henhao
Sign in
Cell BiologyNature Methods
Peer-reviewed journal article

AreTomoLive: Automated reconstruction of comprehensively-corrected and denoised cryo-electron tomograms in real-time and at high throughput

Ariana PeckiD, Yue Eric YuiD, Mohammadreza ParaaniD, Dari KimaniusiD, Utz H. ErmeliD, Joshua HutchingsiD, Jonathan SchwartziD, Daniel SerwasiD, Hannah SiemsiD, Norbert S. HilliD, Mallak AliiD, Julia PeukesiD, Garrett A. GreenaniD, Shu‐Hsien SheuiD, Elizabeth A. MontabanaiD, Bridget O CarragheriD, Clinton S. PotteriD, David A. AgardiD, Shawn Q. ZhengiD

PDFDOIPreprintCode
Post to X
Threads on XPost to BlueskyThreads on Bluesky
In-depth full-text digest available · Jump to digest ↓

Abstract

A high throughput processing pipeline that performs comprehensive corrections is needed to realize the full potential of cryo-electron tomography and subtomogram averaging. The fields fragmented software landscape remains a significant hurdle to this end. Here we present AreTomoLive, an automated real-time pipeline composed of two GPU-accelerated packages. The first, AreTomo3, streamlines tomographic alignment and reconstruction, with new features to fully account for sample geometry, locally correct the contrast transfer function, and curate data for downstream tasks. The second package, DenoisET, is a new implementation of the machine learning algorithm Noise2Noise and runs in parallel with AreTomo3 to perform contrast enhancement. To reduce barriers to routine use, AreTomoLive prioritizes automation: AreTomo3 autonomously pauses and reactivates processing depending on the status of data collection, while DenoisET algorithmically determines when to transition from training to inference. AreTomoLive endeavors to advance cryoET for in situ structural analysis with its comprehensive corrections and full automation.

02 /Specification & Classification

Click any tag to browse matching papers
Citations
23
FWCI
39.51
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (13 Mar 2025) · CC BY 4.0

AreTomo3 and DenoisET form an automated GPU pipeline that reconstructs cryo-electron tomograms with tile-based contrast transfer function correction, screens their quality, and denoises them during acquisition.

Parallel acquisition schemes such as PACE-Tomo yield more than 1000 tilt-series per week on one microscope, but end-to-end cryoET workflows like M, Scipion and TomoBEAR chain third-party tools with clashing coordinate conventions, file formats and repeated disk I/O. AreTomoLive replaces them with one integrated GPU pipeline adding geometry-aware alignment, local CTF correction, metric-based curation and automated denoising to support subtomogram averaging at scale.

Key results

  1. ShownAn integrated scout-worker design keeps data in memory and processes tilt-series concurrently with acquisition. It idles when it gets ahead of collection, resumes when new data arrive, and its throughput scales with the number of GPUs.

    Kept up with acquisition set to 12 targets per stage position using 8 NVIDIA RTX A6000 GPUs on one Linux server. Per-tilt-series timing over >100 tilt-series (1.54 Å data) scaled near-linearly to 4 GPUs, with a slight drop on 8. In a multi-day session, both programs finished within minutes of the session's end. · Introduction; Real-time and autonomous tomographic alignment…; Tables

  2. ShownDenoising after AreTomo3's tile-based hybrid Wiener/phase-flip CTF correction kept intermediate-resolution features better than denoising IsoNet (Warp-style) CTF-deconvolved tomograms, and it avoided high-resolution artifacts.

    Phantom dataset, region-of-interest to background power-spectrum ratio (224 virus-like particles in 45 tomograms vs 135 background regions, 5 Å pixels): on average a 5-fold contrast gain out to the first zero-crossing over weighted backprojection, and 1.5-fold over denoised IsoNet-corrected tomograms. Gains on a cilium/centrosome lamella were qualitative. · Contrast enhancement by 3D denoising; Evaluation of contrast enhancement; Figure legends

  3. ShownTwo-cluster k-means on AreTomo3's six quality metrics partly matched expert curation with minimal per-dataset tuning. It caught fewer bad tomograms than fixed cut-offs but wrongly rejected far fewer good ones.

    Minicells, 789 tomograms (245 kept by experts): k-means flagged 61% of user-rejected (334 of 544) and misclassified 8% of selected (20 of 245). Constant thresholds flagged 77% but wrongly rejected 26%. Phantom and lysosome sets: 45–53% of rejects flagged, 13-25% of selected misclassified. · Comprehensive quality assessment; Figure legends

  4. SuggestedPer-tilt CTF fitting done jointly with the α0/β0 tilt offsets and a tilt-axis ambiguity factor resolves 180° tilt-axis errors, fixes defocus handedness and scales the defocus search to magnification, giving robust tilt-series CTF estimation.

    Testing ta=±1 against the cross-correlation of defocus-rescaled tile spectra corrected the axis by subtracting 180° where needed (shown on example tilt-series). The search range and step are rescaled by the squared pixel size, and only defocus is refined beyond ±30°. The fit was accurate on a modified TYGRESS series with dose and defocus discontinuities. · Tilt-series CTF estimation; Determination of tilt axis directionality and…; Figure legends

  5. SuggestedDenoisET runs Noise2Noise without supervision. It picks training tomograms by AreTomo3 metric thresholds, starts once enough exist, stops when checkerboard artifacts appear, then denoises all tomograms, including new arrivals.

