Moderna mRNA sequence released to GitHub [pdf] (opens in new tab)

The protein they're trying to manufacture is indeed quite simple - AFAIU both BioNTech and Moderna put together their sequences in a weekend. (Though there was a more involved process of winnowing down the sequences for the most effective ones.)

The technically challenging parts are:

- delivery mechanism: you need to take a very unstable molecule, protect it from the environment - both external, and when inside the patient - and insert it into a human cell. (This is called the "platform", and is usually developed independently from the specific payload.)

- manufacturing: both producing the mRNA itself at a large scale, and inserting it into the delivery mechanism, at a large scale and in low-temperature conditions

- testing: the newly-developed payload and the existing platform were integrated at small scales within weeks, but testing the thing for safety and efficacy took months

EDIT: As schoen pointed out, this was not actually released by Moderna, but reverse engineered by third-party researchers. Original text was: "Hence they feel safe releasing this. Their moat is not the gene sequence, their moat is everything else."

Liken it to the 4kb demoscene: it's amazing what can be done with a little bit of information, as long as you don't have to describe the machine running it.

Or the distribution method, or even really invent the thing, since you're mostly just copying someone else's work. Plus it doesn't have to even do anything. In fact, doing anything might be a problem, so best to just sit there and look menacing (and spikey).

It really is that “simple.”

Getting it designed and building it is more difficult.

At its core, it’s a piece of mRNA that creates a protein. That code gets transcribed into a protein (often those are relatively short). That protein then triggers your bodies immune response, which trains it to attack covid19.

Inject this mRNA into a cell and it’ll create the protein. Anything can be injected at this point once the mechanism for injection is developed

Sequencing technologies have improved immensely over the last decade and a half. And, in this particular case, getting the sample RNA is incredibly easy, since its purity and integrity in the vial is quite high.

I think it's a bit like a private key- the difficulty is in finding some combination that works in an absolutely massive space of possible proteins, not necessarily in the length of the protein.

Check out this video by The Thought Emporium to see how far we’ve come in these matters:

https://youtu.be/J3FcbFqSoQY

This should hopefully provide you with some useful perspective.

"but I'm utterly amazed at how simple this _appears_."

Biology is a funny old thing. You can look at that concise description - the orange and so on blocks of a few letters and a few short groupings.

Now ATCG are basic building blocks but they consist of quite a lot of stuff. I think it's a bit more complex than that because this is RNA not DNA so ATCG might not be quite right. Each of those bases are horrifically complicated depending on scale. Search "ATCG" - this is a good start: https://en.wikipedia.org/wiki/Nucleobase

Now dive into one of those bases and decompose it to its constituent atoms. Now look at the maths around this stuff. It gets quite complicated, quite quickly.

That said, the fact that a bloody complicated thingie can be described so concisely is absolutely amazing and as you say it looks so simple.

Here is a breakdown https://berthub.eu/articles/posts/reverse-engineering-source... discussed previously https://news.ycombinator.com/item?id=25538820

Any individual protein doesn't seem that complex since it's just a combination of some 20 amino acids, but the variations are endless:

"Since each of the 20 amino acids is chemically distinct and each can, in principle, occur at any position in a protein chain, there are 20 × 20 × 20 × 20 = 160,000 different possible polypeptide chains four amino acids long, or 20n different possible polypeptide chains n amino acids long. For a typical protein length of about 300 amino acids, more than 10^390 (20^300) different polypeptide chains could theoretically be made. This is such an enormous number that to produce just one molecule of each kind would require many more atoms than exist in the universe."

People tend to think of genetic code as a sort of assembly language which is very verbose, but I wonder if the correct way to view it is in fact a very terse domain-specific language, because it actually depends on the entire complex machinery of the cell to be present in order to work, which in itself contains a lot of information?

> and I can easily fit it on my screen.

...with GATACCA right in the middle, but unfortunately with no GATTACA that I could find.

I'm estimating roughly 90-ish characters in a row, roughly 40 rows encoding the spike protein. So about 3600 base pairs. There are 3 base pairs per amino acid, so That's 1200 amino acids.

For comparison, the smallest chain that they technically call a protein is 100 amino acids that's an arbitrary limit to separate proteins from enzymes. So this thing isn't tiny tiny.

