RNAGenesis: an AI platform for controllable design, structural prediction, and optimization of functional RNA sequences

Background

The field of RNA therapeutics has rapidly expanded in recent years, driven by the promise of RNA-based drugs for treating a wide range of diseases, from genetic disorders to cancer and infectious diseases. Designing functional RNA molecules with specific biochemical properties—such as high translation efficiency, optimal folding, or targeted binding—remains a central challenge, as the relationship between RNA sequence, structure, and function is highly complex and context-dependent. Traditional experimental approaches for discovering and optimizing RNA sequences are labor-intensive, costly, and often limited in scope. As a result, there is a growing demand for computational tools that can efficiently generate, predict, and optimize RNA sequences with desired properties, thereby accelerating the development of novel RNA therapeutics and expanding the range of treatable conditions.

Current computational approaches for RNA sequence design and optimization face several significant limitations. Many rely on heuristic or rule-based algorithms that do not capture the full complexity of RNA structure-function relationships, often resulting in suboptimal or non-functional sequences. Machine learning models trained on limited datasets may struggle to generalize to novel sequence or structural contexts, and most generative models are constrained by fixed input lengths, requiring padding or truncation that can distort biologically relevant information. Additionally, existing methods often lack the ability to directly control or optimize for multiple functional properties simultaneously, making it difficult to tailor RNA molecules for specific therapeutic applications. These challenges underscore the need for more advanced, flexible, and scalable computational frameworks that can accurately model, generate, and optimize RNA sequences across diverse functional and structural landscapes.

Technology

This technology is an advanced deep learning platform for the controllable generation, optimization, and structural prediction of RNA sequences, specifically tailored for RNA therapeutics. It integrates a multi-component architecture comprising a hybrid tokenization encoder, a Querying Transformer-based latent compression module, a latent diffusion model for generative tasks, and an autoregressive decoder for sequence reconstruction. The system leverages a billion-parameter foundation model, pretrained on millions of RNA sequences, and incorporates both gradient-based and beam search optimization strategies to guide sequence generation toward desired biochemical and functional properties. It is capable of inverse folding (designing sequences for specific structures), 3D structure prediction, and de novo structure generation, while also supporting experimental validation pipelines for aptamer binding and CRISPR sgRNA scaffold optimization. The platform’s modular design allows seamless integration of pretrained structural modules and latent reward models, enabling both high-throughput computational screening and robust wet-lab validation.

What differentiates this technology is its unified approach to sequence understanding, structural modeling, and functional design within a single generative framework. Unlike traditional models that struggle with variable-length sequences or require extensive padding and truncation, the Querying Transformer efficiently compresses sequences into fixed-length latent vectors, preserving biological relevance and enabling continuous latent space manipulation. The use of latent diffusion models, adapted from computer vision, allows for precise control over generated sequences and properties, while the hybrid tokenization scheme captures both nucleotide-level and higher-order RNA motifs. The system’s performance is validated through state-of-the-art results on multiple benchmarks and extensive experimental assays, demonstrating superior accuracy, efficiency, and generalizability across diverse RNA modalities. Its ability to directly link computational predictions with experimental outcomes, and to optimize for both structure and function, positions it as a transformative tool for RNA-based therapeutic discovery and research.

Benefits

• Enables controllable generation and optimization of RNA sequences with specific biochemical properties using advanced latent diffusion models.
• Handles variable-length RNA sequences efficiently without padding or truncation through fixed-length latent vector compression.
• Integrates sequence understanding, structural prediction, and de novo functional design within a unified framework.
• Supports gradient-based and beam search optimization for precise tuning of RNA functional properties.
• Demonstrates superior performance in RNA 3D structure prediction, inverse folding, and de novo structure generation.
• Validated experimentally with enhanced aptamer binding affinities and improved CRISPR sgRNA scaffold efficiencies.
• Achieves state-of-the-art results on multiple RNA functional benchmarks, showing broad applicability across RNA therapeutic modalities.
• Provides a scalable, modular architecture with robust training and inference protocols for RNA therapeutic design and research.

Applications

• De novo RNA therapeutic design
• CRISPR guide RNA optimization
• RNA aptamer discovery
• RNA structure prediction
• siRNA/shRNA efficacy prediction

Additional info

This deep learning system generates and optimizes discrete RNA sequences. It uses an auto-encoder with a Querying Transformer to compress variable-length RNA into fixed-length latent vectors. A latent diffusion model then generates new sequences, while a convolutional reward model guides this process based on predicted biochemical properties, enabling unified sequence understanding and functional design.