Multimodal dissection of the autism spectrum скачать в хорошем качестве

Multimodal dissection of the autism spectrum

Alal Eran (Boston Children’s Hospital) https://simons.berkeley.edu/talks/ala... Theory of Computing and Healthcare Autism is currently diagnosed as a singular behavioral entity, yet its profound clinical and genetic heterogeneity suggests distinct underlying etiologies that demand a precision medicine approach. This talk presents a data-driven framework for turning that heterogeneity into actionable subtypes, by combining massive electronic health record (EHR) analytics with deep- phenotyped cohort studies. First, we explore a data-driven phenotypic analysis of 63,673 children with autism derived from a national EHR repository. By applying dimensionality reduction (PaCMAP) and density-based clustering (HDBSCAN) to standardized diagnosis codes (PheCodes), we identified 36 distinct clinical clusters with interpretable comorbidity signatures spanning neurologic, developmental, psychiatric, and metabolic profiles. To translate these findings into clinical practice, we introduced a gradient-boosting prediction model (LightGBM) capable of assigning new patients to these clusters with high accuracy (AUC 0.9) at their first diagnostic encounter, enabling early, individualized intervention trajectories. We then connect the EHR-derived subtypes to mechanistically informative case studies. The first is a dyslipidemia-associated autism subtype that we have identified using multi-dimensional genomic analyses. By integrating whole exome sequence data from 3,531 individuals with spatiotemporal brain gene expression data, we identified a convergence of autism-segregating deleterious variants within lipid regulation pathways. Validation of this dyslipidemia subtype in an independent cohort of 34 million individuals confirmed that ~5% of children with autism have altered blood lipid profiles and a significantly higher prevalence of dyslipidemia diagnoses compared to unaffected siblings and controls. We also used the Simons Simplex Collection to characterize the subgroup of autistic children reported to demonstrate marked behavioral improvements during febrile episodes. We found that fever responsiveness is associated with maternal infection during pregnancy and gastrointestinal dysfunction, pointing to immune–gut interactions as plausible modulators of core autism features. Moreover, we addressed diagnostic timing in the SPARK cohort. Despite improved screening, ~25% of children receive an autism diagnosis after age six, preventing them from achieving optimal outcomes. Moreover, analyzing 23,632 participants in the SPARK cohort, we addressed the paradox of delayed diagnosis (post-age 6). We identified two diametrically opposed late-diagnosed groups: one characterized by camouflaging (lower support needs and fewer comorbidities) and another by clinical overshadowing (severe comorbidities and high support needs that masked the underlying autism). Together, these results show that autism heterogeneity can be partitioned into scalable, interpretable subtypes from routine clinical data and refined with cohort studies. By prioritizing interpretability, we connect machine-learning outputs to clinical reasoning and enable earlier risk stratification, targeted screening, and better-defined cohorts for mechanistic research and clinical