Skip to Content

Phenotypic Screening: The State of the Art

I can recommend this article on phenotypic drug discovery from the latest Nature Reviews Drug Discovery. (For the nonspecialists in the crowd, there are two broad categories of screening for drug leads. One is “target-directed”, where you have an idea from other studies about what protein or pathway you want to affect, and you set up an assay specifically reading out on that mechanism. And the other is phenotypic screening, where you know what eventual effect you want to see in the cells or the whole organism, and set up your assay to read out on that phenotype, with no biases about how any given hit compound manages to do it).

Overviews like this appear every so often in the literature, but this is a rather comprehensive one, and it’s always worth reading the articles that take into account the latest techniques. As has always been the case, the first key to running a successful phenotypic program is to make sure that you have assays that will translate into real-world success. As I’ve said before, a bad phenotypic screen is the worst of both worlds, an amazing waste of time, because you end up chasing illusory activity and trying to figure out things that didn’t need to be figured out in the first place. But when it’s done right:

The fundamental determinant of the potential success of a PDD effort is the ability of the screeninassay to predict the clinical therapeutic response to a drug with a specific mechanism of action. This was described by Scannell and Bosley as the “predictive validity” of a discovery model. Here, we propose the term chain of translatability to describe the presence of a shared mechanistic basis for the disease model, the assay readouand the biology of the disease in humans, as a framework for developing phenotypic screening assays with a greater likelihood of having strong predictive validity.

Whatever you call it, it’s essential. As pointed out in the paper, classic examples of good translatability include infectious disease models, direct effects on known human proteins (such as insulin), and monogenetic diseases. Meanwhile, oncology and neuroscience are at the other end of the scale, which is unfortunate, considering how important they are. The defects in the classic tumor models (such as xenografts) are well known, and the models for neurological indications are generally no better. A real phenotypic assay for Alzheimer’s would be a fine thing to run, but it simply can’t be done under current conditions because there is no remotely trustworthy model for the disease. The move from the current Alzheimer’s animal models to the clinic is already fraught with danger even when you have some hypothesis to hold on to, but going full-on phenotypic would be reckless in the extreme.

The classic phenotypic screen has been in a whole (small) animal or via readouts in cell behavior or morphology, but there’s been a lot of work in recent years to see if high-content assays in cells can deliver finer-grained results. “Molecular phenotyping” is an attempt to get a read on entire pathways, distinguishing the various possible assay hits by differences in a whole list of cellular readouts. You’d have to think that this is where the field is going, but it’s not easy. There are so many things to look at that you run the risk of overfitting a model to your data – it’s the classic von Neumann’s elephant problem. Combining this with modern gene-sequencing techniques is also very promising, but adds to that same so-many-variables problem. Advanced cell culture techniques are also clearly the way to go, giving you (potentially) a much more realistic system to test in. The problem there is that progress in assay development can be slow, since the number of potential variations in culture technique is almost limitless, and their effects on the cells (and the validity of the assays derived from them) can be quite difficult to predict.

The review goes on to a good discussion of the strategic considerations in phenotypic screening versus traditional target-directed projects. Phenotypic work, if it’s going to be any good, is often going to take a lot more time and resourcing up front, and an organization has to be committed to that before someone starts getting testy. Another crucial issue is target ID. Every phenotypic program starts out with a plan to identify the targets of any interesting hits they come up with, as well they should. But that’s not always straightforward, or even possible, and management could well lose patience with the whole effort if the screen itself has already taken longer (and been more costly) than they were expecting.

Honestly, if you’re going to run a phenotypic program, you have be prepared, before you start, for the possibility that you may have to go into the clinic without knowing your target or mechanism. Not every organization is ready for that. It takes (for one thing) a lot of confidence in your assay, but if you don’t have that kind of confidence in it anyway, you might want to rethink the idea of using it to drive a phenotypic effort at all (as harsh as that may sound). At the very least, people should be ready to have to make a longer leap of faith than they would ideally want, because it may well come to that. This calculation is going to vary according to the assay, the disease area, the compounds discovered and their activities, and so on, and it will probably not be easy to arrive at a consensus. But the reward for going through all this is the much greater possibility of finding completely new mechanisms of action, and (if things have been done right), the chance to greatly increase the chances for clinical success.

