An advice for the VC among us. Don't get fooled by all these startups promising to revolutionalize drug-discovery with AI. Go talk to scientists that are specialists but have no stakes in those approaches. I've seen so much money diverted from projects (many that turned out to work once sanity came back) just to pay for some over-confident startupers shitty solution... Yes AI will help drug discovery no doubt about that, but those companies telling you they can learn from existing data are lying, most data available is of really low quality, acquired in conditions adequate only for specific types of questions... One I know closely will promise you a knowledge graph based on a specific slice of drug sources that can supposedly answer every question. It will almost certainly end up like many of those companies that hire humans to do the work in the back and sell it to you as AI... And their source data which I know well is of really low quality so we can't expect much more from their models either...
A lot of times they are not lying, they are genuinely clueless. Part of the problem with academia is that they follow the edict set by George Constanza: “the best liar truly believes in their own lie.” So often they truly believe that their field’s method of choice is how we will use AIDS to cure cancer, because the grants-system ensures that anyone who doesn’t truly believe in their vomit won’t get their grant approved (you need to write them with conviction, not doubt, if you want them awarded). Then mediocre postdocs with an “entrepreneurial flair” decide they’re ready to take their nanobots and cure breast cancer and come to SV to pitch to VCs who are either clueless or are being advised by the same heads that are pouring the Kool-aid in academia to begin with!
Exceptions exist of course, and some of these bets turn good purely because of luck (if you take twenty random chemicals and inject, one of them might just cure cancer; incidentally how stuff like paclitaxel & vinblastine was discovered anyway). But this only seems to reinforce the conviction of these institutions to continue their methods.
One of the most important differences I feel VCs should acknowledge as a contrast between tech and biotech: with tech you can often choose one method of solving a problem and make it work. With biology however, nature will eventually give you a roundhouse kick if you try to force anything but an actual solution down it’s throat.
>Exceptions exist of course, and some of these bets turn good purely because of luck (if you take twenty random chemicals and inject, one of them might just cure cancer; incidentally how stuff like paclitaxel & vinblastine was discovered anyway).
Vinblastine was discovered from folk tea remedies for diabetes. In animal models they saw decreased white blood cell counts.
Paclitaxel was discovered from the NCI plant screening program. The chemical was observed to be cytotoxic in screening studies.
So both of your examples have pre-clinical evidence. These were not randomly injected and developed through scientific research and the clinical trial program.
Vinblastine was discovered when they were testing a folk remedy for diabetes and found cytotoxic effects so it was completely tangential to the study’s goals, and in modern parlance, an instance of a “task failed successfully.”
Paclitaxel was discovered literally from a random spray and pray scan of plant samples from every taxonomic class, here I will quote a literal sentence from a review: “ Barclay collected T. brevifolia, an evergreen, apparently because the strategy was to collect at random. Not much was known about the tree. It belonged to the genus Taxus...” (https://www.acs.org/content/acs/en/education/whatischemistry... )
So I still believe in what I said, which is random screens or complete serendipity often has been the best ways to get t a lot of our therapies, though I’m not advocating for more of that necessarily.
You are right I assigned to malice what is probably just lack of skills and knowledge. Your are also right on the fact that what works is grind and find. You can do it at the scale of one lab or the scale of many labs, but in the end what works is to try a lot of things because we are still really shooting in the dark for a lot of targets or diseases we don't know any druggable target yet.
I’ll propose that we actually kind of know a lot of things about many (not all) diseases, we really just need an objective look at how to address them. The majority of scientists today are mostly quite narrow minded in terms of scientific thinking, every one of them wants to use the method that they are experts at to solve the problem. When the Manhattan project was deployed all the top physicists came and tackled the problem that needed to be solved with the most suitable methods, but there’s been “cancers Manhattan project” or “moonshot” initiatives in biology numerous times and they all fail because the scientists incharge invariably push their own agendas in these.
Take Janelia farms for example- without reading their editorialised website if you try to gander what the hell this institution is trying to do, from just the papers they publish, you would fail miserably. Not saying that institution sucks, well I suppose I am saying that, but only insofar because all of academia seems to suck in the same way.
