PaperlayerView exact revision on arXiv

Source-linked AI summary

ARCHER: Amortized cross-specimen pose estimation for cryo-electron microscopy

Nhan D. Nguyen, Bao Pham

arXiv:2608.22029v1cs.LGmath-phq-bio.BMq-bio.QM

TL;DR

Cryo-EM pose estimation traditionally requires dataset-specific refinement because the specimen is embedded in the estimator. ARCHER conditions an amortized contrastive classifier on a reference volume and applies it across unseen structures without retraining. It achieves competitive reconstruction while preserving downstream conformational signal, with scope bounded by its need for an initial reference.

  • Problem

    Traditional pose estimation restarts per dataset and discards learned comparison computation because the specimen is carried in model parameters.

  • Method

    ARCHER realizes the pose posterior as a contrastive classifier over a discrete rotation grid, using reference-conditioned templates, particle channels, and a jointly trained denoising channel.

  • Results

    ARCHER’s poses preserve downstream conformational signal, retaining 0.323 of the embedding versus 0.311 and 0.248 for compared estimators.

  • Takeaways & Limitations

    Pose assignment can be treated as a reusable operation across structures when the reference volume is supplied as an input rather than stored in the weights.

  • Takeaways & Limitations

    ARCHER requires a reference volume to initialize the loop, although later rounds use a volume rebuilt from its own estimates.

Abstract

from arXiv · show

Single-particle cryo-electron microscopy (cryo-EM) pose estimation is traditionally solved anew for each dataset, where iterative refinement is done from scratch while the estimator learns to store the molecule in its weights. In this work, we show that pose inference is a generalizable, specimen-agnostic operation when conditioned explicitly on a reference volume. We introduce ARCHER, an amortized contrastive classifier that models the pose posterior over a discrete rotation grid. Trained across a variety of protein structures, it operates zero-shot without retraining per structure. This transferability is grounded in Fourier-space information mechanics, where all specimen dependence is captured by the reference structure's power spectrum and spatial extent. ARCHER achieves a median angular error of 5.0° on 100 held-out test structures and 2.5° on experimental particles, matching dedicated estimators within 0.16 Å in 3D reconstruction. Crucially, downstream conformational signal is preserved. The leading conformational coordinate correlates at 0.97 with deposited benchmarks, faithfully reconstructing free-energy basins and mobile domains. These results overall demonstrate that cryo-EM pose estimation can be generalized across different structures.

I. INTRODUCTION

Cryo-EM pose estimation is traditionally recomputed for each dataset because orientation inference and reconstruction are coupled. ARCHER instead conditions pose matching on a supplied reference volume, learning a transferable operation across specimens.

  • The pose-estimation problem: Cryo-EM reconstructs structure by placing Fourier central sections correctly, but unknown particle orientations make placement and reconstruction coupled.Misplacement blurs averaging and drives major computational costs and pipeline failure modes.
  • Existing approaches: Iterative refinement alternates posterior orientation estimation with posterior-weighted volume rebuilding, restarting from scratch for every dataset.Its scoring cost scales with the number of particles multiplied by candidate orientations.
  • Existing approaches: Most learned estimators encode the specimen in network parameters, so a new protein requires new training and prior comparison computation is discarded.Related approaches either amortize within one dataset, use structure-specific supervision, or avoid references for ab-initio determination.
  • ARCHER: ARCHER supplies the reference volume as an input, trains once across structures, and applies the learned matching function to unseen molecules without retraining.The reference separates specimen-specific data from the microscope-governed operation of matching noisy projections to candidate views.
  • Transfer mechanism: Fourier-space orientation information scales with k^2 times spectral signal-to-noise ratio, with specimen dependence supplied by the reference volume’s power spectrum.This shared information structure explains why one encoder can learn across maps and constrains attainable accuracy.
  • ARCHER: ARCHER models the pose posterior with a contrastive classifier over a HEALPix rotation grid, using shared-topology but independently weighted reference and particle encoders.Particle inputs include recorded, phase-flipped, and denoised channels, while the transfer function enters through modulated pixels.

