Skip to main content
Menu

Drug Development

Another AI-Generated Drug?

I see that there’s press coverage today of “the first AI-generated drug” to go into human trials. Some will recall this similar claims have been made before, so what exactly are we looking at?

The compound is DSP-1181, from a collaboration between Sumitomo and the startup Exscientia (out of Dundee). It’s a long-acting 5-HT1a agonist, from what I can see (page 15 of that document). The coverage says that it “was created by using algorithms that sifted through potential compounds, checking them against a huge database of parameters”, and that this took one year to get into the clinic. If that’s accurate, that is indeed a fast path into human trials, but let’s look at what that might get you. Will this be a drug discovery revolution?

The problem is that preclinical drug optimization is not the problem. A more conventional program directed toward this mechanism might have taken two or three years to get to a clinical trial, and it’s important to realize that the idea of 5-HT1a as a target for anxiety and OCD is not a new idea. Here’s a review from 2009, for example. Exscientia has a good deal to say about its target-selection and compound-optimization technology, but in this case they’re going after a target that’s been out there for many years. The publicity for DSP-1181 emphasizes that it’s a long-half-life full agonist for the receptor (as opposed to some partial agonists that have been tried before).

There is indeed a long, long list of compounds with activity at 5-HT1a, and many of these have been in humans (or are already approved drugs with some 5-HT1a activity as part of their profile). The azapirones are a prominent example, although they certainly have other activities as well, and there’s 8-OH DPAT and Addyi (flibanserin) too. Outside of those multi-receptor compounds, there are some more selective ones such as befiradol (which is being studied for Parkinson’s), the related F-15,599 (being looked at in Rett syndrome), osemozotan (which has been into animal models of OCD and other conditions, but has not, as far as I know, advanced into human trials), repinotan (which was taken into the clinic for stroke, and failed), piclozotan (which appears to have also failed in a stroke trial), U-92,016-A, also billed as a long-acting full selective agonist, which has been around since the early 1990s and has not been developed,  GPCR pharmacology is wildly complex, and nowhere more so than in the thick forest of serotonin receptor subtypes, so it’s certainly possible that a new compound could show a different profile. But it’s anyone’s guess as to what that profile might be. Just the list of different indications that such compounds are either used for or have been investigated for is enough to tell you what the field is like.

I have been unable to find a structure for DSP-1181 – patent applications from Exscientia are all directed towards machine learning techniques, and there is nothing I have found from Sumitomo/Dainippon on 5-HT1a. Although if that compound really is such new chemical matter, presumably a patent application hasn’t even published yet. When it does, or when the structure is revealed in a presentation, I will be very interested to see how closely it might resemble existing chemical matter.

Back the the mode of action. I used that word “guess” deliberately. For CNS drug discovery, that’s all we’ve got, AI or no AI. There is simply not enough reliable information to feed into even the greatest artificial intelligence software in the world to allow one to predict what will happen against conditions like OCD, depression, anxiety, and other high-level human psychiatric conditions. And that is the problem. Drugs fail in Phase II because we have not picked the right target, because our biochemical understanding of the disease state is wrong and/or incomplete. They also fail in Phase III for that reason and for unexpected toxicity, and the situation with tox is the same as for CNS efficacy: no amount of artificial intelligence is yet sufficient to tell you whether you’re going to run into such problems. Maybe eventually, but not yet. Those of you who remember Isaac Asimov’s “The Last Question” will recall the phrase “Insufficient data for a meaningful answer”, and that’s exactly the situation.

Here’s my take, then: Exscientia may well have moved a compound along at high speed into the clinic. But this particular example is not going to accelerate drug discovery much, because this is speeding up a minor part of the process and (for this target and indication) one that is nowhere near a rate-limiting step. We already have a whole range of 5-HT1a compounds, and we already have had (for many years) the idea that they might have applications in OCD. This project, at best, seems to me to have saved a few months off the process of sending their compound into the same black-box shredder as every such drug project goes into when it hits human trials. At this point, I believe they are now in the same situation as everyone else, which means a >90% chance of failure – honestly, CNS indications like this are more like 95%. AI has not changed that. I hope it does eventually. But it hasn’t yet.

