Skip to Content

Drug Assays

Experiences With Phenotypic Screening?

Very little blogging time today, but I wanted to throw a question out to the readership instead. I’m at the Keystone conference on Phenotypic Drug Discovery, so here’s a relevant topic: what are your own experiences with phenotypic screening?

Background for those outside the field: broadly speaking, you can sneak up on a drug by two big routes. If you have an idea of a particular target (protein, etc.) you can screen against it directly to see if you have compounds that affect it. That’s the target-based approach. The other way is if you have some assay in cells or animals that shows you improvement in a particular disease state, then you screen compounds on that and see if you find something that helps. If you find a compound, you of course have no idea about how it might work, but there are lot of interesting techniques to try to work backwards and learn something new about the biology (as well as moving towards a new drug). This is the phenotypic approach. There are a lot of ways to sort of be in between these two, but that’s the main point.

I’ll let everyone define phenotypic screening their own way (just mention what that might be), but I’m interested in how many people have been on phenotypic projects and how things worked out. I myself have never been on a pure phenotypic project, but I’ve seen them lead to clinical candidates in the labs right next door to me at different times. At the same time, I’ve also seen them consume a lot of time, effort, and money and not deliver (to be clear, you can say the exact same thing about target-based drug discovery).

My own contribution to the discussion is the observation that if the organization is not committed to making the final leap into the clinic based on the phenotypic model, then you shouldn’t have started the program in the first place. And that does happen. Every such project gets to a leap-of-faith point when you have to go into humans, and that’s when you discover how strongly you believed that phenotypic model in the first place. It would be better to experience that under less trying conditions, i.e. earlier in the program!

41 comments on “Experiences With Phenotypic Screening?”

  1. Curious says:

    Just because you don’t know the target, doesn’t mean you can’t find out. Target deconvolution is difficult but not impossible.

    There is a strong inventive in determining the mechanism. Without knowing the target, it is difficult to rationally set the clinical dose.

    1. AR says:

      Without knowing the target, it is difficult to rationally set the clinical dose.

      Is it? As far as I know you base in on in vitro work initially and adjust based on results in animal models. There are plenty of polypharmacological compounds where we know only some or none of the targets, doesn’t stop us determining therapeutically useful doses.

  2. Delph says:

    In the agrochemical industry, virtually all discoveries arise from phenotypic screening. Target based approaches have traditionally been very problematic and have extremely poor success rates for even basic transitions into the greenhouse. Phenotypic screening coupled with target identification after the discovery of the lead compound has led to most modern pesticides.

    1. Onceachemist says:

      I recall a senior scientist weighing in on the debate during my first job in an ag chem discovery organization. He couldn’t understand the need to try and develop target-based assays for nematocides when “you’ve got a whole worm full of assays right here!”

    2. Red Fiona says:

      Is that because it’s easier to make the leap from phenotypic screening in the (literal) field when you’re working in agrochemicals?

      1. Delph says:

        There are a few reasons but the biggest is simply that for any compound to be an effective pesticide it has to have decent physical and chemical properties – it has to penetrate the leaf/root (or not), be UV stable, survive metabolism and get to where it needs to get within the plant. Lots of compounds can be designed that will hit the target but will never make it through the leaf and so are totally inactive in the greenhouse/field. It’s something that pharma doesn’t have to worry about – imagine if you had to dose a drug by spraying it onto your hand and hoping that it gets to the target inside the body. Add that to the fact that targets are less defined than in pharma – there are a lot less plant enzyme crystal structures available and you’re in a case where phenotypic screening is a lot better than target based.

        1. Red Fiona says:

          Thanks for your answer. Outsides of leaves as barriers wasn’t something I’d thought about.

  3. John Wayne says:

    “Without knowing the target, it is difficult to rationally set the clinical dose.”

    I’ve never heard a statement like this; what is the thinking behind it?

    1. KN says:

      For example, to avoid prescribing two drugs affecting the same target.

    2. Sutemi says:

      Without a target, you can’t measure target engagement. Therefore when you go into the clinic if the compound doesn’t work, you don’t know if it is because you didn’t dose up high enough to test the hypothesis or if the hypothesis is wrong.

      1. Belgian PhD student says:


        1. John Wayne says:

          For the record, these answers seem a bit weak.

          1. AR says:

            Answers to what? Measuring target engagement in vivo is simply just not the way it is done.

  4. Wavefunction says:

    Agreed. I am especially skeptical about a lot of these computational/machine learning approaches that are being used for phenotypic screening. You need a lot of organizational support for follow-up and target deconvolution, especially when you have to chase down a lot of neural net-enabled predictions that may be false positives or weak correlations.

