This post is fifth in a series of posts on the science of antisense oligonucleotides as therapeutics.

So far in this series I’ve introduced ASOs, their mechanisms of action, chemistries, and pharmacokinetics. Today, I’ll dig into safety or, conversely, toxicity: what we know about when and how ASOs can trigger safety concerns, and how any toxicities can be mitigated.

overview

I will first introduce a few terms and concepts useful for thinking about ASO toxicity and safety.

ASOs are nucleic acid therapeutics that are designed to sequence-specifically modulate a particular RNA target corresponding to one gene in the human genome. Therefore from first principles, we could divide their potential toxic effects into three categories:

  1. On-target: problems caused by engagement of the target RNA itself.
  2. Off-target, RNA-dependent: problems caused by hybridization to other RNAs due to their sequence similarity with the ASO.
  3. Off-target, RNA-independent: problems caused by ASOs independent of binding the target RNA or any other RNA — for instance, mediated by ASO protein binding.

It’s not always possible to pinpoint which category a given toxicity falls into, but this framework can be useful for thinking about and discussing toxicology of oligonucleotide medicines. For instance, to me, any on-target toxicities observed for other ASO drugs are the least interesting category, because they are specific to that target and would not be expected to be issues for an ASO against my target of interest.

One can also think of effects that are specific to one exact ASO (its chemistry, its sequence, etc.) or general to many ASOs (say, most or all of those with a particular chemical modification). Effects that are shared by many distinct molecules of a similar chemistry or similar pharmacology are sometimes called class effects, although I have never seen a rigorous definition of what exactly this term means. Depending on the context and who is speaking, a “class” might include all oligonucleotide drugs, all phosphorothiated ASOs, all 5-10-5 MOE gapmer ASOs, or all ASOs against a particular target, and from Googling it there seems no general definition.

A useful, albeit imperfect, framework in which to think about ASO medicinal chemistry is the dichotomy of the pharmacophore versus the dianophore [Khvorova & Watts 2017]. In this proposed framework, a compound’s pharmacophore is the set of molecular properties that determine how it engages its target (pharmacodynamics), whereas its dianophore is the set of properties that determine how it is processed by the body, meaning its absorption, distribution, metabolism, and excretion and so on (pharmacokinetics). The term “pharmacophore” term is commonly used throughout the medicinal chemistry world, including for small molecules, but for small molecules, any atom you change affects everything, so pharmacodynamics and pharmacokinetics are perfectly confounded. For oligonucleotide therapeutics, there is a partial separation, where pharmacodynamics are determined more by nucleotide sequence, and pharmacokinetics are determined more by the backbone and sugar modifications. At first blush, you might imagine that on-target, and off-target RNA-depdendent toxicities, are more sequence- or pharmacophore-related, while off-target RNA-independent toxicities are more backbone- or dianophore-related. There is probably some truth to this, but I have been surprised to learn just how much sequence dependence exists for ASO protein binding, as well as how much ASOs with the same sequence can have different properties depending on their backbone modifications.

Another axis on which to categorize toxicities is acute versus chronic, where “chronic” may not necessarily mean repeated dosing, but can simply refer to issues that arise over a longer period of time. For SOD1 ASOs, for example [McCampbell 2018], if one looks at the patent [WO2015153800A2], one can see that ASOs were screened in rats with a functional observation battery (FOB) score taken after just 3 hours (acute), and also in mice with body weights monitored for 8 weeks and neuroinflammatory markers assessed thereafter (chronic). These different time scales may allow for different types of adverse effects to arise, or not arise.

Finally, we all know the truism that “the dose makes the poison” and every drug has its toxic dose, so what do we mean by toxicity? The salient question for drug development is: what adverse effect will ultimately place a ceiling on how high of doses can be given? In other words, what are the drug’s dose-limiting toxicities (DLTs)? If a drug exhibits a DLT at a dose of X mg, then it doesn’t matter that much what other issues it might have had at yet higher doses.

