ELEN

De novo designed sequences are almost always screened in silico before anyone orders DNA, and structure predictors do that screening well — for ordered regions. Loops are the exception. They are flexible, frequently functional, and prediction accuracy falls off sharply with loop length. AlphaFold's confidence metric, plDDT, conflates local error with local flexibility, so a designer looking at a low-confidence loop cannot tell which they are seeing.

ELEN — the Equivariant Loop Evaluation Network — is a local model quality assessment method built for that blind spot. Released in August 2025 by the Oberdorfer lab at Graz University of Technology, with a co-author at EPFL, it takes an all-atom structure and predicts three per-residue quality metrics at once: lDDT, CAD-score, and RMSD against the experimental reference. Learning all three jointly is deliberate: lDDT and CAD-score are superposition-free contact-based measures that correlate tightly (Spearman ρ = 0.91), while RMSD correlates far less (ρ = 0.76 and 0.69), so it contributes information the others do not.

Architecturally ELEN extends the Equivariant Deep Network framework with 3D equivariant group convolutions over atoms, adding two feature streams the original lacked: per-residue physicochemical descriptors from Rosetta and frozen sequence embeddings from SaProt. It is, to the authors' knowledge, the first model quality assessment method built specifically for loops.

#Key Features

• Loop-focused training: The network trains on extracted "loop pockets" — loop residues plus their surrounding structural context — rather than whole chains, concentrating capacity where prediction is hardest.
• Three quality metrics jointly predicted: Simultaneous lDDT, CAD-score, and RMSD heads give complementary views of local accuracy from one forward pass.
• All-atom equivariance: 3D equivariant group convolutions learn each atom's local geometric environment, so scores are independent of position and orientation.
• Sequence-aware without an MSA: Frozen SaProt 650M embeddings supply evolutionary context; the ablation identifies them as the single most valuable feature added, lifting mean Pearson correlation to 0.44.
• Runs without a language model too: Geometry-only and no-embedding variants ship alongside the full model and are competitive at the per-loop level — useful when SaProt is not installed.

#Technical Details

Training data comes from the Protein Sequence Culling Server, with loops of 2–10 residues identified by DSSP and extracted as pockets that include up to 40 neighbouring residues from the loop's central residue. Ground-truth labels are per-residue lDDT, CAD-score, and all-atom RMSD, computed between AlphaFold2 models and experimental references before loop extraction. Splitting is chain-aware to prevent leakage, yielding 938,000 training, 186,000 validation, and 18,500 test loop pockets. Inputs combine one-hot atom and residue identity and secondary structure, continuous descriptors (Rosetta energy terms, solvent-accessible surface area, hydrogen bond counts), and SaProt_650M_PDB embeddings; atomic features are aggregated at Cα positions and processed by further equivariant convolutions. The final model has roughly 341,000 trainable parameters and a Huber loss over the three metrics.

The benchmark is the CAMEO model quality estimation category: 952 models across 98 targets, 8,681 extracted loops, about 246,000 residues. On loop residues ELEN records the highest local Spearman correlation of any method tested (0.657) against ground-truth CAMEO lDDT, with Pearson r 0.717 and AUC 0.818; ModFOLD9 edges it on Pearson (0.727), and QMEANDisCo_3 is marginally better on per-loop Top-1 loss. Removing the sequence embeddings drops local loop Spearman to 0.583. Matching consensus predictors matters because ModFOLD9 aggregates the outputs of several other quality estimators, whereas ELEN scores a structure standalone.

#Applications

The design loop is the target use: rank candidate loop conformations, flag positions likely to be modelled badly, and decide what to rebuild before committing to synthesis. The authors apply it to three sets of redesigned enzymes. Guiding a redesign of the 180–190 loop of human duodenal cytochrome b raised the mean ELEN score there from 0.812 to 0.839 and mean plDDT from 70.3 to 82.4. On de novo designed retro-aldolases and a designed Kemp eliminase, ELEN separates well-modelled regions from poor design positions, and low scores coincide with regions missing from the crystal structure through absent electron density — real disorder rather than modelling error.

#Impact

ELEN shows that a very small equivariant network, given the right training target and frozen language-model features, can hold its own against established consensus quality estimators on the region of protein structure where those estimators are least reliable. Its practical contribution to design work is a per-residue signal separating "modelled wrong" from "genuinely flexible," which plDDT alone does not provide. Two caveats: the work is a preprint awaiting peer review, and the training code is pending release. Inference code and four trained checkpoints — full, no-SaProt, SaProt-only, and geometry-only — are on GitHub under the MIT license, and the loop dataset is on Zenodo, though the roughly 200 GB of SaProt embedding files are available from the authors on request rather than hosted.