DeepAssembly2

If a crystal structure already exists for each subunit of a complex, an end-to-end predictor that maps sequences to coordinates has no way to use it. It re-derives folds you already measured, and the only question you actually had — how do these pieces sit against each other — falls out of a computation whose memory footprint grows with the whole assembly. That is the gap DeepAssembly2 is built for: hold the monomers fixed, and solve only for their relative placement.

The pipeline treats domains, not chains, as the units being assembled. From the input sequences it extracts sequence, multiple sequence alignment and monomeric structure features and feeds them to DPIC, an in-house network that predicts the distance between every pair of inter-chain residues. Those predicted distances become a knowledge-based energy; a second energy is built from the inter-chain Cα distances read directly off an AlphaFold-Multimer prediction of the same target. A population-based multi-objective optimization then searches rigid-body rotations and translations of the monomeric units against both potentials at once, producing a diverse pool of assemblies rather than one answer, and DeepUMQA-X ranks that pool to choose what is returned.

DeepAssembly2 comes from Guijun Zhang's group at Zhejiang University of Technology and supersedes DeepAssembly, which inferred inter-chain contacts from a network trained on intra-chain domain pairs in multi-domain monomers. That transfer worked but was thin: the interaction patterns available inside a single chain do not span the ones complexes actually use. Training on genuine inter-chain domain pairs is the change that closes most of the gap.

#Key Features

• Domains as assembly units: splitting each chain into structural domains before docking gives a finer-grained rigid-body decomposition than chain-level docking, and keeps large assemblies tractable through a divide-and-conquer strategy.
• Two independent distance sources: the network's own predicted inter-chain distances and those extracted from AlphaFold-Multimer enter the search as separate objectives, so the two signals complement rather than average each other.
• Rigid-body conformational sampling: the search adjusts rotation and translation between whole monomeric structures instead of the linker-dihedral moves used previously, freeing sampling from linker geometry and covering a larger conformational space.
• Interface-aware structural features: per-residue interface propensity from PeSTo, Ultrafast Shape Recognition descriptors placing each residue in the global topology, and intra-chain heavy-atom distances are supplied alongside evolutionary features.
• Built-in model selection: the top five assemblies are returned with per-residue and global lDDT scores, so the ranking that picked them is visible rather than implicit.

#Technical Details

The inter-chain distance network pairs residual CNN blocks with an attention-based triangle interaction module — triangle update, triangle axial self-attention, and linear transformation layers. Inputs are one-hot amino acid identity and physicochemical properties; PSSM, embeddings and attention maps from the pretrained ESM-MSA-1b model over paired MSAs built with HHblits against Uniclust30 and paired by organism; and the three structural features above. Training used a purpose-built inter-chain domain-domain dataset: 36,955 PDB complexes released before June 2022, resolution better than 3.5 Å, two or three protein chains, split into domains by DomainParser, keeping domain pairs with an interface area above 500 Ų and clustered at 40% sequence identity by MMseqs2, leaving 16,897 pairs. On 46 CASP13–15 targets, DeepAssembly2 averaged TM-score 0.769 against 0.568 for DeepAssembly and 0.758 for AlphaFold-Multimer, matching AlphaFold 3 at 0.769, with mean DockQ 0.439 and a correct interface on 56.5% of targets; its i-RMSD of 7.180 Å and L-RMSD of 16.263 Å improved on AlphaFold 3 by 10.1% and 5.7%. On a 247-heterodimer benchmark it reached DockQ 0.438 against 0.415 for AlphaFold-Multimer and recovered the interface on 166 of 247 targets.

#Applications

The server accepts either FASTA sequences or monomeric PDB structures for the chains of a complex, which is what makes it usable in the case it was designed around: experimentally determined subunits that should be preserved exactly while their arrangement is predicted. Groups working on drug targets, vaccine antigens and other large assemblies can supply solved or externally modelled monomers; when only sequences are given, monomeric models are generated with AlphaFold 2 first. Because sampling yields a population rather than a single structure, the output is also a starting point for studying alternative arrangements of the same subunits.

#Impact

DeepAssembly2's contribution is a working demonstration that domain-level assembly guided by learned inter-chain distances stays competitive with end-to-end complex prediction while using the prior information end-to-end methods discard. Its results are complementary to AlphaFold-Multimer's rather than uniformly better, which is the more useful property for a predictor used alongside others. Practical limits are real: it is distributed only as a hosted web server with no downloadable code or weights, its pipeline calls AlphaFold 2 and AlphaFold-Multimer at inference, and its training set was restricted to assemblies of two or three chains.