figure a

Xueyi Yang. Image composite by Erin Dewalt

In January, Atomic AI, with a $35 million A round, joined a short but growing list of companies seeking to drug RNA using small molecules. As the name suggests, the company is applying artificial intelligence to predicting RNA structures. Venture capitalists are not alone in their interest in the RNA-targeted small-molecule approach. Pharma deals have been picking up: Merck & Co., Vertex Pharmaceuticals, Sanofi, Amgen, Celgene, Genentech and Janssen Pharmaceuticals have all recently partnered with biotechs in this space (Table 1). Peter Smith, co-founder and CEO of Remix Therapeutics, which also enjoyed some funding in the last year, calls RNA “the next frontier of small-molecule drug discovery.”

Table 1 Selected companies targeting RNA with small molecules
Full size table

Driving interest is the knowledge that drugging RNAs could, in theory, expand the target landscape by almost two orders of magnitude. Less than 3% of the genome encodes protein, yet 80% of the rest is transcribed and regulated as noncoding RNA, most of it of unknown function. And “roughly 85–90% of the protein targets are largely undruggable,” says Arrakis Therapeutics CEO Michael Gilman. “We want to be able to essentially target any RNA.”

But so far, that goal remains elusive. For decades, drug companies considered RNA too polar and solvent-exposed to accept small molecules, believing compounds would not have enough binding energy to overcome the water barrier and reach the RNA surface. It’s now clear that RNA contains structures that small molecules can and do bind1. But only a few RNA-binding compounds with clear function have been reported. And few, if any, compelling examples exist of investigators taking a compound attaching specifically to a novel, complex target mRNA motif (not already known to bind small molecules) from discovery to disease models. (Evrysdi (risdiplam), approved for spinal muscular atrophy (SMA), targets an RNA–protein complex — a specialized strategy.)

“Are there any publicly available examples where somebody has gone all the way?” asks University of North Carolina chemical biologist Kevin Weeks. “That really hasn’t been done convincingly.” But, he adds, “We don’t have full information.” Innovations in this young field, including discovery platforms, lead compounds and even lead disease indications, are tightly guarded. Companies, holding their cards close, are making very different bets, and it’s far from clear which ones will win out.

These bets reflect several deep divides in the field: whether to target pre-mRNA in the nucleus or mature mRNA in the cytoplasm, whether to conduct phenotypic screening to identify functionally active molecules or instead screen for binders against known targets, and whether to prioritize molecules that bind to certain high-complexity RNA structures or instead to focus on simpler RNAs with more therapeutic potential. Artificial intelligence (AI) and targeted degraders promise disruption but have yet to deliver.

Processing and poisoning

In theory, any stage of the RNA life cycle could be targeted with small molecules (Fig. 1). Many companies are targeting pre-mRNAs. Following the success of PTC Therapeutics’ and Roche’s Evrysdi, RNA processing is now wide open for drug intervention. In eukaryotes, gene transcription generates a pre-mRNA containing introns, which the cell’s splicing machinery removes when ligating exons into mature mRNA. This transcript, after further modification, is exported to the cytoplasm for translation into protein. Evrysdi binds to a pre-mRNA–snRNP (small nuclear ribonucleoprotein) complex to drive the splicing and expression of a missing exon in the SMN2 gene2. In SMA, the SMN1 gene is mutated or missing, and the SMN2 gene is nonfunctional. By splicing an exon into SMN2, Evrysdi enables expression of a functional SMN2 protein that compensates for the loss of SMN1. The drug’s mechanism “opens up a whole new area of therapeutic research on modulating gene expression by targeting splicing,” says PTC co-founder Stuart Peltz, who retired as CEO in March.

Fig. 1: Drugging the RNA life cycle.
figure 1

Source: Concept and graphics courtesy ReviR Therapeutics.

Full size image

Evrysdi “really shows it’s possible to do this,” says Smith. A major Remix focus, besides neurodegeneration, is cancer, where mutations in splicing factors affect roughly 20% of acute myeloid leukemias, 10–15% of chronic lymphocytic leukemias, and smaller percentages of several solid tumors. Remix is seeking RNA-binding small molecules that can correct or compensate for these splicing defects. Skyhawk Therapeutics, like Remix, is working on cancer and neurodegeneration. DNA and mRNA nucleotide repeat expansions are hallmarks of Huntington’s disease and a common form of amyotrophic lateral sclerosis (ALS), among other diseases. Skyhawk’s RNA-binding compounds work by “either reducing or increasing RNA levels, changing the functionality of the proteins,” says Sergey Paushkin, Skyhawk’s senior vice president of discovery biology. PTC, the splicing pioneer, has “ten or fifteen programs that are ongoing where we’re looking for selective molecules that modulate different aspects of splicing,” says Peltz.

