Accelerated long-read variant calling with Clair3 for whole-genome sequencing

The rapid growth of genomic data and increasing adoption of long-read sequencing technologies have rendered variant calling one of the most computationally demanding tasks in genomic analysis. Although deep learning-based methods currently outperform conventional approaches in distinguishing true variants from complex sequencing noise, they impose prohibitive computational and time requirements. To address this limitation, we present a computational framework based on Clair3 that integrates parallelized feature generation, enhanced variant phasing, in-memory read haplotagging, and GPU-accelerated neural network inference to accelerate variant calling. By dynamically optimizing the use of both GPU and CPU resources, our method achieves substantial runtime improvements without compromising accuracy. We evaluated our framework across a range of sequencing depths, diverse samples, and multiple hardware configurations. Our results demonstrate that the optimized pipeline completes variant calling for a 30× whole-genome sequence in 12-20 minutes using standard computational resources (32 CPU threads and one NVIDIA GPU), and in 12-15 minutes on an Apple Mac Studio (32 threads), which is ∼10-20-fold speedup compared with its initial release. In addition to exceptional efficiency, our method maintains state-of-the-art accuracy, achieving SNP F1-scores of 99.32% and 99.70% on 30× ONT and PacBio GIAB HG003 datasets, respectively. This work introduces a rapid, accurate, and scalable variant calling framework that effectively supports large-cohort genomic studies and time-sensitive clinical applications.

The accelerated implementation of Clair3 is open source and available at: https://github.com/HKU-BAL/Clair3/tree/gpu.