It seems kind of odd to blame academia for this. Almost all founders are clueless in all fields, and, additionally, they’re pretty much supposed to be aggressive about their idea in order to get investments. The role of the VC is to invest wisely. If they’re less knowledgeable about tech outside of the niche of computers and they lose money on biotech, that’s their fault.
I know less about founders in computer tech but yes, many founders in biotechs that I know have just had a few years of experience in labs, not really successful scientifically, even if they know how to market it. And thats far from enough to really understand the nooks and crannies of drug discovery projects. Especially since most academic labs are only doing a really thin slice of the whole process. There is a reason why you dont give millions of dollars to a postdoc that submit their first grant proposal, why we do that to postdocs that are founding biotechs is beyond my understanding. The products I've seen from companies that raised $5mil are not worth 1/10 of that, and thats just in dev costs, their stuff could not even scale so they are trying to survive until the big prey comes and swallow them before anybody notices its all smoke and mirrors.
> There is a reason why you dont give millions of dollars to a postdoc that submit their first grant proposal, why we do that to postdocs that are founding biotechs is beyond my understanding
TINA. If you're a fund that needs to put an ass in a seat to say "we have a biotech investment", you go to whatever you can find. You have no way of doing due diligence in your portfolio and even if you hired someone like me to ultrafilter your proposals you'd probably come up empty, because the only postdocs that play this game are those that know they are good at parlaying their social capital, obtaining favorable reviews from their PIs (social validation), and have less scruples.
The heads down brilliant and honest scientist postdocs have already burnt out, and often aren't socially savvy enough to play the VC game, if they were, they'd be gunning for faculty positions instead.
If VC wants a better return they need to have a different heuristic. I'd pick postdocs that did years of grueling work with no results and ask them what they'd really prefer working on instead of their PI's hairbraned idea.
Generally that’s the trajectory computer tech startups take too, except with even less experience. That’s probably a reflection of the difference in the barriers of entry.
I suppose one of the points I was trying to make is that in tech it’s probably okay (maybe even beneficial) to be clueless, but that’s not true for biology or hard science tech - if you float a rocket company, stuff has to fly, and you better get some rocket scientists in there. One can’t just Travis Kalanick their way to Mars because you’re not fighting some cab unions you’re fighting the laws of physics.
I wish I could find it now, but I once read a comment on HN where someone was approached by a VC for technical advice. To make a long story short, the engineer being consulted noted that in order for their growth to continue as the company proclaimed, they would have to break the speed of light before reaching profitability. The VC invested anyway and inevitably the company failed.
VCs chase trends just as much as anyone. In my experience, they are more susceptible to herd mentality than regular people.
Problem is that the big corporations and execs pay far less tax than small to medium corporations and workers. They have extremely distorted view of money and at some point it feels they grow money effortlessly while it is a month to month struggle for workers.
This feeling of ease makes them spend money on a whim because it will grow anyway most of the time and if it won't? Oh well there is plenty of billions in the coffers.
I think once organisations that are supposed to fight corruption pull their heads out of their and start making politicians accountable and make them plug the tax holes.
When those big corporations get a dose of reality, how much effort it costs small business to earn that money, they'll spend it more carefully, but more importantly because of increased tax take it will be possible to lower the tax burden for everyone and level the playing field.
Yup. People act like computer aided drug design is new, but it’s been happening since the 80’s with the same promises that never pan out. People are right to be skeptical.
That said, I could see computational drug design being helpful in particular circumstances solving very specific issues. But the idea it’s going to wholesale change the way drugs are designed and speed up the drug discovery process are very, very optimistic and the landscape is riddled with failed promises over the past 40 years.
I work in drug discovery, we use computers all the time though at at every steps. Drug discovery is already computer aided, and they try to sell it like it is not... They just want to sell their expensive and biased statistical model that doesn't do much more than a regression but use the latest shiny thing (knowledge graphs, rnn, svm, you name it ...)
