processing in autism and intellectual disability — the topic overview this paper’s findings speak to
- Werkman et al. 2022 — ID impact on sensory processing in autism — the sensory processing evidence for the “Mixed ASD with DD” class