When a cryo-electron microscopy (cryo-EM) map is too blurry to read, meaning worse than about 3.5 Å, automated tools that trace a protein chain through the density stop working. Structural biologists then fall back on an old trick from crystallography called molecular replacement: they borrow a related, already-solved structure and fit it into the new data. The trick carries a well-known danger called model bias. A borrowed template can impose its own shape on data that does not support it, and refinement can then "confirm" features that are not there. This is the same failure that retrieval-augmented generation (RAG) has in language models, where retrieved text overrides the model's own evidence and the system cites what it was handed rather than what is true.
We propose to build and test a model builder that fetches structural fragments only where it is uncertain, and accepts a fetched fragment only if an independent half of the experimental data agrees with it. Cryo-EM reconstructions come as two independently reconstructed half-maps. We fit the retrieved fragment against one half-map and score it against the other. This held-out check plays the role that a free R-factor plays in crystallography, and it is cheap. The claim is that most of the completeness gain from template help in low-resolution maps can be kept, that template copying in regions where the template is wrong can be sharply reduced, and that the retrieval cost stays a small fraction of always-on retrieval.
Core Claim
In low-resolution maps, both the benefit and the harm of templates are concentrated in the same places: regions where the map alone cannot decide. A builder that retrieves there, and only there, and that screens each retrieved fragment against held-out density, should match or exceed whole-template docking on residue completeness. On regions where the available templates are genuinely wrong, it should copy far fewer template residues. The contribution is the bias-controlled retrieval step and a measurement protocol for template copying that the field currently lacks as an automated benchmark.
Where This Sits in Prior Work
Automated building in cryo-EM is the starting point. ModelAngelo: Automated Model Building in Cryo-EM Maps and Automated model building and protein identification in cryo-EM maps combine a 3D convolutional network with a graph network that reads local density, sequence embeddings, and geometric priors. They report a floor near 3.5 Å and leave about a quarter of residues unbuilt even on their <4 Å test set. We reuse this system as the base builder and use its per-residue confidence as the retrieval trigger. It does no retrieval of its own.
The template side is supplied by classical practice. Legume lectins were phased from a PHA-L homolog (Remy Loris et al.). Crystal subunits have been docked into low-resolution maps of the HIV-1 envelope spike (Erin E. H. Tran et al.) and of DNA repair complexes (Amanda K. Chaplin et al.). A deep-learning prediction served as the search model for human SRD5A2 (Structure of human steroid 5α-reductase 2 with the anti-androgen drug finasteride). All of these choose whole templates by hand and trust expert judgment to catch bias. We automate the choice at fragment scale and replace judgment with a held-out test.
Two papers constrain the design. Retrieval Augmented Protein Language Models for Protein Structure Prediction shows that learned retrieval over sequence databases can be fast and accurate. Its retrieval is 45 to 90 times faster than alignment search, and it reports large TM-score gains over AlphaFold2 when few homologs exist. It retrieves only 1D sequences and never checks its outputs against experimental data. We retrieve 3D fragments and gate them on density.
AlphaFold Database Debiasing for Robust Inverse Folding shows that predicted structures are geometrically over-idealized. PiFold trained on raw predicted structures recovers 17.16% of sequence on CATH 4.2, versus 35.68% after correction with their autoencoder, DeSAE. The lesson for us is that the fragment library should be experimental PDB coordinates by default. We test predicted and DeSAE-corrected fragments as ablations, because the paper evaluated only inverse folding and never density fitting.
The retrieval-control ideas come from language modeling. RaDIO retrieves only when a hallucination detector fires, and reports EM 0.254 on 2WikiMultihopQA against weaker fixed-trigger baselines. CORAG optimizes retrieval under a compute budget. The attribution-bias audits (Evaluation of Attribution Bias in Retrieval-Augmented Large Language Models) show that generators favor retrieved content for reasons unrelated to correctness. We borrow three things from this work: the gated trigger, the budget framing, and the idea of auditing what the system copies.
The literature on cheap debiasing supplies the cost lens rather than a method. Debiasing Masks and CRISPR show that bias can be removed with tiny interventions instead of retraining. Our analog is to leave the builder frozen and control bias entirely at the retrieval-and-acceptance step.
