Tensorizing Pairwise Structural Analysis: GPU-Accelerated RMSD and FCC Clustering in HADDOCK3 Without Structural Regression.
In the authors' words
Integrative macromolecular docking pipelines generate thousands of candidate structural decoys, relying on downstream pairwise ensemble analysis, including Root-Mean-Square Deviation (RMSD) matrices and Fraction of Common Contacts (FCC) clustering, to identify biologically relevant binding poses. However, pairwise structural comparisons scale quadratically (O(N^2)), and canonical workflows in HADDOCK3 historically rely on disk-bound trajectory serialization and iterative CPU set operations, creating a quadratic computational and storage bottleneck for large ensembles. Here, we demonstrate that these operations can be reformulated as continuous linear algebra on GPU-resident tensors: structural superposition via batched Singular Value Decomposition (SVD) and residue contact overlap via binary matrix multiplication (M = A A^T). Across empirical benchmarks spanning N = 200 to 10,000 models (evaluating up to 49,995,000 decoy pairs), our GPU implementation achieves a 6.41-fold compute speedup on Apple Silicon (MPS) and up to 12.2-fold end-to-end acceleration in an N = 5,000 ablation on NVIDIA CUDA by eliminating trajectory disk I/O. Crucially, the implementation achieves zero downstream structural regression relative to canonical baselines, preserving identical cluster assignments (Adjusted Rand Index ARI = 1.0000, zero differing assignments) with sub-microangstrom RMSD concordance (Delta_max <= 4.07 x 10^-7 Angstrom on single-precision MPS; machine-precision Delta_max <= 8.88 x 10^-16 Angstrom on double-precision CUDA). The accelerated code, test suites, and benchmarking workflows are openly accessible on GitHub under the Apache-2.0 license.
Appeared: Tuesday, October 13. bioRxiv. Preprint, not yet peer-reviewed.