dvdt

BillionToOne | Staff Software Engineer | SF Bay Area | Hybrid or Remote

BillionToOne has developed a DNA molecular counter that increases cfDNA diagnostic resolution by over 1,000x. BillionToOne's first product, UNITY, is the first and only non-invasive prenatal test that directly screens an unborn baby for the most common and severe genetic disorders using only a single tube of blood from the pregnant mother without the invasiveness of amniocentesis.

BillionToOne is ranked at the top 5% of Y Combinator companies and has raised $300M+ in funding from prominent VC firms.

We are hiring a Staff Software Engineer to scale up compute-intensive bioinformatics workflows and build infrastructure tools that enable scientists, bioinformaticians and other technical teams at BillionToOne to robustly write and deploy code.

Tools we use include: python, Django, AWS, Terraform, Sentry, and Datadog.

Apply here: https://boards.greenhouse.io/billiontoone/jobs/4203247005

We used a law firm. The paperwork was quite straightforward because our attorney managed it.

I'm the Co-founder and CTO of BillionToOne, one of the biotech companies profiled in the post.

We're recently opened up a position for a Staff Software Engineer (https://apply.workable.com/billiontoone/j/2F0405817C/).

If you're interested, please reach out to me directly at david+hn@billiontoone.com

BillionToOne (YCS17) | Staff/Senior Software Engineer | Remote (US) or Onsite (Menlo Park CA) | www.billiontoone.com

BillionToOne develops innovative diagnostics tests that can affect the lives of millions of patients. Our QCT technology platform improves the resolution of cell-free DNA testing by >1000x fold and enables novel tests for both prenatal and oncology care. We've raised over $30M and have launched multiple clinical products in the past 2 years, including an FDA authorized COVID-19 test.

We are hiring a senior engineer (5+ years experience) to build internal APIs, bioinformatics procecessing pipelines, laboratory automation tools, and help manage CI/CD. We use python, django, rabbitmq, circleci, dbt, postgres, heroku, aws, and a variety of other tools.

If you have experience in full stack web development, love seeing your work positively affect your colleagues, and thrive in a fast-paced entrepreneurial environment, this could be a great opportunity for you.

Please contact me directly at david@billiontoone.com if interested.

I completely agree with the point about integrated REPL/IDE, and wanted to share some of the combinations I have used in the past, since it can be a concrete getting started point for those who are curious. Some of these are not literally repls, but IMO give a similar experience.

- ClojureScript with Figwheel and the web browser

- Clojure with Emacs Cider, Clojure with Cursive

- R and Rstudio

- Matlab

- ipython jupyter notebook

- Pycharm debug breakpoints that are triggered by unittests (Running the unittest to initiate a python repl at the breakpoint)

A turn around time of 24-36 hours should be easily achievable for the performing lab, depending on the time of day the specimen arrives (morning vs afternoon). It takes about 1 hour for the initial RT-PCR amplification, and sequencing takes about 12 hours.

Our technique would still be affected by shortages in specimen collection (like swabs).

Purely speculative, but I think if swabs remain an issue for too long, alternatives could start coming online, such as even using qtips + saline (no idea if it works, it's just an example). The current swab + Universal / Viral Transport Medium combo is optimized for flexibility; it is designed to work across a very broad range of viruses and bacteria that have different viral loads and shedding characteristics. The current pandemic is pretty much COVID-19 only, so I think it's a priori feasible that a specimen collection procedure can be found that uses common materials. We did try early on to see if saline or other buffers affected the performance of the assay, and it worked fine in those conditions.

We use fairly standard chemicals. I haven't heard from our suppliers about shortages for the chemicals we use. Chemicals and enzymes tend to be relatively fast to scale up for bulk manufacturing.

There's always manufacturing risk that a product will not work as expected. In fact, the first COVID-19 test developed by the CDC did not work as expected, and this delayed testing by several weeks. We de-risk this as much as possible by performing experiments as early as possible, akin to the fail fast mentality of checking for the highest risk failure modes first. Since we don't have a national healthcare system in the US, the manufacturer takes on the vast majority of the risk of a defective product.

