ARCHER

Every single-particle cryo-EM dataset poses the same question — which way was each frozen molecule facing when the electrons went through it — and every estimator built so far answers it by absorbing that particular molecule into its own parameters. Refinement in RELION or cryoSPARC rebuilds the volume from scratch per project; encoders such as cryoAI and cryoFIRE train one network per specimen. The volume lives inside the model, so nothing learned on one specimen carries to the next.

ARCHER — Amortized Reference-conditioned Hierarchical Refinement — moves the specimen out of the weights and into the input. Nhan D. Nguyen at the Pritzker School of Molecular Engineering, University of Chicago, and Bao Pham at Rensselaer Polytechnic Institute posted it to bioRxiv on 21 August 2026. It takes the reference volume as an argument and learns only the operation that survives a change of specimen: matching a noisy, CTF-modulated projection against a candidate view rendered from that volume. The comparison turns on the transfer function, the noise spectrum and the central-slice geometry — not on which molecule is in the ice, which enters only through its power spectrum and spatial extent.

That places ARCHER between two neighbours: ab initio methods such as cryoDRGN-AI need no reference at all, while map-space priors such as cryoFM act after reconstruction and yield no per-particle pose. ARCHER assumes what is usually to hand — a consensus map, a homologue, or a predicted structure.

#Key Features

• Reference volume as an argument: The map is supplied at inference rather than fitted into the parameters, so one set of weights serves proteins that were never in training.
• Contrastive scoring over a rotation grid: The pose is the cell whose template embedding maximizes a temperature-scaled cosine similarity against the particle embedding, on a HEALPix discretization of SO(3) with 36,864 cells at a median spacing of 7.40°. Templates are encoded once per specimen, not once per particle — what makes amortization affordable.
• CTF enters as image content: Each particle arrives as three channels — the recorded image, its phase-flipped copy, and a denoised copy from a jointly trained Fourier-domain diffusion denoiser whose forward process interpolates the identity into the transfer function.
• Refinement below the grid: A derivative-free Newton step along three body axes carries the grid maximum to a continuous pose, cutting median error on EMPIAR-10076 from 4.5° to 2.5°.
• Heterogeneity survives amortization: The same particles embedded with ARCHER's poses and with the deposited ones give a leading conformational coordinate correlating at 0.97.

#Technical Details

Two residual convolutional encoders — roughly 10.4 M parameters per branch — map the particle stack and the rendered templates into the unit sphere of a 512-dimensional embedding space; a U-Net denoiser trains alongside them. Training draws 3,330 structures from the EMDB with 100 held out, cropping every half-map in Fourier space to a common 64³ box, and simulates 96 particles per episode at signal-to-noise ratios of 0.005–0.075 with defocus drawn from experimental star files, over 2 × 10⁶ episodes. On the 100 held-out structures ARCHER reaches a median angular error of 5.0° with 77% of particles within 15°, and on experimental EMPIAR-10076 particles 2.5° with 78% within 5°. Under the CESPED benchmark protocol, EMPIAR-10409 reconstructs to 3.634 Å against 3.625 Å for the classical matched filter and 3.477 Å for the per-specimen estimator cryoPARES — all three within 0.16 Å — while on retained conformational signal ARCHER leads instead, at 0.323 against 0.311 and 0.248. Orientation information per Fourier shell is derived to scale as k²·SSNR(k), and measured at an exponent of 1.89.

#Applications

The model suits pipelines that already have a starting map and need orientations assigned to a fresh particle stack without standing up a training run for it. Scoring is end-to-end — every particle scored, none dropped by an interactive junk-filtering round — so it sits ahead of refinement and heterogeneity analysis. Fitting a per-specimen estimator to four CESPED targets took 90.5 single-GPU hours in the authors' hands, against one training run that serves all four.

#Impact

ARCHER shows that orientation assignment, long treated as irreducibly per-dataset, is a shared operation one set of weights can perform across proteins. The authors are candid about what bounds it: the estimator needs a reference good enough to place the first estimate in the correct basin; where a target's own sampling does not bind it trails a specimen-trained estimator by about 2 Å; and cryoPARES keeps an angular-error advantage on the dataset it was fitted to. The work is a bioRxiv preprint under CC BY 4.0 and has not been peer reviewed; its code-availability statement names a GitHub repository that was not yet public at the time of cataloging, and no trained weights have been released.