AI-Designed Protein Vehicles Could Rewrite the Rules of RNA Delivery

 

TheRNAblog

Scientists have built synthetic RNA carriers from the bottom up—and one of them dramatically outperformed conventional delivery systems in laboratory experiments

For decades, biology has looked to viruses for inspiration.

Viruses are extraordinarily good at one thing: moving genetic material into cells. Their protein shells package nucleic acids, protect them from degradation, recognize cells and ultimately release genetic cargo inside them.

That ability has made viral vectors powerful tools for gene therapy.

But there is an intriguing question we rarely ask:

What if evolution did not produce the best possible RNA-delivery vehicle?

Evolution does not engineer biological systems for our purposes. It optimizes them for survival, reproduction and transmission within particular ecological environments. Features that make a virus successful in nature may therefore be unnecessary—or even undesirable—when our goal is simply to deliver therapeutic RNA into a cell.

A remarkable new study published in Nature takes this idea seriously.

Rather than modifying an existing virus, researchers used artificial intelligence-designed proteins to construct RNA delivery vehicles essentially from the bottom up.

The result was a family of what they call synthetic transfer vehicles, or STVs.

And one particular design—STV-C8—showed striking RNA-delivery performance.

Instead of copying viruses, design something new

Most viruses package their genomes inside highly organized protein structures. Despite enormous viral diversity, many capsids converge on a relatively small number of architectural principles, particularly icosahedral and helical structures.

But modern generative protein-design systems can explore structural possibilities that evolution may never have sampled.

The researchers therefore asked whether AI-designed protein assemblies with unusual geometries could serve as the structural foundation for completely new RNA carriers.

Using protein assemblies produced with RFdiffusion, they combined synthetic protein structures with biological components responsible for several essential tasks:

• binding cellular membranes,

• promoting vesicle budding,

• binding and packaging RNA, and

• facilitating entry into recipient cells.

More than 100 different synthetic transfer vehicle designs were ultimately examined.

The researchers did not simply ask whether a particle formed. Their screening system evaluated several steps independently, including particle release, cellular uptake and successful expression of the delivered RNA.

That distinction is important.

A delivery particle can enter a cell without delivering useful RNA. Likewise, it can package RNA efficiently yet fail to release its cargo where it matters.

The researchers were therefore effectively optimizing the complete delivery process.

The surprise: the best architecture did not look like a virus

One might expect an artificial system that performs virus-like functions to work best when it resembles a viral capsid.

That was not what happened.

Synthetic assemblies containing dihedral and cyclic symmetries often performed better than the more virus-like icosahedral designs.

The eventual lead platform contained an AI-designed protein assembly called HE0690, organized with C8 cyclic symmetry.

After optimizing additional components—including a membrane-binding domain—the researchers created the system they named STV-C8.

Its architecture combines several modules:

membrane binding → budding → RNA recognition → synthetic protein assembly

This modular organization is one of the most interesting aspects of the work. Instead of treating the delivery vehicle as an indivisible biological object, the researchers treated it more like an engineered platform whose components could be independently optimized.

STV-C8 delivered RNA with extraordinary efficiency

When characterized experimentally, STV-C8 showed strong enrichment of its intended cargo RNA and formed membrane-bound vesicles typically around the nanoscale range expected for biological delivery particles.

But its delivery efficiency attracted the most attention.

When STV-C8 was compared with several genetically encoded RNA-delivery systems, it produced higher expression of the delivered RNA across the tested cell types.

The researchers then compared it with lipid nanoparticles containing ALC-0315, the ionizable lipid used in the Pfizer–BioNTech COVID-19 vaccine formulation.

In their experimental system, STV-C8 produced a more than 1,000-fold higher transfection rate than the tested LNP formulation.

Even more strikingly, the amount of STV-C8-delivered mRNA required to obtain comparable protein expression was reported to be more than 10,000-fold lower than the amount required with the LNP system.

Those numbers require careful interpretation. They come from particular experimental conditions and should not be interpreted as proof that STV-C8 is universally 1,000 or 10,000 times better than every clinically optimized LNP formulation.

Nevertheless, they demonstrate something important:

the delivery architecture has considerable potential.

RNA cargo without the usual capsid-size constraint

Conventional viral vectors frequently face another problem: cargo capacity.

There is only so much genetic material that can fit inside a closed viral capsid.

STV-C8 may operate differently.

The researchers proposed that RNA may associate with the outside surface of the synthetic protein oligomer while remaining protected inside the surrounding membrane vesicle.

If that model is correct, RNA does not have to fit inside a rigid protein shell.

Experimentally, the researchers tested RNA cargos ranging from approximately 1 to 10 kilobases and observed effective delivery throughout that range, with no obvious packaging limit within the tested window.

That could be particularly useful for large messenger RNAs encoding genome editors, transcription factors or other therapeutic proteins.

The researchers also programmed where the particles go

Efficiency alone is not enough for RNA medicine.

A therapeutic RNA carrier must ideally deliver its cargo to the right cell.

The researchers therefore added another layer of AI-guided design.

Computationally designed peptide binders targeting EGFR or IL-7Rα were incorporated into the particles. These molecules acted as targeting modules, enabling preferential uptake by cells expressing the corresponding receptor.

Cells expressing both receptors could even respond to vehicles carrying both targeting modules, suggesting that future delivery systems might use combinations of recognition molecules to refine tissue or cell specificity.

This creates a compelling engineering concept:

AI designs the carrier.
AI can help design the targeting ligand.
RNA provides the therapeutic program.

The delivery system becomes increasingly programmable.

Not just reporter RNA

Demonstrating GFP expression is useful for developing a platform, but therapeutic relevance requires considerably more.

The researchers therefore tested STV-C8 with a surprisingly broad range of RNA cargos.

They delivered messenger RNA encoding the transcription factor ASCL1 into primary mouse astrocytes and detected ASCL1 expression in more than 30% of cells.

They delivered Cas9 mRNA and guide RNAs designed to edit the dystrophin gene.

They also packaged the RNA-targeting CRISPR nuclease Cas13d together with a guide targeting SARS-CoV-2.

When this antiviral system was delivered into human iPSC-derived lung cells subsequently infected with SARS-CoV-2-GFP, viral replication was almost completely suppressed in the experimental model.

The important message is not simply that STV-C8 delivers RNA.