  5. Frank says:

    One can argue artemisinin came from a phenotypic screen effort.

    1. Rich says:

      One could make the argument that nearly all approved antimalarials (and most other anti-protozoan agents) were based on phenotypic screening.

      1. Anonymous says:

        And a bunch of other discoveries via unexpected phenotypic outcomes in animals or humans:
        1. Chloropromazine (first antipsychotic?) – developed as an anti-histamine, found to produce sedation w/o narcosis.
        2. Minoxidil – developed as an anti-hypertensive, it produced hair growth as a side effect. “Side effect” became the main desired effect as a treatment for baldness.
        3. Sildenafil – developed as an anti-hypertensive, it produced penile erection as a side effect. “Side effect” became the main desired effect as a treatment for ED.

        (Not to be confused with drug repurposing. Some drugs are studied in vitro for new applications without prior observation of a phenotypic outcome in vivo. E.g., thalidomide for leprosy and cancer.)

  6. MrXYZ says:

    A number of biologic drugs, e.g. rituximab, were initially discovered using a more phenotypic approach. Mice were immunized with cancer cells and the isolated antibodies were tested for their ability to effectively kill tumors in xenograft models. Only later were the targets of the antibodies identified (e.g CD20 in the case of rituximab).

    I don’t think it’s an uncommon approach in the biologics world, particularly in oncology.

  7. cynical1 says:

    Back in the day…………..way, way back in the day (and I am dating myself), I worked on an antipsychotic project. We screened everything in vivo in a Conditioned Avoidance Response in vivo assay in rats. So our screen was the most primitive of all phenotypic screenings. We put it in rats and watched what happened. No PK before dosing and we were required to submit at least 2.00 grams of every analog we made. (You chemists who have only had to make a few mgs of your analog don’t know how spoiled y’all are!) My memory is a bit foggy but I believe they were all dosed orally as well, not i.p. If you think that sounds crude, it is and that’s the way all the old antipsychotics were discovered, as well.

    At the time, the technology was just becoming available to see which 7TMs our compounds were hitting but we really didn’t know which ones were the most important or what a “good” profile would look like. We basically just looked at the in vivo activity. And if you look up antipsychotics on the Wikipedia page and look at the big chart of what 7TMs all of them hit, I would suggest that we, as a scientific community, still don’t know what an ideal profile looks like. It’s 2019 and they still use that in vivo assay because all of the clinically efficacious antipsychotics work in that in vivo model – regardless of the 7TM binding profile. Clozapine really doesn’t hit D2 very well at all but it works in schizophrenia, arguably the best.

    So, the best predictor of antipsychotic activity in humans is still whether your compound keeps a rat from pushing a bar you trained it to push to avoid getting an electric shock.

    As an aside, a very infamous antipsychotic came out of all of the efforts by the company I was working for at the time. The infamy was not related to its efficacy as an antipsychotic. The infamy arose from how it was marketed for everything under the sun.

    1. Paramus says:

      Totally agree, I worked on olanzapine and the development was based on phenotypic screening, or in vivo pharmacology as we called it in those days! When I started lecturing on the compound, it was selective for three receptors, a number of years later it was hitting at least 14 different sites.

  8. Calvin says:

    So in general I’d say that for infection (anti-bacterials and anti-virals) phenotypic screening in the only way to go. Target ID is generally doable and there’s plenty of precedence for taking a product to market with a phenotypic approach even if the target is nailed down. And of course, the direct link to clinical effect is a bonus

    1. Vaudaux says:

      Calvin, did you mean “…even if the target is not nailed down”? If so, I agree with you.

      In infectious disease there are many drugs that have been used for decades for which the mechanism is still not understood. This is particularly true for tuberculosis drugs and antiparasitic drugs.

      Every class of antibiotics that has reached clinical use was discovered phenotypically, using antimicrobial activity (or another effect on the bug, such as morphologic change) to guide chemistry and/or purification from the natural product source). The sole exception I am aware of is trimethoprim, discovered through a deliberate effort to find antimetabolites.

      Starting from a target is intellectually attractive, but has not yet produced an approved antibiotic.

      1. Calvin says:

        Yes, my bad. If the target isn’t nailed down.

  9. Paul Brookes says:

    From the academia side of things, we did this almost a decade ago: Essentally a plate-based method using a cardiac cell line to test for protection against ischemia-reperfusion injury. Tested 2000 molecules (with 20/20 hindsight, a collection of total crap). Got some hits, followed up in perfused rat heart model and saw the method had some predictive power. This led to some interesting follow up studies trying to figure out potential protective mechanisms, which we did for one of the hits (

    As an academic exercise it was informative and we learned a lot of skills that have come in handy for other projects, but as a bona fide drug discovery effort it didn’t really pan out (big surprise!)

