 |
| After the Tertiary Structure |
After the Tertiary Structure
What the next decade of small molecule–RNA research has to solve before it delivers medicines
For twenty years, the case for targeting RNA with small molecules has been made mostly in the subjunctive. Only about 2% of the human genome encodes protein, the argument goes, so the transcriptome ought to be an enormous reservoir of untapped targets — a way to reach the oncogenic transcription factors, aggregation-prone proteins, and structural scaffolds that medicinal chemistry has failed at for decades. The pitch has always been compelling. The delivery has been thin.
There is still, in 2026, exactly one unambiguous success: risdiplam, Roche and PTC's SMN2 splicing modifier for spinal muscular atrophy — orally bioavailable, brain-penetrant, and disease-modifying. Everything else is either preclinical, in early trials, or instructive failure.
That gap between promise and product is the interesting thing about this moment, because the reasons for it are finally becoming legible. Three bottlenecks have held the field back: we could not see RNA structures well enough to design against them, we did not have enough interaction data to learn from, and we could not tell whether a molecule that bound one RNA was leaving the rest of the transcriptome alone. Each is now being attacked directly, and each is yielding at a different rate. Understanding which are actually loosening — and which are not — is the best available guide to where this field goes next.
1. The structure bottleneck is being routed around, not solved
The orthodoxy has been that drugging RNA requires the same thing that drugging proteins requires: a three-dimensional structure with a definable pocket. Docking programs adapted for nucleic acids — AutoDock Vina, rDock, RLDOCK, NLDock — all assume one. So do the recent deep learning entrants like RNAmigos2 and RLaffinity.
The problem is that RNA tertiary structure prediction has not undergone its AlphaFold moment, and the evidence on this is now unusually clear. The CASP16 assessment of nucleic acid structure prediction, published in 2025, found that accuracy still depends almost entirely on whether a closely related experimental structure already exists; without a template, performance collapses. More strikingly, the assessors concluded that despite the surge of interest and the arrival of deep learning methods including AlphaFold3, there had been no notable improvement in nucleic acid modelling accuracy relative to earlier blind challenges. Pseudoknots, non-canonical pairs, and tertiary motifs like A-minor interactions were still not reliably recovered. CASP16 also included, for the first time, blind prediction of RNA–small molecule complexes. Results were generally poor absent templates.
This matters enormously, because the RNAs people actually want to drug — a repeat expansion in a 3′ UTR, an IRES in an oncogene's 5′ leader, a stem loop in a viral genome — are exactly the ones with no solved structure and no close homolog.
Hence the significance of a shift now visible in the literature: skip the tertiary structure entirely. The clearest recent example is SMRTnet, published in Nature Biotechnology in January 2026 by Fei, Wang, Zhang and colleagues at Tsinghua and Peking University. It takes only an RNA sequence, its secondary structure in dot-bracket notation, and a compound SMILES string, fuses a pretrained RNA language model with a chemical language model through an attention-based fusion module, and predicts a binding score. No pocket geometry required.
The results are genuinely encouraging and worth stating precisely. Roughly 0.84 auROC on held-out data from its own PDB-derived training set; 0.72 on an independent benchmark of published interactions; and, in prospective wet-lab screening against ten disease-associated targets, 40 validated binders from 190 predictions — a 21% hit rate, with binding scores that tracked validation rates across the full prediction range. For a class of targets that conventional docking barely handles, that is a real result.
But the ablation buried in the same paper is the more important finding. Substituting computationally predicted secondary structures for experimentally derived ones dropped benchmark performance from 0.72 to 0.66 — most of the way back toward noise. The method does not eliminate the structure requirement; it relocates it from the tertiary to the secondary level, where we happen to have good experimental tools. icSHAPE, SHAPE-MaP, DMS-MaPseq and their descendants can now profile RNA structure transcriptome-wide, in cells, under physiological conditions.
The strategic implication is that the field's structural investment should shift accordingly. The rate-limiting resource for the next five years is not more RNA crystallography — it is chemical probing data at scale, across cell types, disease states, and conditions. Secondary structure is what we can measure well and what current models can actually use. Tertiary structure prediction will improve eventually, but planning a discovery program around its imminent arrival is, on the CASP16 evidence, unwise.
2. The data famine is the real rate limiter
Here is the number that should govern expectations for machine learning in this field. PDBBind, the workhorse resource for protein–ligand affinity prediction, contains tens of thousands of measured protein–ligand interactions. The equivalent figure for RNA is on the order of one hundred. The PDB holds hundreds of thousands of protein structures and roughly a couple of thousand RNA structures.
SMRTnet's training set illustrates the squeeze: 1,061 usable structures from the entire PDB, expanded to 8,672 fragment–ligand pairs by slicing binding regions into 31-nucleotide windows. That is a clever use of what exists, but it is fundamentally an augmentation strategy applied to a small, biased sample — heavy on riboswitches, aptamers, and ribosomal RNA, because those are what crystallizes.
This is why the pretrained-language-model architecture has become near-universal in recent RNA-ligand work (SMRTnet, GerNA-Bind, RLsite, and others all share the pattern). The models are not primarily learning interaction physics from interaction data — there isn't enough. They are importing general representations of RNA sequence space and chemical space learned from hundreds of millions of unlabeled sequences, then fine-tuning on a thin layer of labels. It works better than nothing. It has a ceiling.
Breaking through requires generating interaction data on a scale the field has never attempted. The most promising signals here are industrial rather than academic. xFOREST Therapeutics has described library-versus-library screening across multiplexed panels of more than a thousand RNA structures, measuring selectivity in parallel rather than as an afterthought. Arrakis has built a platform explicitly organized around the sequence–structure–function relationship rather than around one-off campaigns. These are the datasets that will matter, and their concentration in private hands is a genuine structural problem for the field.
If I were placing one bet on what unlocks this area, it would be a precompetitive consortium producing an open, multiplexed RNA–small molecule interaction resource at PDBBind scale. Nothing else on the horizon changes the ceiling as much. Every modelling advance is currently downstream of a hundred-fold data deficit.
3. Selectivity is the unsolved problem, and it is the one that kills drugs
Branaplam is the case everyone in this field should be able to recite. It was developed as an SMN2 splicing modulator for SMA; an off-target effect — inducing a frameshifting pseudo-exon in the huntingtin transcript — looked like a gift, and it was redirected to Huntington's disease. The phase 2 VIBRANT-HD trial, published in Nature Medicine in early 2026, reported that branaplam did what it was supposed to do: it became the first splicing modulator to lower mutant huntingtin in patient cerebrospinal fluid. It also caused peripheral neuropathy in 18 of the 21 participants in the initial dose cohort. The trial was terminated early. The neurofilament light chain elevations reversed on discontinuation, but the drug did not survive.
The mechanism, as worked out in patient-derived neurons, appears to run through nucleolar stress and p53 activation — not through huntingtin lowering. In other words, the therapeutic effect and the toxicity came from different RNAs, and the therapeutic window between them was too narrow.
This is the field's central unsolved problem, and current computational tools barely address it. Essentially every RNA-ligand model, SMRTnet included, is trained and evaluated as a binary, pairwise classifier: given this RNA and this compound, do they interact? But the question that determines whether a molecule becomes a drug is categorically different: given this compound, what does it engage across an entire cellular transcriptome of ~200,000 structured elements, in what proportions, in which tissues?
A model with 0.84 auROC on pairwise classification tells you almost nothing about that. At transcriptome scale, even excellent per-pair specificity produces thousands of off-target engagements. And RNA makes this worse than proteins do, for reasons intrinsic to the molecule: secondary structure motifs are highly recurrent (a 2×2 internal loop flanked by GC pairs appears in innumerable transcripts), RNA is conformationally dynamic, and the same sequence can adopt different structures in different cellular contexts.
The technical responses are visible but immature. Transcriptome-wide small-molecule binding site mapping via covalent crosslinking (Chem-CLIP and successors) gives empirical occupancy maps. Multiplexed selectivity screening builds counter-selection into the primary assay rather than bolting it on. What is conspicuously missing is a computational framework that predicts selectivity profiles rather than binary interactions — and that will not exist until the training data exists. Selectivity is a negative-data problem, and negative data is precisely what curated interaction databases lack.
Expect the next generation of credible RNA-ligand models to be judged on transcriptome-scale precision, not pairwise auROC. Expect most current tools to look worse under that standard.
4. Binding is not function, and the field keeps conflating them
The SMRTnet authors deserve credit for saying this plainly in their own discussion: binding alone does not imply that a molecule will regulate RNA expression or produce biological activity.
Their own data illustrate the difficulty. Of 40 validated binders across ten targets, most had micromolar dissociation constants — respectable for a screening hit, distant from a drug. And the one compound carried forward to cellular work is instructive in an unintended way. Irinotecan hydrochloride trihydrate reduced MYC mRNA and protein in HeLa cells, cut proliferation, and induced apoptosis across three cancer lines. It is also a well-established topoisomerase I inhibitor and an approved chemotherapeutic. When a known cytotoxic reduces proliferation in cancer cells, target engagement at the MYC IRES is not the parsimonious explanation. The authors' luciferase reporter experiment — where the compound suppressed the wild-type IRES construct but not a fully base-paired control — is the right control and provides real evidence for structure-specific activity. But the cell-viability data cannot carry the weight often placed on it.
