Matching Protein Sequences to Low-Resolution Cryo-EM Density

Cryo-electron microscopy (cryo-EM) produces a 3D map of how densely matter scatters electrons. A map is a voxel grid in which a protein appears as a blob of density. At high resolution (better than about 4 Å) individual side chains are visible, and tools can read the amino acid sequence straight off the map. Many real maps are worse, in the 4–10 Å range. These include flexible complexes, membrane assemblies, and in-cell tomograms. At that resolution helices and domain outlines survive but atoms do not. The basic question a biologist asks of such a map, "which protein is this?", then becomes hard. The best sequence-aware builder, ModelAngelo, identifies chains by predicting residue types from the density and searching proteomes, but it degrades sharply past 4–5 Å. The best low-resolution method, CryoDomain, works down to 10 Å only because it ignores sequence entirely. It retrieves a domain fold from a structure database, so it cannot say which of several same-fold paralogs occupies the density.

We propose to train a model that scores a density region directly against candidate protein sequences. A 3D convolutional encoder turns the blurry map into a set of local "density tokens." A protein language model turns each candidate sequence into one vector per residue. A protein language model (PLM) is a network pretrained on millions of sequences that gives each residue a vector summarizing its evolutionary context. The match score lets every density token find its best-matching residue independently. This "late interaction" scheme has already beaten pooled embeddings for remote sequence homology in ProtoCol. Our bet is that this token-to-residue matching is what makes sequence information usable at low resolution. It should tolerate domain motions that break rigid fitting, and it should pick up sequence-specific signatures that fold-level retrieval throws away. Those signatures include helix lengths, loop insertions and segment order.

Core Claim

Our claim concerns the 4–10 Å regime. A density-to-sequence model that keeps multiple tokens on both sides should identify the correct chain from a proteome-scale candidate list more accurately than two existing routes: sequence-free fold retrieval, and residue-reading builders. It should also match or beat the strongest simple pipeline, which predicts a structure for every candidate sequence and then fits it into the map. We expect the gain to concentrate in two kinds of case. The first is maps whose conformation differs from the single state a structure predictor outputs; CryoBoltz documents this "single-state bias." The second is candidate sets dense with same-fold homologs. As a by-product, the argmax of the match gives a coarse alignment of residue segments to density regions. That alignment can seed downstream fitting.

The method differs from its three closest relatives in specific ways. ModelAngelo fuses ESM-1b embeddings with density but needs density sharp enough to call residue types. CryoDomain aligns density to atomic structures, never to sequences. SMARTFold fuses multiple sequence alignments (MSAs) with density for structure determination, assuming the sequence is already known, and its authors name replacing MSAs with single-sequence PLMs as open. Our task is identification rather than building. We use single-sequence PLM tokens, and we target exactly the resolution band where the first two methods split.

Research Program

Density encoder. The density tower reuses CryoDomain's design, a residual 3D U-Net that captures both local voxel texture and long-range shape. We take this as the supplied method because the medical volumetric literature consistently shows that native 3D kernels resolve low-contrast boundaries that slice-wise processing loses (Dou et al., 3D U-Net). The voxel features are pooled into roughly 8 Å local patches, giving a set of density tokens $\{d_i\}$. Voxel grids are not rotation-invariant, and Ornate shows that canonical local frames matter. At low resolution there is no backbone frame, so each patch is expressed in the frame of its local density principal axes, and random rotations are applied during training.

Sequence encoder. ESM-2 produces per-residue tokens $\{s_j\}$. We start with the 35M model that sufficed in ProtoCol, then scale it up. The tokens pass through a learned projection, with the PLM lightly fine-tuned.

Score. For a density region $D$ and a sequence $S$, the score is

$$\mathrm{score}(D,S)=\frac{1}{|D|}\sum_{i\in D}\max_{j\in S}\,\langle d_i, s_j\rangle,$$

where $\langle\cdot,\cdot\rangle$ is a dot product between unit-normalized vectors. Because each density patch picks its own best residue, rearranging two domains leaves the score nearly unchanged. Missing or disordered segments simply go unmatched. A second, symmetric term checks that every residue is matched by some density token, which penalizes candidate sequences that are much longer than the density.

Training data and objective. Training uses a contrastive (InfoNCE) loss over map–sequence pairs. The pairs come from two sources: EMDB maps with deposited models at 4–10 Å, and simulated maps from PDB and AlphaFold DB structures low-pass filtered to random resolutions in that band. Where the PDB holds several conformations of the same sequence, all of them are used. Hard negatives are sequences from the same SCOPe/CATH superfamily as the positive. This forces the model to discriminate beyond fold. We hold out superfamilies, not just chains, using the group-disjoint logic ProtoCol applies on Pfam.

Baselines

The naive reference is a length-and-volume heuristic. It ranks candidate sequences by how well their molecular mass matches the enclosed density volume at the map's contour. Any learned method that fails to beat this has learned nothing about sequence.