41 comments on “Another AI-Generated Drug?”

  1. anon says:

    Can we generate AI human being to test AI generated drug? Just sayin’.

    1. No one says:

      Today in pharma news, Pfizer has partnered with Tamagotchi to develop a new, entirely digital clinical trial patient. They will be running a hybrid human-digital trial beginning in April, while the FDA scrambles to draft a regulation prohibiting taking a patients’ batteries out if they are giving bad data.

  2. S GX says:

    One thing we could find is in this paper:

    https://www.nature.com/articles/nature11691?page=7

    Not in the 5-HT area and don’t know how similar it is.

  3. Druid says:

    Repeatedly claiming that AI will cut decades off the drug development process is an example of OCD. AI-physician – heal thyself!

  4. Pseudonym Smith says:

    If preclinical discovery is easy, why does it take chemists so much time? Why isn’t everyone – AI or no AI – getting candidates to the clinic in one year?

    1. A Nonny Mouse says:

      Because they are not a small start up trying to sell their wares?

      1. anon says:

        Khajit has wares if you have coin

      2. Jeff Kindler says:

        Do you have a real answer to the question? It’s hard to believe that Big Pharma *could* get drugs to the clinic in one year whenever you want, but you just enjoy burning money and patent time. If so, then I guess Pharma Management is right when they decide they can lay off chemists without any downside! Thanks!

        1. Cameron Pye says:

          Well said Jeff.

          While I agree w/ Derek that there’s a lot of overhype around the “AI in drug discovery space”, that doesn’t mean that there aren’t real tangible gains being made here as well. Just because discovery pales in comparison to development costs and timelines doesn’t mean that it’s ever been easy or quick (as the salty med. chemists frequently point out). So even if this is incremental progress, increased efficiencies from new technologies, even against known targets or potentially derivative chemical matter, should be welcomed with open, if rightfully skeptical, arms… This isn’t a magic wand, it’s another tool in the toolbox.

          1. Serotonergic guy says:

            Right… and let’s agree that if your computer is designing serotonin receptor ligands, it starts with the benefit of tens of thousands of compounds already made and with some sort of in vitro data available in the literature.
            That may answer why a project aiming at a novel target would take longer than a year to produce useful leads.

          2. bks says:

            Just collecting, organizing and vetting the data sufficiently to feed into a ML black box probably has a big payoff, even if you never actually use the ML.

  5. Diver Dude says:

    I’m curious about this. The preclinical safety data package to support anything over a single dose exposure in man takes a lot longer than 1 year to put together, even for a well validated target. They may be going with a very lean approach in order to get into man quickly but they’d then have to stop until their chronic safety and tox package caught up. Receptor occupancy PET would give you a surrogate for efficacy after a single dose but it would be brave to bet the farm on this for a NO GO decision.

    I just don’t see the advantage unless you *know* your drug is going to fail. Which, in OCD, is probably the way to bet. But then, why do it at all?

    1. mymagoogle says:

      Yep. Back in my big pharma days, which is to say a few years ago, I had the thrill of leading a project to take two known compounds to the clinic in under a year! Because someone higher up said it should be possible. Really it should be possible. Two known compounds, bulk material was already made, what could go wrong!
      So we set out, lean mean machine team. My job as team leader was to map out all the hurdles, every quality signature, every report (with signature or not), every big long test, every little tinky test, every stage gate management meeting, with pre-read deadlines. This was the first time anyone had done this plotting out. We were on fi-yah! We pulled every favor, for every short cut on every queue we could. We were on the phone at 4am to our European sites begging for some tests to get higher priority.

      We got up to early November, and then, of course, the filing machine had a snafu, exceeded the allowed hold time, open up the quality deviation inquiry, possible CAPA, etc etc, and poof there went the year.