But Titin (also called connectin), a giant protein responsible for passive elasticity in mucles, is ~27,000-35,000 amino acids. So this thing isn't even close to the biggest proteins out there.

> "Instead it's no bigger than a large paragraph of text, and I can easily fit it on my screen."

When I saw it, I thought that it could almost fit in a tweet, so I just did it:

https://twitter.com/weinzierl/status/1376807707957719041?s=2...

The sequence takes 16 tweets, 15 if you don't split at line endings and remove spaces (4175 nucleobases / 280 nucleobases/tweet ~ 14.9 tweets).

"but I'm utterly amazed at how simple this _appears_."

Remind me the joke of the consultant engineer knows where to make X by the chalk. LOL

Not Moderna, but this [1] was a very useful primer on grokking how the Pfizer vaccine works, especially for computer programmers.

[1] https://berthub.eu/articles/posts/reverse-engineering-source...

the way I see it we're just at the beginning, and we're mainly copy/pasting a lot of code, we understand some small parts, and generally in the teenage years of genome programming.

I don't know how long it will be before we get a bit more serious with it, but geneticists have a big obstacle in their understanding, any change might needs a thousand strong lifelong population study to be understood. That's way crappier than dumping the assembly or only having the documentation in Chinese.

I will add that moreover the developers might have been even more conservative in their code because they knew it was going for large scale deployment, they probably avoided the cutting edge as much as they could.

Bravo! Nice execution of the tips from yesterday's article!

https://www.cs.purdue.edu/homes/dec/essay.criticize.html

Great quote from Maurice Hilleman, creator of many (most?) of our childhood vaccines goes something like “Don’t be smart. Instead be careful and accurate”

Lots of these things aren’t complicated. It’s the careful systematic testing and public trust building that’s the hard part.

The genetic code itself is reasonably comparable to ASCII in complexity - every 6 bits is the code for one amino acid in a string, which will fold itself into the required protein.

I remember a lot of features and especially bug fixes where I had to change one line of code, it took hours to figure out how exactly though. I guess this is kinda similar?

The way it reads like source code, truly makes me circle back to the idea we're all living in a simulation.

Mathematical truths about abstract notions of string theory fit in a line.

The New York Times published an article last year with the entire genome of the SARS-Cov-2 virus, with a breakdown of different sections to explain what protein the RNA codes for and what that protein does. Like you said it was amazing that it all fit within an [albeit long] newspaper article. It doesn't surprise me that the RNA for the vaccine, which only targets a single protein, is even smaller than that. Here's the NY Times article I was referring too:

https://www.nytimes.com/interactive/2020/04/03/science/coron...

It appears simple, but a whole lot of work went in to producing that string even pte-COVID. Some of it is generic in the sense that it might apply to any mRNA vaccine. Some is quite specific:

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5584442/

There’s also (IIRC, no citation right now) prior work suggesting that coronavirus vaccines against the spike are likely to be effective and that vaccines against the N protein might be counterproductive.

each one of those letters represents a ~15 atom molecule, so in a way it is a compressed representation

See RadVac: https://radvac.org/

Make your own, open-source. Really cool.

A user on lesswrong made their own (with no prior experience): https://www.lesswrong.com/posts/niQ3heWwF6SydhS7R/making-vac...

It's not really that simple.

Only two companies in the world succeeded, the French company Sanofi which also tried making a mRNA vaccine failed.

so, explain it to me ?

Wouldn't the "compiler" be the bioreactor used to mass-produce it and the "installer" be the lipid encapsulation? :)

But wasn't the whole Pfizer/BioNTech "secret sauce" leaked online after the EMA was hacked?

Meh, they probably just used lipofectamine (which has been around since the 90s) or something very similar.

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

https://www.thermofisher.com/us/en/home/brands/product-brand...

