Skip to Content

Thoughts on An Antibody Failure

Talking with some drug discovery folks the other day, I said “You know, if you don’t hold your breath when your compound goes into tox testing, you haven’t been doing this stuff long enough”. Well, it’s pretty hard to hold your breath across a full tox study, but you know what I mean. There are so many possible problems, and we know so little about them (and so little about what our drugs might do) that you really have to be braced for the unexpected. And that goes for the first two-week rat tox study and all the later follow-ups in other species and for longer times (13 weeks, two years). Even after all that, you still have to be on guard when your compound goes into humans. The animal studies will catch a great number of things that you don’t want to expose the human population to, but they won’t quite catch them all.

That was illustrated yesterday by the news that Amgen’s antibody therapy Evenity (romosozumab) shows an unexpected cardiovascular liability – that is, it seems to increase the risk of heart attacks in its target population of post-menopausal women by up to 30%. The drug was supposed to come up for an FDA decision by mid-July, but that’s not happening any time soon, and it may not ever happen at all.

Romosozumab is a joint project of Amgen and UCB Pharma, and it targets a protein called sclerostin. That’s involved in the Wnt signaling pathway, which I’m willing to stipulate can cause almost anything to happen. There are some very rare bone disorders, characterized by abnormally high bone density, that can be traced back to mutations in the sclerostin gene, leading to a less functional protein. These genetic clues are just what drug discovery people are constantly looking for, and set off the idea of interfering with sclerostin as a possible way to increase bone density in patients that actually need it – those suffering from osteoporosis.

This idea actually worked (not all of them do, of course), and both animal studies and human trials showed increased bone density, as hoped for, and a lower rate of the common fractures in the osteoporosis population (vetebrae and hip). So things really were moving along as hoped for, and a completely new method of treatment was coming along, when the toxicity signal came up. That Matthew Herper piece linked to in the second paragraph talks about how this illustrates the limits of genetics in drug discovery, and also mentions the PCSK9 story in this wise. I see the two as somewhat different, though. What we learned from genetics was that interfering with sclerostin could increase bone density, and this is exactly what happened.

The few dozen people with such mutations, though, are not enough to tell you very much about possible toxic effects (as Herper notes as well). It might be that even a big population would not give you the signal, either, because of compensatory processes that could be acting in people who were born with such mutations, and which don’t have a chance to kick in when you hit someone with the same thing at age 70 – that sort of thing is well known in both human patients and animal models. PCSK9 is similar, up to a point – the genetic data suggested that it would have a big effect on lipoproteins, and so it does. But the PCSK9 mutant humans do also show a much-reduced risk of heart disease, which made that pathway even more exciting. The currently good-but-maybe-not-great human outcomes data for PCSK9 antibodies, then really might illustrate the limitations of genetic signals in developing new drugs, because you really would have expected more (although it’s still possible that longer-term studies will look better).

But sclerostin and romosozumab, at least to me, seem to be the more normal way that drug discovery has always worked. The genetic information did not (and could not) say that much about safety, and it wasn’t a beneficial mutation in the first place. Turning this around and using it for osteoporosis was a good idea, but was always going to be more risky than trying to recapitulate the good phenotype of PCSK9 mutant humans. And here’s the risk, reminding us that it’s still here, reminding us that it never goes away.

16 comments on “Thoughts on An Antibody Failure”

  1. Anon says:

    The devil’s in the details. The much larger FRAME study showed no statistically significant increase in heart risk.

  2. Anon says:

    Humbly, it has been clear for some time that human genetics is not equal to human biology. Just one dimension of it. So, this should not be a surprise – just bad news.
    That is why we do this type of studies, as Derek writes: to keep learning and build a new, hopefully better hypothesis.
    Kudos to those who keep trying!

  3. Newbie says:

    Naive question:
    Given that this is an antibody, does that imply that this is an on-target toxicity issue?
    Or is it possible for an Ab to have toxicity issues that are not target related?

    1. David Borhani says:

      Newbie, your question is a very good one. As for any drug, toxicities caused by mAbs may be on- or off-target. Typically, mAb’s are much “cleaner,” and associated toxicities (or insufficient efficacy) has more to do with unclear or unknown biology surrounding the target as opposed to frank binding to something unrelated.