It appears capable of delivering functional molecular programs.

From cell culture to animals

The study went further.

After intravenous administration of STV-C8 carrying EGFP mRNA in mice, the researchers performed whole-body clearing and high-resolution imaging to determine where expression occurred.

Interestingly, they observed strong expression predominantly in the lung, rather than the liver—the organ that commonly accumulates many conventional lipid nanoparticles.

The study reported no detectable immunological or toxicological effects following systemic administration under the tested conditions.

This natural lung tropism may itself become useful, although whether it remains desirable would depend on the therapeutic application.

For other diseases, targeting will have to be redesigned.

A test case: Duchenne muscular dystrophy

One of the most ambitious demonstrations involved Duchenne muscular dystrophy (DMD).

DMD is caused by mutations that disrupt the dystrophin gene and progressively destroy skeletal and cardiac muscle function.

For some mutations, removal of particular exons can restore the dystrophin reading frame.

The researchers therefore packaged CRISPR–Cas9 mRNA together with guide RNAs targeting regions surrounding dystrophin exon 51.

In cultured porcine fibroblasts, they observed approximately 40% deletion of the targeted region.

They then injected STV-C8 containing Cas9 and guide RNAs directly into pig muscle and confirmed deletion of exon 51.

Finally, STV-C8-mediated editing was tested in skeletal muscle cells derived from patients with DMD, where deletion of exon 51 restored the dystrophin open reading frame.

This does not mean STV-C8 is ready to treat DMD.

The authors themselves point out an important limitation: local intramuscular injection is not a practical treatment for a disease affecting muscles throughout the body.

Systemic muscle targeting remains a major challenge.

Nevertheless, successful delivery of a functional genome-editing system into a large-animal tissue moves the platform well beyond a simple proof-of-concept nanoparticle.

Why this study matters beyond RNA delivery

Perhaps the most important idea in this paper is broader than STV-C8 itself.

Biotechnology has traditionally engineered biological systems by starting with something evolution already produced and then modifying it.

A virus becomes a gene-delivery vector.

A bacterial nuclease becomes CRISPR technology.

A naturally occurring antibody becomes a therapeutic antibody.

Generative AI introduces another possibility:

Start with the function we want and search protein space for structures that evolution never created.

That changes the engineering philosophy.

Instead of asking:

“Which natural protein can we modify?”

we can increasingly ask:

“What protein architecture should exist to perform this function?”

The researchers argue that RNA transfer vehicles are particularly suitable for current generative protein-design systems because these algorithms are already very good at generating stable, symmetric protein assemblies. Designing highly dynamic enzymes remains considerably harder.

Important questions remain

Despite the impressive results, several hurdles separate STV-C8 from clinical translation.

Long-term safety and immunogenicity will need much more extensive evaluation.

Manufacturing at therapeutic scale will have to be developed.

Pharmacokinetics, stability, dosing and repeat administration require careful study.

Targeting must become sufficiently precise to avoid unwanted tissues.

The reliance on additional fusogenic components also needs to be considered when translating the technology into therapeutic formulations.

And most importantly, the exceptional efficiency reported in these experimental comparisons will need validation across additional delivery systems, animal models and clinically relevant dosing conditions.

Even the DMD experiment illustrates the difference between demonstrating molecular editing and creating a practical therapy. The authors explicitly acknowledge that direct intramuscular administration does not solve the systemic delivery problem and identify comparative performance, targeting and cell-type specificity as areas requiring future study.

A new chapter for synthetic biology

The significance of STV-C8 is therefore not that lipid nanoparticles or viral vectors are suddenly obsolete.

It is that the design space for biological delivery systems may be far larger than we previously imagined.

Evolution gave us viruses.

Chemistry gave us lipid nanoparticles.

Synthetic biology is now beginning to give us something different:

delivery vehicles whose fundamental architecture can be computationally invented.

STV-C8 is an early example of what happens when generative AI, protein engineering, RNA biology and gene editing converge.

And perhaps the most provocative lesson from this work is simple:

Nature does not necessarily contain every useful biological machine. Some of them may have to be designed.

---

Original research

Schuhmacher, M. K., Gruber, C., Lang, C. M. R. et al. Creating bottom-up RNA transfer vehicles from synthetic protein assemblies. Nature (2026). DOI: 10.1038/s41586-026-10952-3.

Can We Make Lipid Nanoparticles Without the Inflammation Problem?

 

Charge-switching ionizable lipids

Charge-switching ionizable lipids may separate efficient RNA delivery from one of the biggest liabilities of conventional LNPs

Lipid nanoparticles changed the trajectory of RNA medicine.

They solved one of the fundamental problems facing nucleic-acid therapeutics: how do you move a fragile, highly charged molecule such as mRNA through the body, into a cell, and finally into the cytoplasm where it can function?

But the same chemistry that makes lipid nanoparticles effective can also make them inflammatory.

That trade-off has become increasingly important as researchers try to move LNP technology beyond vaccines toward repeated treatment of cancer, inflammatory disease, genetic disorders and other chronic conditions.

A new study in Nature Nanotechnology proposes an intriguing way around the problem.

Instead of asking how to make conventional ionizable lipids slightly less inflammatory, the researchers redesigned their electrostatic behavior.

Their solution is a class of charge-switching ionizable lipids that are negatively charged under physiological conditions but change their charge as they encounter the acidic environment used during formulation and intracellular trafficking.

The resulting particles—called switchable nanoparticles, or SNPs—were able to deliver mRNA and DNA while avoiding several inflammatory pathways normally activated by conventional LNPs.

If the approach translates beyond the current preclinical models, it could address one of the central engineering constraints in RNA delivery.

The paradox at the heart of lipid nanoparticles

Modern LNPs rely heavily on ionizable lipids.

These molecules are extraordinarily useful because their charge changes with pH.

During formulation under acidic conditions, the lipids become positively charged. That helps them associate with negatively charged RNA and package it efficiently.

After administration, their reduced charge at physiological pH improves tolerability compared with permanently cationic lipids.

Then, after a nanoparticle is taken into a cell, acidification within endosomes helps ionizable lipids interact with endosomal membranes and promote release of the RNA cargo.

It is an elegant system.

But it contains a fundamental contradiction.

The properties needed for efficient nucleic-acid encapsulation and endosomal escape can also stimulate inflammation.

