CNValidatron: accurate and efficient validation of PennCNV calls using computer vision.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Copy number variants (CNVs) are key drivers of genetic variation, evolution, and disease risk.
• Genotyping arrays are a primary source for CNV detection in large cohorts.
• Existing CNV calling methods from array data exhibit high false positive rates, with current validation techniques being inefficient at scale.

Purpose of the Study:

• To address the limitations of current CNV validation methods.
• To develop a scalable and accurate automated approach for CNV validation.
• To enhance the reliability of CNV detection in large genomic datasets.

Main Methods:

• Assembled the largest collection of human-verified CNV calls (nearly 60,000) from 22,500 samples across multiple cohorts and array types.
• Utilized visual validation to assess the accuracy of CNV calls, revealing a high false positive rate with existing methods.
• Trained a convolutional neural network (CNN) using a subset of the visual validation data to automate CNV validation via machine vision.

Main Results:

• Visual validation indicated that over 53% of CNV calls were false positives and nearly 10% were unclear.
• Existing quality control (QC) metrics-based filtering methods proved inefficient in reducing false positive CNV calls.
• The developed CNN model achieved over 90% accuracy in CNV validation, comparable to human analysts, and was validated both within-sample and out-of-sample.
• Orthogonal validation using genome sequencing data confirmed the high accuracy of the visual validation approach.

Conclusions:

• Visual inspection remains the gold standard for validating CNV calls.
• The developed machine vision model effectively automates CNV validation at scale with high accuracy.
• The CNV validation software is publicly available as an R package.