  10. Gareth says:

    As Mick Jagger put it, you can’t always get what you want, but if you try real hard, then maybe, you might get what you screen for.
    My former lab has had some good results with a reporter gene assay that happened to hit a sweet spot in the sensitivity range, and with a high-content screen looking for fluorescent morphology changes.
    There are great tools available for screening and target identification out there now. Although some of them require a lot of expensive toys. The Cravatt group’s diazerine/alkyne moiety for proteomic identification that was mentioned here recently springs to mind. The key, as always, is really solid validation with SAR and orthogonal assays.

  11. HFM says:

    I’m a keyboard jockey. Early in my career, I did the analytics for a rather gnarly pure phenotype screen. It involved high-content imaging of organoids, which were made by mixing several cell cultures together and growing in a particular way. These organoids would, if left to their own devices, develop the phenotype that we wanted to target. So, drug the organoids, take pictures, and then figure out which ones didn’t do the bad thing.

    I don’t think it ever came to anything. Problem #1, the readout was not terribly reproducible. It was a fairly subjective phenotype, and a decidedly complicated assay. Problem #2, this was still a highly artificial system. Even if a drug gave beautiful results for this “phenotype” every single time, there were a million reasons why it could be doing that, and most of them aren’t what we want a drug to be doing. (Usually killing the organoids in creative ways.)

    The nice thing about target-based screening is that it’s generally straightforward to set up your screening assay. Phenotyping at scale is usually hard. Phenotyping something that’s actually relevant at scale is harder. But if the stars align and you can do it…

    Just keep in mind, for the sanity of your keyboard jockeys, that computers cannot magically turn bad assays into good ones. Computers can reason about signal that’s smeared over thousands of dimensions, which is a neat trick, because our puny human minds cannot. This expands the universe of good assays available to you. But a bad assay is still bad.

    1. Derek Lowe says:

      This is a good point – every phenotypic cell assay that has death (or something close to it) as a readout will have the problem of clearing out mechanisms that are too toxic (or too broadly toxic) to be of any real use.

  12. P4j45 says:

    @cynical1: is the infamous one by any chance risperidone from Janssen?

    1. cynical1 says:


  13. JB says:

    There are TONS and TONS of companies that take the phenotype approach. I get that this is more of a chemist/small molecule focused blog, but I work with biologics. A lot of companies are working with stem cells, administering bugs to alter the gut microbiome to treat cancer or have immune engineered cells to do something. Nobody knows how or why altering gut bacteria might be beneficial for treating cancer or autoimmune diseases, so there are no mechanism based approaches to target. That doesn’t stop a lot of companies from existing and trying it. There are hundreds of clinical trials using all sorts of types of stem cells to treat every indication under the sun. The only thing people basically know is that they are like some sort of magic potion that supposed to behave like a live biopharmaceutical factory in vivo that responds to environmental cues. The more complex the products get, the less mechanisms are known to exist and are isolated and tested. The FDA requires no proof of mechanism in CBER for complex biologics. You make a tumor mouse, use the complex biologics, and shrink the tumor to establish proof of concept. The only other thing to focus on is safety.

    1. Derek Lowe says:

      Yikes. Has any of the stem cell stuff gone into humans?

      1. JB says:

        This was written in ’16….there were already almost 500 clinical trials where MSCs were being delivered into humans:

        And that’s just the tip of the iceberg. I mean all it really takes is a quick Google search to show you how many companies are delivering live microbials to alter the gut microbiome to treat things like cancer……literally no one knows how it works except that it should ‘alter the immune system’ in some favorable way.

  14. bacillus says:

    @Calvin. I agree. Have been doing this for 37 years in the vaccine field and have used many thousands of rodents and other species in the process. However, some of the vaccines I’ve worked with have been rodent strain specific in terms of their ability to raise a protective immune response as opposed to a non-protective immune response. Have tried to use this fact for the past several years to try to find a correlate of protection, but without success. I’m not sure to what extent drug candidates are tested in a range of mouse strains to ensure that you’re not throwing the baby out with the bathwater when screening for phenotypic effects.

  15. milkshake says:

    I was on a phenotypic screening medchem project in ocular medicine at Johns Hopkins – the idea was to screen approved drugs and drug candidates that were in clinic, for acting as neurotrophic factors, in preventing spontaneous dying of embryonic retinal ganglion cells.

