Protein-conditioned RNA sequence and structure co-design, refining flow-matched backbones against Lennard-Jones and folding free-energy terms.
No providers recorded yet. Browse all providers
Designing an RNA that binds a chosen protein — an aptamer, a regulatory element, a therapeutic scaffold — is hard for reasons that make it unlike protein design. RNA backbones are far more flexible, the sequence-to-structure mapping is looser, and there is much less structural data to learn from. Generative approaches to the problem have mostly optimized geometry: get the predicted backbone close to the native one, recover the native sequence. That leaves out whether the designed molecule is physically sensible — whether atoms clash, whether the sequence would actually fold into the geometry proposed.
RNA-EFM, from the Bangladesh University of Engineering and Technology with Virginia Tech, adds that physical term. It is a flow matching model for protein-conditioned RNA sequence and structure co-design, in which a Gaussian prior is transported to the distribution of native RNA backbones under time-dependent conditional paths, with Kabsch alignment keeping the objective invariant to rigid motion. On top of the flow matching objective sits an energy-based refinement stage: generated structures are iteratively corrected to reduce both structural error and physical energy, guided by a Lennard-Jones potential and a sequence-derived folding free energy.
The refinement is formulated as an idempotency constraint — the model applied to its own refined output should return that output unchanged. This turns refinement into a convergence procedure rather than a fixed number of denoising steps, and grounds the fixed point in energy minimization rather than in the generative trajectory alone.
RNA-EFM builds on components from prior work: the Noise-to-Seq module from RNAFlow autoregressively predicts nucleotide sequences from noised backbones, and RoseTTAFold2NA serves as the structure predictor generating RNA backbone coordinates. Training and evaluation use protein-RNA complexes from the 2020 PDBBind release under two splits — a structure-predictor-aware split of 1,059 training, 117 validation, and 16 test complexes, and a sequence-similarity split of 1,015 training, 105 validation, and 72 test complexes.
On the structure-predictor-aware split, RNA-EFM reached an RMSD of 10.00 Å against 10.61 Å for the strongest baseline configuration, an lDDT of 0.60 against 0.53, and sequence recovery of 40% against 37%. On the sequence-similarity split, sequence recovery was 35% against 32%. Binding energy improved for 64% of designs, compared with 56% for the best baseline. The reference implementation provides training and inference scripts, but no trained checkpoint is distributed — the inference script points at a checkpoint file that is absent from the repository — so running the model means training it first. The repository ships no licence file either, so reuse terms for the code are not established, although the journal article itself is open under CC BY 4.0.
The intended use is early-stage design of RNA binders against a protein of interest: generating candidate aptamer sequences with accompanying three-dimensional models that can be triaged computationally before synthesis. Because the model conditions on the protein, it fits workflows where the target is fixed and the question is which RNA to make. The explicit energy terms are useful beyond ranking, since designs that clash or that would not fold stably are filtered during generation rather than after it.
Published in Bioinformatics Advances, RNA-EFM makes a focused argument that biophysical priors belong inside the generative loop for RNA design, not only in a post-hoc rescoring pass. The gains over prior flow-based baselines are consistent but incremental — single-digit relative improvements on RMSD and sequence recovery — and absolute accuracy remains modest, with backbone RMSDs around 10 Å reflecting how difficult protein-conditioned RNA design still is. Evaluation is entirely computational, with no experimental binding validation, and the training set of roughly a thousand complexes is small enough that generalization to targets unlike those in PDBBind is unestablished. Related RNA design work in the catalog, such as gRNAde, addresses inverse folding from a given backbone rather than co-designing both.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.