Summary
A medRxiv preprint describes three machine-learning-derived autism profiles in two cohorts of autistic children. The profiles showed different associations with common and rare genetic variants, although the authors say larger-cohort validation is needed.
A medRxiv preprint reports three machine-learning-derived autism profiles among children in two research cohorts, with each profile showing different associations with genetic measures. The researchers based the profiles on clinical characteristics rather than a simple autistic/non-autistic distinction, and found that the profiles’ symptom and medication-use patterns recurred in a second cohort.
The study analysed 6,355 autistic children in the Simons Foundation Powering Autism Research for Knowledge (SPARK) cohort. The authors say the framework also captured distinct genetic patterns, but describe it as needing further validation in larger cohorts. The findings are from a preprint.
How the researchers grouped the children
The team used 14 clinical symptom measures, adjusted for sex and age, to derive three categories and corresponding continuous scores. They used Discriminative Dimensionality Reduction via learning a Tree (DDRTree), a method for representing patterns across multiple measures, to develop the profiles in SPARK.
The high-support-needs (HSN) profile involved broad developmental and functional impairments and very substantial support needs. The emotional-behavioral-support-needs (EBSN) profile was marked by severe repetitive behaviours and elevated psychiatric comorbidity. The low-support-needs (LSN) profile involved milder symptoms and relatively lower overall support needs. The groups also differed in the kinds of treatments and supports commonly used.
To test whether the framework transferred to another dataset, researchers applied a modified k-nearest-neighbour method to 2,717 autistic children in the Simons Simplex Collection (SSC). They report that symptom and medication-use patterns seen in SPARK were also present in SSC. The authors further examined associations with support needs, service use and medication use, accounting for factors including age, sex, genetic ancestry and socioeconomic factors.
Genetic patterns differed across profiles
The researchers estimated SNP-based heritability for each continuous profile score. This measure estimates how much variation in a trait within a study population is associated with common genetic variants measured through single-nucleotide polymorphisms (SNPs); it is a group-level estimate, not a prediction of an individual’s traits. The estimate was highest for EBSN (h²g = 0.224, SE = 0.051), followed by HSN (0.167, SE = 0.049) and LSN (0.064, SE = 0.044).
Polygenic scores for attention-deficit/hyperactivity disorder and major depressive disorder were consistently associated with higher EBSN scores in both cohorts. Polygenic scores combine information from many genetic variants associated with a trait. The study also found that higher de novo variant and rare copy-number variant (CNV) burdens were associated with higher HSN scores in SPARK; higher rare CNV burden was associated with higher EBSN scores in SSC. These are associations across groups, not evidence that the variants determine an individual child’s profile.
The results suggest that studying variation within autism may reveal genetic patterns that a single case-control definition can obscure. These machine-learning-derived profiles are research groupings, not a replacement for clinical assessment. The authors say the framework warrants further validation in larger cohorts.