Poster #13: Haedong Kim PhD

Exome-Wide Copy Number Variation in 142,357 Individuals from Autism Spectrum Families in the Simons SPARK Cohort

Haedong Kim,1,3 Grace Tzun-Wen Shaw,1 Timothy L. Mosbruger,1 Ramakrishnan Rajagopalan,1,3 Laura Conlin, 1,3 Manuel Torres Diz,2 Cécile Poulain, 4,5 Thomas Renne, 4,5,6 Sébastien Jacquemont, 5,6 Tristan J. Hayeck*,1,3. 1.

Copy number variants (CNVs) are a major genetic contributor to autism spectrum disorder (ASD). However, reliable CNV detection from whole-exome sequencing (WES) remains challenging due to high rates of spurious calls and missed variants, inconsistent caller performance, and the difficulty of identifying rare de novo events amid pervasive technical noise. We present a comprehensive CNV resource for the SPARK cohort and an automated machine learning–based framework, Structural Variant Probability Observed from Tree-based Regression (SPOTR), that calculates the probability of validity for each CNV, providing a tool-agnostic alternative to concordance-based filtering. For variant calling, we implemented a kernel-based change-point detection approach that minimizes preprocessing to preserve genuine signal, while also supporting direct integration of CNVs from external calling methods. Model training leverages three types of data sources, cross-platform validated calls, synthetic genuine and putative de novo candidates, and human-in-the-loop manually curated variants. The hyperparameter search was guided by Bayesian optimization, enabling efficient exploration of a high-dimensional parameter space while supporting fully automated tuning and improving model robustness and calibration. The resulting models achieved PR AUCs over 0.9 for both deletions and duplications. Training on synthetic and putative de novo call sets together with family-structure-aware features improved Mendelian precision significantly. Applied to three WES callers across 142,357 SPARK individuals, the probabilistic framework recovered likely de novo calls and rejected invalid calls that concordance-based filtering would otherwise discard or retain, respectively. High-confidence ASD de novo CNVs identified by the model showed enrichment in specific gene superfamilies, including ASD-related groups, and overlapped with loss-of-function–intolerant genes.