Modern drug discovery is built on an implicit premise: drugs act on proteins. Whether small molecules or antibodies, the essence is to find a pocket or epitope on a disease-causing protein, then bind to it and inhibit it. This paradigm has supported the pharmaceutical industry for nearly a century, and it is now approaching its ceiling—only about 1.5% of the human genome encodes proteins, of which only about ten percent are disease-related, and fewer than 700 proteins have actually been targeted by approved drugs globally to date. The vast majority of remaining disease-causing proteins—transcription factors, scaffold proteins, "undruggable" targets lacking active pockets—are beyond the reach of traditional methods. The "low-hanging fruit" has been picked clean, and protein targets are being exhausted.

Since proteins are difficult to tackle, why not move upstream to the RNA that encodes them—this is a natural thought. But what truly blocks this path is an old problem in chemistry and structural biology: RNA is too "soft." It is highly flexible, structurally dynamic, and changes with the cellular environment. Pockets available for small molecule binding are both difficult to find and difficult to validate, and resolving RNA three-dimensional structures is expensive and slow—this has kept small molecules targeting RNA stuck in a stage of "theoretically feasible, practically a matter of luck" for a long time. In the past three years, AI has恰好 provided a solution at this bottleneck: from RNA secondary and tertiary structure prediction, to direct prediction of small molecule–RNA interactions and binding sites, to the accumulation of large-scale RNA structural omics data, a path of "AI-driven RNA-targeting small molecule discovery" is taking shape.

I. Mechanistic Basis: Why RNA Can Be Targeted by Small Molecules
1.1 Target Space: From Fewer Than 700 Proteins to the Entire Transcriptome
The target space of traditional pharmaceuticals is limited by a triple narrowing: only about 1.5% of the human genome encodes proteins; of these, about one to one and a half percent are disease-related; and fewer than 700 have actually been successfully targeted by existing drugs. More troublesome is that most of the remaining unconquered proteins are precisely the "hard bones" lacking ideal binding pockets—this is the structural reason for continuously rising R&D costs and declining success rates.
Shifting the view to RNA, the space suddenly opens up: RNA sequences account for about seventy percent of the human genome. In addition to messenger RNA, a large number of non-coding RNAs (lncRNAs, microRNA precursors, intronic regulatory elements, etc.) have been confirmed to participate in disease occurrence. Targeting RNA means expanding the proportion of the druggable human genome by at least two orders of magnitude, and it can "strike from a higher dimension" at targets that are undruggable at the protein level—without waiting for the protein to fold into a pocket, one can intervene before it is even translated.
1.2 RNA Is Not "Structureless": From Sequence to Three-Dimensional Pockets
A common misconception is that RNA is just a linear sequence. In reality, single-stranded RNA folds through complementary base pairing into secondary structural units such as stem-loops, bulges, internal loops, and pseudoknots, and further stacks into three-dimensional structures with defined spatial conformations. It is precisely these non-double-helical, locally irregular structural regions that form pockets into which small molecules can embed and bind specifically.
This also determines the natural division of labor between two technological routes: oligonucleotides (ASO/siRNA) based on complementary base pairing are suitable for targeting non-structured regions of RNA; whereas highly structured RNA regions are more suitable for small molecules to recognize their spatial conformation. The two are not substitutes, but cover different segments of RNA with different properties.
1.3 Spectrum of Mechanisms of Action: Modulating Splicing, Translation, and RNA Degradation
Small molecules targeting RNA can be divided into three categories by mode of action:

• Splice modulation: Binding to splice sites or regulatory elements of precursor mRNA, altering exon selection, thereby restoring or blocking the production of specific proteins. This is currently the most clinically validated category—risdiplam restores functional SMN protein by promoting inclusion of SMN2 exon 7.

• Translation and stability modulation: Binding to 5′/3′ untranslated regions, internal ribosome entry sites (IRES), or regulatory stem-loops, altering mRNA translation efficiency or half-life, thereby reducing the expression level of disease-causing proteins.

• Induced degradation (RIBOTAC): This is an ingenious design isomorphic to PROTAC—linking a "small molecule that can bind target RNA" with a "heterocyclic group that can recruit and locally activate ribonuclease RNase L," converting an originally inactive RNA-binding molecule into a catalytically active RNA degrader. Its pharmacological significance is consistent with protein degraders: shifting from "occupancy-based inhibition" to "event-driven catalytic clearance."

