Skip to main content

Biological News

A Couple of Ycombinator’s Startups

Last year I mentioned reports that the startup incubator Ycombinator was thinking of getting into the biopharma field. Here’s a look at the current crop of potential companies.
One thing that stands out is that most of these seem to be focused on patient care or some sort of diagnostic. One exception is 20n, which is looking to engineer microbes to produce known pharmaceuticals or intermediates. That’s not at all a crazy idea, but the example given on the site (acetaminophen) is not a particularly compelling example by itself, since it’s extremely easy to make, from cheap precursors, on an industrial scale. And I’m not sure what to make of that “map of every chemical that can be made biologically”. It’s a nice graphic for the middle of the page, but there’s no telling what it means. I like a lot of the ideas kicking around in the synthetic biology field, but I can’t really say what 20n is up to yet.
The other one on the list I noticed is Atomwise, and I’ll let them speak for themselves:

Medicines are getting more expensive to develop. These days, it takes about $1.8 billion and 15 years for a single new drug. Atomwise aims to change that by using supercomputers to predict, in advance, which potential medicines will work, and which won’t. Our tools can tell the difference between great drug candidates and toxic ones, and discover new uses for old medicines.

Actually, what will speak for themselves are the results. If Atomwise can do this, at all, even poorly, then there are billions of dollars waiting out there for them to scoop up. But just the act of saying that you can do things like this makes me suspicious that they really can’t do things like this. Here’s a bit more:

Previous attempts haven’t always met expectations. The techniques of the day were limited by the knowledge and computers available. Today, things are different. We have invented cutting-edge machine learning algorithms that are built specifically for the world’s most powerful computers. We use one of the world’s top supercomputers to analyze databases 1000 times larger than those used in the past. This lets us deliver what many others can’t: precise and reliable medicinal predictions.

I could go on about this for a while, but in the end, these arguments are settled by data. Come on down and try it, guys. There’s plenty of room, and plenty to work on, so let’s see what you can get done. I’ll be watching with interest, and so will others.
Update, from the comments: “Hello Everyone, I’m the CEO of Atomwise and a long-time semi-lurker here. (I’m the anonymous who keeps asking what it would take to convince people that in silico methods work.) I agree that the proof will come from data; that’s one reason why we’re doing a large-scale evaluation with Merck. Send me an email ( and I’d be happy to present our data from previous prospective validation projects to you. Or, if you have some minimally proprietary data against which you’d like to evaluate our predictive capability, let’s run the experiment. Best, Abraham Heifets”

33 comments on “A Couple of Ycombinator’s Startups”

  1. Hap says:

    This might be unduly harsh, but if Atomwise can do what they say, why are they on Ycombinator? Pull up to a big pharma, ask for some minimally proprietary data, and let the computers go to work. It’d be like printing winning lottery tickets.

  2. Mark says:

    Because, as everybody knows, the only thing holding molecular modelling back is the lack of computing power. If only computers were 100 times faster it would all be a solved problem.

  3. Mike says:

    They say they will “analyze databases 1000 times larger than those used in the past”. Considering that essentially every available database of pharmacological and toxicological information has been analyzed in the past, where do they expect to find 1000 times more data? Are they going to use hypothetical data as their input?

  4. Wavefunction says:

    Reminds me of that Forbes headline from 1981…
    Nonetheless, I wish Atomwise good luck. We need all the help we can get.
    #2 Mark: It’s not an either/or proposition. There are some problems like inadequate sampling that can actually be circumvented by a combination of massive computing power and better search techniques. Then there are other problems like inaccurate force fields which cannot be solved simply by increased computing power. It’s a mixed bag.

  5. Anonymous academic says:

    @3: maybe human genomes? not sure what else they could be talking about.
    Startup companies have been making claims like this at least since I was in college 15 years ago* – I feel like one comes along at least every other year if not more often. The only difference in the marketing lingo is that now everyone mentions machine learning. Maybe they really do have a better solution but there are no shortage of reasons to be skeptical.
    (* probably much longer; I vaguely recall seeing a relevant Time magazine cover or similar from the mid-1980s posted on this blog.)

  6. Anonymous says:

    Atomwise has a nice aspiration, but no way this ever works out – not in our lifetimes. Learning algorithms or not, the problem is we know very little about human biochemistry. If we don’t know much then it is pretty hard for a computer to predict events based on data we don’t have. Maybe it discovers some relationships we hadn’t seen before, but predictability will be low. Think how many drugs were discovered serendipitously. Most importantly, is anyone here not going to make a molecule you believe in because a computer told you not to?

