Customizable, Up-to-Date Reporting with AI Assistance

Thank you to everyone who joined our webcast, “Scaling Pharmacogenomics in VarSeq: Customizable, Up-to-Date Reporting with AI Assistance.” Pharmacogenomics applies to everyone, and in some countries a PGx panel or genome is becoming standard of care. That makes it a natural fit for high-volume automation.
But we kept seeing labs hit the same wall: keeping up with current guidelines, customizing what goes on the report, and automating at scale tend to work against each other. Change one, and the others get harder. We set out to support all three on one platform.
More guidance sources, your choice of which to use
In the US, CPIC aggregates widely adopted gene-drug guidance, and the FDA’s gene-drug labeling guide adds its own, sometimes more differentiated, information. In our next release we are adding the Dutch Pharmacogenetics Working Group (DPWG), which has a long history in Europe. These sources often agree, but not always, and some drugs only have guidance from one of them. You can report from one source or combine all three as your starting point.
Knowing a reference call from a no-call
A standard variant-calling pipeline produces a Variant Call File (VCF) with one row per detected change relative to the reference sequence. But PGx workflows need something different; they must know the state of key “known variants” in PGx genes. A standard VCF file can’t distinguish between “this sample has the reference allele” and “there was no coverage here.”
There are three ways to get there. Your secondary analysis can force-call the must-call sites (built in as an option in our Sentieon workflows in VSWarehouse). You can import a gVCF from Sentieon, DRAGEN or another genome pipeline. Or VarSeq can read coverage directly from the matched BAM file. You configure this once in your project template, matched to your secondary analysis strategy, and it applies every time you automate.
Your catalog as the source of truth
The new PGx Catalog Builder is a VSWarehouse plugin that lets you customize what goes on your report. Instead of annotating against our stock PGx sources, you annotate against your own catalog of genes, haplotypes, diplotypes, drugs, gene associations and recommendations. Everything is fair game: add drugs (we have seen people add caffeine metabolism), add local guidance, edit language, or translate.
Staying current is trickier than it sounds. Once you have customized content based on CPIC version 1, version 2 arrives. You want your additions, edits and deletions to be carried forward. To achieve this, we built a powerful upstream source merging system. If you deleted a recommendation, we respect that and keep it deleted. If you changed one field and CPIC changed another, we merge cleanly. Only when you and the source edited the same thing do you need to resolve a conflict, and you can preview the impact of an update before applying it.
Changes are saved as drafts that can be assigned, reviewed, approved and published, with version history and role-based access control to separate who can edit from who can publish. And because this is built on VSWarehouse, the AI infrastructure layer we covered in our recent webcast is right there to translate, summarize, clean up and research, using the AI provider your administrator chooses.
What we demonstrated
We created a VarSeq project from a whole-genome DRAGEN gVCF, subset to the CPIC genes. Sample information, including the CYP2D6 diplotype from DRAGEN’s PGx output, was pulled automatically from the sample catalog. The PGx algorithm annotated against the warehouse’s own catalog, and the rendered report showed FDA and DPWG recommendations side by side, such as clopidogrel for a CYP2C19 intermediate metabolizer.
Saved as a template, the same workflow ran through VSPipeline on four panel samples, going from VCF to PGx reports in under a minute. In the Catalog Builder, we compared drugs where the sources agree, where only DPWG and FDA have guidance, and where they disagree. Then we used the AI assistant to translate the citalopram recommendations into Danish, published the draft, and reran the workflow. The new reports had the Danish interpretations, closing the loop from customization to production.
Questions from the webcast
Can I import my own recommendations and other customizations?
Yes. I edited by hand in the demo, but every catalog behind the Catalog Builder has an API, so bulk imports are a matter of making sure the data is structured correctly. If you have a list of customizations, we will help you automate bringing them in and keeping them up to date along with any custom data sources you are interested in. The AI infrastructure and API layers go hand in hand here and make this much easier.
Can the output PGx data be imported into my LIMS or medical record system?
Yes. The Word report is just one option in the report renderer. Both in VarSeq and in VSPipeline you can choose a backend that outputs a structured file such as JSON, and the report system can even push the data to an API. There are several ways to get structured data out and automate integration.
What software do you recommend using to generate a VCF file? Can you see the differences between PGx results in array data and WGS?
We integrate Sentieon into VSWarehouse to go from FASTQ to VCF with must-call PGx sites built in, and DRAGEN and other standard pipelines work well too. What matters most is that you can distinguish reference calls from no-calls, through must-call sites, a gVCF, or the matched BAM. VarSeq PGx also imports array data (see flexible import of array and NGS data). The difference comes down to what was measured: an array only reports the sites it was designed to probe, while a genome covers every defining variant. For any diplotype call, VarSeq shows the variants it was based on, so you can see what each assay contributed.
Can you upload the FASTQ files from a WES?
Yes. Our Sentieon-powered secondary analysis workflows in VSWarehouse take FASTQ files from panels, exomes and genomes and hand off directly to VarSeq. With an exome, be aware that some PGx-defining sites fall outside the capture targets. With must-call sites enabled, those are reported as missing genotypes rather than assumed to be reference.
Can you call typically hard-to-call regions, for example CYP2D6 or HLA, from 30x short reads? Why not, or if so, how?
For CYP2D6, yes, but not with a small variant caller alone. CYP2D6 has a close pseudogene, CYP2D7, along with gene deletions, duplications and hybrid alleles, so you need to know copy number and structure to get the full range of diplotypes. Our CYPCaller runs in our Sentieon workflow in VSWarehouse, works from the BAM or CRAM to detect these structural variants, and outputs star allele diplotypes ready for the PGx project (see Variant Calling and CNV Import in PGx Genes).
You can also bring in diplotypes from other callers. As in the demo, DRAGEN’s PGx calls can be parsed into the sample catalog and imported automatically. The same diplotype list accepts output from tools such as OptiType or xHLA for HLA (see Importing External PGx Calls and Defining HLA Diplotypes for Pharmacogenomics).
Long reads from PacBio and Oxford Nanopore give you the most complete picture. They can phase variants across a whole gene and resolve structural variation, turning ambiguous diplotypes into precise ones. VSWarehouse supports long-read pipelines and imports calls from long-read tools such as Paraphase (see Paraphase Integration in VSWarehouse 3 and Long-Read Analysis Made Simple). We have also done R&D with Twist on targeted long-read PGx panels, with posters showing very good star allele accuracy. You can find those in the publications section of our website.
Where to go from here
PGx should not force a choice between staying current, customizing your report, and automating at scale. With CPIC, FDA and DPWG sources, a catalog you control with structured updates and governed publishing, AI assistance, and fully automated workflows, you can have all three.
If you would like to see this with your own assay, sample types and level of automation, contact our team to set up an evaluation. You can also watch the webcast recording, or visit us at AMP 2026 in Seattle, booth #742.