DeepAssembly server information

DeepAssembly is designed to automatically construct multi-domain protein or complex structure through inter-domain interactions from deep learning. Starting from the input sequence of multi-domain protein (or protein complex), DeepAssembly first generates MSAs from genetic databases and searches for templates using our in-house PAthreader. Meanwhile, the input sequence is split into single-domain sequences by a domain boundary predictor, and then the structure for each domain is generated by single-domain structure predictor. Subsequently, features extracted from MSAs, templates and domain boundary information are fed into a deep neural network with self-attention to predict rotations, translations between inter-domain residues, and convert them into inter-domain interactions. Finally, DeepAssembly performs the creation of the initial full-length structure using the single-domain structures, followed by multiple steps of iterative population-based rotation angle optimization and atomic-level structure refinements. The domain assembly simulation is driven by the atomic coordinate deviation potential transformed from predicted inter-domain interactions, where the best conformation by our in-house model quality assessment is selected as the final output structure.

DeepAssembly-AA is a computational protocol specifically designed for antibody-antigen complex structure modeling. This protocol integrates deep learning-based inter-chain residue distance prediction with a Monte Carlo conformational sampling method to generate accurate binding conformations. Specifically, the predicted distance maps are employed as knowledge-based energy constraints to guide sampling toward physically plausible configurations. To efficiently explore the conformational space, we designed a two-stage exploration-exploitation strategy. In the exploration stage, diverse binding orientations are extensively sampled under the guidance of the energy landscape. In the subsequent exploitation stage, refinement is performed on the complementarity-determining region loops to enhance geometric and energetic complementarity at the interface. Finally, the sampled conformations are evaluated through an energy-guided confidence selection mechanism, from which the optimal conformation is identified as the final output.

In the CASP16 blind assessment of protein complex structure modeling, DeepAssembly (group name: GuijunLab-Assembly) ranked second in modeling easy and medium targets. The results are available at https://predictioncenter.org/casp16/zscores_multimer.cgi.

In the CAMEO blind test for protein structure modeling, DeepAssembly (group name: ZJUT-DeepAssembly) ranked first in the one-week evaluation on May 28, 2022. The results are available at https://www.cameo3d.org/modeling/.

Please cite the following articles when you use the DeepAssembly server:

• Yuhao Xia^✝, Kailong Zhao^✝, Dong Liu, Xiaogen Zhou, Guijun Zhang*. Multi-domain and complex protein structure prediction using inter-domain interactions from deep learning. Communications Biology, 6, 1221 (2023). DOI: https://doi.org/10.1038/s42003-023-05610-7
• Yuhao Xia, Yilin Pu, Suhui Wang, Jianan Zhuang, Dong Liu, Minghua Hou, Guijun Zhang*. DeepAssembly2: A web server for protein complex structure assembly based on domain-domain interactions. Journal of Molecular Biology, 437, 169128 (2025). DOI: https://doi.org/10.1016/j.jmb.2025.169128
• Suhui Wang^✝, Jianan Zhuang^✝, Xinyue Cui, Zexin Lv, Dongliang Hou, Guijun Zhang*. High-accuracy structural modeling of antibody-antigen complexes. Communications Biology, (2026). DOI: https://doi.org/10.1038/s42003-026-10757-0