While all the above terms and concepts are useful, what I’ve found so far is that the literature on drug safety and toxicity generally is much more prosaic than systematic. If a drug candidate has no toxic effects observed, then great, but if it does have toxic effects, one might work to understand them better, or, one might simply move on in search of a better drug. The “big picture” can often seem incomplete. Therefore, I will structure this post around several separate safety/tox stories I found interesting from the literature, before circling back to as much of a big picture as I can make out.

intrathecally delivered ASOs

My main interest is in the use of ASOs for brain diseases, which relies on bolus intrathecal (lumbar puncture) dosing into cerebrospinal fluid. This dosing route is relatively new as of the last five or so years, which means the data are still limited. So far, bolus intrathecal ASOs appear well-tolerated. Clinical trial results have been published for two bolus intrathecal ASOs: nusinersen for spinal muscular atrophy [Chiriboga 2016, Finkel 2016] and RG6042 for Huntington disease [Tabrizi 2019]. In both cases, human studies began with multiple ascending dose trials, in which no treatment-related serious adverse events were identified. Overall, to date, no DLTs for intrathecally dosed ASOs have been identified. Several other intrathecal ASOs are now in the clinic, in Phase I trials or beyond, and though nothing has been published yet, I assume from the fact that the trials seem to be continuing to advance, that the ASOs are proving well-tolerated.

The toxicities of intrathecal ASOs, if any do emerge with time, might well be different from the toxicities observed for peripherally delivered ASOs. Still, because most ASO drug development over the past few decades has focused on peripheral delivery, and particularly on the liver, that’s where we have the most data. The rest of this post will talk about findings from peripherally delivered ASOs, mostly in the liver, but it’s worth bearing in mind that we don’t yet know how relevant or irrelevant any of these findings may be for the brain.

dose-limiting toxicities of peripherally delivered ASOs across species

Whereas intrathecal bolus dosing is relatively new, most of the data on ASOs over the past few decades comes from ASOs dosed into the periphery, via intravenous or subcutaneous injection. An interesting finding from peripherally dosed ASOs has been that the types of DLTs generally observed can differ a lot between species [Dean & Bennett 2003].

In mice and rats, ASOs can have profound pro-inflammatory effects, and the DLT is usually immune stimulation [Dean & Bennett 2003]. This effect is apparently not specific to any particular chemistry of oligonucleotide, though it has some sequence dependence: CpG dinucleotides are known to be particularly pro-inflammatory, as they seem to trigger some sort of innate immune response that evolved to detect bacterial and viral DNA [Krieg 1995, Rankin 2001]. This response may exist in humans and other primates too, but perhaps it is less sensitive, because it has apparently never been dose-limiting.

In monkeys, a different type of inflammatory response, involving C5 complement activation, is often observed [Galbraith 1994]. This issue seems to be managed effectively by using subcutaneous rather than intravenous dosing to limit peak plasma concentration [Henry 1997a]. Complement activation has not been observed in human trials, and there is some in vitro evidence that it may be specific to monkeys and not humans [Henry 2014].

In humans, immune response and complement activation have never been dose-limiting. Instead, for some peripherally delivered ASOs, inhibition of blood clotting has proven to be a DLT. Effects of ASOs on clotting have been recognized for at least two decades: the earliest references I could find were for ISIS 2302, a former drug candidate for inflammatory bowel disease, with the same effect observed in monkeys [Henry 1997b] and in humans [Yacyshyn 1998]. The effect can manifest as an increase in activated partial thromboplastin time (aPTT), meaning a prolonged lag in an early phase of the clotting cascade, and a molecular basis for this has been mapped out in vitro [Sheehan & Phan 2001]. In some cases, ASO treatment has led to thrombocytopenia, meaning a low platelet count in blood. Thrombocytopenia was observed in a subset of participants in recent human trials of inotersen, an ASO against TTR for transthyretin amyloidosis, and volanesorsen, an ASO against APOC3 for patients with high triglyceride levels due to LPL mutation [Benson 2018, Witztum 2019]. FDA eventually approved inotersen, albeit with a black box warning. FDA has not approved volanesorsen, at least so far, and its committee minutes suggest that thrombocytopenia may be the agency’s major concern. Volanesorsen is approved in Europe, with a warning about thromocytopenia. Although it’s clear that inhibition of clotting is not limited to just one single ASO drug, it does not appear to be a class effect common to all ASOs. A review of clinical trial data from 2,638 patients found that only 7 individuals suffered severely low platelet counts, and only 3 of 16 ASOs tested had at least 10% of patients experience at least a 30% platelet count decrease [Crooke 2017a]. 15/16 of these ASOs had identical backbone chemistry (5-10-5 MOE gapmer), so the thrombocytopenia effect appears at least somewhat dependent on ASO sequence, although as far as I could tell, the exact mechanism is not yet known.

