DeepAssembly

A structure predictor can get every atom of a domain right and still return the wrong protein. For one two-domain human entry in the AlphaFold database, AlphaFold 2 scores TM-score 0.990 and 0.995 on the two domains separately while the full-length model reaches 0.689 — the folds are correct and the angle between them is not. About a third of Protein Data Bank chains are multi-domain, and the bank skews toward compact single-domain proteins, making this failure mode systematic.

DeepAssembly treats the domain models as the solved part, to be reused rather than re-derived: it splits the query into single-domain sequences, builds a structure for each, then solves only for how those rigid pieces sit relative to one another. The placement is predicted by AffineNet, which represents each inter-domain residue pair by the affine transformation between the two residues' local coordinate frames — a rotation and a translation, written as the six-dimensional (6D) vector (α, β, γ, r, θ, φ) and predicted as binned distributions. Those distributions become an atomic coordinate deviation potential, and a population-based evolutionary search over linker rotation angles moves the domains until the conformation satisfies it.

DeepAssembly comes from Guijun Zhang's group at Zhejiang University of Technology and is the predecessor of DeepAssembly2. Its central claim is a transfer result: AffineNet is trained only on intra-chain domain pairs from monomers, yet the same checkpoint, unchanged, predicts inter-chain interfaces once the chains of a complex are treated as collections of domains. The interactions holding two domains of one chain in register are the ones holding two chains together, and what is learned from the first carries into the second.

#Key Features

• Domains, not chains, as assembly units: a rigid-body, local search needs less GPU memory than feeding a whole complex through an end-to-end predictor.
• Affine transformations instead of contact maps: a rotation and a translation between residue-local frames captures inter-domain orientation directly, rather than leaving it to be inferred from a distance map.
• Monomer-trained, complex-applied: the network fit to intra-protein domain pairs is applied without retraining to heterodimers, recovering the interface on 32.4% of a 247-complex test set.
• Remote templates over sequence alignment: templates from the group's PAthreader method rather than HHsearch cut inter-domain distance error by about 14.5% in ablation.
• Assembly from experimental domains: given solved domain structures instead of predicted ones, the pipeline reaches an average TM-score of 0.856 on the 219-protein benchmark.

#Technical Details

AffineNet stacks 8 residual blocks over MSA and inter-domain features projected to 64 channels, routes remote-template features through an axial attention module alternating over rows and columns, concatenates the two, and passes the result through 20 more residual blocks before branching into six softmax heads, one per affine component. Inputs are MSA statistics from HHblits over Uniclust30_2018_08 and BFD, template features from PAthreader, inter-domain contacts from a pretrained MSA Transformer, and a predicted domain-boundary mask. Training used 10,064 of 10,593 multi-domain chains from the group's MPDB database (48,225 entries as of September 2021): up to 50 epochs of Adam in TensorFlow 1.14, roughly 14 days on a single NVIDIA Tesla V100s. Assembly generates 1,000 coarse-grained Rosetta models, runs 500 iterations of crossover and mutation on linker rotation angles, and refines the top 10 with FastRelax; GraphCPLMQA picks the output. On 219 DEMO multi-domain proteins under 30% identity to the training data, DeepAssembly averages TM-score 0.922 and RMSD 2.91 Å against 0.900 and 3.58 Å for AlphaFold 2, with inter-domain distance error 0.560 Å versus 0.724 Å. On 164 low-confidence AlphaFold database entries it lifts TM-score 13.1%, from 0.610 to 0.690.

#Applications

The method is delivered as a hosted server taking sequences for a multi-domain protein or the chains of a complex and returning five ranked full-length models. Its most direct use is repair: re-fitting the arrangement of an AlphaFold database entry whose domains are confident but whose overall shape is doubtful. It also accepts experimentally determined domain structures, so a group with crystal structures of individual domains can preserve them exactly while predicting how they pack — the setting where it separates most clearly from end-to-end prediction.

#Impact

DeepAssembly's lasting contribution is evidence that inter-domain interactions learned inside monomers generalize to interfaces between chains, reframing complex prediction as a domain assembly problem a comparatively small network can address. Nothing is distributed but the web server: no code, no AffineNet checkpoint, no stated software license. Accuracy is bounded by the single-domain models fed in — on one CASP15 target a poor second domain held the assembly to TM-score 0.474, while experimental domains lifted it to 0.949. On the 247-heterodimer set, AlphaFold-Multimer reached a 65.2% success rate against DeepAssembly's 32.4%, though the authors note the comparison is confounded: AlphaFold-Multimer trained on the whole PDB with no redundancy removed against that test set.