A pairwise sequence alignment done with `needle` starts like this:

  BioNTech_Pfiz      1 -----------GAGAATAAACTAGTATTCTTCTGGTCCCCACAGACTCAG     39
                                  |||||.|.|..||||                |||   ||
  Moderna            1 GGGAAATAAGAGAGAAAAGAAGAGTA----------------AGA---AG     31
  
  BioNTech_Pfiz     40 AGAGA----AC-------CCGCCACCATGTTCGTGTTCCTGGTGCTGCTG     78
                       |.|.|    ||       ||||||||||||||||||||||||||||||||
  Moderna           32 AAATATAAGACCCCGGCGCCGCCACCATGTTCGTGTTCCTGGTGCTGCTG     81
  
  BioNTech_Pfiz     79 CCTCTGGTGTCCAGCCAGTGTGTGAACCTGACCACCAGAACACAGCTGCC    128
                       ||.||||||..|||||||||.|||||||||||||||.|.||.||||||||
  Moderna           82 CCCCTGGTGAGCAGCCAGTGCGTGAACCTGACCACCCGGACCCAGCTGCC    131

Us folks in biotech have a special tool just for this :) https://blast.ncbi.nlm.nih.gov/Blast.cgi

Unfortunately, the core algorithm dates back to 1990, so it can be real slow in some cases. Biotech takes a while to improve :(

Wait, you mean you don't extract genomic data from Excel? The MARCH1 gene brings many interesting surprises.

Exactly. FAIR (Findable, Accessible, Interoperable and Reusable) principles are at a loss here [1]. The "Reusable" part seems to be especially problematic as the sequence is buried in a PDF file though all aspects of FAIR are compromised here. Edit: It looks like there is now a PR to address this issue [2]

[1] https://www.nature.com/articles/sdata201618

[2] https://github.com/NAalytics/Assemblies-of-putative-SARS-CoV...

Here you go! https://github.com/brianherman/Assemblies-of-putative-SARS-C...

My thoughts exactly!

Somewhere, Margaret O. Dayhoff is weeping.

The convention of genomic research is to present all RNA sequences as equivalent cDNA sequences. As this will be the output of most common sequencing platforms.

https://bioinformatics.stackexchange.com/questions/11353/why...

DNA is way more stable than RNA. Since you can easily synthesize RNA from DNA, and DNA synthesis technology is much more mature, folks normally synthesize DNA and then derive/make the RNA from it. That makes most researches default to DNA 5' to 3', even when talking about RNA.

You probably mean uracil, not udenine (which doesn't exist AFAIK).

Note that independently of the notation used the mRNA of those vaccines use even more "weird" bases, such as 1-methyl-3’-pseudouridylyl, to make the vaccine mRNA not be detected by the immune system [1].

[1] https://berthub.eu/articles/posts/reverse-engineering-source...

DNA uses base pairs [A,T] and [G,C], this code is for a piece of DNA,. if you keep a DNA sequence in vials for later use, that is much more stable and easier to manipulate, and repair when corrupted.

normally RNA in vivo is complexed with protiens that prevent RNA from folding, and annealing into structure that is not compatible with translation to protien. In the vaccine this isnt happening, this is why RNA is hard to work with and the vaccine must be kept so cold.

This is not to say that DNA is simple to work with, but it solves problems if you dont need direct access to RNA.

RNA uses uracil/uridine rather than thymine, but uridine is actually quite immunogenic. That's what has prevented people from using mRNA as a therapy until recently, when the founders of BioNTech figured out that they could use pseudouridine (abbreviated as Ψ) instead. See [1] for more information.

[1]https://www.wired.co.uk/article/mrna-coronavirus-vaccine-pfi...

Some parts of gene transcription are so straightforward one can almost be tricked into thinking it has the logic of a computer program. It may be an illusion. To stretch the metaphor, TCP parsers don't match probabilistically along the entire length of the packet in parallel, and they don't interpret the same part of a packet as data in some contexts, and a header in others.

I ended up majoring in biochemistry and molecular biology in my undergrad because I was browsing on Wikipedia one day and came across an article written on an E. Coli variant that had sentences like:

01J3 e. Coli has a DNA Polymerase that contains 3k’-5’ proofreading capability and 5’-3’ error correcting with a polymerisation rate of 50bps

I’ve made the above up because I have never been able to find a Wikipedia page winxe that as succinctly pointed out to me that biology was a machine and I was hooked

A given combination of 7 bases has a probability of occurring of 1/16,384. Since the COVID genome is about 22k bases long I guess you have pretty good chance of it appearing in there somewhere. This assumes uniformity, which of course is not true. COVID’s genome is under crazy intense selection pressure!

