Fulcrum Genomics LLC

01/10/2026

11/08/2026

Did you know that most clinical genomics depends on research-grade software?

That’s not a criticism. Many of the tools underneath diagnostic pipelines, clinical trials, drug development programs, and patient reports were built by small teams trying to answer hard scientific questions.

But while the assays built on top of those tools are validated, the shared tools underneath them are often benchmarked once, trusted, and left alone until something breaks.

In Nils Homer, Ph.D.’s latest post, he digs into the gap between benchmarking, verification, and clinical validation, with examples from FLT3-ITD detection, fgumi, bwa-mem3, ferro-hgvs, and htsjdk.

The central point: a benchmark can tell you how a tool performed on one dataset at one point in time. It cannot tell you how that tool behaves across versions, edge cases, architectures, dependencies, and real clinical contexts.
AI-assisted development makes verification infrastructure cheaper to build. It does not remove the need for human judgment, domain expertise, or clinical validation.

Clinical genomics needs more of this boring engineering, such as equivalence tests, regression suites, edge-case datasets, continuous verification, and clear accounting for where tools agree, where they differ, and why those differences are acceptable.

Read the full post from Nils: https://blog.fulcrumgenomics.com/p/benchmarking-is-not-validation

01/07/2026

In Q2’s Fulcrum Field Notes, we’re looking at the practical layer underneath AI-assisted bioinformatics: rewrites, pipeline confidence, data exhaust, sequencing QC, and the old plumbing that still carries most genomics workflows.

Inside this issue:

→ Nils Homer joins OMGenomics to talk about AI-assisted bioinformatics rewrites
→ Tim Fennell introduces Riker, his modern successor to Picard for deep sequencing QC
→ A look at why bioinformatics still runs on old plumbing
→ minibwa, a new aligner from Li and Homer designed as a replacement for BWA-MEM
→ How Fulcrum helps teams modernize pipelines without breaking what already works

Read the Q2 issue: https://resources.fulcrumgenomics.com/q2-fulcrum-field-notes

Want to make sure you don’t miss Fulcrum content and get it directly in your inbox? Subscribe here: https://resources.fulcrumgenomics.com/subscribe

30/06/2026

A new preprint from Heng Li and our own Nils Homer introduces minibwa, a faster mapper for short reads, accurate long reads, Hi-C, and directional bisulfite sequencing data.
The headline is speed: in the reported benchmarks, minibwa is about 4x faster than BWA-MEM and more than 2x faster than BWA-MEM2 for standard WGS short-read alignment.
But the more important question is whether downstream results hold up. In the HG002 benchmark reported in the preprint, minibwa closely matched or slightly improved on BWA-MEM for DeepVariant small-variant calling.
We wrote about why minibwa is worth paying attention to, and why this kind of low-level infrastructure work still matters for teams running genomics at scale.

Heng Li and Nils Homer revisit BWA-MEM with a faster mapper for short reads, accurate long reads, and bisulfite sequencing data.

17/06/2026

Co-founder Tim Fennell has a new post up on a topic that deserves more attention in bioinformatics: old plumbing.

Not old as in useless. Old as in deeply embedded, heavily relied on, and often overdue for another look.

The post gets into HTSJDK, jlibdeflate, chelae, mako, GLIMPSE, verifyBamID, bwa-mem3, and the kind of work that makes existing workflows faster and less painful without forcing teams to rebuild everything around a new method.

A lot of scientific software depends on this layer. Things like file-format code, I/O, sorting, trimming, compression, QC, and utility tools that run constantly in production but rarely get much attention until they become a bottleneck.

That is familiar ground for Fulcrum. We spend a lot of time helping teams modernize pipelines, improve performance, validate behavior, and fix the pieces of bioinformatics infrastructure that build up tax on every run.

Take a look at his thoughts and the tools coming out of that thought process: https://blog.fulcrumgenomics.com/p/bioinformatics-still-mostly-runs

17/06/2026

Deep QC should run on every sequencing dataset.

