Telling Blur from Novelty in Low-Resolution Cryo-EM Enhancement

Cryo-electron microscopy (cryo-EM), which images frozen molecules with electrons and reconstructs a 3D density map, often yields only blurry maps (6–30 Å) for large, flexible or native-state assemblies. At that resolution the overall shape is visible, but the atoms are not. Neural "enhancers" now sharpen such maps by filling in fine detail learned from thousands of high-resolution examples. Every detail they add beyond the measured resolution comes from the network's training prior, not from the experiment. Some of that detail is a good guess about a familiar protein. Some is invented structure that resembles the training data.

Standard uncertainty tools cannot tell these two cases apart. They flag both "this map is blurry" and "this molecule looks like nothing I was trained on" with the same raw disagreement signal. We propose to separate the two. We do this in two steps: we measure model disagreement only in the part of the map the data cannot constrain, and we compare it to the disagreement expected at that resolution for familiar molecules. Novelty is scored separately, on the band the data actually measure. We test this on a controlled benchmark that crosses resolution loss with fold novelty, where ground truth is known.

Core Claim

In low-resolution enhancement, raw epistemic uncertainty is dominated by resolution. Epistemic uncertainty here means disagreement between plausible versions of the model, which is supposed to rise on unfamiliar inputs. At low resolution, however, many different high-resolution structures legitimately fit the same blurry map, so disagreement rises with blur even for perfectly familiar proteins.

We claim that a hallucination detector becomes informative only after two corrections. First, disagreement is restricted to the frequencies the data do not constrain (the "null space" of the blurring). Second, it is normalized by the spread expected for in-distribution molecules at the same local resolution.

We also expect the ensemble to sometimes agree confidently on a wrong answer. From Risk to Uncertainty shows this failure mode: on hard shifts, ensemble disagreement fails and total or aleatoric risk works better. To cover it, we add a second score that does not depend on predictor disagreement at all. It is a density estimate of the measured low-frequency band in the enhancer's encoder feature space.

The contribution is threefold: the 2×2 benchmark that makes blur and novelty separable, the two-score detector, and an honest test of whether either score beats simply trusting the local resolution estimate.

Relation to Prior Work

The closest work is BayesDL-SIM. It samples network weights by Langevin dynamics for 2D optical super-resolution microscopy and reports epistemic uncertainty rising "often by ~400%" on unseen cellular structures. That result motivates this proposal, since it shows epistemic signals can catch hallucination in super-resolution. We differ in three ways. We work on 3D density volumes whose degradation operator (a resolution-dependent low-pass filter) is known and can be projected out. We separate disagreement that is expected at a given resolution from genuine novelty; BayesDL-SIM does not separate them, because its train and test images have comparable information content. We also add a disagreement-free novelty score.

Estimating Epistemic and Aleatoric Uncertainty with a Single Model (HyperDM) separates the two uncertainty types in CT reconstruction using a hypernetwork plus diffusion sampling. Two things constrain our design. Its memory cost scales with backbone size. And Laplacian Segmentation Networks finds that aleatoric and epistemic measures stay strongly correlated in skip-connection networks. That is exactly the architecture family of the SCUNet backbone in CryoLVM. We therefore do not rely on a clean variance decomposition. We rely on an explicit frequency split supplied by physics.

The application gap comes from the cryo-EM tools themselves. CryoLVM is a self-supervised model trained on 7,392 maps at 1–4 Å that sharpens and super-resolves maps. It gives point estimates only, and its authors leave behavior on >6 Å maps open. ModelAngelo builds atomic models automatically but stops working below about 3.5 Å. Generative Priors for Cryo-EM Image Reconstruction enforces data consistency on a diffusion prior but reports no uncertainty. We reuse its data-consistency idea as the projection step.

Two further papers supply real low-resolution motivation. Moscoso et al. show that 20–30 Å HIV Env maps admit multiple contradictory fits. Polley et al. describe 6–10 Å maps lacking side chains. We use such maps only for qualitative case studies, because no ground truth exists for them.

Research Program

The detector. We train an enhancer on Cryo2StructData maps that have been synthetically degraded to 4–30 Å. Degradation uses low-pass filtering, B-factor blurring and noise, with the resolution drawn per map. The enhancer follows an SCUNet-style design as in CryoLVM, or uses CryoLVM weights if they are obtainable. For epistemic sampling we use a five-member ensemble, with a last-layer Laplace approximation as a cheaper variant.

Let $P_r$ be the projector onto spatial frequencies below the local resolution $r$, which is the band the measurement constrains. Let $f_1,\dots,f_M$ be the ensemble members applied to input map $x$. The null-space spread at voxel $v$ is

$$ s(v) = \mathrm{Var}_{m}\big[((I-P_r)f_m(x))(v)\big], $$

and the resolution-normalized score is $z(v) = s(v)/\bar{s}(r(v), \rho(v))$. Here $\bar{s}$ is the expected spread for in-distribution calibration maps at local resolution $r(v)$ and predicted density level $\rho(v)$. It is fit by a simple binned regression.