This review also has a useful section on issues such as library selection and design. Ideally, you probably want as large and diverse a compound set as you can possibly run, but the complexities of the assay may well be working against you. There are surely going to be false positives, perhaps a lot of them, perhaps the solid majority of your hits (depending on how good an assay can be developed), and you’re going to have to work out a way of dealing with those in a timely manner or risk getting bogged down. The authors strongly recommend including a good-sized set of compounds with well-understood mechanisms of action, and I second that. A nice set of known drugs and validated probes (as a reality check) is a very good thing to throw into the screening set when you’re heading off into the unknown. Then you have the issue of optimization and SAR development – are your assays up to task, both in terms of throughput and in sensitivity? You will save yourself much heartache if you address these issues before you get started. Finding yourself with an assay that can’t seem to differentiate any of the med-chem analogs is very unpleasant, since you’re never sure if the fault (as Shakespeare’s Caesar put it) is in our stars or in ourselves. Bad assay or bad compounds?

This mode of drug discovery has provided some of the greatest advances in the field, and it has also (under other circumstances) wasted heaps of time and money. Anyone getting into phenotypic drug discovery in hopes of landing in that first category would be well served by this paper, and experienced workers in the field will find it a concise summary of a lot of expensively obtained wisdom. Have a look!

30 comments on “Phenotypic Screening: The State of the Art”

  1. steve says:

    Here’s my question. Who cares about target ID? If you have a drug that works well in a phenotypic screen, and then works in relevant animal models, and then shows appropriate safety in GLP tox, why spend all the time and headache about target ID? That can be done over time and in parallel but in my opinion the emphasis that large pharma has on identifying the molecular target is silly. I know of several drugs that were perfectly good drugs but abandoned by large pharma because they didn’t know the target. A decade later it turns out that the targets just weren’t known at the time and now that they’ve been discovered the original drugs (or variants thereof) are being developed by some small biotechs. We have no real understanding how any drugs actually work in vivo so the idea that target ID should be a rate-limiting step is just an impediment to drug development and a reason for the continued failure of large pharma research to develop new drugs.

    1. Kelvin says:


      As scientists, it’s unfortunate that we let our egos run wild such that we believe we can figure out and design everything from first principles by reductionism, as if we were all-knowing Gods. Patients, doctors, payers and providers, even regulators don’t care how a drug works, as long as it works. And usually we turn out to be wrong in any case. So we need to figure out the fastest way to find what works without all this hubris, and only then, can we afford the luxury to find out how and why it works, to satisfy our egos and curiosity.

      1. Patrick says:

        Sure, all of this, but understanding the target and mechanism of action is insanely useful as well!

        For example, it lets us predict (some) drug and other interactions, which may be life saving for some populations/individuals. And I’m sure any number of follow up drugs have been developed due to the biological understanding unlocked by finding a target.

        Also, knowing the target may help you decide whether or not to continue the program. A good phenotypic assay is great, but perfect is rare… So it might be hitting something that doesn’t kill your assay animal/cell/whatever… But would kill a human.

        That seems like useful info, too.

        So it shouldn’t be the rate limiting step… but they sure as hell should try.

        1. Kelvin says:

          What you say sounds great in theory, but my point is that nobody actually *knows* the target or *understands* the mechanism until a drug is confirmed by a good predictive phenotypic screen in any case (the ultimate phenotypic screen being a clinical trial), and even then we often find that the drug works by some completely different mechanism than expected.

          So “knowing” the target and “understanding” the mechanism usually remains no more than hubris based on an unproven hypothesis, which is incredibly misleading and value-destroying. This is why success rates are statistically higher with PDD than TBDD.

    2. AQR says:

      Knowing the molecular target will make the clinical development of the molecule more rapid and less risky because of the ability to establish a target engagement biomarker in the clinic. In those cases where the molecular target has been determined (either from a TDD effort or through deconvolution of a PDD effort) one can use an assay of clinical target engagement to establish the doses and time course that the compound is interacting with the target. This forms the basis for the dosing regime that one can use to demonstrate efficacy in the clinic and the denominator for the margin of safety calculation versus any adverse events.

      In the absence of a target engagement biomarker, one is left with using plasma levels or an in vivo or ex vivo equivalent of the phenotypic measure as a surrogate. Using a pharmacokinetic measure will not allow one to identify compounds where there is no clear cut PK/PD relationship, such as where a significant fraction of the therapeutic effect is due to a metabolite (ex. fluoxetine, risperidone).

      This may be one reason that large pharma places a higher premium on target identification than smaller organizations. They are typically responsible for later stage clinical trials and regulatory submissions, activities make much easier when one knows the target.

      1. Kelvin says:

        Ironically, you demonstrate exactly my point:

        What you say sounds great in theory, but my point is that nobody actually *knows* the target or *understands* the mechanism until a drug is confirmed by a good predictive phenotypic screen in any case (the ultimate phenotypic screen being a clinical trial), and even then we often find that the drug works by some completely different mechanism than expected.

        So “knowing” the target and “understanding” the mechanism usually remains no more than hubris based on an unproven hypothesis, which is incredibly misleading and value-destroying. This is why success rates are statistically higher with PDD than TBDD.