The problem is that QC tooling has not kept pace with current sequencing scale. When metrics are slow, teams start making compromises: fewer collectors, fewer samples, less consistency.

In a new blog post, Fulcrum co-founder Tim Fennell explains why he built Riker, a modern successor to Picard for sequencing QC metrics.

His case is simple: QC should do more than raise a flag. It should help explain where signal is being lost, what kind of bias is showing up, and what part of the workflow needs attention. And it should be fast enough that running it every time becomes the obvious default.

Tim also takes on the AI question directly: if AI helps build infrastructure that saves substantial compute across real genomics workflows, that is a trade worth making.

Read the blog: https://blog.fulcrumgenomics.com/p/deep-qc-should-run-on-every-sequencing
Explore Riker: https://github.com/fulcrumgenomics/riker

The case for rebuilding sequencing QC instead of carrying old performance costs forward

17/06/2026

At , Nils Homer, Ph.D. presented our work with Darwin's Ark, Broad Institute of MIT and Harvard, and UMass Chan Medical School on building a high-resolution imputation reference panel for domestic cats.
If you missed the flash talk or poster, we’ve pulled the work into a case study.

The short version:
✅ ~1,000 cats → ~99M variants
✅ 99.5% accuracy imputation down to ~0.2X coverage
✅ Signals show up where they should (FGF5, ASIP, KIT, MLPH)

Case study:https://resources.fulcrumgenomics.com/hubfs/FulcrumGenomics_DarwinsArk_CaseStudy.pdf

What’s been most interesting (besides the data) is the model.

This is a citizen science dataset, built from fur samples collected at scale. That introduces real-world challenges (contamination, uneven data quality) that most pipelines aren’t designed to handle. Solving for that is what makes low-pass sequencing actually usable here.

We’re already thinking about how this approach applies to other species—companion animals, agriculture, and any non-model system where:
-reference resources are limited
-sequencing budgets don’t stretch to deep coverage
-datasets are growing faster than infrastructure

If you’re dealing with something similar, get in touch to see if we can help.

17/06/2026

🚨 Ambient DNA preservation without sacrificing sequencing quality is now backed by whole‑genome data.

Check out our new published collaboration with Cache DNA evaluating ensilication, their silica‑based method that lets DNA be stored at room temperature, head‑to‑head against −80 °C storage across clinical assays and WGS.

Highlights:
-100% concordance on a clinical lung panel from FFPE tumor DNA after 14 days, including variants down to ~2% VAF
-In 32 tumor–normal pairs, frozen storage accumulated up to 65% more artifactual C>T mutations than ensilication, adding thousands of false‑looking variants per million bases
-Ensilication preserved both linear and circular ssDNA across −80 °C to 37 °C while maintaining accurate detection of ~5% variants

As Michael Becich, CEO of CacheDNA, puts it: “The first era of genomics was built on cold storage. The next belongs to elegant biochemistries that preserve truth at the molecular level without ongoing energy or intervention.”

Curious how we quantified these effects at whole‑genome scale, and what they mean for oncology, biomarker development, and global genomics?

Whole‑genome analysis of storage-induced artifacts in tumor and normal DNA

17/06/2026

Fulcrum co-founder, Nils Homer, Ph.D. has been thinking a lot about what AI actually means for bioinformatics tools, and the conclusion he keeps coming back to is pretty simple:

Most tools throw away the exact data machine learning needs.

Pipelines usually emit the final answer — alignments, variants, consensus reads — along with a handful of QC summaries. But along the way they calculate a huge amount of signal: alignment ambiguity, error patterns, read support distributions, intermediate statistics. Much of that disappears once the run finishes.

That discarded information is training data.

In this new post, Nils walks through why tool design needs to change if we expect ML and AI agents to actually help with analysis, and why emitting richer, structured metadata may matter more than the models themselves.

Worth the read if you build pipelines or spend time debugging them.

If you're not building tools that emit rich data for machine learning, you're wasting your compute.