It's a bit like people that want to disrupt the bodega or the breakfast juice or a one drop of blood full test for all diseases. You can pitch it as loud as you want, when you have a shitty product it will not make it better.
The last time I was working with molecular modeling, proteins were still the main target. With the advent of cryoEM, glycobiology of proteins can now be discerned. Now, AI is starting to be able to predict entire secondary structures from sequence alone. It will take massive computational power, but I can see how toxicology may soon be able to be predicted from in silico experiments.
The point of this post (and what I realised during my PhD) was that for the most part this doesn’t matter; choice of targets and predicting the human efficacy of a therapy that works in a tube or mice is far more important and typically more ignored. People don’t focus on that because it’s a hard problem especially for math and computer people who think their absence is the reason cancer isn’t cured yet and come over to try and win the Nobel prize (and don’t; cue relevant xkcd).
AI is just a tool like Excel and not a panacea. It might create some efficiencies but it is not a short cut to the real grind and the too often overlooked serendipity involved.
Serendipity plays such a big role in many “discoveries” but we don’t want to ascribe too much to it because it is just potluck by brute forcing OTHER problems.
Perhaps serendipity should be rephrased as “no method to our madness”.
Unlike Excel, AI truly can deliver a competitive advantage when used judiciously. In the right situation, it can deliver that small catalytic boost that lets you see just a little farther than before. And incremental advancement is the basis for almost all scientific or commercial advancement.
That said, deep CNNs and LSTMs have delivered only a fairly narrow set of catalytic advancements to date, even fewer of which have changed the status quo in big businesses, where R&D ROI really adds up. Accurately identifying the next use of AI that will meaningfully move the needle is beyond hard.
But hasn't that always been true of all forms of innovation? That's why so many are keen on AI today; as hard as it is to wield, it still seems to be the sharpest arrow in our quiver.
I agree with you. I transited from computing to drug discovery. A lot of projects attempt to harness the power of ML but there still seem to be a lot of misunderstanding. A lot of the time the valuable ML methods are supporting / assisting hypothesis generation rather than 'doing it all'. The domain experts will judge the value of ML methods by their usual merits in the drug discovery world. If the more sophisticated methods can predict hit compounds no better than the traditional standard deviation based cut off, or if the models can not explain key insight of any indication of mechanism of action, then it may not be a good method after all. But meanwhile if it opens up a new way (e.g a new phenotype based screening) then it is worth pursuing.
Quick tangential question: Do you see a benefit in incoming medical students knowing programming? If so, which specialty do you see it benefiting from the most (e.g., radiology)?
For practicing day-to-day medicine, programming is useless. However, for a medical student hoping to become an academic it's a highly important asset. Others and even more fundamental assets are to have solid knowledge of epidemiology and statistics. Essentially, if you want to stay in the academic system you'll be slaving under older academics for quite some time. Being able to hack together a dataset from csv with a bit of R in the mix will make you a godsend to those people. Be aware that this advantage comes at a high price, as those people usually understand nothing and will attempt to pressure you into doing impossible/unethical things (extreme p-hacking, etc.).
Now, doing anything more than shitty statistics or code-monkey javascript is almost impossible due to other constraints. In medicine, almost every system is totally locked up. There are of course exceptions as some students will go above and beyond, but the yield in terms of work/achievement will always be poor. It's far easier to just fill in excel spreadsheets and publish lots of shitty papers. Much to my dismay.
As for the sibling comment speaking of abstraction I'm sorry to disappoint, but no. Current medicine is far from being hard science. You cannot approach it like a formal system (yet). These days, docs are still far closer to gardeners than they are to programmers or any kind of engineer.
That said, you can find use cases for programming in any specialty. Sometimes it will be more obvious than others:
Image analysis -> radiology
Statistical genomics -> genetics
Dsp -> anesthesia
...
But anyways, don't get your hopes up too much. Students would usually be much better served by learning to care for people. And that doesn't involve programming.
One hundred percent! I did this, and I'm glad I did. Programming changes your perspective. When you learn to code you learn to value abstraction. That's valuable in medicine because medicine is pathologically reductionist...
