This repository contains a high-performance, two-phase hybrid pipeline designed for the prediction of 3D RNA coordinates. The methodology integrates classical comparative modeling with state-of-the-art diffusion models to achieve robust structural predictions across varying sequence lengths and complexities.
The pipeline is architected to prioritize verified structural data while utilizing generative fallback mechanisms for novel folds.
The initial phase performs global pairwise sequence alignment against a curated database of known RNA structures. When a candidate template exceeds a 50% identity threshold, TBM is utilized to construct a coordinate framework.
- Gap Reconstruction: Sinusoidal backbone reconstruction is applied to alignment gaps to maintain stereochemical continuity.
- Adaptive Constraints: Coordinates are refined using confidence-scaled constraints to preserve template integrity while resolving local geometric clashes.
For target sequences that lack suitable templates or contain unresolved regions, the pipeline utilizes the Protenix-v1 diffusion model for de novo coordinate generation.
- Chunked Inference: To accommodate hardware constraints (Kaggle T4 GPUs), sequences exceeding 420 nucleotides are processed in overlapping tiles.
- Kabsch Stitching: Individual chunks are reintegrated into a global coordinate system using core-trimmed Kabsch alignment, minimizing structural drift at segment boundaries.
| Component | Scientific Rationale |
|---|---|
| Kabsch Alignment | Minimizes Root-Mean-Square Deviation (RMSD) between overlapping structural segments to ensure global connectivity. |
| Diffusion Sampling | Utilizes stochastic denoising to navigate the complex conformer landscape of rare RNA folds. |
| Energetic Ranking | Final candidate models are ranked based on backbone bond length deviation and steric clash penalties. |
The pipeline is implemented using a custom ensemble of structural biology libraries, including Biopython for alignment, Biotite for coordinate manipulation, and a modified Protenix-v1 engine for inference. All dependencies are managed locally to ensure reproducible execution in restricted competition environments.