Research Program
The builder runs ModelAngelo on a map and marks contiguous stretches whose residue confidence falls below a threshold $\tau$. For each marked stretch, we form a query from the flanking built residues, the assigned sequence window, and the local density patch.
Candidates come from two indices over experimental PDB fragments of 5 to 40 residues. One is a sequence index built with standard homology search. The other is a shape index built from a small encoder that embeds density patches and simulated fragment densities into a shared space. That encoder is trained contrastively on the EMDB map–model pairs used by ModelAngelo.
Each candidate is rigidly fit into half-map A, anchored to the flanking residues. It is then scored on half-map B by the local real-space correlation gain over the unmodeled baseline. A fragment is accepted if
$$\Delta_B = \mathrm{CC}_B(\text{with fragment}) - \mathrm{CC}_B(\text{without}) > \delta,$$
where $\mathrm{CC}_B$ is the local correlation with half-map B and $\delta$ is a margin. We calibrate $\delta$ on training maps so that decoy fragments pass at a fixed low rate.
A retrieval budget $K$ caps the number of queries per map. That makes cost a controlled variable, and we report completeness and bias as curves over $K$.
A crystallographic extension, in which assembled fragment models serve as molecular-replacement search models, is deferred. The file offers no benchmark with reported numbers for it.
The template-copy audit is the second deliverable. For a region, we define copying as the fraction of built residues that lie within 1 Å of the retrieved template but more than 2 Å from the deposited model. We measure it under two conditions. One uses deliberately planted wrong templates. The other uses a natural hard case: amyloid polymorphs. Related filament folds exist as tempting templates, yet the true fold differs. Examples include the distinct tau folds of P301L and P301T (Distinct tau filament folds in human MAPT mutants P301L and P301T), the three Aβ40 folds from leptomeninges (Cryo-EM structures of Aβ40 filaments from the leptomeninges of individuals with Alzheimer's disease and cerebral amyloid angiopathy), Arctic-mutant Aβ (Cryo-EM structures of amyloid-β filaments with the Arctic mutation (E22G) from human and mouse brains), and the juvenile-onset α-synuclein filaments (New SNCA mutation and structures of α-synuclein filaments from juvenile-onset synucleinopathy). Template-guided tracing of related polymorphs is already practiced (Yang Yang et al.), so this is the setting where unchecked copying does real damage.
Baselines
The naive reference is ModelAngelo run unchanged on each map. Its reported result on the 177-structure test set of Automated model building and protein identification in cryo-EM maps is 77% residue completeness, and 85% for residues with Q-score above 0.4. We rerun it ourselves on our low-resolution sets. No reported result was found on Lacuna for ModelAngelo on maps worse than 3.5 Å beyond the qualitative floor. Any gain must therefore be shown against our own rerun on identical maps.
The strongest competitor is current practical usage: dock a whole-chain predicted model, from AlphaFold2 or AIDO.RAGFold, into the map, refine it, and merge it with the ModelAngelo build. This is what practitioners already do, as in Qingpin Xiao et al.. It gets its own predicted model, the same refinement protocol, and the same wall-clock allowance as our method. No reported result was found on Lacuna for this pipeline on a cryo-EM completeness benchmark.
Beating it means two things. First, equal or higher completeness and backbone accuracy on ordinary low-resolution maps, within one standard error across maps. Second, a clearly lower template-copy rate on the polymorph and decoy sets.
The third baseline is ungated, always-on retrieval. It retrieves for every low-confidence stretch and accepts the best-fitting fragment on the full map, with no half-map check. This mirrors fixed-interval RAG and shows what the gate and the held-out test buy in cost and bias.
Two ablations isolate the key ingredient. The first keeps gated retrieval but scores acceptance on the full map rather than on held-out half-map B. The second swaps the fragment library between PDB, raw predicted structures, and DeSAE-corrected predicted structures. If held-out scoring does not lower copying relative to full-map scoring at equal completeness, the central idea fails.
Why not just…?