The second score, $d(x)$, is a Mahalanobis or normalizing-flow density of encoder features computed on $P_r x$ alone. It flags inputs whose measured content is unfamiliar even when the ensemble agrees.

The benchmark. We split Cryo2StructData by sequence and structure clustering into in-distribution families and held-out families. We add a third category of maps dominated by non-protein density, such as nucleic-acid-rich or detergent-embedded samples, where these exist in the archive.

Each test map is degraded at five resolutions (4, 6, 10, 15 and 25 Å). This produces a resolution × novelty grid with the high-resolution map as ground truth. A voxel counts as "hallucinated" when the enhancer's null-space content differs from the ground truth's null-space content by more than a threshold. The threshold is set relative to ground-truth density variance, and we sweep it.

Baselines

The naive reference point is local resolution itself: flag every voxel whose local resolution is worse than a cutoff, or score voxels by $r(v)$. This encodes the textbook view that everything beyond the measured band is unsupported. If our score cannot outrank this reference, the enhancer's null-space content is no more trustworthy in familiar regions than in unfamiliar ones. That would be a negative but important finding.

The strongest existing method is a raw deep ensemble with full-band per-voxel variance, the direct 3D analogue of BayesDL-SIM's epistemic map. It is what a practitioner would build first. HyperDM reports that a ten-model ensemble reaches 34.95 dB PSNR against HyperDM's 35.16 dB on LUNA16, so ensembles remain competitive in quality. We also run MC dropout. HyperDM reports it degrading reconstruction to 30.25 dB, and on TopData jets Evidential Deep Learning for Jet Identification reports AUROC 0.887, against 0.890 for ensembles and 0.904 for its evidential variant. We include the evidential variant as a single-pass baseline.

We add total risk and Bayes risk under the log score, following From Risk to Uncertainty, which reports ~91–96% AUROC on CIFAR-10 OOD sets with Bayes risk. We also add a learned per-voxel error regressor trained on in-distribution degraded maps.

On cryo-EM hallucination detection itself, we found no reported result on Lacuna, so no prior number exists to beat. For enhancement quality we will report Q-score on EMReady against CryoLVM's reported 0.444, so that uncertainty gains do not come from a worse enhancer.

All baselines share the same enhancer checkpoints, training data, degradation schedule and calibration set. The ensemble size and GPU-hour budget are matched to our method.

Beating the baselines means two things. First, a higher voxel-level hallucination AUROC than both raw ensemble variance and local resolution, within every resolution stratum and not only pooled. Second, higher map-level held-out-family AUROC at fixed resolution. The ablations isolate each ingredient. We remove the null-space projection (full-band spread), remove resolution normalization, and use $d(x)$ alone.

Why not just…?

Why not just trust local resolution and discard everything beyond it? This is the most serious alternative. It is included as the naive baseline precisely because it might win. The evidence against it is indirect. Moscoso et al. show that maps at identical nominal resolution can hide very different degrees of ambiguity, and CryoLVM's ModelAngelo gain from 92% to 94.3% sequence match on the 28-structure test set shows that prior-supplied detail is sometimes right. The within-stratum AUROC comparison settles the question directly. At fixed resolution, the local-resolution baseline carries almost no information, so any ranking power our score shows there is information that resolution lacks.

Why not just use a deep ensemble's raw variance? Raw variance mixes expected blur-driven spread with novelty. The benchmark will show whether that mixing matters. If raw variance already separates held-out families at fixed resolution as well as $z$ does, the normalization is unnecessary and we will say so. The paradox result in From Risk to Uncertainty further suggests that raw variance will miss confidently wrong outputs on non-protein density. The $d(x)$ score and the non-protein category test exactly that case.

Why not just train a network to predict its own error? A learned error regressor is cheap and could match $z$ on in-distribution families. Because it is fit only on familiar data, we expect it to underestimate error on held-out families. The held-out column of the grid tests this. If it holds up there, it is the simpler method and we would recommend it.

Proposed Experiments

The decisive experiment is the resolution × novelty grid. We report voxel-level AUROC and area under the precision-recall curve for hallucinated voxels, map-level AUROC for held-out families and non-protein maps, and coverage calibration of $z$ thresholds, all stratified by resolution.

The core claim predicts a specific pattern. Raw ensemble variance and local resolution should correlate above 0.8 and perform similarly within strata. The normalized score $z$ should retain its ranking power at fixed resolution. And $d(x)$ should be the only score that catches non-protein maps on which the ensemble agrees. If $z$ does not beat local resolution within strata at 6 Å and coarser, the claim fails for the regime we care about.

A downstream check applies ModelAngelo to enhanced 4–6 Å maps, with and without masking high-$z$ regions. We measure whether masking raises the precision of built residues at acceptable loss of recall. Finally, we run the detector on real low-resolution maps: the HIV Env tomograms studied by Tran et al. and the KMN maps. These serve as qualitative case studies of whether flagged regions coincide with the loops and interfaces those papers report as ambiguous.