Clair3-Trio

Clair3-Trio improves variant calling accuracy from Nanopore long-read sequencing by jointly predicting variants in family trios (child, mother, father) using a Trio-to-Trio deep neural network.

Key Features:

• Trio-to-Trio deep neural network: Integrates sequencing information from all three family members into a single model tailored for trio-based variant calling.
• Joint variant prediction: Produces simultaneous variant predictions across child, mother, and father rather than calling each sample independently.
• MCVLoss Mendelian inheritance encoding: Uses a custom loss function (MCVLoss) that explicitly encodes Mendelian inheritance patterns during model training.
• Nanopore long-read sequencing input: Operates on Nanopore long-read sequencing data as the primary input for variant identification.
• Improved accuracy and reduced Mendelian violations: Demonstrates fewer predicted variants that contradict Mendelian inheritance compared with conventional methods.

Scientific Applications:

• Family trio genomic studies: Enables more accurate variant detection in child-mother-father sequencing datasets for inheritance analysis.
• Inherited disease research: Supports identification of hereditary variants relevant to studies of genetic disease mechanisms.
• Genetic counseling and diagnosis: Provides trio-aware variant calls that inform clinical interpretation of familial transmission.
• Personalized medicine: Improves reliability of variant information used in individualized therapeutic or diagnostic decisions.

Methodology:

A Trio-to-Trio deep neural network trained on Nanopore long-read sequencing data uses the MCVLoss function to encode Mendelian inheritance and generate joint variant predictions for child, mother, and father.