About

RNAInvBench: Benchmarking RNA Inverse Folding in Two and Three Dimensions

RNAInvBench is a robust, open-source Python framework designed to advance research in RNA inverse design by providing unified environments, tasks, curated datasets, baseline algorithms, evaluation metrics, and analysis tools. It enables the training and rigorous testing of models for both two-dimensional (secondary) and three-dimensional (tertiary) RNA inverse folding tasks.

Within this benchmark suite, the complex problem of RNA inverse design is divided into three key challenges:

• Secondary Pseudoknot-Free Inverse Design
  This task focuses on designing an RNA sequence from a two-dimensional secondary structure, expressed in standard dot-bracket notation, excluding pseudoknots. Most algorithms here assume only canonical base-pairing (A–U & G–C pairs) and Wobble pairing (G-U).

• Secondary Pseudoknotted Inverse Design
  In this task, sequences are designed from more complex 2D structures which may include pseudoknots. An extended dot-bracket notation indicates the pseudoknot levels. Algorithms benchmarked for this task handle non-canonical pairings to reflect realistic biological scenarios.

• Tertiary Inverse Design
  This task addresses the reconstruction of an RNA sequence from its full three-dimensional structure. Many 3D RNA structures comprise multiple RNA chains, which must be isolated into single chains before inverse design. The 3D structure, provided as a PDB file, is converted into suitable tensors for algorithmic processing.

Through RNAInvBench, we aim to promote reproducible, comparable, and extensible benchmarking for the global RNA design research community.

---

Our Partners