Skip to Content

Drug Development

Predicting Toxicology

One of the major worries during a clinical trial is toxicity, naturally. There are thousands of reasons a compound might cause problem, and you can be sure that we don’t have a good handle on most of them. We screen for what we know about (such as hERG channels for cardiovascular trouble), and we watch closely for signs of everything else. But when slow-building low-incidence toxicity takes your compound out late in the clinic, it’s always very painful indeed.
Anything that helps to clarify that part of the business is big news, and potentially worth a lot. But advanced in clinical toxicology come on very slowly, because the only thing worse than not knowing what you’ll find is thinking that you know and being wrong. A new paper in Nature highlights just this problem. The authors have a structural-similarity algorithm to try to test new compounds against known toxicities in previously tested compounds, which (as you can imagine) is an approach that’s been tried in many different forms over the years. So how does this one fare?

To test their computational approach, Lounkine et al. used it to estimate the binding affinities of a comprehensive set of 656 approved drugs for 73 biological targets. They identified 1,644 possible drug–target interactions, of which 403 were already recorded in ChEMBL, a publicly available database of biologically active compounds. However, because the authors had used this database as a training set for their model, these predictions were not really indicative of the model’s effectiveness, and so were not considered further.
A further 348 of the remaining 1,241 predictions were found in other databases (which the authors hadn’t used as training sets), leaving 893 predictions, 694 of which were then tested experimentally. The authors found that 151 of these predicted drug–target interactions were genuine. So, of the 1,241 predictions not in ChEMBL, 499 were true; 543 were false; and 199 remain to be tested. Many of the newly discovered drug–target interactions would not have been predicted using conventional computational methods that calculate the strength of drug–target binding interactions based on the structures of the ligand and of the target’s binding site.

Now, some of their predictions have turned out to be surprising and accurate. Their technique identified chlorotrianisene, for example, as a COX-1 inhibitor, and tests show that it seems to be, which wasn’t known at all. The classic antihistamine diphenhydramine turns out to be active at the serotonin transporter. It’s also interesting to see what known drugs light up the side effect assays the worst. Looking at their figures, it would seem that the topical antiseptic chlorhexidine (a membrane disruptor) is active all over the place. Another guanidine-containing compound, tegaserod, is also high up the list. Other promiscuous compounds are the old antipsychotic fluspirilene and the semisynthetic antibiotic rifaximin. (That last one illustrates one of the problems with this approach, which the authors take care to point out: toxicity depends on exposure. The dose makes the poison, and all that. Rifaximin is very poorly absorbed, and it would take very unusual dosing, like with a power drill, to get it to hit targets in a place like the central nervous system, even if this technique flags them).
The biggest problem with this whole approach is also highlighted by the authors, to their credit. You can see from those figures above that about half of the potentially toxic interactions it finds aren’t real, and you can be sure that there are plenty of false negatives, too. So this is nowhere near ready to replace real-world testing; nothing is. But where it could be useful is in pointing out things to test with real-world assays, activities that you probably hadn’t considered at all.
But the downside of that is that you could end up chasing meaningless stuff that does nothing but put the fear into you and delays your compound’s development, too. That split, “stupid delay versus crucial red flag”, is at the heart of clinical toxicology, and is the reason it’s so hard to make solid progress in this area. So much is riding on these decisions: you could walk away from a compound, never developing one that would go on to clear billions of dollars and help untold numbers of patients. Or you could green-light something that would go on to chew up hundreds of millions of dollars of development costs (and even more in opportunity costs, considering what you could have been working on instead), or even worse, one that makes it onto the market and has to be withdrawn in a blizzard of lawsuits. It brings on a cautious attitude.

21 comments on “Predicting Toxicology”

  1. “But the downside of that is that you could end up chasing meaningless stuff that does nothing but put the fear into you and delays your compound’s development, too.”
    Exactly. Hopefully the FDA won’t make this their weapon of choice, and I can see them potentially creating all kinds of trouble (some justified and some not so much) for drug makers, asking them if they tested their drug against this or that target.

