Clair3

Clair3 performs deep learning–based variant calling on long-read sequencing data to detect single nucleotide polymorphisms (SNPs) and other variants with a balance of speed and accuracy.

Key Features:

• Deep learning-based calling: Implements neural network models for variant calling from sequencing data.
• Long-read optimization: Optimized specifically for long-read sequencing data.
• Hybrid inference architecture: Combines pileup-based calling and full-alignment evaluation within a single framework.
• Pileup-based calling: Rapidly processes the majority of candidate sites using summarized read evidence (pileups).
• Full-alignment evaluation: Performs detailed, context-aware inference on complex or ambiguous sites using full-read alignments.
• Selective escalation: Escalates only difficult candidate sites from pileup to full-alignment to balance computational cost and sensitivity.
• Model combination: Integrates lightweight pileup models with high-resolution neural inference for different evidence resolutions.
• Performance characteristics: Delivers state-of-the-art SNP detection, robustness at low sequencing coverage, and improved speed relative to competing long-read callers while maintaining or improving accuracy and optimizing precision and recall.

Scientific Applications:

• SNP detection: Detects single nucleotide polymorphisms (SNPs) in long-read sequencing datasets.
• Complex variant resolution: Resolves complex or ambiguous variant sites by applying full-alignment neural inference.
• Low-coverage calling: Provides robust variant calls at low sequencing coverage where noise and alignment uncertainty are prevalent.
• Production-scale pipelines: Serves long-read variant detection pipelines requiring both throughput and reliable accuracy.

Methodology:

Clair3 uses a hybrid inference approach that applies pileup-based calling for rapid processing of most sites and full-alignment neural inference for complex sites, combining lightweight pileup models with high-resolution neural networks and selectively escalating candidate sites.