      One example serves to illustrate: anti-IL-12 mAbs were developed before we knew there was such a thing as an IL-23. The two cytokines share the p40 subunit but differ in the other subunit (p35 or p19, respectively). Ustekinumab (Stellara) binds p40, and neutralizes both cytokines. ABT-874/briakinumab does the same (need to dig into the patent literature to know this—see WO 2012/094623). By binding both cytokines (~equally), the impact on biology—efficacy and toxicity—is more complex than would be (and I think, is) observed with IL-23-specific mAbs. This is part of why choice of binding epitope is a crucial consideration in the development of an mAb drug.

      Antibodies usually bind tightly and highly specifically for the target at hand (for romosozumab, the Kd is reported to be ~10 pM, not atypical).

      It is usual to approach your question from two directions: toxicology experiments, where you dose animals and look for an adverse event, regardless of an understanding of the mechanism (which you may try to then understand; in fact, the experiment is done such that, by dosing ever higher amounts, you *do* observe *some* adverse effect—which allows you to define the so-called No Observed Adverse Effect Level, NOAEL); and biochemical experiments where you seek to determine, usually in an agnostic fashion, what else the mAb binds to, with what affinity, such that you can then worry about whether that binding might be a problem or not.

      The former approach was certainly done for romo (as for any drug): long-term repeated-dose toxicology (and efficacy) experiments in were done in rats and cynomolgus monkeys. Everything looked good (I couldn’t find the NOAEL nor the dose-limiting toxicity).

      The latter…I would presume they were done, but I’ve not been able to find them in the literature. Typically, bioinformatics checks are done to see what else *might* the mAb bind based on sequence considerations alone, e.g. SOSDC1 in addition to the target SOST (or SOST isoforms, which may have partially different functions). Experimental biochemical checks might include immunohistochemistry (Which tissues bind the mAb? How hard is it to compete away binding with washing, or added SOST?) and “grind [tissues] & bind [mAb]” fishing experiments, etc., aimed at finding where/what else the mAb actually does bind to, and whether it binds with sufficient affinity to be potentially problematic.

      Sorry for the long reply, but it’s an important question. No doubt all this and more was done by teams at Amgen & UCB, and they didn’t find anything that would have clued them into the observed increase in CV risk factor. Otherwise you can bet they wouldn’t have gone to Phase 3.

    2. steve says:

      It is certainly possible for an antibody to have off-target effects. In immunology this is known as cross-reactivity and is usually dealt with by increasing the affinity of the antibody for the specific epitope to which it binds and screening tissue arrays to make sure the antibody is specific. That said, it’s possible that a particular epitope is shared by more than one protein; for example, there could be similar glycosylation, three-dimensional shape or other similarity. This is thought to underly many autoimmune disease where a viral protein looks like a host protein and the immune system begins by attacking the virus and ends up attacking the host. It also resulted in the death of some patients treated with CAR-T directed against MAGE-A3. Of course, you can also get on-target tox as in patients treated with Herceptin who develop cardiac tox. Of course, the same is true of small molecule drugs; the same patients suffer cardiac tox from anthracyclines.

      1. Anon3 says:

        Steve,
        I agree with your comment, and wonder if you have a literature reference including DATA with cross-reactivity of antibodies.
        With small molecules we often run in vitro panels (GPCRs, ion channels, kinases, etc) to assess cross-reactivity – basically to rule out the interactions with the targets tested.
        When I ask colleagues to run appropriate in vitro tests to establish the selectivity of their antibodies, they all came with the same response: “they are very potent, we don’t need to do that.”
        Is that never done, really?

        1. Imaging guy says:

          Use of tissue cross-reactivity studies in the development of antibody-based biopharmaceuticals: history, experience, methodology, and future directions.

          Leach MW, Halpern WG, Johnson CW, Rojko JL, MacLachlan TK, Chan CM, Galbreath EJ, Ndifor AM, Blanset DL, Polack E, Cavagnaro JA

          Toxicol Pathol. 2010 Dec;38(7):1138-66. doi: 10.1177/0192623310382559. Epub 2010 Oct 6.