My guess is that self-swabbing is allowed in the Seattle SCAN study because it is a research study. The SCAN study is super fascinating because it would be crazy unusual under normal times for a research study to return results back to patients; I'm very happy they are able to do that, and it speaks to the severity of this pandemic.

The FDA has recently made clear that no at-home self-collection tests are allowed (for now): https://www.fda.gov/medical-devices/emergency-situations-med...

Thanks for the question! The goal of skipping RNA extraction is to decrease the amount of labor necessary for processing samples and also to eliminate a dependency on RNA extraction reagents that have recently become difficult to find. The FDA is very strict about the specific brand and model of kit you use, so showing that you can swap out one RNA kit for another is actually very useful because you will have alleviated some of the supply chain strain (although I agree at high enough load both supply chains will then become limiting).

The way currently available COVID-19 testing works is by detection of viral RNA. Since the amount of viral RNA in a patient sample is too low to detect directly, we first need to amplify it by PCR. However, this viral RNA is packaged within all sorts of proteins and lipids that could make it inaccessible to amplification unless they are first purified away. Furthermore, the sample is shipped in "viral transport medium", which is essentially a cocktail of chemicals designed to preserve the virus. Unfortunately, these preservatives often have the side effect of interfering with PCR amplification, so these too need to be purified from the sample.

However, since RNA extraction is usually the most laborious part of the assay, there has been a lot of interest in optimizing the amplification so that it is resilient to all of these impurities. The preprint referenced in our manuscript (https://www.biorxiv.org/content/10.1101/2020.03.20.001008v1) gave us the initial idea that this could be possible, and much of it comes down to the choice of amplification method (e.g. choice of enzymes and buffers) that you choose.

However, even when you choose a "good" enzyme and buffer, you will still suffer an amplification penalty, and this will cause you to return a false-negative on some affected samples because there was so little virus in the sample to begin with. The innovation we have is to spike-in a correspondingly low level of DNA to the reaction mixture. That way, if you see the low level of DNA without seeing any viral signal, you can be assured that the amplification still worked and that there truly is no virus in the sample.

Thanks!

The Abbott machines are point-of-care devices that typically sit in doctor's offices. One really interesting use case I've heard of for the Abbott machines is to test all OB patients who are coming in to the clinic for routine care to make sure that they are COVID-19 negative. This allows the clinical staff to conserve PPE and use less burdensome precautions.

I think that where the Abbott machines might hit a wall is that they are one at a time, and they require Abbott's consumable test cartridge and device to run (think printer ink / printer). I don't have any firsthand knowledge, but I would anticipate that it is difficult to scale-up manufacturing of the devices rapidly enough to keep pace with the pandemic growth.

We absolutely need contact tracing to find everybody who needs to be tested. We're not working on scaling up contact tracing, but I think several people in the tech community are working on making that easier to perform at scale.

"When you’re fundraising, it’s AI When you’re hiring, it’s ML When you’re implementing, it’s linear regression"

The core of our machine learning is Ax=b :grins:

More seriously, the main reason why traditional sanger sequencing can't be used for COVID-19 testing is because it would be unclear whether a lack of signal is truly due to lack of virus, or if it is just because the assay failed (happens all the time!)

What we've done is introduce a reference sequencing signal that is biochemically very similar to viral RNA, but produces a distinct vector of electrical signals that is different from the signals emitted by viral RNA. Since we know what both the reference and viral signals look like, we can perform linear regression analysis to fit the linear combination of viral and reference signals that best match our data.

1) and 2) are exactly right.

3) is almost right--these sanger sequencing instruments are already widely available across the country for research use. Here in the Bay Area, I can choose from at least 4 different Sanger sequencing services that will run 10,000 samples at $2/sample in 24 hours. For example, see: https://www.mclab.com/DNA-Sequencing-Services.html

Thank you for your efforts as well!

Sample collection and accessioning (accessioning is unpacking test tubes one by one and aliquoting them into plates in the lab) is definitely going to require a lot of manpower. I'm hopeful that patients "self swabbing" can help alleviate some of the manpower needs. (Self-swabs are not allowed currently under FDA guidance).

Thanks for letting me know about the site!

This is a test to see if someone has a current COVID-19 infection. The antibody tests (serological) tests are also important, but since it is estimated that only ~1% of the US has previously contracted COVID-19, it will be a while before serological testing becomes useful at a population level.