Conventional LNPs can activate pathways including:

• complement signalling,

• Toll-like receptor 4,

• galectin-associated responses following endosomal damage,

• and platelet-activating factor signalling.

These effects become particularly problematic in tissues that are already inflamed.

And many of the patients who could benefit most from RNA therapeutics—people with inflammatory, pulmonary, metabolic or degenerative diseases—may already have substantial baseline inflammation.

So the challenge is not merely:

Can an LNP deliver RNA?

It is increasingly:

Can it deliver enough RNA repeatedly without making the underlying disease worse?

A lipid that changes its electrostatic personality

The new study approaches this problem through lipid chemistry.

The researchers designed what they call S-lipids.

Their head group contains both:

• a carboxylic acid, and

• a tertiary amine.

That combination allows the lipid to switch its charge state across different pH conditions.

At physiological pH, around 7.4, the particles are negatively charged.

Under more acidic conditions, the charge state changes, supporting nucleic-acid encapsulation and intracellular delivery.

This is important because conventional ionizable LNPs are often discussed mainly in terms of their transition between neutral and positively charged states.

Here, the design deliberately introduces a negative surface state at physiological pH.

That apparently changes how the nanoparticle interacts with immune and inflammatory pathways.

The authors therefore attempt to decouple two properties that have traditionally been difficult to separate:

efficient delivery and inflammatory cationic chemistry.

Screening 144 new ionizable lipids

This was not based on a single fortunate molecule.

The team synthesized a library of 144 S-lipids and systematically evaluated their ability to deliver mRNA.

Several structural rules emerged.

The spacing between the amine and carboxyl group mattered.

So did the architecture of the hydrophobic tails.

Lipids with branched tails performed dramatically better than some corresponding linear-tail structures—in certain comparisons, by thousands-fold.

The best-performing particles had charge-transition properties suited to acidic intracellular compartments while remaining negatively charged at physiological pH.

Importantly, many candidates retained strong transfection activity.

Of 18 top-performing S-lipids investigated further, 11 produced nanoparticles capable of transfecting primary fibroblasts approximately as efficiently as LNPs containing highly effective established ionizable lipids.

So reducing inflammatory signalling did not necessarily require sacrificing RNA delivery.

That is the central engineering advance.

The particles still escaped the endosome

Avoiding positive charge alone would not be particularly useful if the RNA remained trapped inside endosomes.

Endosomal escape is one of the major bottlenecks in nucleic-acid delivery.

Conventional LNPs help destabilize endosomal membranes. But membrane disruption itself can also activate cellular stress and inflammatory pathways.

The switchable nanoparticles behaved differently.

According to the study, they were capable of releasing nucleic acids efficiently without producing the same degree of endosomal membrane damage.

This was reflected in reduced activation of galectin-8, a cellular marker associated with damaged endosomal membranes.

The particles also avoided substantial activation of platelet-activating factor signalling.

That suggests a potentially important distinction:

successful endosomal release does not necessarily require extensive destructive membrane disruption.

If confirmed across different cell types and formulations, that concept could influence the next generation of delivery-vector design.

Less complement and TLR4 activation

The negative physiological charge of the switchable nanoparticles also affected extracellular inflammatory interactions.

The particles showed reduced activation of complement and TLR4-related signalling compared with conventional LNP systems examined in the study.

That matters because complement activation is a long-standing concern for nanoparticle therapeutics.

Nanoparticle surfaces immediately encounter proteins when they enter biological fluids, forming a protein corona that affects distribution, clearance and immune recognition.

Changing surface electrostatics can fundamentally alter those interactions.

The S-lipid design therefore does more than modify intracellular trafficking.

It changes the nanoparticle's biological identity before it even reaches the cell.

Human immune cells showed the same trend

A particularly relevant part of the study involved human peripheral blood mononuclear cells.

The researchers compared conventional LNPs and switchable nanoparticles carrying nucleic acids.

Both systems could deliver their cargo.

But traditional LNPs produced substantially greater cytokine responses, whereas the switchable formulations produced much less inflammatory stimulation.

That does not establish clinical safety.

PBMC experiments cannot reproduce the complexity of an entire human immune system.

But they provide an important bridge between standard cell-line experiments and animal studies.

The question now becomes whether this lower immunostimulatory profile persists across different human donors, doses, routes of administration and therapeutic cargos.

The difference became striking in already-inflamed animals

One of the most interesting aspects of the study was its focus on pre-existing inflammation.

This is biologically important.

A nanoparticle that looks reasonably tolerable in a healthy animal may behave very differently in an organism whose immune system is already activated.

The researchers therefore treated mice with lipopolysaccharide, or LPS, to induce inflammation before administering the nanoparticles.

Conventional LNPs substantially increased inflammatory cytokine responses.

The switchable nanoparticles caused much less additional inflammatory stimulation.

That finding gets to the practical significance of the technology.

Future RNA medicines will often not be administered to healthy individuals.

They will be administered to sick patients.

Nanoparticle safety therefore needs to be tested in disease-relevant physiological states—not only under idealized healthy conditions.

A dramatic DNA-delivery experiment

The researchers also examined plasmid DNA delivery.

This is notable because LNP development is usually dominated by mRNA and siRNA, but the underlying platform could potentially carry several classes of nucleic acids.

Switchable nanoparticles successfully delivered nanoplasmid DNA to mice.

The comparison with conventional particles was striking: traditional LNP formulations used for the DNA-delivery experiment produced severe acute toxicity, whereas the switchable formulation was substantially better tolerated.

This result reinforces a broader point.

Nanoparticle toxicity does not depend only on the lipid.

It emerges from the interaction among:

lipid chemistry + nucleic-acid cargo + dose + route + tissue environment.

A delivery system that works well for one cargo cannot automatically be assumed safe for another.

Treating acute lung injury with IL-22 mRNA

The researchers then moved from reporter delivery to a therapeutic experiment.

They used an LPS-induced mouse model of acute lung injury and delivered mRNA encoding interleukin-22 (IL-22).

IL-22 can promote tissue-protective and regenerative responses in epithelial barriers.

The switchable nanoparticles carrying IL-22 mRNA performed better than several conventional LNP formulations.

The proposed reason is especially interesting.

It was not simply that the switchable nanoparticles delivered more therapeutic mRNA.

