IRES-EA merges IRES-LM with an evolutionary algorithm to guide targeted mutagenesis, predicting a 60% success rate for converting non-IRES sequences into functional ones, with 12,000 mutated sequences experimentally validated and 98.4% acquiring IRES activity.

This framework enables precise control of translation for both linear and circular RNAs, supporting scalable design of RNA therapeutics, biosensors, gene circuits, and programmable expression systems, backed by strong experimental validation of computational predictions.

An ensemble language model trained on over 46,000 sequences improves linear mRNA IRES prediction by about 15% in AUC and F1, and extends effectively to circular RNA IRES identification, correctly matching all 21 experimentally verified circular IRES elements.

The AI-driven framework for identifying and optimizing IRES demonstrates sequence versatility, producing biomimetic and diverse, yet conserved variants, with motif analysis revealing key RNA features enriched in functional IRES elements from both natural and AI-designed pools.

In short, researchers present an AI-driven platform to identify, optimize, and de novo generate IRES elements for programmable, cap-independent RNA translation with wide applications in RNA therapeutics and synthetic biology.

IRES-DM, a diffusion-model–based generator, creates de novo IRES sequences; 12,000 AI-generated sequences tested via massively parallel reporter assays showed 99.3% detectable IRES function.

The work combines deep learning with evolutionary methods to streamline identification, optimization, and de novo design of IRES elements, potentially accelerating RNA-based medicine and expanding synthetic biology tools.