  2. Rick Wobbe says:

    It’s one thing to interpolate between points in well-studied chemical space, where you can beat random guessing with practice. But predicting behavior outside that space will tell you more about the limitations of your model than the properties of the compound. That’s the Turing test of these models. Thus far, the models seem more like economic theories, which, to quote the late, great Paul Sammuelson, “predicted 9 of the last 5 recessions”. Brings back fond memories of the days when you could amaze and bemuse people by using dice to “predict” the PK or tox properties of compounds as well as many in silico models. If, as Curious Wavefunction worries, the FDA decided to make this a weapon of choice, I’m going to dust off my magic dice!

  3. OldLabRat says:

    The EPA announced this week that it will start doing the necessary research to implement computational toxicology. The press release says all the right things, but I’m not optimistic that the science will win. I’m sure the FDA will be watching closely.

  4. watcher says:

    Another company that has spent a lot of time on this topic has a similar predictive set of activities. When compounds look good, the go ahead—careful green light.
    When compounds look questionable, they typically go ahead as the predictions with test sets are not reliable enough to give teams enough confidence for definitive decision making after so much time & resource investment, particularly for new compound classes not included in any training sets….cautious green light.
    And so, little changes after big effort, time, and expenditure to take-on predicitive tox.

  5. anon says:

    I thought it was already known that diphenhydramine has some activity at the serotonin transporter. This from wiki “In the 1960s, diphenhydramine was found to inhibit reuptake of the neurotransmitter serotonin.[30] This discovery led to a search for viable antidepressants with similar structures and fewer side-effects, culminating in the invention of fluoxetine (Prozac), a selective serotonin reuptake inhibitor (SSRI).[30][31] A similar search had previously led to the synthesis of the first SSRI, zimelidine, from brompheniramine, also an antihistamine.”

  6. Pete says:

    One needs to be wary of claims of polypharmacology based on IC50 30% inhibition at 10μM). Not sure why they don’t use something a bit closer to likely physiological free levels. Tyrosine kinase inhibition assays tend to be run at different ATP concentrations but ATP-competitive inhibitors all need to deal with the same intracellular ATP concentration.

  7. Anonymous says:

    My comment (#6) got mangled so I’m attempting to re-post. Apologies if the problem is at my end.
    One needs to be wary of claims of polypharmacology based on IC50 30% inhibition at 10μM. Not sure why they don’t use something a bit closer to likely physiological free levels. Tyrosine kinase inhibition assays tend to be run at different ATP concentrations but ATP-competitive inhibitors all need to deal with the same intracellular ATP concentration.

  8. weirdo says:

    Given the number of papers on this very topic over the last few years, I’m very surprised to see this in Nature.
    I also question the premise a bit — the authors’ own introductory paragraph points out that metabolites are often the problem. We do a lot of cross-reactivity testing on the API in the early stages of lead optimization; usually much less so on metabolites. A far more interesting paper would have dealt with in silico work on known metabolites, and tracking those down.

  9. Pete says:

    Comment is still getting mangled so I’ve written micromolar explicitly in case Greek characters were causing the problem. Once again, apologies if the problem is at my end.
    One needs to be wary of claims of polypharmacology based on IC50 30% inhibition at 10 micromolar.
    Not sure why they don’t use something a bit closer to likely physiological free levels.
    Tyrosine kinase inhibition assays tend to be run at different ATP concentrations but ATP-competitive inhibitors all need to deal with the same intracellular ATP concentration.

  10. Anonymous says:

    Comment is still getting mangled so I’ve written micromolar explicitly in case Greek characters were causing the problem. Once again, apologies if the problem is at my end.
    One needs to be wary of claims of polypharmacology based on IC50 less than 30 micromolar.
    One frequently-cited article even uses greater than 30% inhibition at 10 micromolar.
    Not sure why they don’t use something a bit closer to likely physiological free levels.
    Tyrosine kinase inhibition assays tend to be run at different ATP concentrations but ATP-competitive inhibitors all need to deal with the same intracellular ATP concentration.