Rather, conventional LNPs added inflammatory stress to an organ that was already inflamed.

The therapeutic payload was therefore competing against inflammatory effects generated by its own delivery vehicle.

The low-inflammatory nanoparticle avoided much of that penalty.

This illustrates a principle that deserves more attention in drug-delivery research:

The biological effect of a nucleic-acid medicine is the sum of the cargo and the carrier.

An excellent therapeutic RNA delivered by an inflammatory vehicle may produce a weaker therapeutic outcome than expected.

Why this matters beyond one nanoparticle formulation

Most discussion around LNP optimization focuses on parameters such as:

• transfection efficiency,

• organ targeting,

• particle size,

• encapsulation efficiency,

• pKa,

• endosomal escape,

• and biodegradability.

This study adds another design dimension:

What charge should the particle have when it is not actively performing a delivery step?

Instead of optimizing one fixed electrostatic state, the researchers designed a lipid whose charge changes according to its biological environment.

That points toward a broader concept of environment-responsive nanomedicine.

Future delivery vehicles may increasingly be engineered to behave differently during:

formulation → circulation → tissue entry → endocytosis → endosomal acidification → cytosolic release.

The best nanoparticle may not have one ideal property.

It may need a sequence of properties activated at precisely the right biological stage.

Implications for RNA therapeutics

For mRNA therapeutics, the implications are substantial.

Lower-inflammatory nanoparticles could potentially make several currently difficult applications more realistic:

Repeated mRNA dosing.
Vaccines may require only a few administrations. Protein-replacement therapies could require dozens or hundreds.

Treatment of inflammatory diseases.
The carrier should ideally not worsen the pathology it is intended to treat.

Higher therapeutic doses.
Delivery toxicity often constrains the amount of nucleic acid that can safely be administered.

Gene editing.
CRISPR components frequently require potent transient delivery, sometimes at relatively high doses.

DNA delivery.
Improved tolerability could expand nanoparticle delivery beyond mRNA and siRNA.

And perhaps most importantly, lower intrinsic carrier toxicity could widen the therapeutic window.

Could the concept matter for plant RNA delivery?

The study concerns mammalian systems, so direct extrapolation to plants would be premature.

Plant cells present completely different delivery barriers, including cell walls, cuticles, extracellular matrices and distinct endosomal trafficking pathways.

Nevertheless, the design principle is relevant.

RNA-delivery materials do not merely transport RNA.

Their charge state determines how they interact with biological surfaces.

In plant biotechnology, positively charged carriers such as polymers, peptides, carbon-based materials and lipid systems are often attractive because they bind negatively charged RNA efficiently.

But excessive charge can also cause aggregation, poor tissue penetration, membrane damage or phytotoxicity.

A carrier capable of switching charge according to its environment could potentially provide strong RNA complexation during formulation while reducing undesirable interactions after delivery.

Whether the specific S-lipid chemistry described here is useful for plants remains an open experimental question.

The broader principle is nevertheless worth exploring.

Important questions remain

The results are impressive, but these are still preclinical experiments.

Several questions will determine whether charge-switching nanoparticles become a broadly useful platform.

How do they behave after repeated administration?

What are their long-term metabolites?

How does their biodistribution compare with clinically established LNPs?

Does their reduced inflammatory profile persist at therapeutically relevant high doses?

Can their tissue tropism be engineered?

Will the advantage remain in non-human primates and eventually humans?

Can different ionizable-lipid structures reproduce the effect?

And perhaps most importantly:

Can low inflammation and high transfection remain linked across many therapeutic applications—or was this balance specific to the formulations tested here?

These questions will require substantial additional work.

The larger lesson

The success of the first generation of lipid nanoparticles came from learning how to package RNA and release it inside cells.

The next generation may require something subtler.

We need nanoparticles that know when to interact strongly with biology—and when not to.

The charge-switching strategy described in this study is therefore interesting not simply because it produces another high-performing lipid.

It represents a change in design philosophy.

Instead of accepting inflammation as an unavoidable cost of efficient nucleic-acid delivery, the researchers asked whether the molecular mechanisms responsible for delivery and toxicity could be separated.

Their results suggest that, at least in these experimental systems, they can.

And that could be an important step toward a future in which the delivery vehicle becomes almost invisible—doing exactly what is required to move RNA into the right cell, and very little else.

---

Research discussed

Liang D., Qi Y., Murthy N. and colleagues.
Charge-switching ionizable lipids lower the toxicity of lipid nanoparticles.
Nature Nanotechnology (2026).

The study describes ionizable lipids containing both a carboxylic acid and an amine that form charge-switching nanoparticles capable of delivering mRNA and DNA while reducing activation of several inflammatory pathways associated with conventional lipid nanoparticles.

Plants May Fight Viruses by Immunizing Their Neighbors—Just 3–4 Cells at a Time

Plants Fight Viruses by Immunizing Their Neighboring Cells

 A new study reveals an unexpectedly local antiviral defense system built from mobile small RNAs

A plant cell infected by a virus may already be losing the battle.

The virus is replicating. Its suppressor proteins are interfering with RNA silencing. The molecular machinery that would normally destroy viral RNA is being sabotaged from inside.

But the infected cell can still do something valuable.

It can warn its neighbors.

That warning appears to travel in the form of virus-derived small interfering RNAs, or vsiRNAs. These tiny RNA molecules move from infected cells into surrounding uninfected cells, load into ARGONAUTE proteins, and prepare those cells to attack the virus before it arrives.

The protection, however, may be surprisingly local.

A new study published in Cell Host & Microbe proposes that much of plant antiviral RNA interference operates through repeated zones of protection extending only about three to four cells from an infected cell.

Rather than creating one enormous wave of systemic immunity, plants may repeatedly construct thousands of microscopic antiviral barriers as infection moves through their tissues.

That changes how we should think about RNA silencing in plants.

The familiar model of plant antiviral RNAi

RNA interference is one of the central antiviral defense mechanisms in plants.

When an RNA virus replicates, it generates double-stranded RNA. Plant DICER-LIKE enzymes recognize these molecules and process them into small RNA duplexes, particularly 21- and 22-nucleotide vsiRNAs.

These vsiRNAs can then associate with ARGONAUTE proteins to form RNA-induced silencing complexes, or RISCs.

The sequence carried by the small RNA acts almost like a molecular address.

