Questions about multi-batch long reads processing, VCF FILTER field interpretation and 1000 Genomes BAM usage in TRGT

[CaryStar01]

Dear TRGT Development Team,
We are writing to consult three questions regarding genotyping analysis with long reads data:

1. For long reads data of a specific sample from projects like the 1000 Genomes Project (see the attached figure), does such data typically correspond to multiple sequencing batches? When dealing with multi-batch fastq data for genotyping using TRGT, what is the standard workflow? Should we merge all batches into a single fastq file, perform alignment first, and then conduct genotyping, or is it acceptable to directly perform genotyping with data from only one batch?

2. Regarding the filter field in the VCF file generated by TRGT: ##FILTER=<ID=PASS,Description="All filters passed">. Does this field indicate that the corresponding genotyping results have passed all quality filter criteria and can be directly used for subsequent analyses? We would also like to supplement our practical confusion: in the VCF files we generated, the value of the FILTER field for all genotyped variants is . instead of PASS. We wonder if this situation means that the data quality of these genotyped variants is substandard and the results are unreliable?

3. We also have an additional question: In relevant research papers, when analyzing data from the 1000 Genomes Project, we noticed a phenomenon—for short reads data, researchers generally directly use the BAM alignment files provided on the project's official website; however, for long reads data, many researchers choose to abandon the existing official files and instead re-perform sequence alignment to generate new BAM files. If your team has any insights into this, could you please share the reasons behind it?

Thank you for your time and help!

[tmokveld]

Hello, those are great questions, I'll answer them one by one:

1. Indeed, for long-read datasets from projects like 1000 Genomes, a single biological sample typically spans multiple sequencing runs (often multiple PacBio SMRT Cells) that are intended to be aggregated to reach the target coverage and QC. In your example you're looking at HG01457, where HiFi reads are provided across multiple PacBio run folders (each corresponding to a SMRT Cell). In practice, the sample-level dataset is formed as the union of reads across all runs. For genotyping with TRGT, the standard workflow is to use the full sample-level dataset, meaning you merge all runs and genotype on the combined alignments. If you would genotype from only one run you are effectively downsampling, this is mainly useful for quick debugging, per-run QC, or explicitly low-coverage experiments. Now you generally want to stay in BAM space. If aligned BAMs are available, merge the per-run BAMs (preserving read groups) and run TRGT on the merged BAM. If you must realign, avoid a BAM->FASTQ->BAM roundtrip when possible. You should instead strip existing alignment and phase-related tags first, for example with samtools reset, and then realign. The human WGS WDL workflow is a great resource to learn about the recommended practices of working with PacBio data, where this stripping is done as such.

2. In VCF, FILTER=PASS means the record was evaluated against filters and passed them, and FILTER=. means "no filters were applied / not assessed" rather than "failed", i.e., seeing . does not indicate substandard data quality or unreliable calls. In the current TRGT output, the FILTER field is not yet used to encode TRGT-specific quality filtering. Instead, for practical downstream QC you should use TRGT’s per-sample fields that actually quantify evidence and consistency. In particular, inspect allele-level support such as SD (spanning-read support per allele) and AP (allele purity per allele).

3. For short reads, the 1000 Genomes BAMs are usually the canonical, project-standard alignments with a fixed reference genome build using BWA, so most analyses can use those. For long reads, many repositories distribute unaligned BAMs or FASTQs, and even when alignments exist, groups often realign them because mapping choices can be analysis-dependent and have changed over time. Now for your specific HG01457 sample, the header is consistent with an unaligned CCS BAM (generated by ccs, merged by pbmerge, no @SQ reference dictionary shown, no aligner @PG). You can confirm with samtools view -H https://ftp.1000genomes.ebi.ac.uk/vol1/ftp/data_collections/HGSVC3/working/20220831_JAX_HiFi/HG01457/m64119_211210_191732/JAX_HiFi_HG01457_m64119_211210_191732.bam. If it is unaligned, the standard approach is to align with pbmm2, sort/index, then merge all runs for the sample, and run TRGT on the final merged BAM. The PacBio WDL shows clear pbmm2 usage.

Happy to help if you have further questions!