toxicity due to off-target binding of long RNA transcripts in the liver

In my ASO chemistry post, I introduced the concept of conformationally constrained ASOs, such as locked nucleic acids (LNA), ethylene-bridged nucleic acids (ENA), and 2’-4’ constrained ethyl (cEt). All of these are locked in the C3-endo conformation required for DNA-RNA binding, which makes their affinity for the target RNA, or the melting temperature of the RNA-DNA duplex, skyrocket. Unfortunately, many, though not all, such conformationally constrained ASOs seem to come with significant liver toxicity liabilities [Swayze 2007]. A study using RNA-seq to look at the effects of toxic and non-toxic LNA ASOs on the mouse liver transcriptome suggests that this is no coincidence, and that some of the toxicity may arise from high-affinity off-target binding of other RNAs [Burel 2016]. More recent evidence suggests this RNA-dependent mechanism may not be the full story (see next section), but it is still interesting enough to be worth diving into for a moment.

The RNA-seq study found, in short, that toxic LNA ASOs are associated with changes in RNA levels for a large number of transcripts, particularly for very long ones [Burel 2016]. Less toxic or non-toxic LNA ASOs did not have this effect. The interpretation is, it appears that LNA ASOs are so good at binding RNA, that they are even able to tolerate one or two mismatches, and so still bind with pretty strong affinity to off-target RNAs. Very long RNA transcripts are more likely than short ones to contain a site with one mismatch, or multiple sites with >1 mismatch, and so, even though the ASO’s affinity for these sequences is far weaker than its affinity for its intended, perfectly matched, target, there is still enough off-target activity in the aggregate to cause a problem — a figurative death by a thousand cuts. Reducing the level of RNAse H1 — the enzyme reponsible for cleavage of ASO-RNA duplexes — by means of siRNA, or conditional knockout in the mouse liver, seems to mitigate the toxicity of LNA ASOs, consistent with this proposed mechanism [Burel 2016, Kasuya 2016, Dieckmann 2018].

Not all LNA or cEt oligos have this problem, and multiple groups have argued it is possible through cell culture or animal testing to identify LNA oligos that are well-tolerated in the liver [Burel 2016, Dieckmann 2018]. I had to stop for a moment to wonder how this could be true. For while one can make an effort to select preclinical animal models where the on-target sequence is shared with the human drug candidate (for example, a transgenic mouse with the human gene), the list of potential off-targets is likely to be completely different in a mouse or maybe even in a monkey, compared to a human. So if off-targets are what mediate the tox, then how can animal studies be predictive of human safety profiles? One argument is that the amount of off-target cleavage is determined not so much by the available off-target transcriptome, but by the properties of the ASO itself, because RNAse H1 has differential tolerance to mismatches depending upon the exact enzymatic cut site. This difference in mismatch tolerance has been observed in mechanistic studies of RNAse H1 [Lima 2007a, Lima 2007b] and in the development of allele-specific ASOs for HTT [Ostergaard 2013]. So, if you find an ASO with good tolerability in mice, that’s probably because you’ve identified an ASO that is relatively mismatch-intolerant, which will serve it well even when it is dosed into a species with a totally different transcriptome. This explanation is not entirely satisfying, though, as will be addressed in the next section. Alternatively, another group has argued that the goal should simply be to maintain the ASO’s affinity for its intended target in some reasonable range, say, ≤55°C melting temperature (as opposed to the >90°C that is achievable with LNA [Wahlestedt 2000]), so that mismatched targets have a weak enough affinity to minimize off-target cleavage.

This naturally leads to the question of whether bioinformatic prediction could be enough to identify, and minimize, off-target activity in humans. A working group tasked with this question has recommended some guidelines [Lindow 2012], and unfortunately, it appears that only a minority of RNAs that are predicted off-targets of LNA ASOs actually observed to be reduced by those ASOs. That makes sense, because while we can identify all the potential 1- or 2-mismatch sites for an ASO by mining the whole transcriptome, that doesn’t tell us anything about which sites are actually accessible in the cell — many will be occluded by RNA secondary structure. Overall, it does not yet appear possible to predict bioinformatically which transcripts are reduced and which are not, so for now, empirical preclinical safety testing is considered the only solution to minimizing off-target RNA-dependent activity.

toxicities due to protein binding

Although the use of RNA-seq to understand off-target, RNA-dependent toxicity is clever and fascinating, a few gaps made people suspicious that this couldn’t possibly be the full story. As explained eloquently in Stanley Crooke’s blog, toxic ASOs tend to be toxic in a variety of cell models and species, even though their potential off-target transcriptomes are totally different, and rarely has a specific off-target event been identified as critical for mediating toxicity. While one can posit explanations for these phenomena (some of which are explained in the previous section) none are entirely satisfying.

