RNASeek uses a 1.6B-parameter cross-phyla transcriptomic model for RNA function prediction and GRPO-guided design, yielding ribozymes at wild-type activity and 3′ UTRs beyond the training data

RNASeek is a 1.6-billion-parameter generative foundation model built on a DeepSeek architecture and trained on a cross-phyla transcriptomic corpus, supporting RNA sequence representation and generation in one backbone, with natural-language tokens enabling flexible conditional prediction and sequence design. Although large language models have transformed natural language processing and protein design, a general framework connecting RNA foundation models to functional sequence design remained limited; this work places representation, prediction, and generation on a single backbone. The abstract reports model scale (1.6 billion parameters), architectural lineage (DeepSeek), and training corpus (cross-phyla transcriptomes), and states that natural-language tokens let a unified backbone support conditional prediction and design; specific training details and data scale are not given in the abstract.

In a zero-shot setting RNASeek captures species-specific transcript features and intron–exon boundaries; after fine-tuning it predicts ribozyme self-cleavage activity and viral mRNA stability, revealing interpretable sequence features associated with function, including ribozyme loop flexibility and stem stability as well as AU-rich motifs associated with mRNA stability. It links zero-shot transcript-feature recognition to downstream functional prediction and feature interpretation, giving the learned RNA function a readable sequence-level basis. The abstract describes zero-shot capture of species features and intron–exon boundaries and presents two fine-tuned functional prediction tasks with corresponding interpretable features; no specific performance numbers are provided in the abstract.

The authors use the functional predictors as reward models and apply Group Relative Policy Optimization (GRPO) to update RNASeek's generation policy toward sequences with desired properties, producing faster-cleaving ribozymes and stability-enhancing 3′ UTRs while satisfying user-specified IUPAC constraints. It brings reinforcement-learning-style policy optimization into RNA generation, so the generation process can be guided by functional predictors and controlled by user constraints, forming a unified pretrain–predict–optimize pipeline. The abstract explicitly states that functional predictors serve as reward models and that GRPO updates the generation policy, and it reports generation results in terms of cleavage speed, stability, and satisfaction of IUPAC constraints; reward design and training details are not expanded in the abstract.

Experimental validation shows that RNASeek-generated ribozymes achieve wild-type levels of activity, and generated 3′ UTR sequences exceed the performance of the training data and of benchmarked AI-generated 3′ UTRs. It moves generated sequences from computational evaluation to experimental validation and provides comparisons against the training data and existing AI-generated baselines. The abstract reports experimentally validated ribozyme activity at wild-type levels and 3′ UTR performance exceeding the training data and benchmarked AI-generated 3′ UTRs; experimental scale, quantitative metrics, and statistical details are not given in the abstract.

The work targets researchers and engineering teams who need to design and optimize regulatory RNAs at the sequence level, with use cases including ribozyme activity engineering and 3′ UTR design related to mRNA stability, and it can generate candidate sequences under user-specified IUPAC constraints. Its framework strings pretraining, functional prediction, and policy optimization into one pipeline, letting functional predictors act directly as reward signals in generation, thereby turning “learned RNA function” into “controllable de novo sequence design.” The abstract frames this route as a general strategy for engineering regulatory RNAs with desired properties, so its positioning is a transferable methodological paradigm rather than a result for a single RNA type.

Only the abstract has been read; the main text, figures, and supplementary materials were not part of this summary, so several key questions remain to be checked against the original: the specific composition and scale of the cross-phyla transcriptomic corpus, how zero-shot evaluation was judged, the performance metrics of the two functional prediction tasks, the design and training details of the GRPO reward models, and the sample sizes and quantitative comparison standards of the experimental validation. The abstract states that generated 3′ UTRs exceed the training data and benchmarked AI-generated 3′ UTRs but gives no specific values or statistical support, so readers wanting to assess effect size should return to the original. In addition, the abstract does not describe how the framework performs on regulatory RNA types beyond ribozymes and 3′ UTRs, which is a scope for further observation.