That would have been a killer easter egg (possibly literally).

Whats special about this string?

Protein folding is not solved, that headline was overstating the actual achievement by Google's protein folding solution.

If I remember correctly "solving" protein folding was essentially some high probability prediction that state A transform to state B with some reasonably high chance, on a big dataset. Or something like that anyway. It's as far from high level work with genetics as creating nanotubes a few molecules long in lab manually is away from industrial production.

The most fundamental reason for that is that it's just not amenable to human mind. We are quite primitive actually, being able to hold only a handful of "things" in our mind at any one time and relying on abstraction to think of more complex things. However, you can't abstract much in biology; there is no locality or separation of concerns, everything affects everything.

Take that piece of RNA. An intuitive mental model is that it's some form of "instruction" or a bunch of instruction, isn't it? It's also wrong, because it just encodes a protein that acts the way it does only because of its shape (that is, one of its potential energy local minimums) and the shape of other proteins around it. That shape is only weakly local, it can be affected by far-away sections of peptide sequence. So it's almost impossible to systematically break it down, you have to consider and model things as a whole , which is insanely complex both computationally and cognitively.

If you want a good mental model of how it works, imagine you assemble a thing from metal balls and springs. You take a few thousands balls and connect most of them with springs of different strengths. You then take this thing and throw it on the floor; it will assume a shape that is implicitly encoded in spring strengths, its environment, and the way you've assembled it. You can even make it change shape if you poke on it the right way. That's how biology works in a nutshell; it's a nightmare to design anything for systems like that. Again, you can't simplify and break down and encapsulate and abstract like you do in programming.

Because the problem is significantly more complicated than sequencing and folding.

I feel this XKCD describes the situation particularly well: https://xkcd.com/1831/

My thought exactly. So this thing is like a VM with a bunch of primitive opcode, why can’t someone write a higher-level language or at least some gadgets

Because it's a harder problem than it seems at face value.

You are correct. The researchers here sequenced each vaccine starting with the bit of vaccine left in the vial after administration. The goal was to get a raw sequence of the Moderna mRNA component so it can be easily filtered out as being a signal of therapeutic origin. Pfizer's sequence has already been published; it's incldued here to confirm that the result achieved experimentally matches the published sequence.

The authors reverse engineered the sequences of the vaccines, obtaining them from the remaining mRNA present in the vials.

“Assembly” in this case means that they merged several short sequences they obtained, each representing a fragment of the whole mRNA sequence.

They sequenced vaccine leftover remaining in used vials.

So reverse engineering basically.

That gives me the feeling that those reflexion models could do some help for improving our understanding of those microservices

"its pretty interesting that all of human life can be reduced down to 30k editable microservices."

I don't know much about DNA and co, but it sounds as microservice is not the right metapher. Rather just 30k sourcecode?

Because a microservice is something that is already compiled and running..

Sure, but if you took that 30k of data and dropped it on a planet just like earth it would still take 10k years or so for us to build civilizations as we know it again.

It's the 3' (3-prime) UTR (un-translated region). It can affect the translation of the mRNA.

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

The next thing is the poly-A tail:

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

blasting the 3' UTR, we see it ~50% of it was copied from the human mitochondria

tldr, extra regulatory signals (often not well understood)

The 3' and 5' untranslated regions are the parts of the mRNA directly before and after the part that encodes the actual protein. So they are themselves not translated into amino acids.

What they actually do can vary, but essentially they can provide places for other things to bind and influence what happens with the mRNA. There are some fancier cases like riboswitches, but you don't see those in humans. The stuff at the start and end of the mRNA also determines stability of the mRNA.

In the BioNTech/Pfizer mRNA, the 3' (or the latter) half of the 3'-UTR comes from human mitochondrial 12S rRNA.

The Moderna one has the 3'-UTR of the alpha subunit of human hemoglobin.

Among other things that purple region determines the "priority"/"intensity" of the whole sequence.

You want it as high as possible to make as much spike protein as possible.

It's proprietary information, mostly they try various possibilities until they find one with high expression.

That’s the copyright notice.

That's the part that makes you randomly yell, "Hail Bill Gates!"

From that link:

> The injection contains volatile genetic material that describes the famous SARS-CoV-2 ‘Spike’ protein. Through clever chemical means, the vaccine manages to get this genetic material into some of our cells.

> These then dutifully start producing SARS-CoV-2 Spike proteins in large enough quantities that our immune system springs into action. Confronted with Spike proteins, and (importantly) tell-tale signs that cells have been taken over, our immune system develops a powerful response against multiple aspects of the Spike protein AND the production process.

What happens to the "volatile genetic material" at the end of this? Does it just linger in the body indefinitely? Or does it somehow get destroyed (and what does that mean)? From my reading of the above excerpt, it's the produced spike proteins that get destroyed but not the original genetic material that's injected. The reason I'm asking is to understand how the vaccine designers determine if there are any long-term effects of having this artificial material inside your body. They couldn't have tested it over a long time frame given how quickly all this moved.

I imagine we have tools in that direction, but nothing complete. Unlike math and computers, biological systems don’t really go from a uniform set of simple rules to emergent complexity - there is a whole lot of sideways complexity thrown in.

Something that might fit the computation vision of your comment are the various Ontologies for bioinformatics. The Gene Ontology is probably the most complete, although it lags many years behind the literature.

http://geneontology.org/

Checkout https://github.com/clasp-developers/clasp

"Clasp: Common Lisp using LLVM and C++ for Designing Molecules": https://www.youtube.com/watch?v=0rSMt1pAlbE

Answers:

a) From the mRNA you can learn the amino acid sequence of the protein very quickly. You absolutely cannot (yet) learn the function of the protein from that sequence - normally, people just do comparisons with proteins whose functions ARE known. Oftentimes in enzymes there are "domains" or little functional regions that stay consistent over long periods of time, so that's a good way to assign function (given knowledge of other proteins in the same family)

b) Yep. Every virus at some point in their lifecycle use mRNA. You can just sequence the virus and get all that data (I've done that on SARS-COV-2, it's honestly pretty easy). Then you just do homology alignment (as stated above) and you can figure out approximately what each gene does.

The problem of de-novo protein prediction is ONE OF THE HARDEST PROBLEMS IN BIOTECH, but just like getting amino acid sequence, doing homology searches, sequencing viruses, etc, is basic biotech and I'd expect an eager high schooler or undergrad to be able to do them.

a) I don't know if protein-folding software is good enough to figure out the exact structure of the resulting protein given just the gene, but I suspect you could figure out through the equivalent of the strings command - looking for sub-chains of the protein, and looking for matching sequences in the gene

b) Coronaviruses happen to be RNA viruses; that is, their genomes are RNA rather than DNA. DNA viruses also exist and are common. We got full genomes from sequencing early in the pandemic, and continue to use it to monitor the evolution of the virus (see e.g. [1], where the results are available for download). Sequencing is very cheap and easy these days - you take a sample from a patient, use chemicals to break down all the cell membranes and such, sequence all of the DNA and RNA in it, and look through the results for a virus genome (i.e. something that isn't a human chromosome and isn't a known virus or bacterial genome). "m"RNA is more a description of the function than the chemical - tRNA and rRNA are short snippets of RNA used for manufacturing purposes inside the cell, while mRNA is the long chunks that actually carry information from the DNA to the protein manufacturing sites. Virus RNA basically functions as imposter mRNA, getting those manufacturing systems to make more viruses.

[1] https://www.ncbi.nlm.nih.gov/datasets/coronavirus/genomes/ - SARS-CoV-2 is the COVID-19 virus. As of my fetch, there are 71,509 full sequences of the virus, reflecting slight mutations over time and space.

a) Yes, you can translate a mature[1] mRNA sequence, codon by codon, from the start until the stop codon, and it will give you the sequence of the protein it encodes.

b) Coronaviruses have a RNA genome. Researchers extracted it from wild-type viruses and then sequenced it.

[1]: mRNAs can undergo several maturation steps, such as splicing, which removes regions that won’t be translated into protein.

