• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Thursday, November 19, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

CryoBlob Detects Nanoparticles in Low-Dose Cryogenic Electron Microscopy Without Prior Training

by
November 19, 2026
in Technology
Reading Time: 6 mins read
0
CryoBlob Detects Nanoparticles in Low-Dose Cryogenic Electron Microscopy Without Prior Training

CryoBlob Detects Nanoparticles in Low-Dose Cryogenic Electron Microscopy Without Prior Training

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Cryogenic electron microscopy has emerged as a cornerstone technique for resolving the three-dimensional structures of vitrified biological macromolecules, but researchers are increasingly turning to the platform to investigate beam-sensitive synthetic nanomaterials, polymers, and soft-matter formulations. Characterizing these systems, which include polymeric nanocarriers and lipid-based assemblies used in targeted therapeutic delivery, requires preserving hydrated particle morphologies and internal architectures under tight operational constraints. Because beam-induced damage can rapidly degrade delicate synthetic matrices, data collection must be conducted under strictly limited electron exposures, typically well below one hundred electrons per square angstrom. The resulting micrographs inevitably exhibit extremely low signal-to-noise ratios, rendering small or weakly scattering nanoparticles exceedingly difficult to distinguish from background vitreous ice and carbon support structures.

To overcome the operational bottlenecks associated with identifying non-biological nanoparticles in noisy imaging fields, researchers have introduced CryoBlob, an open-source, reference-free computational tool built specifically to analyze low-dose cryogenic micrographs without requiring pre-trained neural network weights or hand-crafted geometric templates. In traditional structural biology workflows, automated particle detection predominantly follows one of two paradigms: reference-free blob detectors configured around fixed globular protein dimensions, or supervised convolutional neural networks that demand extensive labeled training sets. Outside of macromolecular reconstruction projects, however, target specimens frequently display broad, uncharacterized size distributions and variable visual contrast across relatively small collections of images. CryoBlob directly addresses these constraints by leveraging gradient-based scale tuning and an analytic scale-space framework that extracts compact objects without user-supplied coordinates or prior annotations.

Beyond eliminating the reliance on pre-existing labels, the software resolves a longstanding architectural barrier in computational microscopy by establishing direct interoperability with modern differentiable programming environments. Established computer vision and image processing libraries, including implementations found in NumPy, C, and scikit-image, are structurally decoupled from modern automatic differentiation engines, interrupting gradient propagation whenever integrated with specialized deep learning models. CryoBlob circumvents this limitation by formulating its continuous computational stages, including image preprocessing, adaptive Wiener filtering, and Laplacian-of-Gaussian scale-space filtering, as pure, host-independent functions within the JAX numerical framework. This architecture allows scale-space responses to undergo just-in-time compilation via jax.jit, execute natively across graphics processing units and central processing units, vectorize over image batches via vmap, and preserve end-to-end differentiability alongside downstream optimization frameworks such as cryoJAX without transferring intermediate data off accelerated hardware.

The underlying software package, engineered for Python environments running version 3.12 and above, organizes its image analysis routines into modular components that address each phase of the detection pipeline. Micrograph onboarding supports standard scientific formats, including MRC, TIFF, and DM3 files, which are validated against user-specified parameters using beartype and jaxtyping type-checking tools to prevent silent computational errors. Within the adapt module, raw image data undergo adaptive Wiener deconvolution wherein local noise-power estimates are dynamically updated through adaptive thresholding. These continuous preprocessing stages are defined as differentiable functions of multidimensional arrays, permitting iterative parameter optimization through gradient descent rather than exhaustive, computationally prohibitive grid searches across arbitrary parameter spaces.

Candidate particles are subsequently localized across continuous spatial scales using an analytic Fast Fourier Transform transfer function implemented within the Laplacian-of-Gaussian detection module. By convolving preprocessed micrographs with normalized kernels across a user-defined scale interval, the software identifies local extrema within the synthesized scale space, outputting candidate particle centers and corresponding radial dimensions. Conducting scale-space evaluations directly within the frequency domain enables the rapid processing of full-resolution 4096 by 4096 pixel micrographs without requiring spatial binning or compression, thereby avoiding the memory constraints that typically bottleneck spatial-domain convolution kernels. While scale-space continuous response fields remain entirely differentiable, discrete detection operations, including intensity thresholding, host-side connected-component labeling, and centroid extraction, are handled through discrete SciPy algorithms.

Because low-dose formulations and drug-delivery preparations frequently contain crowded aggregations where neighboring particles touch or partially overlap, standard blob-finding algorithms frequently merge adjacent entities into spurious, oversized detections. To address this failure mode, CryoBlob incorporates distance-transform watershed segmentation to separate adjacent boundaries that simple intensity filtering conflates. Furthermore, the multi module provides an experimental ridge-detection routine that evaluates the local eigenvalues of the smoothed image Hessian matrix, seeking signatures where one eigenvalue exhibits a large absolute magnitude alongside a near-zero counterpart to isolate elongated, filamentous features. Although Hessian boundary refinements and ridge functions are bundled within a unified multi-method dispatch function, evaluations indicate that the ridge pipeline does not yet achieve reliable recovery for elongated structures, marking robust anisotropic particle picking as a target for future development.

