{
  "id": 10399559,
  "title": "RNASeek: A Cross-Phyla Generative Foundation Model for Multipurpose RNA Modeling and Reinforcement Learning-Based Design",
  "url": "https://urgent.news/2026/09/28/rnaseek-a-cross-phyla-generative-foundation-model-for-multipurpose",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-28T00:00:00.000Z",
  "source": {
    "name": "bioRxiv",
    "slug": "biorxiv",
    "url": "https://www.biorxiv.org/content/10.64898/2026.09.24.754173v1?rss=1"
  },
  "original_language": "en",
  "account": "RNA is essential for regulating information flow and serves as a flexible platform for designing biological functions. Although large language models have revolutionized natural language processing and protein design, a comprehensive framework bridging RNA foundation models to functional sequence design remains scarce. In this study, the researchers introduce RNASeek, a 1.6-billion-parameter generative foundation model constructed upon a DeepSeek architecture and trained on a diverse transcriptomic dataset spanning various species for RNA sequence representation and generation. By incorporating natural-language tokens, RNASeek facilitates flexible conditional prediction and sequence design through a single backbone. Remarkably, RNASeek can identify species-specific transcript features and intron-exon boundaries in a zero-shot setting.\n\nThe researchers then fine-tune RNASeek to predict ribozyme self-cleavage activity and viral mRNA stability, uncovering interpretable sequence features linked to functionality. These features include ribozyme loop flexibility, stem stability, and AU-rich motifs tied to mRNA stability. Utilizing these functional predictors as reward models, the researchers apply Group Relative Policy Optimization (GRPO) to adjust RNASeek's generation policy towards sequences with desired properties. GRPO-guided generation yields faster-cleaving ribozymes and stability-enhancing 3' UTRs while adhering to user-specified IUPAC constraints.\n\nThe experimentally validated RNASeek-generated ribozymes exhibit wild-type levels of activity, while the RNASeek-generated 3' UTR sequences surpass both the performance of the training data and AI-generated benchmark 3' UTRs. In summary, RNASeek establishes a unified pretrain-predict-optimize framework that connects learned RNA function to controllable de novo sequence design, offering a general strategy for engineering regulatory RNAs with specific desired properties.",
  "summary": "RNA plays central roles in regulating information flow and provides a versatile substrate for engineering biological functions. While large language models (LLMs) have transformed natural language processing and protein design, a general framework connecting RNA foundation models to functional sequence design remains limited. Here, we present RNASeek, a 1.6-billion-parameter generative foundation…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}