[Paper Review] Designing RNAs with Language Models

Milan Gautam, Ning Dai+4 more|arXiv (Cornell University)|Feb 12, 2026

RNA and protein synthesis mechanisms0 citations

TL;DR

The paper reframes RNA design as conditional sequence generation with a pretrained autoregressive language model, using supervised learning on random-induced structure–sequence pairs and reinforcement learning on a small, selected set of structures to outperform state-of-the-art methods while being faster.

ABSTRACT

RNA design, the task of finding a sequence that folds into a target secondary structure, has broad biological and biomedical impact but remains computationally challenging due to the exponentially large sequence space and exponentially many competing folds. Traditional approaches treat it as an optimization problem, relying on per-instance heuristics or constraint-based search. We instead reframe RNA design as conditional sequence generation and introduce a reusable neural approximator, instantiated as an autoregressive language model (LM), that maps target structures directly to sequences. We first train our model in a supervised setting on random-induced structure-sequence pairs, and then use reinforcement learning (RL) to optimize end-to-end metrics. We also propose methods to select a small subset for RL that greatly improves RL efficiency and quality. Across four datasets, our approach outperforms state-of-the-art systems on key metrics such as Boltzmann probability while being 1.7x faster, establishing conditional LM generation as a scalable, task-agnostic alternative to per-instance optimization for RNA design. Our code and data are available at https://github.com/KuNyaa/RNA-Design-LM.

Motivation & Objective

• Reframe RNA design (inverse folding) as conditional sequence generation conditioned on target structures.
• Develop a reusable neural approximator (language model) to map target structures to sequences.
• Enforce biochemical validity through constrained decoding during generation.
• Combine supervised learning on structure–sequence pairs with reinforcement learning to optimize folding metrics.
• Show that random-induced SL data can yield competitive transfer to test sets and enable scalable RL with efficient structure selection.

Proposed method

• Frame RNA design as conditional seq. generation using an autoregressive LM conditioned on target structure.
• Introduce a constrained decoding mechanism that enforces base-pairing rules during generation to ensure valid designs.
• Adapt a pretrained GPT-style decoder LM (Qwen2.5–0.5B) via minimal architectural changes for RNA tokens and structure prompts.
• Train with supervised learning on a large random-induced structure–sequence dataset generated via SAMFEO to create yX_train (10M pairs).
• Fine-tune with reinforcement learning using a group-relative policy objective (GRPO) on selected structures to optimize a thermodynamics-based reward.
• Use a data-driven subset selection to pick RL targets based on sample diversity (AoN and NSD thresholds) to improve RL efficiency and quality.

Figure 1: RNA design is the inverse problem of RNA folding.

Experimental results

Research questions

• RQ1 Can RNA design be effectively solved as conditional sequence generation with a reusable neural solver?
• RQ2 Does constrained decoding plus pretrained LM adaptation yield valid, high-quality RNA designs for given target structures?
• RQ3 Does supervised learning on random-induced structure–sequence pairs transfer to test datasets and enable effective RL?
• RQ4 Can a small, diverse RL subset outperform training on larger, noisier RL sets in terms of efficiency and quality?
• RQ5 How do Boltzmann probability and ensemble-defect-based metrics behave under SL and RL regimes?

Key findings

• The best SL model achieves a best-of-N Boltzmann probability of ~0.55 at N = 10^4 samples.
• RL applied to a carefully filtered RL subset yields substantial efficiency gains (≈2.9x speedup) compared with using a larger, unfiltered set.
• Across four test sets, the approach outperforms state-of-the-art systems on key metrics and is 1.7x faster in sampling.
• A large SL training set of 10M random-induced structure–sequence pairs enables competitive performance on test data.
• Constrained decoding guarantees valid designs, enabling high-throughput, structure-conditioned generation with preserved biochemical constraints.

Figure 2: We convert a general-domain LLM into an RNA designer by keeping the pretrained transformer backbone and shrinking the input and output layers. The original embedding and LM head are downsized and reinitialized to support RNA tokens.