  11. MolecularGeek says:

    The horse has already left the barn on the use of predictive models for regulatory activities. REACH in the EU is predicated on using QSAR and related technologies to triage industrial compounds for laboratory risk assessment. They have, however, also promulgated a set of best practices in model development that includes guidance on issues like interpolation vs. extrapolation and applicability domains.

  12. FDA lurker says:

    @ #1:
    As a FDA employee, I personally believe that industry kills too many potential drugs too quickly based on early preclinical results (even more potential drugs are probably killed before they get through the door). Often before human ADME is investigated. There are lots of reasons for this, and most are not scientific, IMO.

  13. partial agonist says:

    Predictions based on two-dimensional structural similarity are always going to have huge error bars, given all of the examples where two enantiomers or two diastereomers have vastly different toxicity profiles.
    This is a favorite area where the PETA people love to say that animal testing is not needed. Unfortunately, the animal toxicity (and then the human toxicity) is not something a computer is going to tell you much about, with high confidence. That is unless we get one of those computers from Star Trek or from the cheesy Batman TV show.

  14. barry says:

    ” That split, ‘stupid delay versus crucial red flag’, is at the heart of clinical toxicology”
    That split of course is at the heart of the current $billion price for a new drug. Everyone wants a tool that will kill a loser earlier but no one wants his project killed. So we build and fund the new tools and the new departments, but we don’t trust them enough to kill the projects early.

  15. DCRogers says:

    It’s beside the point if the predictions are good: let’s even say they’re perfect. How would you use them? From the paper:
    “[T]he 656 drugs considered here each modulated an average of seven safety targets, sometimes across several classes, and more than 10% of the drugs acted on nearly half (45%) of the 73 targets.”
    So these aren’t even close to filters commonly used to detect reactive substructures, where even one is a killer. Here, the “red flags” are so common as to give a boy-who-cried-worlf numbness.

  16. dvizard says:

    “REACH in the EU is predicated on using QSAR and related technologies to triage industrial compounds for laboratory risk assessment.”
    But ecotox is a whole different story in terms of regulatory impact than pharmacological toxicology is. Damaged ecosystems don’t sue you, unlike patients might.

  17. Morten G says:

    It is a very good paper though and the chlorotrianisene story is quite compelling. Recommended read.
    From here of course they need to show whether they could retrospectively separate compounds that failed clinical trials because of adverse events from those that didn’t.
    Two bits of this paper that Derek didn’t mention which are especially interesting to the medicinal chemists:
    1. In fig 3 31% of the tested _approved_ drugs ding hERG in vitro. So ask yourself how many adverse events have been put down to hERG inhibition simply because it is a very promiscuous target and how many compounds were unjustly killed because they dinged the hERG assay.
    2. SEA could be used to jump scaffolds. Take a library of a couple of millions compounds that you like (in theory at least – you probably haven’t met them yet) and mash it against the current binders of your target. Buy/synthesize hits and take it to the biologists.

  18. Anonymous says:

    they already published a similar paper in Nature:
    http://www.bkslab.org/publications/keiser_2009.pdf
    When did Nature start to come out so easiy? quite surprising

  19. Anonymous says:

    Most of the newly identified targets are only new for those that doesn’t have access to a database like BioPrint
    I checked the results in the Nature 2009 paper with the BioPrint database and could confirm what they found was correct, however, the compounds investigated were also active on several other targets in the database that were apparently not picked up by their method
    Still an interesting paper, though.

  20. Anonymous says:

    Most of the newly identified targets are only new for those that doesn’t have access to a database like BioPrint
    I checked the results in the Nature 2009 paper with the BioPrint database and could confirm what they found was correct, however, the compounds investigated were also active on several other targets in the database that were apparently not picked up by their method
    Still an interesting paper, though.

  21. Jonadab says:

    > diphenhydramine turns out to be active at the serotonin transporter
    That might go a long way toward explaining what happens when somebody takes an entire box of Benadryl. Isn’t serotonin a neurotransmitter? [Checks.] Why yes, yes it is. That could definitely be relevant.

Comments are closed.