Our initial data show no false-positives and no false-negatives out of all specimens assayed. However, it is early days still and none of the leading tests have real-world data on false-positive and false-negative rates. The crucial parameter here to compare test performance is limit of detection (LOD). We showed we could detect as few as 10 molecules of virus, which is on par with the best RT-qPCR tests.

Cost is definitely an important consideration for roll-out of a widespread test. We anticipate that the cost will be about $15 per test.

I'm the co-founder and CTO at BillionToOne. I'm happy to answer any questions here. I've also posted a slightly more technical explanation of how the test works and why it can scale here: https://twitter.com/dtsao/status/1247642005510873088?s=21

Edit: Since our site seems to be overwhelmed at the moment, here's a recap:

We’ve been working hard at BillionToOne on a new COVID-19 test that scales testing to everyone in the US. Our test (1) re-purposes existing infrastructure, (2) eliminates time-consuming RNA extraction, and (3) enables a distributed system for COVID-19 testing.

We need 1 million tests per day to end the stay-at-home orders. Schools are still open in Iceland because they test 15x more than the US does, per capita (https://www.washingtonpost.com/world/2020/04/02/free-coronav...).

The first thing we figured out is how to run COVID-19 tests on existing automated Sanger sequencers. One sequencer can process up to 3840 samples per day. There are hundreds of sequencers of excess capacity because they were built for the Human Genome Project over 20 years ago.

It would take only 2 sequencers to surpass the current test capacity for all of California. There are far more than 2 sequencers in California (some individual labs have 10 or more).

We tweaked the protocol so COVID-19 could be detected from sequencing data using linear regression. Basically, we add ~100 copies of a known DNA sequence to help us calculate how much virus nucleic acid is in the specimen. It works just as well as gold-standard RT-qPCR.

Lab workflow for COVID-19 testing is traditionally 1. Specimen accessioning, 2. RNA extraction, 3. RT-qPCR 4. Reporting. RNA extraction, in particular, has been a huge bottleneck in terms of reagent shortages and labor-intensiveness.

We showed that we can skip RNA extraction entirely without affecting test sensitivity and limit of detection.

By skipping RNA extraction and using automated Sanger sequencers, we think we can get to an additional 200,000 samples per day test capacity in existing clinical labs.

A distributed system is often the only way to operate at massive scale. A fully distributed system could have different sites and labs responsible for each process and dynamically re-allocate resources based on availability and capacity.

The Broad institute COVID-19 lab has already started doing this. They are asking for specimens to be submitted in a standardized tube format and pre-barcoded. They have essentially distributed the specimen accessioning work.

Because there is a highly developed service industry for Sanger sequencing with <24 hour turnaround, there is an opportunity to further scale up testing by distributing the work to their (currently) idle sequencers.

Distributed testing could scale from 200k to >1 million tests per day, but would require a change in regulations that currently prohibit it.

Thanks to the BillionToOne team for pulling this work together! Next step is to start manufacturing test kits and obtain Emergency Use Authorization from the FDA. We’re eager to work with clinical Lab Directors and contract kit manufacturers.

Edit 2: Link to scientific manuscript: https://www.dropbox.com/s/07esyehsvfpmllc/A%20Highly%20Scala...

BillionToOne | Senior Software Engineer | Full-time | Onsite | Menlo Park, CA

Do you want to develop prenatal diagnostics that can affect the lives of millions of expecting parents? BillionToOne (Y Combinator S17) is looking for a Senior Software Engineer. We transform diagnostics to be truly grounded in quantitative principles and improve resolution of cell-free DNA testing by >1000x fold. As engineer #1 you will work directly with the CTO to build backend infrastructure, bioinformatics data processing pipelines, laboratory automation tools, and web-based tools to communicate genetic results to patients. This is a highly impactful position with the opportunity to own engineering end-to-end from internal prototypes to widely deployed products directly affecting patients.

Email me at david@billiontoone.com or apply at https://billiontoone.workable.com/jobs/1143752

BillionToOne (YC S17) | Senior Scientist | Palo Alto, CA | Full-Time, Onsite | billiontoone.com

BillionToOne detects genetic disorders in the baby through a simple blood test of the mother. Our first prenatal test for beta-thalassemia and sickle cell disease is already in clinical trials. Over 100 million people are carriers for these disorders.

