In a breakthrough that could revolutionize RNA therapeutics and synthetic biology, researchers have unveiled a powerful artificial intelligence (AI) framework that enables the precise control and design of internal ribosome entry sites (IRES) for programmable RNA translation. IRES elements are specialized RNA sequences that facilitate cap-independent translation initiation, a process crucial for the production of proteins from RNA molecules. Historically, the intricate interplay between IRES structure and function has posed significant challenges, rendering their rational design and optimization difficult. This new comprehensive AI-driven approach not only overcomes these obstacles but also opens up unprecedented possibilities for scalable RNA-based therapeutics.
From the outset, the study highlights the limitations that conventional methods face in controlling protein expression through RNA constructs. Since protein production is a pivotal step in therapeutic efficacy, the ability to fine-tune translation initiation independent of the cellular cap-binding machinery offers a strategic advantage. The research calls attention to IRES as versatile molecular tools because they enable translation initiation without relying on the 5’ cap structure typically required by eukaryotic ribosomes. Yet, the complexity of natural IRES sequences, which interact with a myriad of RNA-binding proteins and possess elaborate three-dimensional conformations, necessitates sophisticated computational tools for reliable identification and design.
Central to this advancement is IRES-LM, an innovative language model ensemble comprised of two deeply trained natural language processing architectures. Trained on an extensive dataset of over 46,000 sequences, IRES-LM surpasses previous benchmark methods by achieving a 15% improvement in critical performance metrics such as the area under the curve (AUC) and F1 score. This improvement is not merely incremental but signifies a robust leap in the capacity to accurately predict linear mRNA IRES elements, which are pivotal in therapeutic mRNA design. Impressively, IRES-LM also showcases remarkable versatility by demonstrating strong cross-applicability to circular RNA IRES identification, correctly pinpointing all 21 experimentally verified circular RNA IRES elements—a feat that existing tools struggled to achieve.
Building upon this predictive prowess, the research team integrated an evolutionary algorithm with IRES-LM, resulting in the creation of IRES-EA. This synergistic approach harnesses targeted mutagenesis guided by AI predictions to drive the conversion of non-IRES sequences into functional IRES elements. The scale of this approach is staggering: computational analyses of over 37,000 sequences initially lacking IRES functionality predicted a 60% success rate in functional conversion. These computational predictions were backed by experimental validation through massively parallel reporter assays involving 12,000 mutated sequences, revealing a remarkable 98.4% acquisition of IRES activity. This convergence of in silico prediction and wet-lab validation underscores the framework’s ability to induce precise functional transformations efficiently.
Extending the frontier even further, the researchers introduced IRES-DM, a diffusion model designed to generate novel IRES sequences de novo. Unlike evolutionary optimization, which works incrementally, IRES-DM creates entirely new sequences from fundamental principles encoded in the trained model. This generative capability has major implications for synthetic biology, where creating unique, tailored RNA elements can circumvent natural sequence limitations. Validated by another extensive massively parallel reporter assay involving 12,000 AI-generated sequences, 99.3% exhibited detectable IRES function, thereby establishing de novo generation as an effective and reliable avenue for RNA element design.
A compelling feature of IRES-DM’s generative capacity is its ability to produce a diverse range of sequence variants. It can generate sequences that mirror natural IRES candidates as well as structurally conserved sequences that diverge significantly at the nucleotide sequence level. This balance between biomimicry and innovation is crucial for applications that demand both predictability and novelty in RNA design. Structural conservation is particularly significant because it underlies the functional integrity of IRES elements, highlighting the model’s sophisticated grasp of structure-function relationships.
The study also delves into motif analysis to dissect the essential building blocks underpinning IRES activity. By mining both natural and AI-generated sequence pools, researchers identified motifs highly enriched in functional IRES elements. Some motifs are prevalent in naturally occurring sequences, while others emerge predominantly in AI-designed sequences with high IRES activity. This insight not only aids in understanding the molecular grammar of translation initiation but also guides future rational design and synthetic biology efforts by pinpointing key RNA features to embed in engineered constructs.
The fusion of deep learning and evolutionary algorithms presented in this work exemplifies the potential of AI to accelerate biomedical discovery. The framework’s integrated strategy—from identification to optimization, and finally de novo generation—offers a scalable solution to one of the biggest hurdles in RNA therapeutic development: modulating translation with precision. It effectively transforms the longstanding challenge of deciphering IRES’s complex structure-function interplay into a programmable, user-driven process.
Moreover, the broad applicability of this framework to both linear and circular RNA modalities expands its utility across diverse RNA therapeutic platforms. Circular RNAs, which are gaining traction due to their enhanced stability and translational potential, previously suffered from limited tools for IRES characterization. By addressing this gap, the AI framework paves the way for next-generation RNA therapeutics with superior efficacy and durability.
This research also holds promise beyond therapeutics, offering synthetic biologists a new suite of tools for the design of custom RNA elements for biosensors, gene circuits, and synthetic protein expression systems. The ability to program translation through AI-directed sequence design could transform the speed and accuracy with which synthetic biological systems are engineered and optimized.
Emphasizing the interplay between computational and experimental sciences, the extensive massively parallel reporter assays employed in this study validate the models’ predictions at an unprecedented scale. This experimental rigor ensures that the AI-generated sequences are not just theoretically appealing but functionally robust in cellular contexts, addressing a common pitfall in computational biology.
As RNA-based medicines continue to expand their footprint in treating cancers, genetic disorders, and infectious diseases, the technological leap demonstrated by this AI framework offers a transformative platform. By bridging the knowledge gap in IRES biology and enabling precise, scalable control over RNA translation, it sets the stage for a new era of personalized, programmable RNA therapeutics.
Ultimately, this work exemplifies the power of artificial intelligence to decode and harness complex biological information, translating it into practical tools that can reshape biomedicine. The researchers’ accomplishment in unifying IRES identification, optimization, and de novo design into a single, cohesive framework heralds a new chapter in RNA science, where computational ingenuity accelerates discovery and innovation.
Subject of Research:
Article Title:
Article References:
Chu, Y., Yin, D., Yu, D. et al. Programmable RNA translation through deep learning-driven IRES discovery and de novo generation.
Nat Mach Intell 8, 559–574 (2026). https://doi.org/10.1038/s42256-026-01213-z
Image Credits: AI Generated
DOI: April 2026
Keywords: RNA therapeutics, internal ribosome entry sites, IRES identification, IRES optimization, de novo RNA design, deep learning, language models, evolutionary algorithms, diffusion models, massively parallel reporter assays, synthetic biology, programmable translation, circular RNA, artificial intelligence