  7. Anonymous says:

    There are over 300 known post translational modifications that can occur on proteins which can radically alter their biochemistry from ligand binding, signaling, half lives, to cell localization– you name it. When will models start incorporating the entire PTMome to truly reflect what’s going on physiologically? Imagine the combinatorial nightmare it would be to model a the proteome with 5 different PTMs all at once that are rapidly changing. Then expand that to over 300. Even modeling a single protein with 5 different PTMs that can occur all at once and that are rapidly changing becomes difficult to model when each combination of PTM on a protein may change something the protein’s conformational shape and signaling properties. This is why screening phenotypically in a wetlab works.

  8. Dick Cheney says:

    Derek and readers might want to check out Ginkgo Bioworks — they produce high value chemicals using microbial fermentation, including complex mixtures like rose essential oils. They were the first biotech to go through YCombinator (last year) and recently went through a Series A round.

  9. Chris Ing says:

    I wouldn’t expect the YC investors to know about the field of drug docking, but this isn’t one of those problems like “recognizing digits” or “detecting cat pictures” that you can just throw machine learning at it to get results. The difference is that molecular modelling and docking data just isn’t reliable enough. Even doing our best with state-of-the-art molecular models (which Atomwise can’t afford to do, even with a slice of BlueGene/Q) and eliminating most of the statistical error (like the Pande/DE Shaw labs) you don’t have an iron clad approach.
    Doing machine learning the “cheminformatics-way” like Google/Stanford here ( seems more useful in comparison.

  10. Anonymous says:

    Derek, what’s your take on Notable Labs? Their CEO, Matt de Silva, answered some questions here:

  11. @1 Hap. It’s not unduly harsh, it’s what we do and we are profitable. It’s not like printing winning lottery tickets though, it may come as a surprise to the readers of this blog but selling the concept of automated design is surprisingly hard.

  12. bank says:

    Atomwise’s claims are grounded in the expectation that the data pointing to which medicines will “work” actually exists in these databases.
    For many indications this will simply not be true, for those where it is true, there are likely previous extensive efforts in that particular space.

  13. jd says:

    Atomwise will not approach a big pharma, because it’s unlikely that a big pharma will seriously consider working with a bunch of ‘kids’ straight out of college.
    Let me be clear here: i think it’s pretty unlikely that these computational approaches will work now.
    But what if they do?
    There is a point to made: you need some level of resources, larger than academy, to try some of these approaches fairly.
    Humans are never keen to update their views of the world. Once burned by the computational promises of the past, the same manager will be super-unlikely to try again. Many of these methods may sound like something you’ve seen before but history only happens once and every time you need to re-evaluate the circumstances and details.

  14. pete says:

    @6 Anonymous
    I wish them the best. But Douglas Adams comes to mind here.
    “In the radio series and the first novel, a group of hyper-intelligent pan-dimensional beings demand to learn the Answer to the Ultimate Question of Life, The Universe, and Everything from the supercomputer, Deep Thought, specially built for this purpose. It takes Deep Thought 7½ million years to compute and check the answer, which turns out to be 42. Deep Thought points out that the answer seems meaningless because the beings who instructed it never actually knew what the Question was. When asked to produce The Ultimate Question, Deep Thought says that it cannot; however, it can help to design an even more powerful computer that can.”
    (credit for quote: )

  15. Anders says:

    If paracetamol is made by a bug, you may claim it is a natural product and a new marked awaits you: The eco-fascists will love it even though it is identical to the old synthetic product.

  16. matt says:

    There’s a old Biblical saying that applies here, paraphrasing: “One who puts on his armor should not boast like one who takes it off.”
    But, to be fair, pitches for investment funding almost are forced to violate that good sense. I doubt there is any funding available for someone offering Churchill/Roosevelt’s “blood, toil, tears, and sweat” honest assessment of reality.

  17. Anonymous says:

    @9 The same can be said about deep learning:

  18. Hap says:

    My assumption was that the ability to predict the effects of drugs would be worth lots of money to lots of people (so that funding shouldn’t be short). Of course, having never run a business or started one, I don’t know the pros and cons of different funding methods – maybe people don’t want to get eaten by big pharma, or lose lots of equity to VC funding, or want to make the products widely available as a general tool which wouldn’t work with conventional funding.
    It seems like bragging (though Dizzy Dean comes to mind) but lots of people would be happy if Atomwise can do what it says.

  19. Biotech Capitalist says:

    I agree with Hap, this is unduly harsh. Small biotechs or whatever these companies are need a little breathing room to flout their capabilities and goals, even ambitiously stated ones. Absolutely we should judge them on their results and data, but as a practical matter, it is not a good business practice to put up on your website: “We are tackling very intractable problems. We have no money. We have no results. We have no pharma experience. We know that 99% of similar startups fail but for some reason think we are in the 1%. So if you give us several million dollars, we promise we will give you back more.”
    To be honest, the biggest thing these startups need are you. They need (crave?) expertise from seasoned drug discoverers. Crazy new theoretical ideas need to be atom-smashed into seasoned pharma experience. Hopefully the more visionary, optimistic and constructive readers of In the Pipeline have already reached out to these companies and offered their help (free, paid consulting, equity, SAB, …) and it is only the cranky that post negative comments.

