Table 1. Performance of small variant calling with CCS reads.
Precision, recall, and F1 of small variant calling measured against the Genome in a Bottle v3.3.2 benchmark using hap.py. Bold indicates the highest value in each column. Underline indicates a value higher than the GATK HaplotypeCaller run on 30-fold Illumina NovaSeq reads. Coverage is 28-fold for PacBio CCS and 30-fold for Illumina NovaSeq. Rows are sorted (“^”) based on F1 for SNVs.
| Variant caller (training model) | SNVs | Indels | |||||
|---|---|---|---|---|---|---|---|
| Platform | Precision | Recall | F1 ^ | Precision | Recall | F1 | |
| Illumina (NovaSeq) | DeepVariant (Illumina model) | 99.960% | 99.940% | 99.950% | 99.633% | 99.413% | 99.523% |
| PacBio (CCS) | DeepVariant (CCS model) | 99.914% | 99.959% | 99.936% | 96.901% | 95.980% | 96.438% |
| PacBio (CCS) | DeepVariant (haplotype-sorted CCS model) | 99.904% | 99.963% | 99.934% | 97.835% | 97.141% | 97.486% |
| Illumina (NovaSeq) | GATK HaplotypeCaller (no filter) | 99.852% | 99.910% | 99.881% | 99.371% | 99.156% | 99.264% |
| PacBio (CCS) | GATK HaplotypeCaller (hard filter) | 99.468% | 99.559% | 99.513% | 78.977% | 81.248% | 80.097% |