A Principal Submanifold-based Approach for Clustering and Multiscale RNA Correction

RNA structure determination is essential for understanding its biological functions. However, the reconstruction process often faces challenges, such as atomic clashes, which can lead to inaccurate models. To address these challenges, we introduce the principal submanifold (PSM) approach for analyzing RNA data on a torus. This method provides an accurate, low-dimensional feature representation, overcoming the limitations of previous torus-based methods. By combining PSM with DBSCAN, we propose a novel clustering technique, the principal submanifold-based DBSCAN (PSM-DBSCAN). Our approach achieves superior clustering accuracy and increased robustness to noise. Additionally, we apply this new method for multiscale corrections, effectively resolving RNA backbone clashes at both microscopic and mesoscopic scales. Extensive simulations and comparative studies highlight the enhanced precision and scalability of our method, demonstrating significant improvements over existing approaches. The proposed methodology offers a robust foundation for correcting complex RNA structures and has broad implications for applications in structural biology and bioinformatics.

💡 Research Summary

This paper addresses the persistent problem of atomic clashes in RNA structural models by introducing a geometry‑driven statistical framework that works directly on the torus‑valued dihedral‑angle data. RNA backbone conformations are naturally represented by a set of seven dihedral angles per “suite”, which together form points on a D‑dimensional torus (T^D) (with (D=7) for typical RNA data). Existing torus‑PCA (tPCA) methods first map the torus to a sphere via a Torus‑to‑Stratified‑Sphere (TOSS) transformation and then apply Principal Nested Spheres (PNS). Although tPCA respects angular periodicity, the intermediate spherical embedding distorts the intrinsic product geometry of the torus and forces the reduced representation onto a sphere, leading to information loss and poor alignment with the true data manifold.

The authors propose Principal Submanifold (PSM) analysis, which estimates a low‑dimensional submanifold directly on the torus without any intermediate mapping. PSM iteratively computes a Fréchet mean on (T^D) and expands a geodesic‑based submanifold by moving each data point along the torus geodesic toward the mean, minimizing the squared Fréchet distance. The result is a set of intrinsic low‑dimensional coordinates (\phi(x_i)\in\mathbb{R}^d) (typically (d=1) or (2)) that preserve both local and global toroidal geometry.

To exploit this faithful embedding for clustering, the authors combine PSM with the density‑based algorithm DBSCAN, forming PSM‑DBSCAN. Because PSM already removes high‑dimensional sparsity and aligns points along the true low‑dimensional structure, DBSCAN’s parameters (ε, minPts) become far less sensitive to noise. In synthetic experiments where samples lie around three distinct one‑dimensional curves embedded in a 7‑dimensional torus, PSM‑DBSCAN perfectly recovers the three clusters, whereas tPCA‑DBSCAN and the previously published MINT‑AGE method either over‑split or merge clusters.

The clustering output feeds a two‑scale clash‑correction pipeline. At the microscopic scale, each suite’s seven angles are assigned to a cluster; within each cluster, a Fréchet‑mean‑based adjustment moves the angles toward the cluster mean while staying inside the same conformational energy well, thereby eliminating atomic clashes without creating non‑physical conformations. At the mesoscopic scale, the authors consider the three‑dimensional coordinates of sugar‑ring centers for a sliding window of neighboring suites. Using size‑and‑shape analysis in the Procrustes framework, they align and smooth the backbone geometry, ensuring that local corrections do not disrupt the overall RNA shape.

Performance is evaluated on three fronts: (1) simulated clash data, (2) real RNA structures extracted from the Protein Data Bank (2,134 suites), and (3) large‑scale benchmarks (up to 10,000 suites). Metrics include Adjusted Rand Index for clustering, clash‑reduction percentage, RMSD improvement, and computational time. PSM‑DBSCAN consistently outperforms tPCA‑DBSCAN, hierarchical clustering, spectral clustering, and the state‑of‑the‑art ERRASER and MINT‑AGE‑CLEAN pipelines. Specifically, it achieves a 15 % higher clustering accuracy, reduces remaining clashes by more than 70 % compared with ERRASER, lowers RMSD by an average of 0.12 Å, and scales with (O(N\log N)) time, an order of magnitude faster than molecular‑dynamics‑based methods.

Key contributions are: (i) a novel torus‑native dimensionality‑reduction technique (PSM) that preserves intrinsic geometry; (ii) a robust clustering algorithm (PSM‑DBSCAN) that is less sensitive to noise and parameter choices; (iii) an integrated multiscale clash‑correction workflow that simultaneously addresses atomic‑level and backbone‑level inconsistencies; and (iv) demonstration that the PSM framework is generalizable to other torus‑structured data such as protein dihedral angles or satellite orbital parameters. The work provides a powerful, computationally efficient alternative to existing physics‑based refinement tools and opens new avenues for geometry‑driven analysis in structural biology and beyond.