II. Divergence from Similar Modalities: Small Molecules vs. Small Nucleic Acids
RNA therapeutics itself is a large family, including ASOs and siRNAs targeting RNA sequences, aptamers targeting proteins, and mRNA therapeutics encoding antigens or proteins. The global small nucleic acid drug market has grown from approximately $2.7 billion in 2019 to approximately $5.7 billion in 2024 (compound growth rate of about 16%), and is widely expected to reach the scale of twenty billion dollars around 2029.
In the goal of "downregulating a certain disease-causing protein," small molecules and small nucleic acids constitute two parallel routes, and their differences directly determine indications and business models:

The truly scarce value of the small molecule route lies in "accessibility": oral administration, no delivery system needed, ability to cross the blood-brain barrier, room-temperature storage and transport, and low production cost—the superposition of these points gives it irreplaceable advantages over small nucleic acids in scenarios such as central nervous system diseases and chronic disease-style long-term administration. The price is a sharp increase in design difficulty: you must first know what the RNA looks like before you can design a molecule that can embed into it. This is precisely AI's entry point.

III. AI Breakthrough: Three Layers of Tools Unlocking the "Structure" Lock
The historical bottlenecks of RNA-targeting small molecules can be summarized as three: scarcity of RNA structural data, complexity and dynamics of RNA structure, and the fact that the same RNA has different conformations in different cellular environments, making specificity difficult to guarantee. AI's intervention corresponds precisely to three levels of tools.
3.1 First Layer: RNA Structure Prediction
In the protein field there is AlphaFold, but the RNA field long lacked a counterpart—even though AlphaFold3 has extended its capabilities to nucleic acids, its prediction accuracy for RNA remains significantly weaker than for proteins. This gap has spawned a batch of specialized RNA structure prediction efforts, among which Chinese teams have made significant contributions: Professor Zhou Yaoqi's team has successively produced SPOT-RNA (secondary structure prediction), BRiQ-RNA (tertiary structure prediction), RNA-MSM (RNA language model), and the MARS database, and is one of the international leaders in this direction.
3.2 Second Layer: Small Molecule–RNA Interaction and Binding Site Prediction
Even with structures, "which small molecule can bind to which site on RNA" remains an independent problem, and traditionally must rely on expensive three-dimensional structure determination. In January 2026, the team of Zhang Qiangfeng at Tsinghua University and the team of Wang Yangming at Peking University published the AI method SMRTnet in Nature Biotechnology, providing a key breakthrough: without needing RNA three-dimensional structure, using only secondary structure information, it can precisely predict small molecule–RNA interaction relationships and binding sites. According to its published data, compared with traditional methods, screening efficiency is improved by about 200 times, with an average hit rate of 21.1%—this directly pushes "luck-based screening" toward a "predictable, iterable" engineering path.
3.3 Third Layer: Data Infrastructure—RNA Structural Omics
AI's upper limit is determined by data, and this is precisely the weakest link in the RNA field: the scale of publicly available RNA structural data is far smaller than that of proteins. Therefore, whether one can self-build a large-scale, high-quality in vivo RNA structural omics database is becoming the hardest barrier in this track—unlike algorithms that can be reproduced from papers, it requires long-term investment in experimental platforms to "produce" data. This is also the fundamental difference between this niche and most AI pharmaceutical companies: the focus of competition is not on model parameters, but on data exclusivity.

IV. Clinical and Investigational Evidence
4.1 Proof of Existence: Risdiplam
Risdiplam (brand name Evrysdi / Aimaxin, originally developed by PTC Therapeutics, developed by Roche) is the foundational asset of this route: it is an oral small molecule that restores functional SMN protein by acting on the splicing process of SMN2 precursor mRNA, used to treat spinal muscular atrophy (SMA), and has been approved for marketing (PharmCube annotates its mechanism as "SMN2×U1 snRNP non-degradative molecular glue / SMN2 splicing modulator").
Its significance goes beyond one drug: in the disease of SMA, three modalities—gene therapy (Zolgensma), ASO (Spinraza/nusinersen), and oral small molecule (Evrysdi)—compete on the same stage, and the oral small molecule, with its convenience of administration and accessibility, has captured a considerable market share—this provides a commercial model precedent for "RNA-targeting small molecules," not just scientific feasibility.
4.2 Case Study: SCA3 (ATXN3)—Six Modalities Competing on the Same Stage
Spinocerebellar ataxia type 3 (SCA3, also known as Machado-Joseph disease) is the best window to observe modality competition. According to PharmCube data, around the same target ATXN3, at least six technological routes are currently being advanced:

This table illustrates two things: first, rare neurological diseases are a common testing ground for various new modalities because targets are clear and mechanisms are well-defined; second, on the same target, the direct competitors of the small molecule route are international players like Skyhawk and PTC, while the ASO camp's BIIB132 has entered Phase I—for small molecules to win, it is not by being "earlier," but by structural advantages such as oral administration, brain penetration, and long-term medication accessibility.

