Deep Learning in Variant Calling

Variant calling—the process of identifying differences between a sample genome and a reference—was traditionally done using statistical models like GATK HaplotypeCaller. Deep learning has changed the game.

DeepVariant: Treating DNA as an Image

Google's DeepVariant treats the pileup of sequencing reads not as text, but as an image. It uses a Convolutional Neural Network (CNN)—the same technology used for facial recognition—to "see" the variants.

Clair3: The Hybrid Approach

Clair3 combines a pileup model for speed with a full-alignment model for accuracy, achieving state-of-the-art performance on long-read sequencing data (Nanopore, PacBio).