DeepAccNet

Structure refinement is a search problem, and what it lacks is a map. Rosetta can sample thousands of conformations around a starting model, but with no idea which parts are wrong it either drifts away from regions that were already correct or stays too timid near a backbone that needs to move. Earlier accuracy estimates mostly reported where an ensemble disagreed with itself — a statement about sampling, not about error — and refinement gains between CASP rounds had stalled below 2%.

DeepAccNet supplies the missing map. From a single input structure it predicts three things at once: a per-residue Cβ l-DDT score, a native contact mask marking which residue pairs genuinely lie within 15 Å, and — the part nothing else provided — an "estogram", a 15-bin histogram of the signed Cβ–Cβ distance error for every residue pair. A signed error says not merely that two residues are misplaced but whether they sit too far apart or too close, and by how much — which converts directly into a distance restraint, so one network can both grade a model and tell a sampler what to do about it.

It came from Naozumi Hiranuma, Hahnbeom Park and colleagues in David Baker's group at the Institute for Protein Design, University of Washington, released in July 2020 and published in Nature Communications in February 2021. Two variants share the main repository and one network, selected with a --bert flag: DeepAccNet-Standard, and DeepAccNet-Bert, which adds ProtBert-BFD100 embeddings from ProtTrans. A third, DeepAccNet-MSA, adds a trRosetta distance prediction as a further 2D input; it lives in a separate repository with its own TensorFlow implementation and checkpoints, and was later vendored into RoseTTAFold as its accuracy-estimation module.

#Key Features

• Signed pairwise error, not a scalar confidence: the estogram gives a direction and a magnitude for every residue pair, so downstream code learns which way each pair has to move.
• l-DDT derived, not regressed: per-residue Cβ l-DDT is computed from the predicted estogram and contact mask through the l-DDT definition itself, keeping it consistent with the pairwise predictions underneath.
• Errors converted into restraints: estogram bins become log-odds distance potentials in the Rosetta energy function, while predicted l-DDT decides which regions get rebuilt and which structure is finally selected.
• Useful on experimental structures: predicted Cβ l-DDT tracks resolution across 20,710 X-ray entries (Spearman-r 0.48), 64 membrane-protein structures (0.64) and 65 cryo-EM structures (0.87).
• Single-sequence operation: the Standard and Bert variants need no alignment, letting them score de novo designs, where evolutionary information does not exist.

#Technical Details

Voxelized atomic coordinates around each residue are convolved in a residue-centred frame, making the local description invariant to rotation and translation. The result is flattened per residue, concatenated with 1D features (amino acid properties, backbone torsions, intra-residue Rosetta energies, secondary structure), tiled into a 2D tensor and joined with the Cβ distance map, residue-residue orientations and pairwise Rosetta energies. That stack feeds a residual trunk of 20 blocks with dilation rates cycling through 1, 2, 4 and 8, branching into two arms of four blocks for the estogram and mask heads. Training used roughly one million decoys — about 150 for each of 7,307 non-redundant X-ray structures (40% maximum sequence identity, 2.5 Å cutoff, 50–300 residues), split 6,749 / 280 / 278 across train, validation and test — generated by RosettaCM comparative modelling, native perturbation and trRosetta-guided folding. On 73 refinement targets, guidance from DeepAccNet-Standard raised all-atom l-DDT by 10% over the starting models and 4% over the previous unguided Rosetta protocol, with the fraction of targets improving by more than 10% rising from 27% to 47%. Feature generation uses PyRosetta by default; the repository also ships PyRosetta-free distance-plus-3D variants.

#Applications

The immediate use is triage: hand it one PDB file and the per-residue profile says which loops and termini to distrust — what a refinement protocol or an experimental follow-up needs to know. Refinement carries a crystallographic payoff: across 41 X-ray datasets, the fraction yielding a robust molecular-replacement solution rose from 0% with the starting models to 20% with the earlier protocol and 37% with DeepAccNet guidance.

#Impact

At CASP14 the predictors, entered blind as BAKER-experimental and BAKER-ROSETTASERVER, were the top single-model methods for global quality assessment, and the refinement protocol was the only entry that consistently improved targets longer than 200 residues. The estogram idea — signed, binned pairwise error rather than a confidence scalar — carried into later work, and the model became the standard baseline for estimators that followed, including DeepUMQA. Code and the four pretrained checkpoint sets are MIT-licensed in a public repository, alongside training code, the dataset splits and the full decoy corpus. Two limitations are stated plainly in the paper: the network overestimates the quality of models its own protocol has heavily optimised, and estogram accuracy degrades on larger proteins, where long-range errors accumulate beyond what a locally sensitive Cβ l-DDT score captures.