V. Industry Landscape: Four Major BD Deals and China's Technology Cluster
5.1 MNCs Cast Votes of Confidence with Four Major BD Deals
Although RNA-targeting small molecules are overall at an early stage, multinational pharmaceutical companies have continuously placed bets at the level of hundreds of millions to billions of dollars (PharmCube transaction data):

Within four years, four MNCs, cumulative outbound collaborations exceeding $5 billion, and the transaction structure is mostly "platform access + multi-target," indicating that what big pharma is buying is not a single molecule, but platform discovery capability—this is precisely the value of AI-driven companies.
5.2 China: A Real Technology Cluster Has Formed
This niche in China is not following, but starting simultaneously with the world, and has gathered a group of teams with hardcore academic backgrounds:

• Xunjing Biotech (founded 2023): Co-founded by Professor Zhang Qiangfeng of Tsinghua University (Director of the Center for AI in Life Sciences, Tsinghua University) and Professor Wang Yangming of Peking University (Director of the Institute of Molecular Medicine, Peking University), taking the "AI + RNA structural omics" route, self-building a large-scale RNA structure database and multimodal RNA large models, with core algorithm SMRTnet published in Nature Biotechnology in January 2026; pipeline layout in neurological rare diseases (SCA3/ATXN3, candidate drug XGEN-049) and oncology (Myc). Angel round led by Cowin Capital, with participation from Shuimu Tsinghua, Inno Angel, and Capital Technology Development Group; Efung Capital participated in its Pre-A round investment.

• Libobio (founded 2022, incubated by Shenzhen Bay Laboratory): Scientific founder Professor Zhou Yaoqi has深耕 structural computation for nearly thirty years and is an international leader in RNA structure prediction (MARS database, RNA-MSM language model, SPOT-RNA, BRiQ-RNA); completed nearly 100 million yuan Pre-A round in August 2025, co-led by Tasly Capital and Panlin Capital.

• ReviR Therapeutics: Advancing the HTT-PMS1 dual-target project for Huntington's disease (RTX-317) with the VoyageR AI platform, attempting to simultaneously reduce mutant huntingtin protein and inhibit CAG repeat expansion, received $4.6 million funding from the California Institute for Regenerative Medicine (CIRM) in November 2025.

• Kaiyue Life Sciences: Oral small molecule KY2 targeting RNA helicase DHX33, advancing in dual US-China filings.

• HitGen: Using DNA-encoded compound libraries (DEL) to directly perform affinity screening against RNA targets, entering this track from the chemical tools end.

Looking at global and Chinese players side by side, a clear differentiation can be observed: at the algorithm level, methods for RNA structure prediction and interaction prediction are rapidly becoming public (top journal papers, open-source models); while what is truly difficult to replicate are two things—large-scale in vivo RNA structural omics data, and the wet-lab closed loop that turns prediction results into real lead compounds. This also explains why MNCs' collaboration targets are all "companies with platforms" rather than "laboratories with algorithms."

VI. Summary and Outlook
From targeting proteins to targeting RNA, this is essentially an expansion of druggable space: moving the intervention point up from protein to RNA theoretically expands the proportion of the genome that can be reached by two orders of magnitude. Risdiplam has already proven this path is viable, four MNC deals cumulatively exceeding $5 billion prove the industry is willing to pay for it, and AI—especially interaction prediction that no longer relies on three-dimensional structure—is turning this path from "luck" into "engineerable."
But soberly viewed, several key issues remain unresolved:

1. Specificity is the greatest scientific risk. Similar stem-loop structures abound in the human transcriptome, and the same conformation may appear on multiple RNAs; how to ensure a small molecule acts only on the target RNA without off-target effects is a more thorny problem than protein targets—proteins at least have relatively unique three-dimensional folds.

