Every few weeks someone asks why we don’t just let people upload their genome to a chatbot. It’s a fair question with a boring answer and an interesting one.
The boring answer is arithmetic. A 15× whole genome comes off the sequencer at 38 to 47 gigabytes. Strip it down to variants only and you are still two orders of magnitude past what fits.
At current input pricing, one question against the raw file would cost more than the sequencing did.
The interesting answer is that even if it fit, the answer you want isn’t in there.
The answer is computed, not looked up
This is the part that surprises people who assume a genome is a lookup table. The genuinely novel, genuinely personal insights are not sitting in a database waiting to be joined against. They are model outputs, and most of those models are less than a year old.
98% of your genome does not code for protein. That fraction was politely called junk for decades, and then turned out to be the regulatory layer: switches, dials, timing. The overwhelming majority of variants that matter for common traits live out there, and no lookup table will tell you what they do.
So the frontier moved from annotation to prediction. Models like Google DeepMind’s AlphaGenome take a stretch of sequence and predict what a variant does to expression, splicing and chromatin accessibility in a particular tissue. What that produces is a mechanistic profile rather than a bibliography. Not “this SNP appears in a paper” but “this variant likely dampens this gene in this tissue.”
No database contains that. It does not exist until you compute it.
Most traits refuse to be simple
That is the easy part. The harder reality is that nearly everything you actually care about is highly polygenic.
Take ADHD. Twin studies put heritability somewhere around 70 to 80%. The largest GWAS to date finds on the order of 27 genome-wide significant loci, and common-variant heritability lands near 14%. There is no ADHD gene. There is a diffuse signal smeared across thousands of variants, most individually meaningless, interacting with an environment nobody sequenced.
Each dot is a locus; the amber ones reach genome-wide significance. The missing heritability is the gap between what twins tell us is inherited and what common variants currently explain.
Anyone who hands you “you carry the ADHD variant” is selling a horoscope with a citation attached.
We think accepting that complexity is the entire job. The honest output is a distribution, a confidence, and a mechanism where one actually exists. That is less satisfying than a verdict and considerably more true. It is also why we keep benchmarking and evaluation pipelines running continuously against the models we ship: a score you cannot audit is a guess with a decimal point on it.
Which is why the file never moves
Two more reasons, briefly.
Confidence. At 15× mean depth, coverage is not uniform. GC-rich promoters, homopolymer runs and segmental duplications all read shallower, and a heterozygous site with four reads can present as homozygous. A model handed a bare genotype will state it as fact. A model that can ask about read depth can say it is not sure.
Consequence. A genome is not a credential you can rotate. Leak a password and you change it. Your genome also discloses roughly half of each parent, sibling and child, people who never consented to your prompt. GINA covers health insurance and employment. It does not cover life, disability or long-term care insurance. There is no undo.
So the file stays where it is. What moves is a question and an answer: five scoped tools, raw export off by default, every call audited. The model does language. The pipeline does genomics. None of which is a sequencing problem, because the gap was never technical.
The industry keeps asking how to get the genome into the model. It is the wrong question. The genome should never be in the prompt. The answer should.
Read your own genome.
Order your genome