Tensorizing Pairwise Structural Analysis: GPU-Accelerated RMSD and FCC Clustering in HADDOCK3 Without Structural Regression.

Macromolecular docking pipelines typically generate a large number of candidate structural decoys, which are analyzed using pairwise ensemble methods such as Root-Mean-Square Deviation (RMSD) matrices and Fraction of Common Contacts (FCC) clustering to identify relevant binding poses. However, the computational and storage demands of these methods can become a bottleneck for large datasets.

The HADDOCK3 software has traditionally relied on disk-bound trajectory serialization and iterative CPU set operations, which create a quadratic computational complexity issue.

This study demonstrates that the aforementioned operations can be restructured as continuous linear algebra computations on GPU-resident tensors. Specifically, structural superposition can be achieved through batched Singular Value Decomposition (SVD), while residue contact overlap can be accomplished via binary matrix multiplication (M = A A^T). The GPU implementation was evaluated against empirical benchmarks involving model sizes ranging from 200 to 10,000, encompassing up to 49,995,000 decoy pairs.

Across these benchmarks, the GPU-accelerated approach achieved a significant 6.41-fold speedup on Apple Silicon (MPS) and up to 12.2-fold end-to-end acceleration on NVIDIA CUDA during an N = 5,000 ablation test. Notably, the GPU implementation maintained zero structural regression relative to the traditional baselines, ensuring that the cluster assignments remained identical (Adjusted Rand Index = 1.0000) and the RMSD differences were minimal (Delta_max).