  20. Hunter says:

    The argument for Atomwise is similar to the arguments made for conducting a high throughput screen; if the cost is low, why not give it a chance. One major issue though, as discussed on this site several times, current databases are riddled with unreproducible data. So similar to screening a low quality compound library, considerable resources may be spend on following up hits that lead nowhere.

  21. Hap says:

    No, I was the one being kind of of harsh in 1 – sorry. I was assuming that YCombinator would be a less preferred method of funding for something with those capabilities, but as noted, I could easily be wrong in that.

  22. Biotech Capitalist says:

    @Hap, ah my mistake. I can answer you: the dirty little not-so-secret of all startups is that they cannot do what they say they can. … BUT!, they may be able to one day.
    One of the other commenters is exactly right, they need to get some data from a pharma, show what they can do with it. But one of the other commenters said no pharma would engage them without experience. I do not believe that is true, my dealings with big pharma from a tiny startup are they are very engaging. They may not move mountains for you, but it is not unreasonable to convince them to give you data for a pilot project, especially if they blind Atomwise and get to see the accuracy of their results first. Any initial validation, such as YCombinator residency, gives these startups at least a tiny platform to declare themselves a real company and to flout themselves to pharma/investors.

  23. SteveM says:

    Re: Atomwise. The proof is in the pudding. There is so much historical data available, Atomwise should be able to apply it’s algorithms retroactively and compare the predictions with the actual known outcomes and post those results for comparison.
    Anybody who pays Atomwise for predictions without the obvious validation is buying smoke and mirrors.

  24. genius says:

    Hey Ycombinator, get in on the ground floor with me! I’m in the early stages if inventing something I call “the wheel”…

  25. alig says:

    Re: Atomwise. Another bad assumption is that drugs that are already approved could get approved now. If you submitted a drug like acetaminophen to the FDA today, it would fail to be approved because of toxicity. So the historical data may be of no use predicting the future.

  26. Hitchhiker says:

    @6, 14 IC50 42 nM, solubility 42 uM, bioavailability 42 percent – that’ll do nicely sir.

  27. Anon2 says:

    As a biologist I poke structural chemistry with a very long stick. Meaning I call up an old buddy of mine, spend 5 minutes prefacing how little I know about the topic, and then proceed to ask my questions is a very cautious manner. I applaud their goals, but their confidence discredits them more than they realize.

  28. Hello Everyone,
    I’m the CEO of Atomwise and a long-time semi-lurker here. (I’m the anonymous who keeps asking what it would take to convince people that in silico methods work.) I agree that the proof will come from data; that’s one reason why we’re doing a large-scale evaluation with Merck.
    Send me an email ( and I’d be happy to present our data from previous prospective validation projects to you.
    Or, if you have some minimally proprietary data against which you’d like to evaluate our predictive capability, let’s run the experiment.
    Abraham Heifets

  29. P.S. If you want more details about what we’re doing, I answer some questions here:

  30. Some idiot says:

    @28: Abraham, I wish you the very best… And I really mean that from the bottom of my heart. I am extremely sceptical, but I really, really hope that you prove me wrong. And I think that this goes for the vast majority of posters on this site too….!

  31. Morten G says:

    @8 Evolva has been making high-value chemicals via fermentation for a long time now. Kinda has the drop on those guys. Link in name.
    Maybe a round-up Derek? Are the microbiologists trying to make the scale-up org synths redundant?

  32. Christophe says:

    Compare the numbers and the statements of the leaders of atomwise
    CEO Abraham Heifets 184 citations, h-index 7
    statement on Atomwise website: “Abraham created SCRIPDB, one of the largest public databases of
    patented chemical structures, and LigAlign, a protein analysis tool used by researchers in 70 countries.”
    CTO Izhar Wallach 110 citations, h-index 6
    statement on Atomwise website: “Most notably, he invented the widely-used Virtual Decoy Set technique.”
    COO Alexander Levy name too common to figure out stats
    statement on Atomwise website: “”
    Scientist Misko Dzamba 459 citatios, h-index 7
    statement on Atomwise website: “Misko was also instrumental in the development of several
    leading research tools, including SHRiMP, VARiD, and MDUST-P.”

  33. DrSnowboard says:

    That’s OK, Dzamba has his research bets hedged with PetBot….

Comments are closed.