      Yeah, when were they going to do that big long test? It needs doing at some point. Then there are all the other things that the TRD people might not do themselves that needs doing – Clinical packaging, takes a month. Identity testing, takes another week, every time. Clinical sample testing methods transfer, ooh, start that one early. QP release – gotta give that guy 2 weeks.

      I wish them the *best of luck*.

    2. Spike says:

      I don’t think that it took 1 year from discovery to clinic. Going back and reading the Exscientia press release
      https://www.exscientia.ai/news-insights/sumitomo-dainippon-pharma-and-exscientia-joint-development
      what is actually stated is “This project was delivered by the strong synergy of the joint research, requiring less than 12 months to complete the exploratory research phase, just a fraction of the typical average of 4.5 years using conventional research techniques.” That doesn’t necessarily mean that it took 12 months from the computer saying “Eureka!” to starting the clinical trial (despite what BBC news reported). Also, this compound is mentioned as a project on the Evotec website under their “Pipeline of Product Opportunities”. Not sure what to make of that

  6. Molmechanic says:

    In this particular case, AI may not make much of a difference because there is so much inherent risk in the selection of the target. But it does seem unfair to detract from Exscientia’s contribution. Shaving a few months in lead optimization is huge. It frees up resources to go work on other life-saving projects. And drug optimization CAN certainly be the problem, depending on the lead series one is working in. Poor physical properties can present a huge impediment for progression of a drug candidate. AI can help by proposing ideas and evaluating a large of number of compounds thus narrowing down the list of compounds that need to be synthesized and tested. I’ll take AI over my shoulder any day!

    1. anon3 says:

      This is a weird statement. How would they possibly know that AI shaved a few months off the timelines? You would need a large number of controlled studies to determine this…3 months basically noise for the average LO project. It can easily take a month just to arrange a time, to have the meeting, to decide if a compound is ready to leave the LO phase.

  7. heliophobicHypergol says:

    I wonder how long it will take for computers to be able to accurately predict what a chemical will do in the body.

  8. Cb says:

    please show me that with machine learning/AI we can predict just one of the essential properties of a potential Lead Opt compound: solubility. I feel many med chemists would be very happy…..if this works perhaps we may announce discovery of clinical candidates at the end of the day….be humble for the time being….

  9. Retro Sinner says:

    From a CMC perspective, cross-nation transport of materials e.g. into the US requires an assigned IND number before customs will release. A phase 1 IND has a turnaround time of maybe 30 days at FDA and must be written in advance to include released batch data so maybe another month if you don’t care about quality. It’s unlikely that drug product is released from customs then being dosed to a patient the next day

    Is the idea that the AI says compound X on day one 1 and this is scaled and formulated (horribly) then proceeds through necessary safety, toxicity etc. I can understand manufacture, formulate and clinical in the same country saving some time but there are many risks about jumping into a trial – unless your small biotech model is WC Fields Pharma.

  10. John W says:

    This is how they got into humans in a year. Many companies could develop a drug and get to a human clinical trial within a year if you’re doing it Japan.

    From the BBC article: ‘The first drug will enter phase one trials in Japan which, if successful, will be followed by more global tests.’

  11. ChairmanMao says:

    I’m with Diver Dude- Where’s the preclinical studies and animal tox come in? I’d hate to be the 1st patient in Phase I.

  12. Ffghhszxfgg says:

    If Ai generates drugs, scripps is out of business for their grad program and here comes a bunch of lawsuits. If Ai can generate drugs there is not much need for students hence firings and defensive lawsuits led by Paul Plevin. As a Scripps Prof, I just hope we can eliminate the possibility of students getting the smarts to fight back.

  13. Neo says:

    Nice analysis. Same 95% of failure that any of the molecule entering clinical trials in this area. But only require 1 instead of 3 years. This also means that many expensive wet-lab tests over those 2 extra years could be avoided by using machine learning (I wish people stop calling AI to what always has been ML, I admit it could be a lost battle at this point…). Sounds a like a good first step, but in no way the major milestone is sold to be (that would be approval).