The detection performance and localization precision of the software were initially benchmarked against synthetic micrographs generated via physics-based simulation tools in the ACORN software suite, establishing direct comparisons against exactly defined ground truth. Synthetic benchmarks modeled spherical poly(lactic-co-glycolic acid)-like particles spanning diameters from 50 to 250 nanometers on 4096 by 4096 pixel grids sampled at 5 angstroms per pixel, simulated at 300 kilovolts with an accumulated dose of 40 electrons per square angstrom. Two distinct simulation regimes were evaluated: a dispersed set containing 20 separated particles per image and a crowded field containing 30 overlapping particles per image. True-positive identifications were rigorously defined as detections whose measured centers fell within 0.5 times the true particle radius relative to an unmatched ground-truth target, directly establishing comparative precision, recall, and harmonic F1 measures.

On the non-overlapping simulated dataset, CryoBlob attained an overall precision of 0.77, a recall of 0.98, and an F1 score of 0.86, demonstrating an average spatial localization error of 3.3 nanometers across the full 4096 by 4096 field. Under these dispersed conditions, its accuracy proved competitive with established reference-free pickers, including DoGpicker and the RELION Laplacian-of-Gaussian autopicker, which achieved F1 scores ranging from 0.81 to 0.86, while substantially outperforming naive scikit-image detectors that accumulated hundreds of false positives. In the crowded, overlapping simulated condition, standard Laplacian-of-Gaussian detection scored an F1 of 0.68, featuring a precision of 0.66 and a recall of 0.70. Activating the optional distance-transform watershed pass elevated recall from 0.70 to 0.80 while preserving baseline precision, raising the overall F1 score to 0.72. By comparison, RELION attained an F1 score of 0.77 in the overlapping condition, outperforming CryoBlob slightly on crowded fields but trailing it in non-overlapping configurations, while requiring substantially heavier computational infrastructure compared to CryoBlob’s pip-installable execution profile.

To evaluate whether synthetic performance gains translate effectively to authentic laboratory imaging, the algorithm was tested on experimental micrographs of PLGA-LA nanocarriers acquired at 300 kilovolts with calibrated spatial samplings of 0.2229 nanometers per pixel. Two full-scale 4096 by 4096 micrographs containing 112 independently expert-annotated nanoparticles, contoured and measured using equivalent circular diameters without prior reference to algorithmic predictions, served as the primary experimental standard. When evaluated alongside DoGpicker and a scikit-image Laplacian-of-Gaussian baseline using a matched 30 to 200 nanometer search range, all three tools captured nearly the entire target population, with CryoBlob recovering 109 true particles, DoGpicker recovering 108, and the baseline recovering 102. The critical operational difference resided in the non-target detections generated by each method.

Across the experimental evaluation, the scikit-image baseline returned 404 total detections to yield an operational precision of 0.25 and an F1 score of 0.40, driven by extensive false positives across featureless ice and along the carbon support grid, alongside repeated detections of single particles. DoGpicker yielded 238 total detections, corresponding to a precision of 0.45 and an F1 score of 0.62. In contrast, CryoBlob produced 171 total outputs, comprising 109 accurate particles, 13 duplicate identifications, and 49 false-positive background coordinates, resulting in an experimental precision of 0.64, a recall of 0.97, and an F1 score of 0.77. Average localization errors measured 5.5 nanometers for CryoBlob, compared to 7.4 nanometers for DoGpicker and 7.8 nanometers for the Laplacian baseline. By drastically suppressing redundant detections and spurious background signals, CryoBlob extracted apparent particle size distributions that aligned closely with ground-truth curves, preventing sizing distortion from propagating into downstream formulation statistics.

Parametric robustness experiments conducted across 2048 by 2048 pixel grids demonstrated that CryoBlob preserves high quantitative accuracy across substantial variations in image degradation and target morphology. The detection pipeline remained resilient across an approximately fourfold reduction in signal-to-noise ratio, preserving F1 scores between 0.97 and 0.98 as per-pixel signal-to-noise diminished from 1.7 to 0.40, before dropping to 0.88 under severe noise conditions of 0.23. The tool proved similarly robust across varied physical scales, recording F1 scores from 0.86 to 0.98 across particle diameters ranging from 50 to 130 nanometers. However, structural overlap remained the dominant driver of algorithmic failure, with F1 measures dropping from 0.98 in uncrowded conditions to 0.54 under high-density regimes containing 80 overlapping targets per field. Hollow vesicles were successfully identified with a high recall of 0.96 but exhibited an attenuated precision of 0.59 due to multiple centroid responses generated by ring-like cross sections, while elongated objects recorded poor detection performance with an F1 score near 0.10.

Hardware benchmarking on an NVIDIA V100 graphics processing unit and central processing unit architectures demonstrated practical computational efficiency for high-throughput laboratory settings. Warm execution runtimes for single micrographs on the V100 scaled from 0.25 seconds for 1024 by 1024 pixel inputs to 6.62 seconds for full 4096 by 4096 arrays evaluated across 25 distinct scales, with peak device memory consumption expanding from 419 megabytes to 4.3 gigabytes. Initial cold invocations required between 2.6 and 7.7 seconds due to one-time just-in-time compilation overheads. Because the computational workload is dominated by Fast Fourier Transform operations, GPU acceleration yielded a modest 1.2-fold speedup over CPU processing at full 4096 resolution. Nevertheless, native JAX distribution primitives such as pmap allow multi-image batches to be sharded effortlessly across multi-GPU or high-performance computing clusters, providing materials scientists with an accessible, mathematically rigorous framework for quantifying complex nanocarrier formulations directly from low-dose micrographs.

Subject of Research: Automated reference-free nanoparticle detection in low-dose cryogenic electron microscopy using differentiable scale-space algorithms

Article Title: CryoBlob – a JAX-based, reference-free nanoparticle detector for low-dose Cryo-EM

Article References: Mukherjee, D., Brown, S. R., Jones, A. L., Brantley, C. L., Gonzalez, M. T., