Stanford Medicine COVID-19 test now in use | Hacker News Reader

Stanford Medicine COVID-19 test now in use

med.stanford.edu

128 pointsdivbzero35 comments

How test works

The Stanford test uses a technique called reverse-transcriptase polymerase chain reaction, or RT-PCR, to rapidly identify the presence of viral RNA in swabs from the noses of potentially infected people. RT-PCR uses short stretches of DNA called primers that bind tightly and specifically only to matching sequences in SARS-CoV-2 RNA. An enzyme called reverse transcriptase then converts the viral RNA into complementary DNA, and as the reaction continues, an enzyme called polymerase is used to generate billions of DNA copies that can be detected by fluorescently tagged molecules called probes.

Thanks!

What I don't get from this description is hos they get the virus to release the RNA?

I thought virii only release their RNA after entering a cell? Is there some sort of artifical cell wall involved?

You can just destroy the virus, e.g. by grinding it down or via a certain chemical and then you have a resulting crude mixture of RNA + other virus parts. In the past you had to separate the RNA first, but with newer methods you can do the PCR directly on the crude mixture.

Why can't RT-PCR tests be used to detect virus in a sample mixed from a large number of individual samples? This would enable cheap and accurate mass testing, since you can find P positives from S samples with O(P*log(S/P)) tests. What's the catch?

The main cost isn't per sample, it's the amount of time the machine takes to sequence enough fragments of RNA to achieve statistical significance. If you combine samples and want to not miss anything, you need to run it for longer.

Realtime PCR isn't sequencing. It increases the concentration of the targeted sequence exponentially, so it "should" take only logarithmically more temperature cycles to handle a smaller concentration of virus . If that is actually a big contributor to the testing cost, it would make the log in my analysis into a log^2.

There's no catch. It's called "pooling" and is used, e.g., to keep costs down for early detection of primary HIV infection.

Just curious, can you expand a bit on the analysis? I see that by binary search you get O(log(S)) to identify a single positive, and naively rerunning search from the beginning P times yields O(P*log(S)). One can certainly do better by avoiding redundant tests but at some point one needs to start reasoning about the distribution of the positives in the samples and I'm not sure how to do that.

See https://youtu.be/GqAdVQeIXA0 for an explanation: starting at around 30% probability pooling stops making sense.

Thought about it overnight and this is what I have:

To reason about the distribution of positives in the sample, we consider two limiting cases, then interpolate between them. The cases are:

1. All positives are clustered exactly together. We start with log(S/P) queries to find the parent node of all positives, then test every node in the subtree with 2P-1 queries to identify all positives. In total, this is O(p + log(S/P)).

2. All positives are uniformly distributed. By the pigeonhole principle, the lowest level where every node tests positive must occur at log(P). Above this level, every node tests positive, so we must perform 2P-1 tests. Every P nodes on the level must be binary searched; each node has a subtree of height log(S/P), for Plog(S/P) operations. In total, this is O(P + Plog(S/P)) = O(Plog(S/P)).

It is not immediately obvious that the worst case distribution is either case 1 or 2. To address this, we define an interpolation variable K in [0, 1] which describes a distribution where all positives begin clustered together, then KP positives are adversarially moved elsewhere. We claim that K = 0 is case 1 and K = 1 is case 2.

Consider the case where KP = 1. As in case 1, we need 2P-1 + log(S/P) tests to fully identify the cluster; we also need a single binary search starting at the 2nd level corresponding to the other half of S, which takes log(S) - 1 operations, for a total of 2P-1+log(S/P)+log(S)-1. As KP increases, the level at which we transition from testing every node to initiating binary search on a node will decrease; as an example, when KP = 4, we fully test halves, then quarters, and then begin binary search. In general, we will need KPlog(S/KP) queries for all binary searches. Finally, no matter what KP is, we need no more than O(P) queries for subtrees where we need to test every node. This yields O(P + KPlog(S/KP))

This allows us to navigate around an adversary choosing the worst distribution of positives possible by taking the upper bound where K = 1, yielding O(P + Plog(S/P)) = O(Plog(S/P)), though this bound is not expected to be tight.

Looks like it's a RT-PCR. How different are their primers compared to the IDT sets (https://www.idtdna.com/pages/landing/coronavirus-research-re...)?

There are also a couple of papers on biorxiv with primers for RT-LAMP tests, which should be much faster, easier to perform than RT-PCR. And it requires non-sophisticated equipment.

It all comes down to the false negative rates though.

To everyone questioning the false positive and false negative rates: your point is well-taken, and it is the reason why it was necessary to push the FDA to allow non-CDC labs to perform COVID-19 testing based on a fast-tracked version of the clinical validation process.

To those that point out that there are newer/faster/better ways to actually achieve the same results: you're right but that doesn't mean that the machines, reagents, collection media, etc. are ready for prime time in this type of setting. Remember: the entire process needs to be clinically validated in order to earn the trust of doctors and public health officials who are literally making life vs. death decisions.

TL;DR - Your points are valid, and there are real challenges that make certain solutions impractical at this point in time. Consider this a call to action to entrepreneurs who are serious about helping patients at home and around the world.

Do we have any idea of what the false-positive rate of these types of tests generally are? It sounds pretty thorough but some actual data on that would be nice.

Also, given that the prevalence of the virus is still very low, I am assuming (and the article seems to support this) that the test would only be used on people that have some pre-existing risk factor, like exposure to a known infected person, or travel to/from one of the high risk locations. Otherwise (let's assume a FP rate of 1%) you would still be getting a large number of false positives. Even with the assumption that there are a large number of very mild undiagnosed cases.

The reason I bring this up is because I have seen many people calling for mass testing of everyone in the United States, regardless of symptoms or risk factors. I think doing that would be a very big mistake because then you end up with a few possible scenarios:

1. You tell everyone who tested positive without any symptoms/high risk factors that they probably don't have it actually.

2. You end up treating massive amounts of people with potentially risky treatments, or quarantine people for no reason.

3. You have to do lengthy/expensive follow up tests of millions of people and waste valuable time and resources.

I guess 1 isn't a big deal, but then how do you prioritize who to treat given the symptoms alone don't provide a very useful criteria for judging who has it?

Obviously the doctors and scientists working on this know all of this already, but the general public and politicians seem to be demanding mass testing of the entire population, which worries me.

False negatives may actually pose a greater public-health problem. A suspected case, assayed as a false negative can be released back into the community without quarantine.

That's very true. From what I can tell they are focusing on making the false negatives very unlikely for that reason.

Do we know how easily it spreads without any symptoms? If someone is exhibiting flu-like symptoms then I'd expect/hope they self isolate and take precautions regardless of whether they test positive or not.

Right now we are releasing everyone who has non-extreme flu symptoms without testing whether it’s flu or covid-19, so even a small false negative rate would be a great improvement.

It’s generally hard to have a “false positive” in a PCR based test unless the sample was contaminated with the material trying to be detected. False negatives would be much more common.

My understanding - from the context of criminal forensics - is that false positives from RTPCR are quite problematic, actually, due to the tendency to pick up non-target DNA that comes close enough to catch the primer. It's an issue in DNA-rich environments (e.g., nasal swabs swarming with local flora). I'm open to learning to the contrary if anyone has a decent citation on the topic.

Human DNA is extremely similar to any other human, but coronavirus is not very similar to other contaminants. So you need fewer samples of DNA to determine a match.

Also, even a few percent chance of error in forensic DNA matching is a huge problem. If a virus test is wrong a few percent of the time it's still useful.

Wouldn't it be relatively similar to other coronaviruses? With coronavirus causing 15-30% of common colds and it being cold season, that's a lot of potential contaminants.

[Actually, the article contains the answer to my objection: they tested for this. The test was developed by explicitly looking for primers that give positive results for known-positive COVID-19 patients but give negative results for other test subjects with common colds or other respiratory illnesses. So assuming they had enough training data, this should already be factored in, and the test should remain accurate. Posting this anyway in case other people are thinking along the same lines...]

I thought the PCR tests in Asia had a specificity of around 70% which is why they needed two positive tests in a row (or two negative tests in a row to confirm the patient had recovered).

What is "specificity" and do you have a source to back up your claim?

(ie, if an individual is infected does it detect it 70% of the time? Or just a true negative 70% of the time? What about the chance of a false negative? A false positive? Etc)

Not op but specificity is

TN/(TN+FP)

https://en.m.wikipedia.org/wiki/Sensitivity_and_specificity#...

RT-PCR is the typical way to test viral infection. Anyone knows if Stanford Clinical Virology Laboratory has invented anything new or just a press release to note that they are prepared to do SARS-CoV-2 viral testing since CDC has finally authorized all private and university labs can perform tests just recently?

The last two paragraphs mention that they look for two gene while the CDC test looks for a different gene. So I don't think anything fundamentally new but having multiple tests let them crosscheck each other.

Is nobody else bothered by the logic in the 3rd paragraph?

> As with all currently available tests, it’s not yet clear how long a person needs to be infected before testing positive, or whether someone who's infected could be identified by the test before displaying symptoms.

I suppose they haven't yet had a chance to test it fully in the field, so they don't know the error rates. Doesn't necessarily mean the test is useless.

A recently saw a paper calling that COVID-19 was easy to detect early in the course with an oral swap. But later in the disease course it was easier to detect with a fecal test. Turns out a lung scan was more accurate than either.

It bothers me too reading about the testing methodology on various countries. In China anal swabs tested positive in released infected patients, while they were repeatedly testing negative with oral swabs. This has come up in their "reinfection" investigations, which may have really been partial recoveries appearing as full recoveries according to oral swab tests.

What exactly is it that bothers you? Were you hoping for a test that could flag a person as soon as the first body cell was invaded by a copy of the virus?

Isn’t China using rtpcr for the test from the beginning?

Hug of death.

I hope the publish their own (anonymized) stats and don't let the whitehouse tamper with hiding numbers like they did with the CDC.

NY State has their own test too right? Are they going to publish stats?

> As with all currently available tests, it’s not yet clear how long a person needs to be infected before testing positive, or whether they could be identified by the test before displaying symptoms.

Kind of made me sad. We're celebrating that we are able to test something, in very... precarious? .... results. Of course it's amazing that we're able to known the genetic sequence in so few days. But as travels are much faster around the globe, it takes so few days to spread. Everything being so fast isn't exactly helpful to settle emotions.

Reply on news.ycombinator.com