Germline Variation Calling from Long Reads' Alignment Data through Spatiotemporal Attention

Although Oxford Nanopore Technologies (ONT) long-read sequencing with the Q20 chemistry has reduced the raw error rate to 1%, this erring rate in base insertions, deletions, and substitutions remains significantly inferior to that of NGS short-read sequencing (0.1% error rate). This limitation poses substantial challenges for variant calling in complicated genomic regions from deeply sequenced long-read whole-genome alignments. Current deep learning methods exhibit 20,000-30,000 false variant calls per chromosome in Q20-calibrated ONT longread data. For Guppy v5.0.14 basecalled data, the number of errors or false calls exceeds 30,000 in structurally complicated regions. We present Attdeepcaller, a spatiotemporal attention based deep learning model that dynamically disentangles genuine sequencing errors from true germline variants in the alignment data, thereby significantly improving the prediction robustness at the complicated regions. On the HG002 chr1 Q20 data, the false calls are decreased from 26,043 to 22,739 (12.69% improvement); on the HG003 and HG004 datasets, there are 16.49% and 23.58% reductions in misidentifications, respectively. In addition, Attdeepcaller has significantly improved performance on crossversion data benchmarking tests. On the Guppy v5.0.14 datasets, the accuracy on the complicated areas has increased by 3%, and the recall rate has increased by 1%. On the Guppy v3.4.5 datasets, the accuracy is improved by 16%, and the recall rate is increased by 10%, proving the algorithm's strong adaptability to low-quality sequencing data.