yakRNA: A Multimodal RNA Language Model Ushering in a New Era of Nucleic Acid Sequence Design

yakRNA is a deep learning-based multimodal RNA sequence generation model with 110 million parameters, supporting RNA design under multiple conditional constraints such as secondary structure, consensus sequence, and Gene Ontology (GO) terms. This open-source tool provides strong support for bioinformatics and synthetic biology research, ushering in a new era of RNA sequence design.

RNA molecules play key roles in biological systems (e.g., transmitting genetic information, catalyzing protein synthesis, regulating gene expression). With the development of synthetic biology and RNA therapeutics, the demand for precisely designing RNAs with specific functions and structures is growing. Traditional methods based on physicochemical simulation or experimental screening are time-consuming and labor-intensive, while artificial intelligence (especially large-scale language models) brings revolutionary possibilities to this field.

yakRNA is a multimodal language model specifically designed for RNA sequence design. Unlike ordinary text generation models, its training goal is to understand and generate RNA sequences that comply with biophysical constraints. Its core capabilities include five generation modes: unconditional generation (target length only), secondary structure-constrained generation, consensus sequence-constrained generation, GO term-constrained generation, and sequence infilling. These modes can be used individually or in combination to achieve multimodal conditional generation.

• Secondary structure constraint: Supports the dot-bracket notation (e.g., "((((....))))") to specify the target structure, and provides five constraint strengths (strict/classic/classic+clipping/classic+common/relaxed) to adapt to different application scenarios.
• GO term constraint: Innovatively supports GO terms (e.g., "GO:0075523" corresponds to viral transcription inhibition) as generation conditions, directly using biological terms to describe the target function.
• Consensus sequence constraint: Integrates evolutionary conservation information to generate new sequences that retain family functional characteristics, and can be used in combination with secondary structure constraints.

Application Scenarios: Multimodal combined generation can meet complex needs (e.g., RNA drug design requires simultaneous consideration of structural stability, functional conservation, and targeting); it can be used to design riboswitches, aptamers, ribozymes, or optimize mRNA vaccine stability and reduce immunogenicity. Deployment and Usage: Model weights are hosted on Hugging Face, supporting CLI and Python API; Google Colab notebooks are provided (usable even without GPU resources); environment requirements: Python 3.10, 16GB memory, NVIDIA GPU + CUDA recommended; cross-platform support (corresponding Conda configurations for Linux/macOS).

yakRNA combines deep learning with biophysical constraints to provide a powerful and flexible tool for RNA design. Its MIT open-source license encourages wide contributions, and it is expected to play an important role in RNA therapeutics and synthetic biology. For researchers in related fields, this is an open-source project worth paying attention to and trying.