We are hiring a senior scientist to lead R&D activities. BillionToOne integrates nucleic acid biochemistry and machine learning to make DNA testing more affordable and accurate. The ideal candidate has DNA wet lab bench experience and is comfortable devising custom analysis pipelines in Python or R.

More here: https://billiontoone.workable.com/

One fun application: I used Hy to control my liquid handling robot,

https://twitter.com/dtsao/status/891038101706989568

Someone please correct me if I'm wrong, but it sounds like the secret is to use countercurrent exchange to rapidly cool incoming air? Incidentally, countercurrent exchange is the same principle used in the kidneys to concentrate waste products into urine.

https://en.wikipedia.org/wiki/Countercurrent_exchange

Thanks =). I'm planning on putting it up on GitHub as soon as I clean the code up a bit.

Wanted to second the NICTA course as a top-notch resource for learning Haskell. It was only by working through the NICTA exercises that I finally understood functors, applicatives, monads, etc. and more importantly strengthened my intuition for when I could apply these functional tools to practical problems.

About a year ago, I got a free email address with Google Apps on my custom domain by doing the Google App Engine workaround (https://groups.google.com/forum/#!msg/google-appengine/pVZfd...). Does this not work anymore?

This is awesome work. I just want to point out that it's pretty much a marketing lie to say that the lens gives you 160x magnification. Sort of like saying how digital zoom gives you magnification on a camera.

A non-linkbait title would also report the numerical aperture of their microscope. From their paper, it looks like the new lens performs comparably to a research grade 10x 0.25NA objective.

I can't remember the last time I ran 'git log.'

Rather, I use tig to get a quick glance at commit history and staged / unstaged changes.

Can someone explain the relative upsides and downsides to using tig vs git lg?

Looks like The Atlantic decided to preserve the 1996 layout of the "prefab hell" article. It is refreshingly minimalistic -- no ads, no links to "Just In" or "Top 10" articles. I like it much better than the modern theatlantic.com

Cvx is basically a convex solver that is called from matlab. Boyd wrote it. I'm not sure how easy it would be to divorce it from matlab, but there is a lot of existing instructional material built off of cvx, including examples in Boyd's book on convex optimization.

If you think you might be interested in this topic, it's worthwhile getting a matlab license. A lot of published academic code in this field is implemented with cvx.

I'm confused. Is python really so far ahead of the competition to be considered the "language of choice for data science?"

My impression (and I am not a statistician/data scientist in my day job, so I would very much like to hear opposing perspectives) is that the R ecosystem is far more mature and widely adopted than scikit-learn for things like regression, classification, clustering etc.

The author also cites expensive MATLAB licenses as a driving force behind python adoption, but here too I'm skeptical. As a grad student, I get MATLAB for free. But I switched to R/pandas for data analysis because R has a native data structure for working with multidimensional datasets (i.e. data.frame).

To illustrate the utility of this, let's say you asked developers all over the US for their zipcode and salary and recorded the results in salary.poll.data. Here's an interesting question: what is the mean salary in each zipcode? In R, all sorts of libraries make this computation concise and highly readable. Using the excellent data.table package, you would do `salary.poll.data[, list(mean.salary.by.zip.code=mean(salary)), by=zipcode]`.

No such libraries exist in popular usage for MATLAB. You'd have to roll your own, or more likely, write a lot of crufty loops and conditional statements. (Or use higher order functions with map/reduce/filter, which, by the way, you would have to implement yourself).

For me at least, just having the right data structures for working with data makes R/pandas a clear winner for doing statistical analysis of data.

Demanding President Obama's resignation by comparing the current NSA scandal with Watergate is an outrageous false equivalence. The Nixon administration was spying on the DNC for political gain. PRISM was built in the name of national security, and with congressional oversight to boot.

I'm as outraged as anyone about the NSA's activities, but this is an overreach that dissipates righteous anger down an unproductive path.

> ... the elite players were significantly more relaxed than the average players, and the best of the best were the most relaxed of all.

Presumably the best players also KNOW that they are the best--doesn't that make it easier to relax? How do you control for this?