    20-40 high-quality tomograms generally sufficed across datasets. It trains on paired 96-pixel cubic even/odd-frame subvolumes with kernel size 3. The authors report the checkerboard metric at subvolume borders is more sensitive to overdenoising than training loss; if it never triggers, the epoch with the highest denoised standard deviation is used. · Contrast enhancement by 3D denoising; Automation of 3D denoising

Methods, data and reproducibility

AreTomo3 merges MotionCor2-style motion correction with AreTomo alignment. It adds joint tilt-series CTF and tilt-offset fitting, cross-correlation-based thickness estimation, tile-based hybrid CTF correction, algebraic or weighted-backprojection reconstruction and six quality metrics. DenoisET trains Noise2Noise on even/odd-frame tomograms in parallel.

  • Cryo-ET tomograms of Mycoplasma mycoides JCVI-Syn3A near-minimal cells (minicells), expert-curated for subtomogram averaging and used to test k-means screening · Mycoplasma mycoides JCVI-Syn3A (bacterium) · n = 789 tomograms (245 user-selected) · CryoET Data Portal CZCDP-10312
  • Purified synaptosomes from hippocampus of 10-week-old rats, used to show how misestimated thickness degrades alignment · rat · CryoET Data Portal CZCDP-10313
  • Phantom dataset of virus-like particles used for contrast quantification and k-means generalization · n = 224 virus-like particles in 45 tomograms; 135 background regions · CryoET Data Portal CZCDP-10310
  • Affinity-captured lysosomes from HEK293T cells with knock-in mEGFP-tagged TMEM192, some gold-labelled; used for k-means generalization · human (HEK293T cell line)
  • FIB-milled lamellae of cilia and centrosomes from serum-deprived hTERT RPE-1 cells expressing HTR6-HaloTag3; used for the denoising comparison · human (hTERT RPE-1, ATCC CRL-4000)
  • FIB-milled lamellae of S. cerevisiae expressing Myo5-GFP, Myo3-GFP and Ab1-RFP, imaged at 1.94 Å pixel size · yeast (Saccharomyces cerevisiae)
  • Tomograms of coronavirus-induced double-membrane vesicles illustrating thickness estimation with α0 correction and ice confounding
  • CryoET Data Portal tilt-series being reprocessed with AreTomoLive into standardized, well-aligned tomograms · n = ∼17,000 tilt-series
Methods & tools
AreTomo3 (motion correction, CTF estimation, alignment, reconstruction), DenoisET (Noise2Noise 3D denoising), MotionCor2, AreTomo, IsoNet CTF deconvolution (Warp-based comparison baseline), Warp, RELION-5 (downstream subtomogram averaging), Topaz-Denoise (architecture template), k-means clustering, Simultaneous algebraic reconstruction technique (SART) and weighted backprojection, Tomography 5 (TFS) acquisition software, Krios G4 with Falcon 4i detector and SelectrisX energy filter
Code
http://github.com/czimaginginstitute/AreTomo3https://github.com/apeck12/denoiset
Assay conditions & compute
A Linux server with 8 NVIDIA RTX A6000 GPUs keeps up with collection at 12 targets per stage position. Timings over >100 tilt-series show near-linear scaling to 4 GPUs and a slight drop on 8. DenoisET is not parallelized.
Reproducibility gaps
Only minicell, synaptosome and phantom tomograms are deposited (CZCDP-10312, CZCDP-10313, CZCDP-10310). Availability of lysosome, lamella and raw tilt-movie data is not stated, and no environment or version details are given.
Limitations (6)
  • RELION-4/5 pseudo-subtomogram workflows currently use only AreTomo3's global beam-induced motion correction. They ignore its local alignments and α0 offset correction, so those gains do not yet reach subtomogram averaging or polishing. · stated by the authors
  • Thickness detection assumes the sample is level in the tomogram, so tilted lamellae need α0 correction. Ice contamination can also inflate the measured thickness, and a wrong thickness degrades the local-alignment reference. · stated by the authors
  • Agreement between k-means and experts dropped on phantom and lysosome data (45–53% of rejects caught), possibly because of biology-driven manual criteria or different reconstruction software. The metrics also miss content such as ribosome abundance. · stated by the authors
  • Pretrained denoising models generalize poorly, so retraining per dataset is often preferred. DenoisET does not yet support parallelization, and AreTomo3 scaling dips slightly on 8 GPUs because of overhead. · stated by the authors
  • The IsoNet baseline's filter parameters were chosen by visually inspecting denoised output, and contrast was quantified on a single phantom dataset. No gain in subtomogram-averaging resolution and no alignment-accuracy comparison with other pipelines is reported. · noted in this summary
  • Handedness resolution, CTF robustness and the effect of thickness errors are shown with example tilt-series rather than error rates across many datasets, so how often each fails is unknown. · noted in this summary

Who should read:Read this if you run high-throughput cryoET acquisition and want one GPU pipeline for real-time alignment, local CTF correction, automated tomogram screening and Noise2Noise denoising.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISZheng '22Yee '23Zhang '26Kishore '25Wang '24Costa-Filho '25
82% semantic matchCell Biology
S. Zheng, G. Wolff, G. Greenan+6 more·bioRxiv·Feb 15, 2022