Everyone else has had good answers, but I'm also going to note that we knew a ton about covid's general molecular biology well before it ever came into existence. Covid (more properly, SARS-CoV-2) is a cronovirus. Cronoviruses have been studied for some time since some of them cause common colds, and studied very intensely since 2002 when SARS showed pandemic potential. So when Covid showed up, there was a ton of prestablished information and expertise avaliable to help every element of the pandemic response.

Their codon compositions were optimized differently, but both coding regions translate to the same amino acid sequence.

> Why haven’t viruses evolved to have a lipid container?

They have. https://en.wikipedia.org/wiki/Viral_envelope

I was going to say you can't get very far without a protein vehicle, but then I remembered that's quite incorrect:

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

The injection is important, however, as it gets the genetic material past a whole lot of nucleases that cover your epithelia.

Note that this is a sequencing result, so it will lack a lot of nonstandard RNA tricks that these companies are or might be using, like pseudouridines, or fluorobases. I think those would have to be disclosed in the patent.

We aren’t that different from machines, we just need to know more about the CPU and all the co-processors and how the logic gates interact

But for now we can inject code to trigger protein configuration via the immune system

Just because you see the RNA/DNA sequence on Github doesn't mean anything, DNA sequencing has been around since at least the early 70s [0]. Many pharmaceutical drugs already employ such techniques.

[0] https://en.wikipedia.org/wiki/DNA_sequencing#History

There is a system that transports protein fragments to the cell surface and “presents” them to the immune cells: https://en.m.wikipedia.org/wiki/Antigen_presentation

We do know that bit in the middle encodes the structural spike protein of the virus

What does "hacked" mean here? The article makes it sound like this wasn't something illegal:

> Fire and Shoura told Motherboard that they had received permission from the FDA to collect scraps of vaccines that wouldn’t have otherwise been used from empty vials and that they’d notified Moderna in advance of their plans to publish the sequence without receiving any objection in turn.

Also:

> The research team told Motherboard that they didn’t “reverse engineer” the vaccine, they simply “posted the putative sequence of two synthetic RNA molecules that have become sufficiently prevalent in the general environment of medicine and human biology in 2021.”

I'm not familiar enough with how these sequences to work to understand what's being discussed. Is it simply that they took a sample of the vaccine and studied its composition using some standard machine/process?

I think memorizing any of these two is pretty much totally useless.

Maybe one could compare it to having different recipes for the same cake. Or different source code to solve the same problem.

Different codons can encode the same amino acids, so different sequences can encode the same protein.

I didn't check it was the only explanation, but the DNA -> protein encoding is surjective.

Well we do with anti-malware stuff, when a 0-day comes out and we know there are exploits in the wild and beta software is all we've got.

They're the same. It's just a second dose as a booster.

I believe only Sputnik vaccine has different first and second dose, but their vaccine is of different category (it belongs to the same as AstraZeneca and Johnson and Johnson). The reason is that these vaccine use a vector (adenovirus) and there's a risk that body will develop antibodies for the vector and the second dose might not be as effective.

"Some vaccines require two doses because the immune response to the first dose is rather weak. The second dose helps to better reinforce this immune response." - I would have to think over time that could be optimized somehow to just require one w/ ML and test results, etc.

Proteins are written in standardised IUPAC amino acid codes that carry some semantic meaning, e.g. Alanine: A, Glycine: G etc. Also viral genomes often have overlapping transcription with shifted open reading frames. Biology is not as simple as you think.

If you have the time, it would be nice to transfer the data to the commonly used Genbank format.

No. The RNA still needs to be fit in a lipid nanoparticle which is Moderna and BioNTech’s secret sauce.

No reason you can't make your own NFT of it. Heck, if you promise to pay at least 1.337 ETH I'll figure out how to do it myself and make it for you. :-P

read the article linked in the thread. it actually does not work against all of the covid virus. it works against the spikes.

so, the virus is sort of like a ball with these spikes on top (that’s where the corona name comes from) and the vaccine helps your body develop antibodies against the spikes. so when the virus gets in your body, it actually receives a “haircut” which leads to it no longer being able to enter the cells and hijack their internals to produce more viruses.

it’s extremely clever, but it also means that your code is wrong ;))