  14. yf says:

    Anti TNF was initially designed to treat sepsis and failed. Anti PD1 was developed circa 2004-2007 and was shuffled among companies many times. it was sitting at the bottom of a long list of candidates for development. Both drugs become landmark break-through either by serendipity ( anti TNF) or by a giant leap of faith ( someone pounded the table to invest millions of dollar to go to a clinical trial with anti PD1). A.I. can not help in either case.

  15. A Nonny Mouse says:

    Seems that the work on the molecule is being done by Evotec.

  16. Big Nose says:

    Can AI smell BS?

    1. Baldoni's Nose says:

      Yes and no – as in the case of humans, AI can only detect the BS of other systems. It is incapable of smelling its own BS, in the sense of a negative control. In fact, like humans, AI often reinforces its own BS, to the detriment of others, i.e. it thinks its BS doesn’t stink.

      1. loupgarous says:

        “Human, all too human”.

        Intelligent people can trick themselves, in the same way ignorant people do, by mistaking their prejudices for facts. It’d be odd if AI didn’t mimic that flaw of human intelligence, failing to set aside “rules” that lead it astray. So the evolution of AI may be right on track, but we’ll still have to be vigilant for when AI does just what humans do when trying to solve difficult problems.

        1. metaphysician says:

          I recall several stories where AI/ML has already discovered some very human ways of misbehaving: when looking for a solution to some problem, discovering the “solution” that fills the criteria in the most minimal technical sense possible, that is the easiest to do. In particular, if you can cheat by accessing outside data, do so.

          It does inspire hope that the robot revolution may not happen, because the AI will be too lazy to bother conquering humanity. . . *ahem*

  17. loupgarous says:

    The real rate-limiting/project killing step in drug development is, as usual, the unpredictability of human toxicity and poor choice of a drug target to treat a complex disease. The latter is the troll under the bridge that’s eaten every Alzheimer’s disease drug to hit human trials, while the former slew drugs like odanacatib because their mode of action didn’t cause toxicity signals in healthy volunteers, but in actual patient populations, once the cohort got big enough (Phase III or postmarketing, like the fluoroquinolone antibiotics), yep, they were toxic to a significant group of patients.

    That suggests to me that the AI guys, having snagged the low-hanging fruit of drug design, ought to move on to building models of human response to various molecules so they can predict with greater reliability which new drugs are apt to cause problems, or how likely a given drug is to actually help treat complex diseases. There are well-documented failures and successes out there, waiting to be placed in AI datasets for analysis.

  18. AlloG says:

    There should be a contest between a team of hardcore med Chemists and a group of AI schlubs. Dey can both have da same target and got to make a compound dat don’t exist yet and da winner is da first into a Phase sponsored by Bill Gates.

    Like those cooking shows you Americans like so much, except of using spiked headed chefs Derek can lead da med Chemists, and da AI schlubs can hire R2D2 or Ray Kurzweil.

    1. Dave Kielpinski says:

      As an ML engineer, I find R2D2 far more plausible than I find Ray Kurzweil and his minions.

  19. Todd says:

    Let there be light!

  20. still_here says:

    One year is an impressive speed to go from a computer hit to modest scale synthesis and tox testing and formulation studies and finally a GMP synthesis to support FTIH.

  21. Paramus says:

    From the company press release, the compound completed the EXPLORATORY RESEARCH PHASE in less than 12 months. That presumably means the development phase – GLP Tox, GMP material, stability etc. is still is needed. The BBC article is the usual media misinformation.

  22. MoBio says:

    Although there are certainly many 5-HT1A agonists out there (mainly partial agonists) many/most have significant off-target actions (frequently against D2 and 5-HT7). Obtaining a *selective 5-HT1A FULL agonist* this way is certainly impressive enough.

    1. Derek Lowe says:

      True, but they can impress everyone with actual profiling data for maximum effect. Looking forward to seeing some!