      2. Carol Gebert says:

        I agree with AQR. In situations of clinical clarity, the efficacy question trumps the MOA details. However, what about when the clinical situation is borderline or confusing? Knowing the MOA makes all the difference in being able to segment patients apriori, and thus design a clear clinical trial around a known MOA. The classic example would be KRAS mutants for lung cancer. Knowing that MOA enables focused drug design during early discovery and patient selection for clinical studies.

  2. Curious Wavefunction says:

    It need not be either/or. I know of at least one project where PDD was used to discover a hit that was then handed to the target deconvolution team who found the target and then optimized it all the way to a candidate. Some of the advantages of TBDD over PDD include cleaner readouts and less possibility of close analogs causing the same phenotypes through non-specific activity. In addition there’s the cost and setup complications for PDD which are not trivial, especially for smaller companies. I think that in specific cases where the readout has a very solid correlation with the disease state across several chemical series and the assay is relatively easy to set up, PDD would indeed be preferred, but your general portfolio would probably still need to consist of a mix of TBDD and PDD.

  3. John Wayne says:

    “Expensively obtained wisdom” is something that should be valued by managers of R&D, and it often isn’t. It has to be balanced by people who gather baggage (instead of wisdom) as they go through their careers. A senior scientist with wisdom will speak softly about assumptions and design good experiments, while a person with baggage will just tell you that B is crap because of A.

    I’m a veteran of a lot of phenotypic screening, and there certainly are caveats (see above). In addition to the usual artifacts, you also have to know what you don’t want. For example, if you are hoping for a new mechanism of action you should put assays that can determine this right at the beginning. I appreciate Derek’s comments about making sure that you have to will to push something with an unknown mechanism forward. I have seen situations wherein the scientists are okay with it, but the business people want to know.

    1. Some idiot says:

      Interesting…I would have gotten that wrong, assuming that the business guys wanted to push it on, but the scientists were cautious…!

      Any idea why in this case? Risk/benefit analysis, or just harder to argue for a good price if the mechanism is unknown…? Or something totally different?

      1. John Wayne says:

        I think it is a mix of things:
        1. A lot of business folks are familiar with the target-based approach to drug discovery; when something appears different they ask questions.
        2. Knowing the target could give you a competitive advantage over others if you keep it a secret and get really far ahead of everybody else.
        3. Knowing the biochemical target (or targets) of a lead derived from a phenotypic assay makes it feels like it is de-risked. I don’t think that this is always true, but it can be in some situations (Example: a phenotypic hit has the same biochemical mechanism of action that is likely to lead to rapid generation of resistance.)
        4. In my observation you can’t effectively patent biological targets that you deconvolute, but maybe you can burden others so greatly it represents a good strategy to create and temporarily maintain a monopoly.

        1. Some idiot says:

          Thanks… Yes, they feel like they are in the right direction, especially 2 and 4.

  4. Marcin says:

    What is the purpose of “optimizing” a compound in TBDD for one of the targets that initial compound might hit ??
    I think it should have been accepted by now that a lot of “good drugs” work because of their pleiotropic effect, which could be picked up in a PDD, but not necessarily in a single-target assay. The more I study Anatomy and Physiology the more I realize how interconnected tissues are. It is continuous drive to “de-risk” vs. “courage to develop” makes the bigger organizations less and less productive.

    1. qazakh says:

      Slight correction, I think it is well established that many drugs bind to multiple targets or at least are active in bioassays, which is not the same thing at all. What’s not always so clear is that any of this is the reason for their in vivo activity. Quite a difficult thing to prove, that…

      1. Barry says:

        Several companies followed Novartis’ Gleevec with more selective inhibitors of BcrAbl. Their PK and PD were good, but their efficacy was less. At least in this case, the “dirty drug” was better. Other readers will supply other examples.

    2. Dr CNS says:


      I agree with your comment.
      My follow up question is: how do you know all these pleiotropic effects are identical in preclinical species and in human?
      Or in healthy human and the disease state?
      IMHO, most of the time, you don’t know.

  5. luysii says:

    There is a rather frightening article (about which I plan to post) on two different phenotypic screens giving different wildly different targets, with little commonality between the hits. Off to play Faure this afternoon however

  6. Bunsen says:

    Render unto Caesar the lines that are Caesar’s, and unto Cassius the lines that are Cassius’s.

  7. 10 Fingers says:

    That’s a great article, well worth the (long) read. Thanks, Derek!

    In my (somewhat chemocentric) view, a lot of small molecule drug discovery boils down to two questions, which one can usually address up front, prior to initiating a program:
    1) How do we find a starting point for a therapeutic approach?
    2) Assuming what we find isn’t good enough to be a drug yet, how do we decide what to make next?