2. The dynamics of RNA structure still lack a mature characterization method. RNA conformation changes in different cell types and different metabolic states, and "the structure measured in vitro" may not equal "the structure that works in vivo," which directly affects the transferability of predictions.

3. Data scarcity remains AI's upper limit. Compared with the accumulation of protein structure databases, the scale of RNA structural data is still several orders of magnitude smaller, and model capability is limited by this—this is both a bottleneck and a moat for those who build their own data.

4. The common problem of AI pharmaceuticals: value realization comes later. AI currently mainly acts at the earliest stage of drug discovery, while the bulk of R&D costs lies in clinical stages. Platform capability ultimately still needs to be validated by differentiated molecules and clinical data.

5. The dilemma of indication selection. Rare neurological disease targets are clear and suitable for validating platforms, but the commercial ceiling is limited; to move toward large indications, one must directly face more complex target biology and more crowded competition.

Looking ahead two to three years, three clues are worth tracking: First, RIBOTAC moving from concept to clinic—upgrading "binding" to "catalytic degradation," whose pharmacological gain is similar to that of PROTAC versus occupancy inhibitors; Second, moving from rare diseases to major disease types, especially central nervous system neurodegenerative diseases where the advantage of brain penetration can be best realized; Third, whether the "AlphaFold moment" for RNA structure prediction will arrive—if RNA three-dimensional structure prediction accuracy undergoes a leap similar to the protein field, the efficiency of the entire track will be redefined.
RNA-targeting small molecules are essentially answering a more fundamental question: What exactly determines the objects drugs can act on? The past answer was "proteins with pockets," and now this answer is being rewritten.

Disclaimer: This article is a science and industry-level popular science sharing, for industry exchange and learning reference only, and does not constitute any investment advice or offer, nor does it represent Efung Capital's investment views, investment strategies, or investment decisions; the companies and project information involved in the article are all from public channels or disclosed information, and do not recommend any specific target.

References and Data Sources

1. PharmCube innovative drug pipeline and transaction database, search date 2026-09-01.

2. Yang Y, Zhang Q, Wang Y, et al. SMRTnet: AI-based prediction of small molecule–RNA interactions without 3D structure. Nature Biotechnology, 2026-01 (Tsinghua University Zhang Qiangfeng team × Peking University Wang Yangming team).

3. Nature Biotechnology 2023 Top Ten Biotech News of the Year: Small molecule targeting RNA ranked third.

4. Small molecule approaches to targeting RNA. Nature Reviews Chemistry, 2024-01 (University of Nice, French National Centre for Scientific Research, etc.).

5. Related research on RIBOTAC (ribonuclease-targeting chimera) design, Nature, 2024.

6. PharmCube pipeline database: Risdiplam (Evrysdi, PTC/Roche, SMN2 splicing modulator, approved for marketing).

7. PharmCube pipeline database: Global investigational landscape of ATXN3 target (BIIB132, ARO-ATXN3, AMT-150, AAVhfCas12Max-gATXN3, Skyhawk and PTC's ATXN3 splicing modulators, etc.).

8. PharmCube transaction database: Roche × Remix (2024-01), Ipsen × Skyhawk (2024-04), GSK × Rgenta (2024-12), Merck × Skyhawk (2025-08).

9. Public reports: Xunjing Biotech angel round financing (led by Cowin Capital); Libobio nearly 100 million yuan Pre-A round (led by Tasly Capital, Panlin Capital); ReviR Therapeutics received $4.6 million CIRM funding.

10. Zhou Yaoqi team RNA structure prediction series work: SPOT-RNA, BRiQ-RNA, RNA-MSM, MARS database.

Recommended Reading

• Efung Hardcore Sharing "From Rewriting Sequences to Rewriting 'Software': The Persistent Silencing Paradigm of Epigenome Editing"—belongs to the same "intervention at the regulatory layer without changing the genome," and is mutually referential with this article.

• Efung Hardcore Sharing "From Rewriting DNA to Editing RNA: The Druggable Paradigm of Programmable A-to-I RNA Base Editing"—both take RNA as the object of action, one uses oligonucleotides to rewrite sequences, the other uses small molecules to recognize structures.

• Efung Hardcore Sharing "From Cytotoxic Payloads to Targeted Protein Degradation: The Paradigm Evolution of Functional Payloads for Antibody-Drug Conjugates (DAC/MAC)"—the isomorphic design of RIBOTAC and PROTAC can be read in comparison.