      1. cynical1 says:

        From page 14 of your link: “DSP-1181 is a novel compound created by Sumitomo Dainippon Pharma using Exscientia’s AI technologies. In contrast to conventional serotonin 5-HT1A receptor partial agonists (non- benzodiazepine anxiolytics), DSP-1181 has a potent full agonistic activity for serotonin 5-HT1A receptors and is expected to have a long half-life, therefore it is suggested that DSP-1181 has strong efficacy over a long period of time. In Obsessive compulsive disorder (OCD) model mice manipulated OCD-related neural circuit, DSP-1181 is expected to have an earlier onset of efficacy than a standard
        medication, a selective serotonin reuptake inhibitor (SSRI).”

        That doesn’t say that it is selective or that it doesn’t hit dopamine or 5HT7. The only thing it says is that it is a full agonist and not a benzodiazepine.

  23. Todd W says:

    Even though it only took around 12 months to identify the first advanced molecule that fit the development quality criteria, it appears it took DSP-1811 around 5-6 years to get into clinical trials.

    WO2018168738 was filed by Sumitomo Dainippon Pharma in March 2018 (https://patentscope.wipo.int/search/en/detail.jsf?docId=WO2018168738&tab=FULLTEXT). Several of the statements in this patent seem to line up with the details shared in the “what I can see” document as well as the prior Excientia press release from September 2015 on reaching the first delivery milestone in their collaboration with Sumitomo Dainippon Pharma (this press release also states that it took 12 months to achieve this milestone).

    “The present invention relates to a 2,6-disubstituted pyridine derivative or a pharmaceutically acceptable salt thereof which has dual agonism for serotonin 5-HT1a receptor and dopamine D4 receptor; and a medicament for treating symptoms of anxiety-related disorder.”

    “A method for treating anxiety disorder, major depression, obsessive-compulsive disorder, Parkinson’s disease, Rett syndrome, attention deficit hyperactivity disorder, autism spectrum disorder, or dementia.”

    “Further, in a preferred embodiment, excellent metabolic stability, loss of human half-life (T1 / 2) is long, and dopamine D is another GPCR 2 receptor (hereinafter, D 2 receptors) and to hERG channel Weak inhibitory effect.”

    “From the above pharmacological point of view, by simultaneously stimulating 5-HT 1A receptor and D 4 receptor simultaneously and controlling neural circuit functions involved in anxiety from multiple directions, it is stronger than existing 5-HT 1A agonists. It is expected to create drugs having a wide range of anxiolytic effects”

    Based on what I could find in SciFinder, this patent includes many novel compounds that were first reported in this patent. For example, Reference Example 1 in this patent is CAS RN 2244686-21-1. Also, many of the compounds included in this patent are on the same scaffold as Haldol (haloperidol) which was approved by the FDA in 1967. Haloperidol was listed on your long, long list of 5-HT1a active compounds.

  24. Condescending ape says:

    A few unknowns that need answering before anyone can declare victory for AI over humanity.

    How much proprietary information was provided in advance that hasn’t been mentioned in the press release? How advanced was the starting point before AI got involved? How many chemists were assigned to this program to accelerate it because it was a collaboration in a hyped area? I was involved myself with a similar collaboration with this company on another target. Folks on the team saw nothing much out of the ordinary but that didn’t stop other, wiser, more senior people declaring it a great success.

  25. MoMo says:

    That’s what this industry does best Condescending Ape- They Hype everything on the business level, misrepresent themselves on the science while the scientists and clinicians cringe, then wonder why they are caught up in bad press, revenue loss, market cap implosion, and SEC and class action suits.

    But once in a while a great drug emerges and human suffering is alleviated for a time.

    AI is hyped now as other innovations have been in the past, and its the cycle of life in this Industry.

Leave a Reply

Your email address will not be published. Required fields are marked *

Time limit is exhausted. Please reload CAPTCHA.

This site uses Akismet to reduce spam. Learn how your comment data is processed.