RNA design moves to discrete diffusion: structure-matched candidates without a folding oracle, and far more sequence variety
Synopsis
RNA inverse folding has long relied on iterative optimization scored by an external folding algorithm, which is sensitive to initialization and prone to local optima. This work defines multinomial diffusion directly over nucleotide sequences, guides denoising with a target secondary structure, and compares a 2D contact-map projection with a purely 1D pairing-aware architecture. On the in-distribution test set, the 2D and 1D models reach median normalized RNAforester similarities of 0.82 and 0.79, while generated sequences stay close to random-sequence diversity.
Interpretation
RNA sequence design can be modeled generatively directly in discrete nucleotide space, without first mapping sequences to continuous vectors and adding Gaussian noise. Earlier inverse-folding methods mostly relied on heuristic search, reinforcement learning or iterative optimization, with a folding algorithm evaluating candidates at each step. On the in-distribution test set, the 2D and 1D models reach median normalized RNAforester similarities of 0.82 and 0.79, with means of 0.69 and 0.70, indicating good matches for most targets but some failing samples.
Structural guidance enters the denoising network in two ways: one projects sequence embeddings into a 2D representation aligned with the contact map, the other keeps a 1D sequence representation and modulates attention with base-pairing partner information. Multinomial diffusion replaces Gaussian noise with categorical transitions, so the diffusion process acts directly on the discrete alphabet of A, C, G and U. Evaluation uses 1,148 bpRNA test sequences screened by CD-HIT-EST to share no more than 80% identity with training or validation data, with structural agreement measured as normalized RNAforester similarity between MXfold2 predictions and the target.
Generation needs no external folding oracle, and inference time grows approximately linearly with sequence length. LEARNA and SAMFEO embed thermodynamic folding information directly in their optimization loops, and their runtime grows non-linearly with sequence length, which the source says likely reflects dependence on target-structure complexity. On target structures of 50 to 165 nucleotides, median time per sequence rises approximately linearly with length, with the 1D architecture giving the highest throughput; because the baselines run mainly on CPU and the diffusion models on GPU, absolute times are not compared on equal hardware.
Perspective
Evaluation covers a sequence-identity-screened in-distribution bpRNA test set and the expert-curated Eterna100 benchmark, with structural agreement referenced to MXfold2 predictions; partial sequence completion experiments were run only on the 1D model.
On Eterna100, the diffusion models' median normalized RNAforester similarities fall below LEARNA and SAMFEO; structural information is learned mainly from sequence-structure associations in training data rather than an explicit thermodynamic objective; randomly fixed positions can be incompatible with the target structure.