B. Conformational Heterogeneity

Cryo-EM heterogeneity makes pose assignment harder because flexible molecules are mismatched by a single consensus volume. ARCHER’s reference-conditioned, discretized approach is positioned within broader efforts to transfer reusable components across specimens.

  • Existing approaches: Heterogeneity methods range from discrete class assignment and multi-body refinement to continuous latent spaces and deformation-field models.Examples include cryoDRGN, 3DFlex, and RECOVAR.
  • Conformational heterogeneity: Flexible molecules can produce projections that no single rigid volume explains, creating systematic consensus-map mismatch during pose scoring.This is the regime in which pose estimators differ most.
  • Existing approaches: Per-structure encoders replace per-particle search with a forward prediction but amortize training only across particles from one specimen.Search-based corrections can refine initial encoder predictions.
  • Existing approaches: Structure-specific supervised orientation classifiers learn the matching operator for one molecule, with training cost recurring per specimen.Training four CESPED targets took 90.5 single-GPU hours in the reported comparison.
  • Existing approaches: Reference-free methods estimate poses and structure without a starting model, but this is described as the hardest setting and the one where heterogeneity costs most.CryoFastAR and cryoDRGN-AI represent reference-free alternatives.
  • ARCHER’s design: ARCHER uses a HEALPix discretization of SO(3) with continuous refinement to recover sub-cell pose precision while retaining near-uniform coverage.This design makes rotation-grid spacing controllable.
  • Transfer across specimens: Reusable cryo-EM components already transfer for particle picking, preprocessing, regularization, restoration, density priors, and particle features, but pose assignment had remained per-dataset.ARCHER addresses this gap by treating reference-conditioned noisy-projection comparison as shared across specimens.

III. METHODS

The methods derive how Fourier-space information controls pose accuracy and establish why reference-conditioned inference can generalize across specimens. They also distinguish local precision from global pose-basin detection.

  • Information geometry: Pose perturbations move Fourier content along arcs whose displacement grows with spatial frequency, linking angular error to attainable resolution.The displacement magnitude is kδ sin ψ.
  • Information geometry: Orientation information scales as k2 SSNR(k), concentrating useful evidence in a finite high-frequency band before noise and damage dominate.The useful band maximizes k2 SSNR(k).
  • Information geometry: Real-particle curvature follows Λk/ck ∝k^1.89 over shells k = 4 to 32, agreeing with the predicted scaling within a few percent rather than exactly.The autocorrelation-robust uncertainty is 0.042, placing the measurement 2.6σ below 2.
  • Information geometry: A per-particle Cramér–Rao bound sets a specimen-level precision ceiling, but its simplified form assumes the other two rotation components are known.The corresponding inverted 3 × 3 information matrix is less optimistic by a factor of order unity.
  • Information geometry: The local precision bound does not address assigning particles to the wrong basin, which requires a separate detection criterion involving competing orientation candidates.The matched-filter statistic must clear the 2 ln|G| threshold, conservatively adjusted because neighboring grid scores are correlated.
  • Generalization: Reference-conditioned inference separates specimen-dependent power spectrum and spatial extent from a matching operation determined by transfer function, noise spectrum, and slice geometry.The reference supplies Pk and Rg at inference time, while the common operator is trained across specimens.

C. Reference-conditioned posterior

ARCHER scores candidate rotations by contrasting particle and reference-derived template embeddings on a discrete rotation grid. The design trades some CTF-weighted information for reusable template encoding and uses refinement to improve precision.

  • Reference-conditioned posterior: Templates are rendered from the reference on a HEALPix rotation grid and encoded once per training step, while particles are encoded per image.Both residual convolutional encoders map inputs into S^511 ⊂ R^512.
  • Reference-conditioned posterior: Pose logits are temperature-scaled inner products between particle and template embeddings over candidate rotations.The learned temperature falls from 0.07 to 0.0212, sharpening the posterior by a factor of 3.3.
  • Reference-conditioned posterior: Cross-entropy uses geodesic soft targets with σq = 6°, providing gradient to neighboring grid cells while respecting SO(3) geometry.The target is proportional to exp(−dgeo(Ri, R⋆)^2/2σq^2).
  • CTF handling: Phase-flipped particles retain ζ = 0.542 ± 0.009, or 54% of available orientation information, at a factor 1.36 angular-standard-deviation cost.The trade-off permits templates to remain defocus-independent.
  • CTF handling: Encoding templates once per specimen rather than per particle makes amortization affordable by avoiding defocus-dependent template embeddings.This produces one template bank per specimen instead of one per particle.