When RISC encounters complementary viral RNA, ARGONAUTE can direct its destruction.

Viruses, of course, have evolved countermeasures.

Many plant viruses encode viral suppressors of RNA silencing, or VSRs, which interfere with different stages of this defense. Some bind small RNAs. Others attack ARGONAUTE proteins or disrupt downstream silencing.

This creates a molecular arms race inside infected cells.

But plants possess another important property: RNA silencing is not necessarily confined to the cell in which it begins.

Small RNAs can move through plasmodesmata into neighboring cells and can also participate in longer-distance signaling through vascular tissues.

For years, this mobility suggested an appealing hypothesis:

Perhaps vsiRNAs move ahead of an invading virus and immunize cells before infection reaches them.

The idea made biological sense.

Directly proving it was much harder.

The experimental problem

Virus movement and RNA-silencing movement often use the same cellular highways.

A virus can move from cell to cell through plasmodesmata. RNA-silencing signals can do the same.

If both are moving simultaneously, how can researchers determine whether a neighboring cell became resistant because mobile vsiRNAs arrived first?

Another problem is viral suppressors.

If a virus strongly suppresses RNAi, it becomes difficult to determine how much of the apparent antiviral effect comes from silencing within infected cells and how much comes from protection of neighboring cells.

This distinction is crucial.

Intracellular antiviral RNAi tries to reduce viral replication after infection has occurred.

Intercellular antiviral RNAi attempts to prepare other cells before the virus reaches them.

Albertini and colleagues found an unusually useful system for separating these processes.

A virus that naturally stays in the phloem

The researchers studied turnip yellows virus, or TuYV, in Arabidopsis thaliana roots.

TuYV is particularly useful because it is naturally restricted largely to vascular tissues.

The researchers used GFP-tagged TuYV so infected cells could be visualized at cellular resolution.

In wild-type Arabidopsis roots, the infection did not spread uniformly through the tissue. Instead, viral GFP was concentrated mainly in particular phloem-associated cell files, often with conspicuous interruptions.

Meanwhile, TuYV-derived siRNAs could travel beyond the virus's own infection domain.

This effectively separated the movement of the virus from the movement of its small-RNA products, allowing the researchers to ask a much cleaner question:

What happens to viral spread when those mobile vsiRNAs disappear?

The answer was striking.

Remove the small RNAs, and the virus occupies far more territory

The researchers examined dcl234 plants lacking DCL2, DCL3 and DCL4 activities and therefore severely deficient in normal antiviral vsiRNA production.

The virus expanded dramatically.

In wild-type roots, viral occupancy of the vasculature was approximately 30%. In dcl234 roots it increased to approximately 85%.

Young lateral-root infection similarly increased from roughly 30% to 75%.

The viral RNA phenotype was even more dramatic.

Across independent infections, TuYV RNA levels were more than 20-fold higher in dcl234 roots than in wild-type roots.

At first glance, one might conclude that removing antiviral RNAi simply allowed the virus to replicate much more efficiently inside each infected cell.

But that is not what the researchers observed.

The amount of GFP within individual infected cells was similar between wild-type and RNAi-deficient roots.

What changed most was how many cells became infected.

The virus was not primarily becoming stronger inside each cell.

It was gaining access to more cells.

That distinction leads to the central conclusion of the study:

The dominant antiviral effect of vsiRNAs in this system is spatial containment.

In other words, RNAi functions less like a drug reducing the viral load inside an already infected cell and more like a molecular firewall surrounding that cell.

The three-to-four-cell antiviral firewall

The authors propose the following model.

An infected cell generates vsiRNAs.

Some of those small RNAs escape into nearby uninfected cells.

There, they associate with ARGONAUTE proteins and establish antiviral RISCs before the virus arrives.

When viral particles subsequently enter one of these pre-immunized cells, the incoming viral RNA encounters an antiviral system that has already been programmed against it.

The infection therefore stalls.

But small RNAs do not remain indefinitely at high concentrations.

As the mobile signal becomes diluted, its protective effect eventually decreases.

The virus can then establish another infection.

That newly infected cell begins producing additional vsiRNAs.

A new protective zone forms.

The process repeats.

infection → vsiRNA production → local movement → neighboring-cell immunization → viral containment → signal dilution → renewed infection

This repeated cycle could explain the interrupted pattern of TuYV infection seen along Arabidopsis root cell files. The authors describe recurrent short-range immunization as a major determinant of viral containment.

The effective antiviral radius was approximately three to four cells.

That number may prove to be one of the most interesting findings of the paper.

Small RNA movement is not the same as functional immunity

There is an important twist.

Using a GFP-based RNA-silencing sensor, the researchers could detect functional primary TuYV-derived siRNAs at least 15 cells away from where they were produced.

Yet when the researchers measured actual protection against viral invasion, effective immunization extended only around three to four cells.

The distinction is critical.

A small RNA can apparently travel far enough to silence an easily accessible reporter transcript without necessarily being present—or active—at sufficient levels to stop a replicating virus.

The authors suggest several possible reasons.

Viral RNAs can be highly structured. Replication intermediates may also be partially protected inside membrane-associated replication compartments. Such viral RNA could therefore be considerably harder for RISC to access than an exposed cytoplasmic reporter mRNA.

Consequently:

detectable RNA movement ≠ equivalent antiviral protection.

The study explicitly warns that reporter-based assays may overestimate the biologically effective range of antiviral RNA silencing.

That lesson extends well beyond this particular virus.

ARGONAUTE1 emerges as the neighborhood defender

The study also examined which ARGONAUTE proteins execute this mobile immunity.

Both AGO1 and AGO2 can participate in antiviral silencing within TuYV-infected cells.

But TuYV encodes the suppressor protein P0, which interferes with ARGONAUTE-based defense. The study suggests that this makes sustained intracellular RNAi relatively ineffective against P0-proficient TuYV.

Neighboring cells are different.

They contain mobile vsiRNAs but do not yet contain the virus—or its suppressor.

Those recipient cells therefore provide a much more favorable environment for antiviral RISC formation.

The genetic evidence particularly points toward AGO1 as the main effector of this intercellular immunization.

TuYV accumulation changed little in an ago2 mutant but increased roughly fourfold in an ago1 mutant. The authors argue that compensatory increases in AGO2 probably reduce the apparent magnitude of the AGO1 phenotype, reinforcing the conclusion that AGO1 is the major effector of mobile immunity.