This is not classic rational design — splicing is probably too complicated for that to be practical yet, says Paushkin. Both companies screen for compounds using cell-based assays to read out the splicing outcome they seek, then optimize hits for efficacy, potency and other qualities using structural biology tools to visualize the drug–RNA interaction. They then try to nail down the mechanism, determine specificity (using RNA sequencing to identify altered transcripts, and sometimes proteomics), and further optimize chemical leads for druglike properties, before going into animals for efficacy and safety studies.

The biology is complex, but not intractable. RNA splicing is performed by an RNA–protein complex called the spliceosome. It contains snRNPs that initiate the process when they bind to the pre-mRNA. PTC and Roche have reported that Evrysdi acts as a molecular glue that stabilizes the interaction between the 5′ splice site of the SMN2 mRNA’s exon 7 and the U1 snRNP, enhancing splicing. This general mechanism can, say Peltz and Paushkin, be exploited in mRNAs implicated in other diseases. That’s “the best of all worlds,” says Weeks. “You’re targeting a pretty well-defined complex in splicing, you are targeting a specific RNA. And then in addition splicing is the kind of molecular event that lends itself to functional assays.”

Another splicing strategy is to insert pseudoexons, or ‘poison exons’, into mRNA. These sequences, which encode premature stop codons, can be found in introns and, if somehow spliced into coding regions, engage nonsense-mediated decay pathways to degrade the mRNA. This story gained traction when Novartis reported that branaplam, an experimental small-molecule drug for treating SMA, induces a pseudoexon in HTT (huntingtin), the gene that, when mutated, causes Huntington’s disease3. This pseudoexon reduced levels of HTT in patient-derived cells. Novartis suspended its Huntington’s clinical last August because branaplam “might be causing peripheral neuropathy,” and has since discontinued the program. PTC is testing its own small-molecule pseudoexon inducer, more specific than branaplam, in a global phase 2 Huntington’s trial. (The FDA paused the US portion of this trial last October pending more data from PTC.) Several other companies are in the hunt for poison exon inducers. Some are also looking for small molecules capable of boosting, instead of reducing, expression of proteins in diseases where poison exons are already causing transcript degradation. “It could be either way,” says Paushkin. “You could induce or inhibit poison exon inclusion.”

The potential of small molecules to remove transcripts this way extends beyond the suite of rare diseases caused by aberrant splicing. “We’re looking for being able to do this in common diseases as well,” says Peltz.

Whereas rare diseases are a clear opportunity, there’s doubt about the ultimate efficacy of splicing drugs in common diseases with splicing mutations, like cancers. One hypothesis posits that such mutations are not individually driving the disease but that accumulated abnormal splice products create a state of ‘splicing sickness’. If that’s the case, targeting individual splicing defects may not be that effective. “It’s still an open question,” says Smith. “Are [splicing mutations] truly driving the disease, or are they contributing in a broad sense?” The fact that many mutations are ‘hot spot’ clustered mutations and occur early in the disease, he says, suggests that they’re driver events and that drugs that induce poison exons should be effective. “All of the signals to me point towards it being an important event that should still be targeted therapeutically,” he says.

Beyond splicing

Targeting mature mRNAs rather than pre-mRNA processing awaits clinical proof of concept. Few programs have even reached the disclosure stage. “We are exploring other things in addition to splicing, additional mechanisms, that would involve the mature RNAs,” says Paushkin, who declines to elaborate. Ribometrix is primarily targeting mature mRNAs, again unspecified. Arrakis has disclosed its lead program, which targets the mature mRNA encoding the myc transcription factor in cancer. Gilman says the compound, currently at the lead optimization stage, binds and stabilizes a structure in the 5′ UTR and probably interferes with ribosome scanning. It’s important to the company to go beyond splicing applications, Gilman says. “Given any RNA of therapeutic interest, we want to be able to find drug-like chemistry that hits that RNA,” he says. “Our ambition from the beginning was to bring the entire transcriptome into play.”