Why not just dock a whole AlphaFold prediction and refine? This is the real threat. On well-behaved single-domain proteins it may match us, and if it does across the ordinary low-resolution set we will say so. The proposal predicts it loses in two places. The first is multi-domain targets where the predicted arrangement is wrong, the rigid-body failure noted in classical molecular replacement. The second is amyloid polymorphs, where the prediction or the nearest template carries the wrong fold. The polymorph copy-rate experiment settles this directly, because a whole-chain dock has no mechanism to reject a wrong region short of manual inspection.
Why not just retrieve everywhere? RaDIO's evidence (RaDIO) suggests fixed retrieval wastes calls without improving answers, but that is text, not density. The budget curve over $K$ tests the claim here. If always-on retrieval gives higher completeness with no rise in copying, gating is only a cost optimization, and we will report it that way.
Why not just debias the templates offline with DeSAE? DeSAE corrects idealized bond geometry (AlphaFold Database Debiasing for Robust Inverse Folding). The bias that harms map interpretation is mostly a wrong fold, loop, or domain placement, and smoothing geometry cannot fix that. The library ablation measures whether corrected fragments help at all once the half-map test is in place.
Why not just fine-tune ModelAngelo on low-resolution maps? This could raise the 3.5 Å floor without any retrieval. We will fine-tune it on blurred versions of the EMDB training maps with a matched GPU budget. If the fine-tuned builder closes most of the gap, retrieval's remaining value is limited to regions where no local signal exists, and we will quantify exactly that residual.
Proposed Experiments
The data are the EMDB map–model pairs from the ModelAngelo papers, split by release date and fold to avoid template leakage. Fragments from the target's own deposited entry and from its close sequence homologs above 90% identity are excluded from the index.
Low-resolution test maps are built in two ways. One is reconstruction from subsets of particles. The other, where only maps are available, is low-pass filtering of half-maps to 4, 5, and 6 Å. We add every deposited map in the test period worse than 3.5 Å that has both half-maps and a deposited model. The polymorph set uses the amyloid structures cited above. For each target, the "nearest wrong polymorph" is supplied as an available template.
The metrics are residue completeness, sequence recall, and Cα RMSD to the deposited model. Alongside these we report the template-copy rate, retrieval calls per map, and GPU-minutes per map.
The decisive checks are three. First, at 4 to 5 Å, gated and verified retrieval must beat unchanged ModelAngelo on completeness. Second, on the polymorph set, it must show a lower copy rate than both whole-chain docking and full-map-scored retrieval. Third, it must reach its completeness with a small fraction of the always-on retrieval calls.
If the half-map test does not separate true fragments from decoys at 6 Å, the approach has a resolution limit of its own. We will report where that limit lies.
References
- ModelAngelo: Automated Model Building in Cryo-EM Maps
- Automated model building and protein identification in cryo-EM maps
- Carbohydrate binding, quaternary structure and a novel hydrophobic binding site in two legume lectin oligomers from Dolichos biflorus
- Structural Mechanism of Trimeric HIV-1 Envelope Glycoprotein Activation
- Cryo-EM of NHEJ supercomplexes provides insights into DNA repair
- Structure of human steroid 5α-reductase 2 with the anti-androgen drug finasteride
- Retrieval Augmented Protein Language Models for Protein Structure Prediction
- AlphaFold Database Debiasing for Robust Inverse Folding
- RaDIO: Real-Time Hallucination Detection with Contextual Index Optimized Query Formulation for Dynamic Retrieval Augmented Generation
- CORAG: A Cost-Constrained Retrieval Optimization System for Retrieval-Augmented Generation
- Evaluation of Attribution Bias in Retrieval-Augmented Large Language Models
- Debiasing Masks: A New Framework for Shortcut Mitigation in NLU
- Mitigating Biases for Instruction-following Language Models via Bias Neurons Elimination
- Distinct tau filament folds in human MAPT mutants P301L and P301T
- Cryo-EM structures of Aβ40 filaments from the leptomeninges of individuals with Alzheimer’s disease and cerebral amyloid angiopathy
- Cryo-EM structures of amyloid-β filaments with the Arctic mutation (E22G) from human and mouse brains
- New SNCA mutation and structures of α-synuclein filaments from juvenile-onset synucleinopathy