This creates an elegant division of labor.

Inside an infected cell, the virus can sabotage antiviral machinery.

Outside that cell, the same viral suppressor is absent.

Mobile small RNAs can therefore exploit the virus-free neighboring territory.

The plant may lose the molecular battle inside one cell while still winning the spatial battle across the tissue.

The phenomenon may extend beyond TuYV

TuYV is an unusual virus because of its strong vascular restriction.

Could this three-to-four-cell immunity simply be a peculiarity of that system?

The researchers tested an unrelated virus, turnip mosaic virus (TuMV), in Nicotiana benthamiana leaf epidermal tissue.

Those experiments again supported short-range vsiRNA-mediated protection around infected regions.

Importantly, the behavior depended on the type of viral suppressor involved. Suppressors that sequester small RNAs can interfere not only with intracellular silencing but also with the supply of mobile vsiRNAs available to protect neighboring cells.

The study therefore argues that short-range immunization is neither exclusive to Arabidopsis roots nor unique to TuYV.

That raises an intriguing evolutionary possibility.

A viral suppressor that completely prevents vsiRNA movement may help a virus invade surrounding tissues.

But a suppressor that mainly disables ARGONAUTE activity inside infected cells while allowing vsiRNAs to escape could unintentionally promote the formation of antiviral barriers around the infection.

For some viruses, that might even be advantageous if remaining within particular tissues improves transmission or persistence.

The paper therefore turns the classical virus-versus-RNAi arms race into something more spatially complex.

Why this matters for RNA-based crop protection

The study was designed to understand natural antiviral immunity. It did not test spray-induced gene silencing, nanoparticle delivery, or field-scale RNA biopesticides.

Nevertheless, its mechanistic implications for these technologies are difficult to ignore.

Many exogenous RNA-delivery strategies are judged primarily by questions such as:

• Did dsRNA enter the plant?

• Can siRNAs be detected?

• How far did the signal move?

• How much target RNA decreased?

This paper suggests another question may be equally important:

How many neighboring cells become functionally protected against an invading pathogen?

For antiviral dsRNA or siRNA technologies, successful delivery may require more than maximizing total RNA uptake.

An effective formulation may need to optimize several sequential processes:

RNA stability → cellular entry → DCL processing → escape or movement from treated cells → plasmodesmatal transport → loading into AGO proteins → persistence in recipient cells → accessibility of the viral target RNA.

A carrier that delivers enormous amounts of RNA into one cell but traps the RNA there might perform very differently from a carrier that generates a smaller but highly mobile population of active siRNAs.

Likewise, simply demonstrating fluorescence several hundred micrometers from an application site may not prove that the delivered RNA creates antiviral immunity at that distance.

This study therefore provides an important conceptual distinction for RNA-delivery research:

Physical delivery range

How far can the RNA molecule or carrier travel?

Silencing range

How far away can a reporter or endogenous transcript be suppressed?

Protective range

How far away can a biologically relevant pathogen actually be stopped?

Those numbers may be very different.

A more spatial view of RNA immunity

For decades, antiviral RNAi has often been described as a molecular pathway:

viral dsRNA → DCL → vsiRNA → AGO → viral RNA degradation.

That pathway remains correct.

But it is incomplete without spatial information.

Which cell produces the small RNA?

Which cell receives it?

Where does ARGONAUTE loading occur?

Is a viral suppressor present in that cell?

How accessible is the viral RNA?

How many cells separate the source from the protected tissue?

And how quickly does protection decay?

At cellular resolution, antiviral RNAi begins to look less like a single biochemical pathway and more like an evolving landscape of infected, transmitting, protected and susceptible cells.

That may be the most important conceptual contribution of this work.

Some important caveats

The study should not be interpreted as proving that every plant virus is contained by an identical three-to-four-cell barrier.

TuYV has unusual vascular biology, and Arabidopsis roots provide an exceptionally structured experimental system.

The authors also acknowledge that other barriers may contribute to TuYV restriction. Even in RNAi-deficient roots, the virus did not invade every possible tissue, suggesting additional RNAi-independent restrictions or residual silencing activity.

The difference between ≥15-cell reporter silencing and three-to-four-cell antiviral protection also emphasizes that experimental readouts matter enormously.

Different viruses, tissues, developmental stages and viral suppressors may produce different effective immunization distances.

That is precisely why this work opens so many interesting questions.

The next questions

Can mobile antiviral RISCs be inherited during cell division?

Could this help explain why plant meristems often remain virus free?

Do different 21- and 22-nucleotide vsiRNA populations have different mobility or protective ranges?

How do plasmodesmatal states influence the antiviral radius?

Can RNA carriers be designed specifically to enhance the population of mobile, AGO-competent siRNAs?

Could external dsRNA treatments establish overlapping three-to-four-cell protective zones before infection begins?

And can we measure these phenomena at single-cell resolution in crop plants rather than relying solely on whole-leaf measurements?

These questions connect fundamental RNA biology directly with the emerging field of RNA-based crop protection.

The bigger picture

One infected plant cell does not necessarily need to defeat the virus itself.

It may instead sacrifice that battle while transmitting molecular information that protects the cells surrounding it.

Those neighbors then form a temporary antiviral boundary.

The virus eventually advances.

Another boundary forms.

Again and again.

What appears at the whole-plant level as RNA-mediated disease resistance may therefore emerge from innumerable microscopic events occurring only a few cells apart.

The fascinating possibility raised by this study is that plant antiviral RNAi is fundamentally a spatial defense system.

Small RNAs are not merely molecular scissors.

They are also messengers.

And sometimes the most important message they carry is simple:

The virus is coming. Be ready.

---

Original study:

Albertini, D., Devers, E. A., Schott, G., Rajeswaran, R., Jacquemettaz, M., & Voinnet, O. (2026). Mobile virus-derived siRNAs drive plant antiviral silencing through reiterated 3–4 cell immunization. Cell Host & Microbe. DOI: 10.1016/j.chom.2026.08.005. The article was published online September 2, 2026, and is open access under a Creative Commons license.