D. Diffusion denoising channel

ARCHER adds a jointly trained Fourier-domain denoising channel to sharpen pose estimates where single-particle noise overwhelms orientation-bearing shells. The channel improves fine precision while preserving coarse assignment.

  • Denoising channel: The diffusion channel supplies a denoised particle copy as a third encoder input, targeting shells where noise dominates orientation information.The operation is motivated by restoring amplitude in the orientation-informative band.
  • Denoising channel: The corruption schedule places the denoiser at a known terminal measurement, making evaluation cheap enough for differentiable training-loop integration.Equation (13) evaluates the denoiser once at t = T − 1 rather than sampling a reverse trajectory.
  • Denoising channel: A U-Net predicts noise using the complex image, transfer function, and time embedding, while the total loss combines pose cross-entropy with ε-prediction.The ε-prediction loss is weighted by inverse noise variance across shells.
  • Denoising channel: Wiener inversion estimates the clean image with a bounded regularized gain that avoids inventing content in transfer-function blind bands.Direct inversion diverges at zeros of H, whereas the regularized gain is bounded by 1/(2λ).

E. Hierarchical refinement

Hierarchical refinement starts from the grid maximum, rebuilds a volume, rescoring particles against rendered templates before applying a continuous derivative-free Newton update. The procedure stays on SO(3) and preserves independent half-set processing.

  • Refinement procedure: Given current poses, ARCHER reconstructs a volume, rescored particles against low-shell shell-normalized templates, and polishes each pose with a derivative-free Newton step.The pipeline proceeds from discrete grid inference to a continuous estimate.
  • Refinement procedure: The Newton update parameterizes nearby rotations with the exponential map and applies R ← R exp(δ⋀), keeping estimates on SO(3).The stationary point is δ⋆ = −Λ−1g.
  • Refinement procedure: Central differences estimate gradients and diagonal curvature without differentiating through the scorer, allowing the same refinement routine for learned and classical scores.Wrong-sign curvature falls back to a clamped gradient move for stability on flat plateaus.
  • Training: Training samples specimens and rotations from a reference-volume bank, with the reference treated as data rather than a learned parameter.The training algorithm encodes candidate templates and computes pose cross-entropy jointly with denoising loss.
  • Training: The procedure uses Titer = 3 outer refinements and Tnewton = 4 inner Newton steps, while independently processing half-sets against opposite-half volumes.This preserves gold-standard separation throughout reconstruction and scoring.

F. Data

Training uses particles and volumes from 3,330 structures, with simulated noise and Fourier-cropped half-maps standardized to a 64^3 box. Hierarchical inference initializes poses on a rotation grid, alternates independent half-set volume updates, and applies Newton refinement.

  • Training data: Training episodes sample volumes and 96 particles from 3,330 EMDB structures using known poses.
  • Inference: Hierarchical inference predicts an initial volume, scores grid rotations, updates independently processed half-sets, and iterates refinement for T_iter = 3.
  • Inference: NewtonRefine updates three body-axis directions through four iterations after grid-based pose selection.
  • Training data: Particles receive additive noise colored by real-micrograph power spectra, with signal-to-noise ratios from 0.005 to 0.075.
  • Volume preprocessing: Deposited half-maps are Fourier-cropped to a common 64^3 box, preserving whole molecules while discarding high frequencies.

G. Evaluation

Evaluation separates generalization to unseen proteins from domain shift by using held-out simulated particles and experimental EMPIAR data. Pose, map, and conformational agreement are measured with explicit benchmark protocols.

  • Evaluation design: Held-out structures are evaluated on simulated particles to isolate generalization to unseen proteins from domain change.
  • Evaluation design: Experimental evaluation uses EMPIAR-10076 for pose accuracy and heterogeneity, while five CESPED targets are used for reconstruction.
  • Metrics: Angular error is the geodesic distance on SO(3), summarized by the median and the fraction within a threshold.
  • Metrics: Map quality uses independently reconstructed half-sets compared with deposited structures at the 0.143 Fourier shell correlation criterion.
  • Metrics: Conformational landscapes are compared by matching RECOVAR basins and evaluating mobility fields weighted by basin accessibility.