  4. Chrispy says:

    Well, Amgen seems to be having some surprises in antibody-land. Recall Siliq, the anti-IL17R antibody that increased suicidality. It is fair to say that no one saw that coming, and Amgen and AZ backed away from it — bottom-feeding Valeant scooped it up, just in time to compete with anti-IL17 ligand antibodies that did not show suicidality. Then we had Repatha, the anti-PCSK9 antibody that showed a remarkable effect on cholesterol but less of an effect on survival (we are all learning here). At least they didn’t have an issue like Pfizer, where immunogenicity forced them to pull the plug on bococizumab, their PCSK9 candidate. (Industry scuttlebutt says that Pfizer mutated the Fc in that antibody to extend half life, which may account for the immunogenicity.) I wonder what has been taking so long with romosozumab — the deal with UCB was done in 2002! Amgen has been quite good at evergreening old drugs with extended patent life (lookin’ at you, Enbrel!), but one has to wonder if they are running up against patent expiration on this antibody. To your question @Newbie: yes, the toxicity seen from antibodies is typically “on target,” but with a pathway like Wnt it is difficult to predict a priori all the buttons you might press. In fact, there is a letter in the NEJM (linked to by Herper) that states that the role of sclerostin in vascular pathology is becoming more and more apparent. Of course, it is studies like this one which will enable us to make the connections, even if they’re a bit disappointing.

  5. Imaging guy says:

    In 2010, Genentech published a paper in Nature Biotechnology screening for many medically relevant phenotypes (bone density, blood chemistry, immunology and etc.) in 472 secreted and membrane genes knockout mice. When they plotted the ratio of knockout mice’s values to wildtype’s ones as a Gaussian curve, a few ratios were at the extreme ends. I noticed that sclerostin (Sost) knockout mice showed the most extreme value compared to all other knockout mice lines in “volumetric bone mineral density” (Fig.2H in the article). I don’t know whether Genentech pursued further research in this area.
    A mouse knockout library for secreted and transmembrane proteins – Nature Biotechnology, 2010 (PMID: 20562862)

  6. steve says:

    Off topic question. Why can’t I open Forbes articles (like Matthew Herper) in Safari on my Mac when they open perfectly well with Chrome? This is the only website where I have this problem.

  7. Andy II says:

    Another naive question. I remember that human body will generate antibodies against these therapeutic antibodies, Therefore these therapeutic antibodies (or protein-based drugs) have certain elimination half-lives. So, my question is: would the antibodies against those therapeutic antibodies would cause any unexpected side effects?

    1. Sarah says:

      Anti-drug antibody (ADA) response is a complex biological process that’s dependent on the delivered antibody drug as well as the target. A typical human or humanized antibody therapy will generally have a long half life – something in the order of few weeks to month scale – and show minimal ADA. Target and mechanism of action impact clearance rates and more importantly your ADA profile. A typical IgG-based therapy can escape degradation by FcRn. Protein therapies and some bispecific molecules show much shorter half life – in the order of hours. ADA can at best render the therapy ineffective.

    2. Druid says:

      Anti-drug antibodies against antibody drugs have the effect of increasing the clearance substantially (shortening the half-life) so that effect is lost over most of the normal dose interval (and continued dosing amplifies the ADA) so that the disease symptoms return. The situation for protein drugs which are not antibodies but analogues of endogenous proteins is more complicated because sometimes epitope spreading leads to neutralization of the endogenous protein as well as the drug, so that the disease symptoms can return worse than before treatment irrevocably.

  8. Chris Phoenix says:

    Anon wrote, “The much larger FRAME study showed no statistically significant increase in heart risk.”

    Basic probability says that, if you study 20 random variables, one of them on average will show significance at the p<0.05 level – regardless of population size. Given the number of bad results that could disqualify a drug, times the number of trial stages, it must happen fairly often that a phantom toxicity shows up by sheer random chance.

    When the toxicity is something apparently serious, like an increase in heart attack rate, it would be awfully difficult to justify running another trial to see if the effect goes away. But without that, how can a drug's record be cleared of spurious adverse health results?

    Note carefully: I'm not saying "What if it's toxic, but maybe should be accepted anyway?" I'm saying "What if it's _not at all_ toxic, but appears toxic (in one of 20 dimensions) due to predictable bad luck?"

    What statistical, policy, and/or research tools are in place to deal with this problem?

    1. Sophist says:

      Now is the point in drug discovery and development where economics can eclipse statistical methods in the decision making calculus. The question of whether the compound is “toxic at all” is one that must be answered to advance this molecule. Addressing the specter of potential cardiac risk to the satisfaction of regulatory agencies will likely require significant investment. The question likely is not “can it be done”, rather, “is it worth the investment to do it”?

      1. Eric says:

        You are right, but I’d qualify your comment slightly and say that economics trumps statistical methods at every stage of development. Throughout the industry early discovery research in osteoporosis has been dead for nearly a decade because the cost of these late stage trials is so prohibitive. There’s no reason to do the early, cheap preclinical work if you aren’t willing to pay for phase 3 later on.

Leave a Reply

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

Time limit is exhausted. Please reload CAPTCHA.