But we must prioritize carefully, says the UF (University of Florida) Scripps chemist Matt Disney. “Companies can articulate that they can drug every RNA in every disease, or every disease with any modality, or any target they want. But at the end of the day you’ve got to deliver,” he says. For the company Disney co-founded, Expansion Therapeutics, that means selecting indications where patient genetics identify an RNA target — for example, small-molecule splicing modulators that bind to the CUG repeat expansions driving myotonic dystrophy4. Expansion Therapeutics is also targeting the GGGGCC expansions in an inherited form of ALS and frontotemporal dementia (FTD)5 and mutant tau RNA in the disease FTDP-17 (FTD and Parkinsonism linked to chromosome 17)6. “The most urgent thing right now is to take on targets that have a biological de-risk, which means we know that that target is driving a disease,” says Disney. “Those [compounds] are the ones we should be front-loading to get into patients. And then, you know, we can expand after that.” Expansion is building on Disney’s genreral approach, writes CEO Renato Skerlj in an e-mail: “We use structural biology to inform drug design and develop orally bioavailable molecules that cross the blood–brain barrier for each of the programs.”

Overcoming bottlenecks

The second big divide in the field is over drug discovery strategy. The main choice is between phenotypic screening, which identifies functional compounds without knowing the target, and screening compounds for binding against known targets. Each has advantages and drawbacks. PTC’s early phenotypic screens yielded hits that evolved into Evrysdi, and the company continues to do them. In the original phenotypic screen, PTC made an SMN2 minigene fused to a luciferase reporter coding sequence that glowed only with full-length SMN2 exon 7 expression. It then transfected that minigene into a standard immortalized human kidney cell line and screened roughly 200,000 compounds, adding them to the transfected cells looking for signal. Hits were then confirmed by running them through PCR assays designed to detect truncated and full-length SMN2 exon 7. Peltz says PTC’s high-throughput screening platform can now test a library of potential small-molecule splicing modifiers against hundreds of different splice sites at the same time.

But phenotypic screens yield hits that don’t even bind RNA, plus other false positives. “Almost any screen you run is going to generate hundreds or thousands of hits,” says Gilman. “There may well be a needle in that haystack, but it’s going to take you a lot of work to find it.” The key to making phenotypic screening work efficiently, says Skyhawk’s Paushkin, is a specialized compound library. “If you just take a commercial library and do the screen, the chance of finding molecules like [Evrysdi] would be almost zero,” he says. Skyhawk, PTC and other companies are constantly enriching their libraries for likely binders. Companies now routinely use machine learning to iteratively sort through screening hits for common chemical features that increase the likelihood of binding RNA and RNA complexes. Then they add compounds with these features to the library and remove those that don’t have them.

But another bottleneck with phenotypic screening is figuring out what RNA the compound is hitting. “In order to do med chem on that stuff, you really have to have a meaty target deconvolution part of the platform,” says Disney. Phenotypic screening “is exciting, because you get directly functional molecules at the beginning, but you’ve got to find what the targets are.”

However, screening against a known target RNA has some big problems of its own — nonspecific binders, for one. “There’s plenty of chemotypes that bind to RNA nonspecifically,” says Gilman. He says Arrakis has “ruthlessly screened out” nonspecific binders, such as intercalators (which insert between base pairs) and compounds with more than one positive charge (which can bind indiscriminately to the strongly negatively charged RNA). But even many specific binders have no useful effect on the target, so they’re worthless as leads. “Most of these hits that we find are … legitimate binders; we can prove that with a variety of orthogonal methods,” says Gilman. “But more often than not these compounds don’t have activity. And in retrospect I think it’s easy to understand why: which is because these [target RNAs] are not enzymes. These pockets are not active sites.” Ruling out the biologically silent hits can require laborious cell-based functional assays.

Smith says that Remix has removed that bottleneck by using a high-throughput screening platform with a functional, instead of a strictly binding, readout, thus greatly speeding drug discovery. It’s target specific, not phenotypic screening, but Smith offers no details. Ribometrix CEO Mike Solomon says his company has largely solved the problem in the last year. “Now we can zero in pretty quickly on the real actionable chemical matter that comes out of the screen,” he says. But he also declines to elaborate.

Another bottleneck is figuring out how a compound that does bind the RNA and alters biology is actually working. The mechanism may not even be related to the binding event. Even if it is, “is it directly interfering with the translation machinery? Is it affecting the half-life of the mRNA? Have you turned on some decay pathway? Is it splicing?” asks Ribometrix co-founder and VP of biology Katie Warner. “The list goes on and on.” Standard anti-enzyme drug discovery doesn’t have this problem. “With an enzyme, you have a functional biochemical assay,” says Solomon. “With the RNA, you don’t have that. There isn’t an active site that you’re going after.”