IV. RESULTS

ARCHER combines reference-volume embeddings with a discrete HEALPix rotation grid and achieves accurate pose estimates across experimental and held-out settings. Comparisons show strong transferability, while remaining errors concentrate in orientation selection.

  • Pose estimation: ARCHER embeds particles and reference templates on a shared unit sphere and selects poses by temperature-scaled cosine similarity.
  • Pose accuracy: A 2.5° median error is achieved on experimental particles, with 78% within 5° of deposited poses.
  • Comparisons: A 4.2° median error for cryoPARES is more accurate than the full ARCHER pipeline's 7.5° median across four targets, but requires fitting per specimen.
  • Transferability: Applying cryoPARES models off diagonal yields 80° to 135° median errors, showing that its single-specimen training does not transfer across targets.
  • Error structure: ARCHER residual errors concentrate in a band of viewing directions rather than spreading diffusely over the sphere.

B. Reconstructions quality

ARCHER produces reconstructions comparable to dedicated estimators on some targets, but reconstruction quality depends strongly on sampling limits and pose-error shape. Its poses also preserve conformational signal and support transfer across specimens through shared orientation-information scaling.

  • Reconstructions: On EMPIAR-10409, ARCHER reaches 3.634 Å versus 3.625 Å for the classical filter and 3.477 Å for cryoPARES, all within 0.16 Å.
  • Reconstructions: Across four targets, ARCHER trails cryoPARES by 0.16 Å and 0.22 Å on two targets but by 2.11 Å and 2.06 Å on two others.
  • Sampling limits: Across five CESPED targets, reconstructions lie within factors of 1.00 to 1.39 of the 64^3 working-box Nyquist limit.
  • Error shape: Uniform 10° pose blur degrades maps to 6.5 Å, whereas heavy-tailed errors with the same mean remain at the 2.94 Å Nyquist limit.
  • Conformational signal: The leading conformational embedding correlation is 0.97, remaining consistent from two to twenty latent dimensions.
  • Conformational signal: On EMPIAR-10409, ARCHER retains 0.323 embedding agreement versus 0.311 for the classical filter and 0.248 for cryoPARES.
  • Conformational landscapes: Matched conformational basins show map correlations of 0.92–0.99 and mobility-field agreement of 0.90–0.94.
  • Information scaling: Measured orientation-information scaling has exponent 1.89 with autocorrelation-robust standard error 0.042, close to the predicted k^2 law.

V. DISCUSSION

ARCHER shifts pose estimation from a per-dataset operation to a reference-conditioned, transferable one. Its main benefit is generality across structures, while performance depends on reference quality and downstream conformational preservation.

  • Conditioning on a reference volume lets one parameter set serve proteins absent from training, with accuracy competitive with per-dataset methods.
  • ARCHER’s transferability is supported empirically rather than implied by the shared Fourier-space functional alone.
  • ARCHER occupies the reference-available regime, complementing ab-initio and supervised per-dataset estimators.
  • 0.323 of the conformational embedding is retained, versus 0.311 and 0.248 for comparison estimators.The authors argue that conformational signal should be reported alongside resolution for flexible assemblies.
  • 4.54° versus 4.53° at the grid stage and 2.60° versus 2.54° after refinement when using self-generated versus independent references.The agreement holds within 0.1° over three further rounds.
  • A sufficiently good initial reference is required; a deliberately crude reference causes accuracy to fall as iterative errors compound.The stated suitable starting points include a consensus map, homologue, or predicted structure.
  • ARCHER reconstructs unseen structures within competitive error margins using conditionals obtained from reference structures.
  • When target sampling is not binding, ARCHER trails a specimen-trained estimator by about 2 Å; generality, not accuracy, is its main advantage.When sampling binds, conformational signal remains the distinguishing downstream measure.

Appendix B: Architecture and objective

The architecture uses separate residual convolutional encoders to compare particle and template representations, with deliberate restrictions that keep pose inference non-generative. Optimization and channel ablations define the reported configuration and its scope.