FL0445: A Better Lipid for a Circular-RNA Delivery

In brief

A new branched ionizable lipid, FL0445, is presented as the vehicle that finally makes circular RNA deliverable in vivo. The delivery data are strong. The circular-RNA framing is not what the data support — and the claim underneath the framing is bigger than the one on the cover.

Kimura and colleagues at Nagoya University report FL0445, a multi-branched ionizable lipid formulated with DOPE into a lipid nanoparticle that outperforms MC3, SM-102 and ALC-0315 across an unusually broad cargo panel: linear mRNA, capped circular RNA at three sizes, an antisense oligonucleotide, plasmid DNA, and a model mRNA vaccine. The in vivo capstone is GLP-1-encoding capped circular RNA in obese ob/ob mice.

Breadth like this is rare. Most ionizable-lipid papers optimise against one cargo and one endpoint. Benchmarking a single formulation against three clinical-standard lipids across four modalities, with cell-level transfection mapping in Cre-reporter mice and a transcriptomic safety readout, is a real contribution regardless of how the individual claims settle.

But the paper's framing and its data pull in slightly different directions. The gap is worth naming precisely, because the claim the data actually support is the more interesting one.

The stated premise isn't supported by the paper's own measurements

The motivating argument is that circular RNA's rigid topology imposes steric constraints on encapsulation, and that lipids optimised for flexible linear mRNA therefore fail on it. FL0445's branched scaffold is proposed to solve this by generating a more flexible internal particle architecture.

The characterisation data don't show the problem. Encapsulation efficiency exceeded 90% for capped circular RNA and matched linear mRNA. Dynamic light scattering found no substantial size difference across lipid composition, payload size, or topology — and that includes SM-102, the benchmark supposedly ill-suited to rigid cargo.

If the incumbent lipid encapsulates circular RNA at over 90% and produces particles of the same size, the bottleneck the introduction invokes is not visible at the level the authors measured it.

What the data do show is that FL0445-LNP delivers everything better or comparably: 10- to 100-fold higher reporter expression in vitro with linear mRNA, more than threefold with circular RNA, superior ASO-mediated factor VII knockdown versus MC3, and plasmid DNA delivery on par with SM-102.

That is a cargo-agnostic advantage, not a topology-specific one. The honest headline is that FL0445 is a better general-purpose ionisable lipid whose advantage carries over to circular RNA — a broader and more useful claim than the one on the cover.

The structural hypothesis underneath remains untested. Cryo-TEM observations are reported qualitatively — circular-RNA formulations "tended to show more bleb-like or phase-separated structures", FL0445 particles appeared more angular — without quantification or blinding. Given that bleb morphology has been repeatedly linked to mRNA accessibility and expression, this is the observation most in need of numbers. The authors flag the absence of SAXS and SANS in their limitations, which is appropriate, but it leaves the central mechanistic premise as an interpretive frame rather than a finding.

Circularity is a small-payload, slow-route strategy

The most durable result here may be the one that cuts against the modality's usual pitch. For the 604-nucleotide NanoLuc payload, linear mRNA beat circular RNA in vitro at every measured timepoint, and after intravenous dosing produced both a higher peak and a higher two-week area under the curve.

Circular RNA only won at 200 nucleotides in vitro, and after subcutaneous administration — where it crossed over linear mRNA at roughly 30 hours and peaked at 72.

This is a genuinely useful design heuristic. The fractional stability gain from closing the loop is largest when the loop is short, and the depot kinetics of local injection give a slow-building, long-tailed expression profile somewhere to accumulate. It points circular RNA toward small secreted peptides delivered subcutaneously — precisely the GLP-1 case — and away from large protein-replacement cargo delivered systemically, which is where much of the field's enthusiasm currently sits.

The authors report this cleanly. The Highlights state it without the two conditions that make it true.

The mechanism is the weakest link, and the authors mostly say so

FL0445-LNP showed lower cellular uptake than SM-102 and ALC-0315 while achieving roughly tenfold higher expression. That dissociation is the paper's most interesting cell-biological observation, and it rests on three pieces of evidence that each carry caveats.

Methyl-β-cyclodextrin inhibition

Near-complete block of FL0445-LNP uptake is read as a distinct cholesterol-dependent entry route. But MβCD depletes plasma-membrane cholesterol globally, altering fluidity and inhibiting clathrin-mediated endocytosis as well; it is not a caveolae-specific tool. It can also extract cholesterol from the particles themselves in the medium. The authors name the single-inhibitor limitation explicitly, which is the right call.

The amiloride and chlorpromazine results

Uptake increased under both. The parsimonious reading — cargo redistributing to the remaining route — is consistent with the authors' model. A less flattering reading is that DiD, a lipophilic membrane dye that reports lipid position rather than RNA position and can transfer to serum lipoproteins, is not a reliable uptake tracer across chemically dissimilar formulations. The paper does not distinguish these.

Lysosomal colocalisation

No significant difference between formulations. This is the one direct measurement of the trafficking step the model depends on, and it does not support differential escape.

With the direct assay null, the "bypasses degradative routes, reaches ER-adjacent compartments" hypothesis rests on bulk liver and spleen RNA-seq showing lower phagocytosis-associated gene expression at three hours — a tissue-level transcriptional correlate standing in for a single-cell trafficking claim. The authors call it a working hypothesis. Readers should hold it as one.

Low inflammation and a stronger vaccine: a tension worth testing

The paper argues that FL0445's low immunostimulation explains its high transfection efficiency, citing the TLR4 → PKR → eIF2α translational-shutdown literature. It then reports that FL0445-LNP produces higher ovalbumin-specific IgG1, more antigen-specific CD8⁺ T cells, and greater antigen-specific killing than the more inflammatory SM-102.

If real, decoupling reactogenicity from immunogenicity is a significant result on its own — arguably more commercially consequential than anything else in the paper. But it sits awkwardly against the mechanistic argument in the preceding sections, where inflammatory activation is cast as the thing suppressing output.

The likely resolution is that antigen dose dominates adjuvanticity here, possibly helped by the enhanced lymph-node accumulation seen after intramuscular dosing. The paper asserts both halves and reconciles neither. A dose-matched antigen-expression comparison would settle it.

GLP-1: restraint worth crediting, comparator worth questioning

The authors repeatedly and correctly frame the ob/ob result as a functional proof of concept rather than efficacy. The glucose AUC did not reach significance (p = 0.056), body weight showed only a non-significant trend, and the positive finding is a single timepoint — 30 minutes post-challenge — among several, uncorrected for multiplicity. In a field where mouse GLP-1 data are routinely oversold, this restraint deserves saying out loud.

