6 min read

I Sequenced My Family. We Still Don't Have an Answer.

experiments genomics bioinformatics homelab

A personal story, not medical advice — I’m a data engineer, not a doctor or a geneticist.

For a few years now a question has hung over my family and my sister’s family: one of the kids has a developmental disability and a seizure disorder, and after years of specialists, scans, panels, and tests, there’s still no diagnosis. No name for it, no cause — just a growing folder of normal and inconclusive and variant of uncertain significance, and a question that doesn’t go away.

I’m a data engineer, not a geneticist — most of what follows is a layperson wandering respectfully through a field professionals spend whole careers in. But at some point it struck me that the labs and I were staring at the same thing. DNA is just data — very strange, very dense, very consequential data, but data — and moving and interrogating data is what I do all day. I couldn’t not look.

So I sequenced both families and built a pipeline to search for an answer. I did not find one. This is the honest story of doing it anyway, and of what the doing taught me.

What I actually did

The setup was an extended-family study — my family and my sister’s family, parents and kids, and later more relatives — because the most powerful thing you can do with a genome is compare it to the genomes it came from. Sequence everyone, align the reads to the reference, call the differences, and then start asking the questions clinical genetics asks. Is there a change in the child that neither parent carries — a new mutation? A pair of changes inherited one from each side? A missing or duplicated stretch? A piece of DNA that stutters and expands? Something in the mitochondria? A quirk of which parent a chromosome came from?

I built that out as a long series of analysis passes — dozens of them — each one a different hypothesis about the shape an answer might take, annotated against the public knowledge bases that record what’s known about human variation, and ranked so the most plausible candidates floated to the top. The heavy lifting — aligning billions of reads, calling variants across everyone — I burst to the cloud, because it’s embarrassingly parallel and my homelab is not a data center. The iterative part, where you form a hunch and chase it, I did locally, over and over, for months. (If you want the engineering underneath all of this — the file formats, the data-quality checks, the free tools, and how I used an AI to assemble each pass without letting it draw the conclusions — I wrote that up separately.)

It was the most familiar project I’ve ever done

And here is where it turned into something I recognized completely, in the least familiar setting I’d ever met it.

The first wall wasn’t biology. It was data movement. Genomes are enormous — the raw reads, the alignments, the intermediate files between every step — and underneath everything, the whole project was a problem of where the data lived and how it moved. The compute was cheap and parallel; feeding it, and shuttling that much data between stages without losing track of any of it, was the actual job. I have written that exact sentence about pipelines at work more times than I can count. It is precisely as true when the pipeline is your own family’s DNA.

The second wall was that every tool in genomics ships a beautiful tutorial that runs perfectly on a clean sample, and the tutorial is not the system. On real family data you fight reference builds that disagree, coordinates that are off by one, a tool version that changes an answer, two respectable databases that flatly contradict each other. The demo runs in a minute. The real thing is months of making the boring parts line up.

The third wall was the one that mattered most: a ranked list of candidates is not an answer. It is a list of hypotheses. Annotation and ranking are wonderful at surfacing possibilities and useless at confirming them — automation is good exactly at the work you can check. Every promising candidate had to be verified by hand: pulled up in a genome browser to look at the actual reads, checked against who in the family carried it and who didn’t, weighed against whether it even fit the phenotype. The surfacing was the easy five percent. The checking was everything, and the checking is where most of the candidates died.

The mystery held

After all of it — the passes, the extended family, several independent variant callers, the honest resets when a promising framework fell apart under scrutiny — I did not find a clean answer. There were leads. There were things that made my heart jump for a day, until I checked them properly and they dissolved. But nothing conclusive. The most honest label for the result is the same one on all those clinical reports: still unknown.

I’ve had to sit with that, and I’ve decided it’s a real result, not a failure. Roughly half of cases like this stay undiagnosed even in the best hands, and the reason is humbling: we can now sequence a genome cheaply and we still cannot read most of it. The sequencing is solved. The interpretation is not. A “variant of uncertain significance” isn’t a rare edge case — it’s the ordinary state of the map. We have the book, and we mostly can’t read the language yet.

What it taught me

This became the purest version of a lesson I keep meeting in smaller forms everywhere else I work: having the data is not the same as understanding it. A pipeline can be perfectly built — every stage green, terabytes flowing correctly — and still not hand you the one thing you actually wanted to know, because the knowledge required to interpret the output doesn’t yet exist, anywhere, in anyone.

And it tested something about honesty that I thought I already believed. When you’re doing this as a parent, the strongest pull in the world is to let hope turn a maybe into a yes — to round a suggestive variant up into “this is it,” because you want it to be, because a wrong answer somehow feels less unbearable than no answer. The discipline I’ve written about in calmer contexts — distrust the result you most want to be true — I have never needed more, or found harder, than at two in the morning looking at a variant in the genome of a kid in my family that I desperately wanted to be the explanation. It wasn’t. Checking it honestly, and then letting it go, was the hardest data work I’ve ever done.

The data is patient

I don’t have an answer to offer. What I have is a pipeline I’m quietly proud of, a great deal of new respect for the clinical geneticists who do this for a living and carry its weight professionally, and a kind of patience I didn’t have before.

Because the data is still there — sequenced, aligned, stored, waiting. The reason we can’t read it yet is that the science isn’t ready, and the science moves. A variant that means nothing to anyone today could be the answer in a paper five years from now. So the job was never really to solve it. The job was to capture it faithfully, keep it honest, and be ready for the day the reading catches up to the sequencing.

The data is patient. I’m learning to be.

Sources & further reading