Telomere-to-telomere CHM13 reference reveals missing truth variants and improves deep learning-based variant calling in long-read sequencing data.

Although long-read sequencing enables comprehensive genomic analysis, its potential is hindered by its reliance on the incomplete GRCh38 reference genome. GRCh38's unresolved bases, structural errors, and haplotype mosaics introduce reference biases, leading to missed variants and false-positive calls. This limitation creates a circular dependency because benchmarks derived from GRCh38 exclude about 15% of complex genomic regions, thereby restricting the training and evaluation of state-of-the-art variant callers. The complete telomere-to-telomere (T2T-CHM13) assembly resolves these gaps and provides an accurate genome-wide coordinate system. To bridge this transition, we present a variant training framework that integrates high-confidence variants from both GRCh38 and T2T-CHM13. Our hybrid training strategy combines rigorous variant filtering and multi-caller aggregation to produce more reliable variants for model training. This mixed-reference model demonstrates superior accuracy across sequencing technologies, coverage depths, and diverse samples. Notably, the model trained solely on a single CHM13-specific sample outperformed models trained on multiple GRCh38-specific samples on the Oxford Nanopore Technologies platform. Our findings not only confirm the advantages of complete reference genomes but also reveal new challenges and opportunities for variant discovery in complex genomic regions.