The first residue-reading baseline is ModelAngelo's own identification route: build into the map, convert predicted residue probabilities to HMM profiles, and search the candidate proteome with HMMER (ModelAngelo). It is the strongest existing sequence-aware identifier. On its 177-map test set at ≤4 Å it reports 77% residue completeness, and 85% for residues with Q-score above 0.4. No reported result was found on Lacuna for its chain-identification accuracy at 5–10 Å, so we will measure it ourselves on the same low-pass-filtered maps.

The second baseline is CryoDomain, the strongest low-resolution method. Its paper reports an 84% relative gain in top-1 accuracy and a 67% relative gain in mAP over prior tools on about 200,000 domain maps at 4–10 Å. The file gives no absolute values. We will retrain or rerun it on our splits. To make it answer the sequence question, every candidate sequence's AlphaFold DB structure is embedded with its AtomTower, and retrieved structures map back to their sequences. This is the fairest possible version of "fold retrieval as identification."

The most dangerous baseline is fold-then-fit. Every candidate sequence is predicted with a structure predictor. Each prediction is simulated as a density at the map's resolution and rigidly fit into the map. Candidates are ranked by cross-correlation. We will also run a variant that splits each prediction into domains and fits them independently.

All learned systems get the same training maps, the same resolution augmentation and the same hyperparameter-search budget, and are evaluated on identical candidate lists.

Four ablations isolate our ingredients. The first replaces late interaction with mean-pooled single vectors on both sides. The second replaces ESM-2 tokens with one-hot residues, which tests whether evolutionary context carries the signal. The third removes same-superfamily hard negatives. The fourth removes the local-frame orientation of density patches.

Beating the field means two things. Overall, we need higher top-1 identification accuracy than every baseline at 6–10 Å. On the conformational-mismatch and dense-homolog subsets, the margin over fold-then-fit must hold with confidence intervals excluding zero.

Why not just…?

Why not just fold the whole proteome and fit predictions into the map? This is the real threat, and we treat it as a co-equal competitor rather than a straw man. Its failure mode is documented: sequence-to-structure models output one state. On SMARTFold's single-chain benchmark AlphaFold2 alone reaches median ChainMatch 0.800, against 0.956 once density is used. CryoBoltz needed density guidance to recover both inward and outward states of STP10. Rigid fitting of a wrong-state prediction should lose correlation exactly where late interaction does not. Domain-split fitting may close much of that gap, which is why it is a baseline. If fold-then-fit with domain splitting matches us on the mismatch subset, we will report that our contribution reduces to speed and alignment output, not accuracy.

Why not just extend ModelAngelo to lower resolution? ModelAngelo depends on recovering per-residue identities from side-chain density. Its own reported limits, together with CryoDomain's report of dramatic decline past 5 Å, indicate that this signal is physically absent in the target regime. Our resolution sweep from 4 to 10 Å will show where its HMM route stops working relative to ours.

Why not just use CryoDomain and map the fold back to sequences? Fold retrieval collapses paralogs that share a fold, and this is precisely the limitation its authors accept by being sequence-free. The same-superfamily candidate sets directly measure how much accuracy that collapse costs. If CryoDomain-plus-AlphaFold-DB ties us there, then sequence tokens add nothing beyond predicted structure.

Why not just a pooled single-vector embedding? That design is cheaper, but ProtoCol found that pooled ESM-2 embeddings trailed late interaction by about 7 points in SCOPe Recall@1 (0.946 for the late-interaction model). The pooled ablation tests whether the same holds across modalities.

Proposed Experiments

Benchmarks. The primary benchmark is ModelAngelo's 177-map protein test set, low-pass filtered to 5, 6, 8 and 10 Å. Candidate lists are the full source-organism proteome. Simulated maps can mislead: Aligning Proteins and Language fell from 35.18% top-1 on PDB-derived inputs to about 16–18% on real EMDB maps. We therefore build a second test set of experimental EMDB maps natively at 4–10 Å, deposited after our training cutoff, with chain masks taken from deposited models. Both sets are scored by top-1 and top-10 identification accuracy and by mAP. We also run CryoDomain's low-resolution domain benchmark so that its numbers are directly comparable.

Decisive checks. Three checks decide the claim.

The first is the conformational-mismatch subset. It contains maps whose deposited model differs from the AlphaFold DB prediction by more than a set TM-score margin, and it includes the transporter and antibody targets CryoBoltz studied. Our method must beat rigid and domain-split fold-then-fit here, or the core claim fails.

The second is the same-superfamily candidate setting. Distractors are restricted to homologs below 30% identity sharing the target's fold, so that fold retrieval cannot succeed by construction.

The third is a falsification check on the alignment by-product. For correctly identified chains, the argmax residue segments assigned to density patches should fall within about one secondary-structure element of their positions in the deposited model. If identification succeeds while alignment is random, the model is exploiting global shortcuts such as size, not reading sequence from density.

Finally, the resolution sweep should show our accuracy degrading smoothly while ModelAngelo's collapses near 5 Å. Below 4 Å we expect ModelAngelo to remain better, and we will report that honestly.