Enhancing RNA 3D Folding Prediction via Transformer and Equivariant Architectures under Resource Constraints
Mehra, R.; Mehra, A.
Show abstract
Predicting the three-dimensional structure of RNA molecules is a fundamental problem in computational biology, with applications in understanding gene regulation, viral function, and drug design. Existing RNA structure prediction methods often rely on large models or heavy computation, as exemplified by protein folding solutions like AlphaFold2 [2], which makes them impractical under tight resource constraints. In this work, we present an efficient neural architecture that integrates a Transformer encoder enhanced with Rotary Positional Embeddings (RoPE) and a small residual MLP block, along with an optional E(n)-equivariant refinement module. We train our model on the Stanford Ribonanza RNA 3D dataset [11] using BF16 precision on a single NVIDIA H100 (80 GB) GPU. Our model achieves lower RMSD than simpler baselines, demonstrating that modern Transformer and equivariant techniques can be effectively applied to RNA folding under practical compute constraints.
Matching journals
The top 3 journals account for 50% of the predicted probability mass.
Similar papers in this journal
- ConsAlign: simultaneous RNA structural aligner based on rich transfer learning and thermodynamic ensemble model of alignment scoring 96%
- Graph neural representational learning of RNA secondary structures for predicting RNA-protein interactions 95%
- FlowPacker: Protein side-chain packing with torsional flow matching 95%
Similar papers in this journal
"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.