DIY genomics
The human genome project was the at-the-time largest collaborative project in biology. Completed in 2003, it took 13 years and (in today’s money) about $6.5 billion USD. Last month, I sent a swab of the inside of my cheek to be fully sequenced (30x redundant reads) for $400 USD. That means the cost to sequence a human genome fell by more than a factor of two each year for the last 23 years.
When I mention that I’d gotten my genome sequenced, people have asked what I expect to find out about my ancestry. The thing they’re thinking of is a cheap genotyping array of known sites, covering ~0.02% of the full genome, or ~5 MB compressed. My full genome sequence (with what turned out to be closer to ~52x redundancy) came in at just under 80 GB compressed. The title image is a visualization of my sequenced genome.
Now, the service which sequenced my genome helpfully provided pages and pages of auto-generated reports of medical risks, potential medical risks, medication interactions, etc. So why would I spend storage to keep an 80 GB file, with backups, let alone 250 GB of additional genomics software, caches, references?
- First, genomics is not static. My genome remains the same, but the amount of useful information I can get from it increases every year. 🧬💖
- Second, and more immediate: the average person is not well equipped to assess the quality of a gene sequencing service’s auto-generated reports, and generating high-quality reports costs compute time (I ended up running DeepVariant for 5.5h), genomics expert time (keeping databases of flags up to date with latest medical expertise), and software engineering expert time. So, responding to those incentives, the quality of the generated reports is abysmal.
Poor quality reports
Sequencing.com provides a large number of health reports, once they’ve sequenced your genome. Of course you’ve got your AI Reports Premium in nine different categories, your Health Scan, your Genome Explorer, your Next-Gen Disease Screen (NGDS Premium). You can use AI report credits to generate new reports, with options like Profession Reports ("designed for specific professionals, such […] naturopaths"). 🙄
Seems pretty comprehensive! What more could we want?
The immediate issue - all of these reports appear blind to genotype. Despite the thoroughness of the genome sequencing itself, the generated reports told me my single copy of GJB2 c.35delG was a highest-evidence highest-risk of hearing impairment, despite that being a recessive trait. Radial Aplasia-Thrombocytopenia Syndrome is on my report, because ruling that out requires checking two places in the genome, and their pipeline only checks one. Von Willebrand Disease Type 2 is on there. There are no polygenic risk scores, and even single-gene risks aren’t accounted with any accuracy.
Doing it myself
Open source to the rescue.
My machine minds and I used DeepVariant, VEP with ClinVar and gnomAD annotation, PharmCAT (drug genetics), ExpansionHunter, bwa-mem2, samtools, mosdepth, etc. So, not so much “doing it myself” as “doing it with the help of so many incredibly dedicated people and institutions, and AI minds distilled from the expertise of millions more”.
Some personal results
In a win for the assumption that genomics improves over time, my Von Willebrand Disease Type 2 mutation was rated as “likely benign” just last year! ☺️
Less encouraging, what Sequencing.com only listed as APOE (possible risk of Late-Onset Alzheimers), turned out to be APOE ε4/ε4, that is, I inherited one copy of the ε4 allele from each of my parents. About 2% of the population has two copies of ε4, and having two copies significantly increases the risk of Alzheimer’s (MCI or dementia) to ~60% (lifetime), vs 11–14% in the general population.
It’s actually just a single base pair (T->C) in each of my copies of the APOE gene that may decide my fate. Here’s a quick visualization:
click chromosome 19 to zoom in
click APOE to zoom in
click codon 112 to zoom in
| type | 112 | 158 | |
|---|---|---|---|
| ε2 | Cys | Cys | |
| ε3 | Cys | Arg | GRCh38 (reference) |
| ε4 | Arg | Arg | both my copies |
But in a further win for deeper analysis, my polygenic non-APOE risk for Alzheimer’s puts me in the bottom third of measured cohorts. With a nice genome-wide risk reduction, plus LDL and blood-pressure control, and a lot of exercise, some back of the envelope estimates bring my risk down to something more like ~26–32%. So in the counterfactual where we’re all still alive in 50 years (re: the impending singularity), and Alzheimer’s hasn’t already been cured for decades (re: the impending singularity), there’s only about a one-in-four to one-in-three chance that I’ll have to kill myself (rather than lose my mind) - not so bad. 😊✨
I’m looking forward to repeating this analysis in the years to come.