Predicting the three-dimensional structures of large protein complexes remains a key challenge. While recent deep learning approaches such as AlphaFold and RosettaFold have advanced single-chain and some multimer predictions, their accuracy and scalability drop for very large assemblies because of size and complex multi-chain interactions. Traditional docking-based methods can handle larger proteins but struggle with the enormous combinatorial binding possibilities when many chains are present. Assembly-based strategies can recreate large complexes from subcomponents but their success is tightly linked to how accurate those subcomponents are.
HDOCK-Multimer (HDM) is presented as a hybrid framework that integrates ab initio docking with combinatorial assembly to address these complementary limitations. By combining docking-driven pairwise interaction exploration with an assembly pipeline, the authors aim to reduce reliance on perfect subcomponent models while still exploiting the interaction information embedded in subcomponents.
The HDM framework fuses two principal ideas: first, using ab initio docking to explore pairwise orientations and interfaces between chains, and second, applying a combinatorial assembly strategy to stitch together pairwise solutions into full multi-chain models. The docking component is used to identify probable interactions even when individual subcomponent models are imperfect, thereby reducing the downstream impact of subcomponent errors on the assembled complex. The assembly component leverages pairwise docking results to guide combinatorial search through possible ways to connect chains into a complete complex.
The preprint frames this integration as a balance: docking mitigates the need for highly accurate subcomponents, while assembly exploits pairwise interaction evidence to constrain the exponential combinatorial space inherent to multi-chain complexes.
According to the authors, HDM provides practical advantages where existing approaches have shortcomings:
Deep-learning approaches deliver high accuracy for many cases but are challenged by very large complexes and intricate multi-chain interactions.
Docking methods are scalable but can be overwhelmed by combinatorial possibilities as the number of chains grows.
Assembly methods can build large structures from parts but are vulnerable when subcomponent predictions are inaccurate.
By integrating docking and assembly, HDM is positioned to combine the scalability of docking with the structural context provided by assembly, aiming to improve overall prediction quality for large complexes.
The method was validated across multiple benchmarks reported in the preprint. These include three test sets: a benchmark of 35 large heteromeric complexes, a broader set of 172 large protein complexes, and a set of 7 CASP15 targets. These datasets were used to evaluate HDM’s capacity to reconstruct large assemblies and to compare its performance with competing methods.
The authors compared HDM with several state-of-the-art approaches: MoLPC, CombFold, AlphaFold-Multimer (AFM), and AlphaFold3 (AF3). On the reported benchmarks, HDM is described as substantially outperforming these methods. The preprint asserts consistent performance gains across the evaluated datasets; specific numerical metrics, per-target examples, and statistical details are reported in the original manuscript and supplementary material.
An additional capability highlighted in the preprint is HDM’s reported ability to predict complex stoichiometry and to model complexes without explicit stoichiometry input. The authors state that HDM can both infer the subunit composition and construct structural models in the absence of provided stoichiometry, which could be useful for systems where subunit counts are uncertain. The preprint contains the authors’ evaluations of this capability across their benchmarks.
The HDM package is provided openly under the repository: https://github.com/huang-laboratory/HDOCK-Multimer. The preprint notes that the package is freely available and that supplementary material accompanies the manuscript. The copyright statement indicates a CC-BY-NC 4.0 International license for the preprint version.
This work is presented as a preprint posted August 6, 2026, and has not undergone peer review. The authors declare no competing interests. Funding sources include grants from the National Natural Science Foundation of China. The manuscript reports comparative advantages for HDM, but because this is a preprint, readers should consult the full paper and any future peer-reviewed version for complete methods, metrics, and limitations.
Note: numerical results, per-target performance metrics, and algorithmic implementation details are contained in the source preprint and supplementary files; specifics beyond the summary presented here were not reproduced in the abstract and therefore are not claimed in this summary.