Two design points constrain what the experiment can show.

First, the dosing schedule conflates durability with cumulative dosing. Three doses at 0.4 mg/kg every other day, with the glucose tolerance test on day 17. The clean durability data come from a different experiment, with a different payload, after a single dose.

Second, the comparator is linear mRNA, not the standard of care. Against semaglutide, a marginal glucose shift from three RNA doses is not an efficacy argument. The modality's case has to be made on dosing interval and manufacturing rather than effect size — which this design does not yet test. The paper's own reference list indicates the GLP-1 circular-RNA space is already contested; direct comparison will come quickly.

The manufacturing footnote that isn't a footnote

The internal-cap design is elegant. Dispensing with a roughly 600-nucleotide IRES is what makes a 256-nucleotide GLP-1 construct possible at all, and small constructs are exactly where circularisation pays off.

But the synthesis route is a chemically synthesised 54-nucleotide capped oligonucleotide at 39.8% yield, splint-ligated to in vitro transcribed RNA, then PAGE-purified. For the GLP-1 construct: 28.9% ligation, 8.6% isolated yield. Those numbers compound, and denaturing PAGE does not scale.

Against IRES-based permuted-intron-exon circularisation from a single transcription reaction, this is a materially heavier process. For a modality whose pitch includes manufacturing simplicity, that belongs in the discussion rather than the supplement.

Relatedly, the low-cytokine claims for circular-RNA formulations would be strengthened by explicit quantification of residual nicked linear species and double-stranded RNA in the circular preparations — the standard confounder for circular RNA immunogenicity, and one the paper addresses for linear mRNA but not visibly for circular.

Six claims most exposed to challenge

Each pairs the load-bearing assertion with what the paper measured.

Claim 01

Rigid circular-RNA topology creates an encapsulation problem that FL0445 solves.

Contradicted by the paper's own greater-than-90% encapsulation efficiency and matched particle sizes across all lipids tested, including SM-102.

Claim 02

Branching creates internal flexibility that accommodates rigid cargo.

Asserted, with only qualitative cryo-TEM support. No SAXS or SANS, no quantification, no blinding.

Claim 03

A distinct cholesterol-dependent entry route explains the potency gain.

Single-inhibitor evidence, a null result on the one direct trafficking assay, and indirect transcriptional corroboration.

Claim 04

Capped circular RNA gives more durable expression than linear mRNA.

True only at small payload size and subcutaneous route. False for intravenous dosing at 604 nucleotides, by the paper's own two-week AUC.

Claim 05

The in vivo comparison is like-for-like.

FL0445 received composition-ratio optimisation in vivo; the benchmarks did not. Credit where due — the phospholipid comparison was symmetric, with all five ionisable lipids tested against both DOPE and DSPC.

Claim 06

The safety profile supports chronic dosing.

The tolerability study is single-dose. Anti-PEG responses, accelerated blood clearance and cumulative hepatic effects are untested — and repeat administration is the paper's own stated motivation.

What would settle it

• Genetic validation of the entry pathway — caveolin or flotillin knockouts, or an arrayed CRISPR screen of the kind already applied to MC3-LNPs — instead of pharmacological inhibition.
• Direct cytosolic-release quantification via galectin-8/9 recruitment or split-luciferase complementation, rather than lysosomal colocalisation area ratios.
• SAXS or SANS on matched linear- versus circular-payload particles, to test the flexibility hypothesis directly.
• A repeat-dose study over months, in the chronic-disease setting the paper invokes.
• A head-to-head against IRES-based circular RNA at matched protein output, with yields reported.

One practical caveat for replication: FL0445 and FL2266 are FUJIFILM materials covered by an existing patent application. Independent benchmarking will be gated by material transfer, which matters for a lipid being positioned as a benchmark-beating platform.

Bottom line

None of this diminishes what the paper does well. The cargo breadth is genuine, the payload-size dependence of circularisation is a finding the field can use tomorrow, and the authors' framing of the GLP-1 data as proof of concept rather than efficacy is the kind of restraint that should be more common. The overreach is in the packaging, not the bench work — and the claim underneath the packaging is the stronger one.

Frequently asked questions

What is a branched ionisable lipid?

An ionisable lipid is the component of a lipid nanoparticle that carries positive charge at acidic pH — letting it bind RNA during formulation and destabilise the endosomal membrane after uptake. "Branched" refers to the hydrophobic tail architecture: instead of straight chains, the tails split into multiple arms, which disrupts tight lipid packing and is thought to favour the non-bilayer phases associated with endosomal escape.

Why is circular RNA harder to deliver than linear mRNA?

The usual argument is that a covalently closed loop is topologically rigid and packs differently inside a nanoparticle. This paper's own data complicate that: encapsulation efficiency and particle size were essentially identical for linear and circular payloads across every lipid tested. Whatever advantage the new lipid confers appears to happen after encapsulation, not during it.

Does circular RNA always outlast linear mRNA?

No — and this paper is a useful corrective. Circular RNA won only at small payload size and after subcutaneous injection. At 604 nucleotides given intravenously, linear mRNA produced both a higher peak and a higher two-week cumulative expression.

Is this a viable route to a long-acting GLP-1 drug?

Not on this evidence. The glucose result is a single significant timepoint with a non-significant AUC (p = 0.056), no significant weight change, and a dosing schedule that can't separate durability from repeat administration. The authors themselves label it a proof of concept. The real test is a dose-response against an approved peptide agonist, on dosing interval rather than peak effect.

Source

Kimura S., Tsutsumi S., Nakamura N., et al. "A branched ionizable lipid nanoparticle platform for versatile in vivo delivery of nucleic acids: Validation from mRNA to capped circular RNA." Cell Biomaterials, 100555 (2026). Published online 19 August 2026.

What the next decade of small molecule–RNA research has to solve before it delivers medicines

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:

1. 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.
2. Open interaction data at scale. Watch for a multiplexed RNA–ligand dataset released precompetitively. Its absence caps every model in the field.
3. 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.
4. Chemical probing displacing crystallography as the structural foundation, with prediction models trained on in-cell secondary structure rather than PDB-derived idealizations.
5. 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/