“The fundamental problem at the end of the day is that we don’t have a playbook for structure–function relationships in RNA like we do in proteins,” says Gilman. “We know that RNA has structure, we know that RNA has function, but we don’t know a lot about how those things line up.” That makes it hard to build platforms that can compete on efficiency with the protein-targeting world. “That’s what we’re trying to shoot for here,” says Solomon. “Create the rules, the assays, the know-how, how to do drug discovery on RNA so that ultimately you’re doing it just like you would do it on an enzyme.” Companies (including Solomon’s) claim that they’ve done this or are close. “We’ve built the foundation,” Solomon says. “We can routinely screen RNAs and identify potent, selective binders. And now on to the next stage, creating drugs.”

The target debate

Then there’s target selection, probably the most contentious divide. Should companies go after structurally complex targets for their pockets? Or simpler targets with clear therapeutic impact? RNA is transcribed as a single chain that then folds up into a variety of secondary structures (helices). These in turn interact to form tertiary structures, including knot-shaped ‘pseudoknots’. When selecting an RNA structural element to target, “we think that complexity, local complexity, is important,” says Weeks. “So we’re fans of multi-helix junctions, like three- and four-helix junctions. Regions that form pseudoknots are likely to be good choices.” That’s because such complex motifs are apt to contain, or be adjacent to, pockets in the RNA. Also, Weeks says, the more complex the RNA tertiary structure, the more unique, and the more likely a binding compound will be selective.

“A lot of effort has been devoted towards basically poor targets,” argues Weeks: “clearly biologically functional targets, but ones that are never likely to bind small molecules with high selectivity.” In a deliberately provocative 2018 paper7, Weeks singled out helices with CAG and CUG repeat sequences (hallmarks of Huntington’s disease and myotonic dystrophy, respectively). He also noted that RNA motifs composed primarily of secondary structures, including microRNAs, have small or shallow pockets. “It’s possible that these kinds of motifs can be targeted,” Weeks wrote, “but doing so will probably prove challenging in much the same way as has targeting the shallow grooves that characterize many protein–protein interactions.”

“The field is moving towards accepting the view that was proposed in that review,” says Weeks. But that’s not clear. Expansion Therapeutics, for example, is targeting the CUG trinucleotide repeats in RNA that cause myotonic dystrophy. “We have a ton of data to show that you can target repeat expansions or oncogenic RNAs pretty selectively with small molecules, with high affinity and specificity,” says Disney. Weeks doesn’t rule out this approach. “If you could find a motif that recognized an individual element of the repeat, or a couple elements in a row, and then you would link those together, maybe cooperativity would help you,” he says, even in the absence of pockets.

Besides directly targeting repeats, Disney’s group has also targeted mRNA regulatory sequences8 to disrupt translation. Regardless, in his view, the field should prioritize genetically validated targets and not limit itself to certain complex RNA structural motifs. “All these repeat expansions have genetic basis,” he says. “I don’t think all RNAs are going to have a druggable structure; I think a subset are. … The bigger question becomes not this debate of what’s the perfect situation; I think we have to be practical, given the amount of money that’s gone into the area, and go after things that you can translate into a patient.”

Enter artificial intelligence

AI is beginning to inform this debate by predicting the best RNA tertiary structures for small molecules to bind to. Chemical methods that work at nucleotide-level resolution can only suggest or rule out potential binding sites. Nuclear magnetic resonance structural studies provide atomic-level information, but work best on small RNAs, which are unlikely to harbor pockets. And RNA X-ray crystal structures are few because RNA is unstable and less likely to form crystals than proteins. Inspired by the spectacular success of the AI-based protein structure predictors AlphaFold and RoseTTAFold, a group at Stanford, including PhD student Raphael Townshend, used machine learning to improve on existing RNA structure prediction tools. In May 2021 Townshend founded the startup Atomic AI to refine the model and use it in drug discovery.

Townshend’s 2021 Science paper9 was “a big leap forward, but it was a leap forward from a field that really had struggled,” says Weeks. “By no means does the published work look anything like what you need to really understand RNA tertiary structure.” New models now posted to the bioRxiv preprint website are impressive, Weeks says, but still don’t go far enough. In his hands, “essentially none of the groups could even get the pattern of base pairing consistently correct,” he says. “If they’re having a problem getting the base pairing pattern correct, there’ll probably be some problems getting the three-dimensional structure consistently right.”