    The cell based assay was nice but getting the super-fragile cells from retinas of newborn mice pups was a chore, there was just one competent guy (brain surgeon by training) who could do the cell harvesting under a microscope.

    The problems for a medicinal chemist working on this project were twofold: 1) we got mixed SAR, and the original leads (Sutent and another first-generation oxindole kinase inhibitor) were notoriously nonselective. We got better compounds from screening later on but too late to make a difference in the project, which was limping due to the PI screwing up the collaboration 2) The assay output was not in typical format – it showed just bars of cell survival at different concentrations – and it was not a neat sigmoid curve but rather a sigmoid with a plateau of varying height followed by a downslope (as the compounds started to be toxic at higher concentration). The data format tremendously frustrated our chemistry superiors in the collaboration – as medicinal chemists they were used to having a neat table of IC50 or EC50 values, and now they had to deal with a multiparameter curve (a plateau at lower concentration is better, but what if another compound gives higher and broader plateau albeit at higher concentration?).

    Anyway, the collaboration ended because the biology PI was not able to hold this kind of project together, and he was used to people doing favors for him for free (or for being added to a publication) and he turned to be out surprisingly inept at getting his medicinal project funded. And unfortunately he was a bit of a bullshitter too, had no clue about what is needed for the project, and so on. A typical academia medchem screwup that did not do much good to anyone involved in the project.

  16. 10 Fingers says:

    Two comments:

    With phenotypic screening, you get to really explore the courage of your convictions – how much do you believe in your readout and its translatability to human biology? There is often a big gap between that assay and anything you can see in a whole animal, much less a human – until you get to a phase 2 trial. If you have an acute or proximal readout of activity that translates, the approach can be very fast and powerful – as long as you can evaluate compound SAR accurately enough to drive to a potential drug.

    Second, how well do you think that you can evaluate the window between that readout and various measures of potential toxicity, on the way to the clinic – this, so that you can set doses reasonably expected to provide a window of efficacy without toxicity.

    The first generation glitazones are a famous example of whole animal screening of compounds (on glucose lowering) that led to a new class of drugs for diabetes. Initially, there was no clarity on the target of these compounds. Many of the benefits and potential problems of the approach are apparent if you spend a little time on the history of this class of compounds.

  17. John Wayne says:

    My opinion is that if you could reasonably be doing some phenotypic screening, you probably should be. It certainly isn’t a cure-all and it doesn’t work all the time, but I have found it to be generally useful.
    1. You do have to be honest about translation from your screening strategy to the clinic
    2. Typical phenotypic screens are lower throughput and are more expensive to set up than biochemical versions – buy in from others can be hard to get (especially if you cowboy up and screen in animals)
    3. It is easy to end up with garbage hits (especially if death looks like a lead;) tough counterscreens are usually needed
    4. Do use lot of positive and negative controls from the literature. Testing material issues (cells loosing phenotype, etc) are more common in phenotypic screening. In one project I worked on, we had six controls (four positive, two negative) on each plate – only way to be sure
    5. Phenotypic screens are great are rediscovering known leads. You need to know what they are. As an additional control, add some of these to your screening samples – if everything is working you should find them again.
    6. If your program goals are to find a lead that is differentiated from what is currently available, screen directly for that property. Setting a high bar will save you tons of time validating leads you may not be interested in.

  18. Gffffghjvv says:

    Phenotype is king…..PIs long ago understood that mice experiments gave poor phenotype for humans, genotype from dna ( a weaker predictor, ) cannot be had easily had. So, they started with phenotype experiments on the easily accessible staff and students. Physical experiments are obviously prohibited by lawyers, but on the contrary lawyers encourage psychiatric experiments, since they dismiss anything that is not in text because it doesn’t make money. Think I’m lying, try to explain the insane human experiments that went on in world world II by multiple countries and explain how all those sources are lying.

  19. cytirps says:

    Just a reminder that both Zetia and Zolina were approved by FDA years before their targets were identified.

  20. Grad student tips says:

    Do not mess with the Paul Plevin law firm in San Diego. The business model of this firm is to harass former students of San Diego research institutes and sue them.

  21. Fabio Gasparri says:

    “Without knowing the target, it is difficult to rationally set the clinical dose.” Absolutely not. There are many significant examples of drugs approved before the target and the mechanism of action were identified (paclitaxel, cyclosporine, thalidomide in multiple myeloma, just to mention some).

  22. Kate says:

    New assays are making it easier to deconvolute the target following a phenotypic screen. One of the scientists I work with presented a poster on Capture Compound Mass Spectrometry, which has applications in this area. This is the poster link:

Leave a Reply

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

Time limit is exhausted. Please reload CAPTCHA.