Atom-in-the-Loop Reconstruction for Cryo-EM with Missing Views

When a generative model fills in parts of a 3D object that no camera ever saw, it tends to paint in what it has seen most often. In text-to-3D this is the "Janus problem": a model trained mostly on frontal photos puts a second face on the back of a dog. Single-particle cryo-electron microscopy (cryo-EM) builds protein structures from thousands of noisy 2D projection images taken at unknown angles, and it has the same weakness. Molecules often settle on the grid in a few preferred orientations, so some viewing directions are barely sampled. A learned prior asked to fill those directions has every reason to reuse density patterns from the dominant views. Per-image scores such as PSNR and SSIM cannot catch this, because the hallucinated regions are exactly the ones no held-out image covers. What can catch it is an atomic model: real proteins have fixed bond lengths, no overlapping atoms, allowed backbone angles, and a single handedness.

The proposal is to put a fast atomic model-building and validity-scoring step inside the reconstruction loop of a diffusion-prior cryo-EM pipeline. A diffusion model is a generator that learns to turn random noise into realistic samples by gradually removing the noise. Every few denoising steps, we fit atoms to the current volume and check them for steric and stereochemical validity. Projections of that atomic model then constrain the poorly sampled directions. The claim has two parts. First, under controlled preferred-orientation subsampling, existing generative-prior reconstructions show a measurable Janus-like failure: their image metrics stay good on observed angles while atomic accuracy degrades in unobserved ones. Second, atomic validity is a better source of consistency in the missing views than reprojection consistency in image space.

Where this sits

The method most directly extended is Generative Priors for Cryo-EM Image Reconstruction. It trains a diffusion prior on EMPIAR micrographs to reconstruct from compressed or masked projections. It also argues that PSNR and SSIM mislead in cryo-EM, and that tracing atoms with ModelAngelo is the real test. That paper motivates our evaluation choice, but it uses atomic tracing only at the end, as a judge. By the file's own account it does not address preferred orientation or multi-view consistency, and it does not apply stereochemical refinement during the reverse diffusion steps. We move the judge into the loop and study the orientation-starved regime it leaves open.

The loop structure is borrowed from text-to-3D. Carve3D reduces Janus artifacts with a reward computed by lifting generated views into a radiance field, re-rendering them, and scoring the result with LPIPS. Its consistency signal lives purely in image space. DreamCS argues that such 2D rewards are biased and instead scores explicit 3D meshes, cutting MVDream's Janus rate from 52% to 30%. That result is the main evidence that a signal defined in 3D, rather than on renders, is what removes multi-face artifacts. Our atomic model plays the role of DreamCS's mesh. Unlike a learned mesh preference, its validity is defined by chemistry, not by human ratings.

The molecular side constrains the design. PoseBusters and PoseX show that models scored by coordinate RMSD produce clashes, warped rings and inverted chirality. They also show that post-hoc force-field relaxation repairs much of this: SurfDock's validity rose from 41.7% to 73.6% after OpenMM relaxation. However, relaxation cannot rescue wrong topology on novel targets. This tells us two things. Restrained minimization of the kind used in HiQBind-WF is a serious baseline. The cases where in-loop correction should help are those where the volume itself, not the atoms fitted to it, is wrong.

Protein model accuracy estimation in CASP14 shows that image metrics become informative when they are applied to geometric objects such as distance maps rather than raw pixels. This supports the idea that scores tied to geometry track structural truth. Finally, AtomGS and Laplace-Beltrami Operator for Gaussian Splatting show in vision that geometric error keeps changing after PSNR has plateaued. That is the same decoupling we expect to measure here.

Research Program

The first step is diagnostic. Starting from EMPIAR-10076 (50S ribosome) and EMPIAR-10648 (PKM2), the two datasets used by the generative-priors paper, we create orientation-starved subsets. We remove particles whose estimated poses fall inside chosen cones of viewing directions, with cone sizes increasing until large regions of the sphere of directions are empty. The removed particles are kept as held-out views.

For each reconstruction method we then report three kinds of quantities. The first is image metrics on observed angles. The second is image metrics on the held-out angles. The third is atomic accuracy, split into residues whose local density depends mainly on the missing directions and residues that do not. Atomic accuracy here means ModelAngelo backbone RMSD to the deposited model, clash score, and the rate of backbone-angle outliers.

The central diagnostic statistic is the rank correlation, across methods and cone sizes, between observed-view image quality and atomic accuracy. We predict that this correlation weakens as cones grow, which would make the metric misalignment quantitative. As a check on the "canonical bias" hypothesis, we also test whether density in the missing directions resembles a rotated copy of dominant-view density more than the true held-out density does.