Just replicating the AlphaFold method isn’t enough for RNA, says Townshend, because there are far fewer RNA than protein data available for the training and validation sets. He says Atomic AI, by feeding its own RNA data into the algorithms, has cut its model’s. RMSD (root mean square deviation, a measure of the similarity between known and modeled structures) from roughly 6.5 Å in the Science paper to 2 Å depending on application. “We’re still validating a lot of it as we go,” he says. “At least on the face of it, we are actually hitting the point where you can leverage these structures for docking [of small molecules].”

This would be very valuable because, to have something to screen against, companies must identify a bindable RNA structure. It then can be synthesized outside the cell for screening purposes. “We’ve talked to major pharmaceuticals that have said that that’s their critical need,” Townshend says. “They have a [target] transcript, and they don’t even know where to start.”

Most biotechs in this space, though, don’t use AI for this purpose, at least for now. (Atomic AI has not published its algorithms.) They do use it to verify other structural findings and to enrich their chemical libraries for RNA-binding chemical matter. “There’s a lot of hype in AI,” says Disney. “We’re going to have to really see what that can deliver on. But it’s promising.”

“We’re obviously keeping a close eye on the field,” says Warner. “If for some reason they were able to get us to some early actionable structural information faster than we can get to with our other methods, we want to be the first to know. But right now I don’t think anything can replace the experimental ‘get your data in the cell type you care about’ type of information.”

Atomic AI intends to go even further by building a map of the ‘pocketome’, an atlas of every single tertiary RNA structure across the transcriptome. Then “you really can target those that look the most promising, hit those first,” says Townshend, instead of just focusing on a single transcript. But first the company, like others, must validate its platform through successful drug development.

RNA degraders: potential and pitfalls

One theoretical way to bypass the ‘best target structure’ debate altogether is to design small molecules to remove any RNA they bind, pocket or no pocket — for example, convert a binding compound into a degrader that recruits a host effector protein to remove the target RNA. PROTACs famously exploit this ‘induced proximity’ idea to degrade protein targets. Cells can also remove RNA with mechanisms that can be hijacked by small-molecule drugs. “There’s lots of ways to destabilize RNAs,” says Gilman.

Arrakis is developing small-molecule degraders called NUTACs, or nucleic acid-targeting chimeras. Like PROTACs, they are small molecules that link an RNA binder to a host effector molecule that triggers degradation. Gilman won’t say which host effectors Arrakis is employing. “There’s definitely a bit of a footrace to try to find them first,” he says.

Disney’s group reported the first targeted RNA degraders, called RIBOTACs10 or ribonuclease-targeting chimeras. These recruit and activate RNase L to eliminate a microRNA implicated in cancer, as shown in cells and in a mouse model of breast cancer metastasis. Disney’s group later showed that RIBOTACs can degrade the C9orf72 mRNA guanine-cytosine repeat expansion that’s responsible for many cases of ALS and FTD.

Because degraders can in theory destroy anything they bind, companies need very specific mRNA ligands, more specific even than their mRNA inhibitors. “Lots of molecules bind RNAs and don’t do anything to them,” Gilman says. “But a degrader, if it binds to five RNAs, could well degrade those five RNAs.” Such promiscuity threatens toxicity. Fortunately, RNA sequencing enables companies to look at how a given compound affects every transcript in a cell to weed out nonspecific binders.

But off-target degradation remains a concern. “We really need to be paying attention to this,” Disney says. One reason Disney used RNase L was for its substrate specificity. Not all targets it engages are degraded. Regardless, Weeks expects that companies will test RNA degraders largely in lethal cancers, at least to start, because of the generally higher acceptance of side effects.

As with PROTACs and lysosomal degraders, there is also the concern that hijacking the RNA degradation cellular machinery will interfere with normal biology. “The ligands that we’ve identified do not bind to the active sites of these [host effector] proteins, to the extent that they have active sites,” Gilman says. “So we don’t expect these compounds to broadly inhibit the activity of these effectors.” But Arrakis is developing a second degrader approach that doesn’t rely on the endogenous degradation machinery. Instead, small molecules place a covalent adduct on the target RNA that interferes with translation. But these compounds also must be specific to work safely. And another worry is feedback regulation of genes. “Particularly for highly regulated genes, my concern is that if you degrade the RNA the cell will just respond by upregulating transcription of that gene,” Gilman says. “We just don’t know.”

Now that field of RNA-targeted small molecules is attracting more attention — and money — questions like these should ultimately be answered. Despite its divisions and disagreements, the field’s growth heartens Disney, who’s been working in this space for two decades. “It’s been very nice to see large pharma come in to be thinking about this,” he says. “It speaks to the promise of it. But at the end of the day, you’ve got to fulfill that promise.”