ASOs are also intensely protein-binding — some chemistries moreso than others — and they rely on protein binding to enter cells and reach an active compartment [Crooke 2017b]. Thus, toxic ASOs could well also exert their effects through protein binding. For example, one study a few years ago found that LNA ASOs with particularly bad hepatotoxicity were also more intensely protein-binding — they also pulled down a wider variety of proteins, and more of each protein, from liver homogenate [Burdick 2014]. That effect exhibited some sequence dependence, with TCC and TGC trinucleotides more often being found in bad ASOs [Burdick 2014]. A bit of sequence dependence is not inconsistent with the proposed protein-based mechanism. For example, one in vitro study showed that several ASOs with the same chemistry but different sequence bound a particular protein (P54nrb) with Kd values varying more than two orders of magnitude (from 0.51 to 191 nM) [Vickers & Crooke 2016].

A recent study provides evidence that a lot of ASO toxicity in the liver is driven by protein binding, specifically by ASOs binding proteins that should be in paraspeckles and causing them to move elsewhere [Shen 2019]. Paraspeckles are regulatory ribonucleoprotein particles, 0.2 - 1 μm in diameter, found in the nucleus. Through a series of immunocytochemistry experiments — staining different compartments of cells — the authors show that many toxic ASOs (but not non-toxic ASOs) cause important proteins (including P54nrb, mentioned above) to relocate out of their usual homes in the paraspeckle, leading to apoptosis. Their exact fate depends on ASO chemistry. 2’-fluoro ASOs tended to make paraspeckle proteins migrate into the cytoplasm and get degraded by the proteasome, and this occurred regardless of whether RNAse H1 was present. In contrast, cEt and MOE ASOs tended to move paraspeckle proteins into nucleoli, but only in the presence of RNAse H1. Regardless, the effect was driven by protein binding, and because different chemistries, particularly different 2’ sugar modifications, can greatly control ASO protein binding, the authors explored how they could tweak these 2’ modifications to mitigate the toxicity. They found that adding one 2’-O-methyl (OMe) sugar ring at the 2nd position of the “gap” (the normally non-2’-modified sequence at the center of the oligo that allows RNAse H1 cleavage) dramatically reduced paraspeckle protein binding and mitigated the toxicity in cell culture and in vivo in the liver [Shen 2019]. We still do not know whether, or to what extent, this mechanism is applicable in the brain, but at a higher level, it’s an interesting story about how a specific mechanism mediated by protein binding gives rise to ASO toxicity (still with some sequence dependence), and how this can be mitigated through clever medicinal chemistry efforts.

the big picture

If there is a big picture for ASO safety, it’s that, like any other class of drug, designing potent and tolerable compounds is a non-trivial, iterative process. Years of research have identified some specific liabilities that are, if not universal “class effects”, are at least common to more than one oligo. Medicinal chemsitry offers ways to mitigate some of these liabilities, while others may simply need to be avoided by careful tolerability screening in preclinical models. At the end of the day, any new drug will always need to undergo initial clinical testing starting at low doses in small numbers of closely monitored people.

When I first started learning about ASOs years ago, I mistakenly imagined that the “pharmacophore” and “dianophore” properties of the drugs were completely separated. I therefore wondered why new ASO drugs that were comprised of novel sequences, but with tried-and-true backbone chemistries, should have to go through all the same preclinical and clinical safety testing that is expected for completely novel small molecules. I now understand that the pharmacophore/dianophore separation is only partial, and that while even this incomplete separation greatly accelerates the pace at which new drugs can be developed, the separation is still not enough to abrogate a great deal of empirical testing to achieve safety as well as desirable pharmacokinetic properties. As we’ve seen in this post, ASOs with identical backbone chemistry and differing only in sequence can have very different tolerability profiles, whether mediated by protein binding, off-target RNA hybridization, or other processes we don’t understand yet.

Almost all of the research cited here comes from animal and human studies of ASOs dosed peripherally (mostly subcutaneously) for the liver. ASOs for the brain and spinal cord are a relatively new beast, and it remains to be seen which lessons from the liver will apply. Promisingly, no dose-limiting toxicities have yet been identified for any intrathecal ASOs, and the number of patients on drug is no longer tiny — nusinersen, the approved ASO drug for spinal muscular atrophy, is now being taken by over 7,500 patients worldwide as of July 2019. I am eagerly following the news on other ASOs now in clinical trials to see what, if any, safety issues are identified. In the meantime, the data so far suggest that even if identifying well-tolerated ASOs is not trivial, it is achievable.