  • Architecture: The image and template branches map inputs to separate 512-dimensional, ℓ2-normalized embeddings without shared weights.
  • Channel assembly: The reported image branch uses two channels, while the template branch uses one and the denoiser can provide a third.
  • Stem: A stride-2 stem reduces activation memory fourfold while costing little at L = 64.
  • Residual trunk: The residual trunk uses four stride-2 blocks with widths [128, 256, 512, 512] and a 2 × 2 output at L = 64.
  • Projection head: The projection head applies global average pooling and two linear layers before ℓ2 normalization onto S511.
  • Objective: The learned temperature converges to τ = 0.0212, making the posterior 3.3 times sharper than at initialization.
  • Design constraints: The pose network maps two images to vectors and scores a fixed candidate set with a softmax, while volumes are recovered by a closed-form linear solve.
  • Input representation: The CTF is supplied through pixel channels rather than as a parameter vector; raw and phase-flipped channels correlate at only 0.047.

Appendix C: Full ablation and sweep tables

The appendix documents evaluation settings, training sweeps, detectability, confidence-based junk rejection, and ambiguity-limited classical scoring. Together, these analyses identify calibrated operating conditions and show that pose confidence separates real particles from multiple junk classes.

  • Evaluation protocol: The reported held-out accuracy uses an order-3 grid with 36,864 cells, while ablation and curriculum tables sweep training choices separately.Table S3 distinguishes the main-text evaluation configuration from training-choice sweeps.
  • Annealing and endpoint sweeps: Hard-noise performance improves monotonically across five re-annealing stages, whereas the legacy metric improves through stage four before flattening at 0.6130.The stage trend supports a dose–response interpretation rather than a single-point effect.
  • Annealing and endpoint sweeps: The calibrated endpoint leads the legacy endpoint at all 32 matched steps under otherwise identical schedules, differing only in SNR bounds.The controlled comparison holds checkpoint, stage count, caps, patience, learning-rate schedule, and plateau metric fixed.
  • Particle detectability: A real particle lies 1.7× above the 2 ln M = 16.9 detectability threshold, and its true pose ranks first among 4,608 candidates.This measurement establishes substantial pose information in experimental particles at the tested box size.
  • Confidence and junk rejection: The normalized negative entropy pose confidence separates real particles from all three junk classes at AUROC 0.963–0.969.Phase-scrambled images preserve the real amplitude spectrum but are rejected like pure noise, indicating sensitivity to structure rather than image power alone.
  • Ambiguity and operating point: Classical scoring reaches 0.8222 within 15° for non-symmetric proteins at SNR 0.04, while pooled-protein performance is limited by pose ambiguity.The non-symmetric result matches the real-particle anchor of 0.8217 within 0.0005.

7. Symmetry: the ambiguity audit and its point-group verification

This section audits pose ambiguity caused by point-group symmetry and verifies that the reported comparisons use controlled, directly comparable settings. It also separates symmetry-aware corrections from reconstruction and implementation checks.

  • Point-group verification: All 26 symmetric proteins quantize to integer point-group orders, with none falling between orders.The observed dominance of C2 and D2/C4 matches the expected composition of an EMDB-derived oligomeric-complex corpus.
  • Symmetry correction: Symmetry-aware correction changes the D2 result while moving the non-symmetric control almost nothing, supporting a symmetry-specific rather than inflationary effect.Table S12 scores both checkpoints on the same 100-protein set and pinned noise range.
  • Controlled comparisons: Across four targets, median angular errors are 9.5° for the network, 7.5° for the pipeline, 6.6° classically, and 4.2° for cryoPARES.The corresponding fractions within 15° are 0.689, 0.714, 0.809, and 0.805, producing different orderings.
  • Training cost: ARCHER is trained once across 3,330 structures, whereas cryoPARES requires a separate optimization for each target.For a new target, ARCHER incurs an inference pass rather than recurring per-specimen training.
  • Validation checks: Reconstruction and implementation checks compare against independent references, including a mean FSC of 0.9325 against an external reconstructor.These checks cover geometric, transfer-function, and reconstruction conventions.
Loading 2608.22029v1…