Reproduce “Designing RNAs with Language Models” on OpenResearch

About this paper

Milan Gautam, Ning Dai, Tianshuo Zhou, Bowen Xie, David Mathews, Liang Huang

A framework utilizes a pretrained language model with constrained decoding and a combined supervised and reinforcement learning approach to design RNA sequences that fold into specific target secondary structures. The method achieves state-of-the-art performance on ensemble-based folding metrics and demonstrates a 1.74x speedup over previous state-of-the-art heuristic solvers.

Key insights

• Supervised learning on a large, randomly generated dataset of structure-sequence pairs provides a strong initial model, surprisingly yielding competitive performance even when unrelated to test data.
• Constrained decoding is essential for guaranteeing biochemical validity of generated sequences, reducing invalid outputs and significantly improving effective sample efficiency and training stability despite a modest 30% computational overhead.
• Leveraging a pretrained general-domain language model backbone (Qwen2.5 –0.5B) is highly effective, demonstrating successful transfer learning and outperforming models trained from scratch on RNA-specific tasks.
• Careful selection of a high-quality, filtered dataset for reinforcement learning is crucial for maximizing RL efficiency and preventing performance degradation on unseen structures.

Summary by alphaXiv.