RiboBridge: protein-conditioned de novo RNA binder structure design
RiboBridge addresses the challenge of designing RNA molecules that bind to target protein structures when neither the native RNA sequence/structure nor the binding interface is known. Its core idea is to separate RNA conformation generation in a pretrained discrete structural space (RiboSphere) from protein-relative assembly via predicted cross-frame interaction geometry and rigid placement. This decomposition matters because high-quality protein–RNA complex data are scarce and existing methods entangle conformational sampling with interface discovery.
RiboBridge offers a principled, interpretable decomposition that improves relative performance on a 124-complex temporally split benchmark, achieving the strongest reported interface recovery and complex geometry among compared methods. However, absolute metrics remain far from solved: only 20 of 124 targets ($16.13\%$) reach $\text{DockQ} \geq 0.23$, $\text{fNAT}$ is below $0.06$, and the authors acknowledge that reliable interface ranking and high-fidelity sequence realization remain unsolved problems.
The stage-wise diagnostic framework is a highlight: replacing ground-truth RiboSphere tokens with generated canonical RNA drops $\text{DockQ}$ from $0.854$ to $0.464$, and predicted interaction maps lower it further to $0.253$, while rigid placement under exact constraints adds almost no error. This cleanly identifies conformation generation and interaction prediction as the true bottlenecks, and the ablations show that matched predicted-pocket conditioning genuinely helps RNA generation beyond length-only controls.
Absolute performance is weak: a $\text{DockQ}$ of $0.13$ and an $\text{fNAT}$ of $0.059$ mean the generated complexes rarely recover native-like interfaces. The evaluation protocol uses post-hoc interface correspondence, i.e., native information selects the best-matching predicted pocket after generation, bypassing the harder real-world problem of ranking hypotheses. The self-correction objective shows no statistically significant benefit, with target-paired $95\%$ confidence intervals all including zero, and the model produces more steric clashes ($\text{ClashScore}=14.96$) than RFDpoly ($9.66$). Sequence realizability is likewise poor, with only $8.06\%$ of targets designable below an AF3 scRMSD of $2$ \AA.
Baseline comparisons are transparent: each method is evaluated with its supported conditioning pipeline on the same temporally split, redundancy-controlled benchmark. Nevertheless, the post-hoc interface-correspondence protocol advantages RiboBridge because native overlap is used to select among pre-generated predicted-pocket candidates, whereas in practice the model would have to rank them without native knowledge. The authors explicitly note this gap, stating that candidates are generated before the native interface is consulted and that the native interface is used only afterward to establish evaluation correspondence.
Hyperparameters and architecture details are extensively documented in appendix tables, and token decoding is deterministic for fixed inputs, which reduces sampling variance. However, no code repository, model weights, or curated data-split files are linked in the provided text, and the pipeline depends on the separate RiboSphere autoencoder (arXiv:2603.19636), adding an external dependency that is not packaged with this work.
Designing RNA molecules that bind to target protein structures requires generating both an RNA conformation and its bound complex geometry, without assuming a predefined binding site. Existing approaches generally model these coupled processes within a unified representation, entangling RNA conformation generation with protein-relative placement. In this work, we introduce RiboBridge, a structure-first framework that separates RNA conformational generation from protein--RNA assembly through cross-frame interaction geometry. RiboBridge first proposes multiple candidate RNA-binding interfaces on the protein surface, generates canonical RNA conformations within a pretrained RNA structural latent space (RiboSphere), and then predicts protein--RNA interaction constraints to bridge the independently defined RNA and protein coordinate frames. The resulting interaction geometry is converted into a geometrically consistent protein-relative pose through rigid placement. We evaluate RiboBridge on a temporally split, sequence-filtered, and redundancy-controlled benchmark of 124 protein--RNA complexes. Under post-hoc interface-correspondence evaluation, where native information is used only after generation to select among already generated predicted-pocket candidates, RiboBridge achieves the highest fNAT and DockQ and the lowest interface and ligand RMSDs among compared methods. Component ablations and stage-wise diagnostics further characterize the contributions of RNA conformational generation, interaction prediction, and rigid placement, identifying the main sources of complex-recovery error.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.