This is not a criticism peculiar to one paper; it is endemic. Screening natural product and metabolite libraries returns compounds that are pharmacologically loaded by construction. Attributing phenotype to a newly discovered RNA interaction requires ruling out the activity the molecule was already known for, and the field's standard evidence package rarely does.
The methodological fix is not mysterious: structure-matched RNA controls (as used above), rescue with binding-site mutants, transcriptome-wide occupancy mapping, and dose-response concordance between binding affinity and cellular effect. Programs that adopt these as standard will be distinguishable from those that don't, and the distinction will show up in the clinic.
5. Where the value actually accrues: modality, not just binding
The most underrated development in this field is that the binder itself is increasingly not the drug. It is a recognition element.
Ribonuclease-targeting chimeras (RiboTACs) couple a weak, non-functional RNA binder to a moiety that recruits RNase L, converting occupancy into degradation. This solves the field's most awkward problem — that micromolar binding to a structured loop usually does nothing — by making the binding event catalytic rather than occupancy-driven. It is the same logic that transformed protein degradation with PROTACs, and it dramatically relaxes the affinity requirement. Amgen's 2025 deal with Arrakis for RNA-directed degraders, reportedly $75 million upfront, signals that large pharma has priced this in.
The related insight is that RNA-targeting small molecules compete with oligonucleotide therapeutics, not with protein-targeting small molecules. ASOs and siRNAs already hit RNA with exquisite, programmable specificity. What they do not offer is oral dosing, broad biodistribution, blood-brain barrier penetration, and conventional supply chains. That is the entire value proposition — and it means the field's north star should be selective, orally bioavailable, CNS-penetrant molecules, not merely potent ones. Arrakis's DM1 program, aimed at CUG repeat expansions with an oral compound and an IND targeted for 2026, is the clearest test case. Remix's REM-422, a splicing modulator reducing MYB in adenoid cystic carcinoma and AML, tests whether the approach reaches genuinely undruggable transcription factors.
What to watch
Five markers will tell you whether this field is compounding or stalling:
- A second approval. Risdiplam remains a sample size of one. A splicing modulator or repeat-expansion binder clearing phase 2 with a clean neurological safety profile would change the field's risk calculus more than any methodological advance.
- Open interaction data at scale. Watch for a multiplexed RNA–ligand dataset released precompetitively. Its absence caps every model in the field.
- Selectivity as a primary endpoint. Papers reporting transcriptome-wide occupancy alongside target binding — rather than pairwise auROC alone — mark the transition from method development to drug discovery.
- Chemical probing displacing crystallography as the structural foundation, with prediction models trained on in-cell secondary structure rather than PDB-derived idealizations.
- RiboTACs reaching the clinic. If catalytic degradation works in humans, the affinity bar drops by orders of magnitude and the addressable target space expands accordingly.
The honest summary is that computational methods have recently gotten good enough to generate credible hypotheses at scale — that is what SMRTnet and its contemporaries demonstrate, and it is not nothing. What they have not done is solve the two problems that actually determine whether RNA-targeting small molecules become a therapeutic class: getting from micromolar binder to selective drug, and proving that engagement causes the biology. Those are experimental problems, they are expensive, and no amount of architectural cleverness substitutes for the data.
The field's next decade will be decided in assay development, not model design.
Sources
- Fei, Y. et al. "Predicting small molecule–RNA interactions without RNA tertiary structures." Nature Biotechnology (2026). https://doi.org/10.1038/s41587-025-02942-z
- "Assessment of nucleic acid structure prediction in CASP16." Proteins / PMC12248019. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12248019/
- "Oral splicing modulator branaplam in Huntington's disease: a phase 2 randomized controlled trial." Nature Medicine (2026). https://www.nature.com/articles/s41591-025-04117-4
- Krach, F. et al. "RNA splicing modulator for Huntington's disease treatment induces peripheral neuropathy." iScience (2025). https://pmc.ncbi.nlm.nih.gov/articles/PMC12059699/
- Carvajal-Patiño, J.G. et al. "RNAmigos2: accelerated structure-based RNA virtual screening with deep graph learning." Nature Communications 16, 2799 (2025).
- Childs-Disney, J.L. et al. "Targeting RNA structures with small molecules." Nature Reviews Drug Discovery 21, 736–762 (2022).
- Arrakis Therapeutics DM1 program updates. https://arrakistx.com/pipeline/
- "RNA-targeting small molecules: a new frontier of drug discovery." Pharmaceutical Technology (2026). https://www.pharmaceutical-technology.com/features/rna-targeting-small-molecules-a-new-frontier-of-drug-discovery/
No comments:
Post a Comment