    A lot of the success or failure of projects hinges on the answer to the first question, in that many studies have shown that the liabilities of a chemical series are baked into the initial hit. Depending on how the phenotypic approach is set up, there can be some real world benefits in compound properties (permeability, bioavailability, etc.).

    However, there are often challenges in getting structure-activity data from a phenotypic model – particularly one that integrates several different effects. Even with only one target responsible for the activity this can be complicated (say, if distributed in different cell compartments or tissues), but if the molecule is truly polypharmaceutical then changes at the atomic level can have conflicting (even dramatic) effects that cancel out in the SAR.

    If the bar for success of the screen wasn’t high enough, and there is a lot of work to do on the molecule, this can really be challenging. It is very easy to get caught in the “plateau of despair” – the region where lots of changes to the molecule yield pretty modest changes in bioactivity. Since a lot of phenotypic assays can be low-throughput and slow to turn around, this can add to the challenge. Not everyone has the patience to finish what they start in this niche, much less the foresight to plan for it in the first place.

    Pragmatically, it is possible to use PDD in pretty powerful ways, either as part of a nascent TBDD effort or as a fairly direct path to a drug. However, if the answer to the questions above is something like “we are going to look for anything that moves our really great endpoint in our novel assay and then the chemists will do their magic and we will go into the clinic,” that’s probably not going to work.

  8. Anon says:

    PDD-derived drugs have a better success rate in the clinic, but also fare better in the market once approved, simply because you don’t get 10 competitors following the same target.

  9. qazakh says:

    Add to that the unfortunate truth that high throughput phenotypic screens themselves are only linked to disease by some hypothesis that could very well turn out not to be important when it hits the clinic. Screening direct in a phenotypic assay may accelerate the route to cell-active hits, but in terms of clinical success there’s no reason to believe these will do any better than hits found some other way and subsequently optimized to have activity in the same phenotypic screen.

  10. JB says:

    Yup. We used to never care about target ID or even a lot of the principles of drug design back in the old days when we were cranking out molecules left and right. Reductionism in biology is dangerous. And many molecules that are used every single day in the clinic would have been thrown away by Pharma is we followed the dogma that’s around today for how to properly design a good drug.

  11. Marcin says:

    We should really ask someone how discovery was done before the advent of single target assays. 1980s? 1970s? Earlier? Maybe once the potency in the PDD is “good enough” the SAR should be geared towards PK/ADME only with only cross-checking if the initial bioactivity is not lost. Maybe too much “optimization” is not necessary??

  12. Istvan Ujvary says:

    This could be relevant.
    “The history of drug discovery spans approximately 200,000 years”
    is the first sentence of a suberb oldie by Enna & Williams. The last 2 sentences conclude that there is life the beyond “targephilia”:
    “Moving animal model assays to an earlier stage in the CNS drug discovery process to compliment and enhance the “targephilic” approach can provide data enriched with an element of calculated serendipity. It will also yield an earlier indication of a relationship, or lack thereof, between a molecular target and animal behavior, thereby accelerating the discovery of novel drugs for the treatment of psychiatric and neurological disorders.”

    1. Anon2 says:

      And which psychiatric or neurological animal models did they so highly esteem? Putting a stronger emphasis on animal models only makes sense if the animal models are predictive of activity in the clinic. Alzheimer’s? Schizophrenia? Depression? Where are the wonderful animal models?

      1. Istvan Ujvary says:

        Yes, proper animal models are indispensable (they do no need to be wonderful ;-). Of course, they need to be developed.
        Anyway, risperidone, a best-seller, comes to mind.
        Discovering risperidone: the LSD model of psychopathology
        perhaps memantine and fluvoxamine could be another examples of research project relying heavily (solely?) on in vivo / model studies. But these are old success stories.

        1. Dr CNS says:

          … and how do you know these models had anything to do with the approval of these drugs?
          How many other compounds work in these models, and have failed (or would have failed, which we would never know) clinical development?

  13. Anon says:

    If only there was a cure for wrongtargetitis. Or at least a faster and less painful death!

  14. steve says:

    @Istvan – Thanks a lot for the article; it makes my exact point. I love the term “targephilia”. I also love the conclusion, which says it better than I did: “Although the modern empirical approach is more informed than that employed by prehistoric scientists, the objective remains the same: to identify as quickly as possible whether a particular chemical agent displays drug-like characteristics of clinical importance. Although knowledge of site(s) of action is critical for fully characterizing and exploiting a drug class, the primary aim of a drug discovery program must be the identification of new therapeutic agents, not the synthesis of high-affinity ligands for molecular targets of unknown clinical value.” To mangle Shakespeare, “There is more to biology than is dreamt of in your philosophy.”

Leave a Reply

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

Time limit is exhausted. Please reload CAPTCHA.