The second step is the method. We run the diffusion-prior reconstruction as before. Every K denoising steps, a fast tracing pass fits a backbone, plus side chains where the density supports them, to the current volume estimate. The fitted model is then minimized against a stereochemical energy (bond, angle, clash and torsion terms) with harmonic restraints to the traced positions. This is the same division of labor as post-docking relaxation: the network finds the rough structure, and physics cleans up the local geometry.

Next we simulate projections of the relaxed atomic model at the missing angles and add a guidance term that pulls the volume's projections at those angles toward them. This term is weighted by a per-residue confidence taken from the tracer. Handedness is fixed early, because projections cannot tell a molecule from its mirror image while L-amino acids and right-handed helices can. This is a cheap sanity check on the loop, not a contribution. Because tracing is the expensive part, K and the size of the traced region are explicit compute knobs, and we report wall-clock cost at every setting.

The third step is a stress test on heterogeneity. On EMPIAR-10076 we check whether atom-in-the-loop guidance merges distinct conformational states, which would be a real risk of pulling every volume toward one atomic template. We measure this with the cluster-agreement metric of the generative-priors paper.

Baselines

The naive reference point is standard reconstruction from the orientation-starved subset with no learned prior. This shows how bad missing views are without hallucination of any kind. Its blur in the missing directions is honest, and a method that gives better image scores but worse atomic accuracy than this reference is failing in exactly the Janus-like way we describe.

The strongest existing method is the diffusion-prior reconstruction of Generative Priors for Cryo-EM Image Reconstruction, run with its released settings on identical particle subsets. That paper reports a ModelAngelo backbone RMSD of 2.34 Å on EMPIAR-10648 at 2× compression and 87.9% heterogeneity cluster agreement on EMPIAR-10076. We found no reported result on Lacuna for any method under preferred-orientation subsampling, so we must establish this baseline ourselves. Beating it means lower backbone RMSD, together with fewer clashes and backbone-angle outliers, on the residues controlled by the missing directions. This must hold at matched total GPU time, with the baseline given extra denoising steps or ensemble samples to equal our tracing cost, and without losing accuracy on well-sampled residues or heterogeneity agreement.

The post-hoc baseline traces atoms on the baseline's final volume and applies restrained force-field minimization, as in PoseX and HiQBind-WF. Using the same minimizer as our loop makes the comparison isolate timing, in-loop versus after the fact, rather than the energy function.

The key ablation replaces atomic guidance with Carve3D-style image-space consistency. In this version, the guidance term pulls held-out reprojections of the current volume toward agreement with a re-reconstruction from those reprojections, with no atoms involved. A second ablation keeps the atomic loop but drops the stereochemical minimization, so we can tell whether the gain comes from validity or simply from passing the density through a protein-shaped bottleneck.

Why not just…?

Why not just relax the final atomic model with a force field? PoseX shows that relaxation repairs local chemistry very effectively. If the hallucinated density in missing directions is merely blurry, post-hoc refinement could match us. PoseBusters, however, shows that relaxation fails when the topology is wrong, and a duplicated or misplaced density feature is a topology error that minimization cannot move across. The post-hoc baseline above settles the question directly. If it matches our method on missing-direction residues, the in-loop claim is wrong and we will say so.

Why not just train the prior on orientation-balanced data? This is a real threat to the bias explanation. If the prior's hallucinations come from orientation imbalance in its training micrographs, rebalancing might be enough. We will include a prior retrained with pose-balanced sampling. Rebalancing still does not supply information about angles that are absent for the target molecule, so we expect it to reduce the bias without removing the ambiguity. If rebalancing alone closes the gap, the paper becomes a diagnostic paper with a cheap fix, which is still worth reporting.

Why not just enforce reprojection consistency, as Carve3D does? Image-space consistency is cheaper and needs no tracer. DreamCS's finding, that 3D-defined rewards outperform render-based ones on Janus artifacts, suggests it will fall short. The reason is that a self-consistent but wrong volume reprojects consistently. The image-consistency ablation tests this directly.

Why not just collect tilted or more data? Tilted collection is the experimental remedy and our held-out particles act as its upper bound. Where extra collection is affordable, it should be done. The method targets the many deposited and in-progress datasets where it was not.

Decisive Checks

The idea succeeds if three things hold together under matched compute. On the largest cones, atomic accuracy on missing-direction residues beats both the diffusion-prior baseline and its post-hoc-relaxed version. The image-consistency ablation recovers clearly less of that gain. Heterogeneity cluster agreement on EMPIAR-10076 does not fall below the baseline's. The idea fails if image-space consistency or a rebalanced prior matches atomic guidance, or if guidance collapses distinct conformations toward one template. Either failure would be informative: it would tell us whether the cryo-EM Janus effect comes from prior bias or from a lack of 3D structure in the consistency signal.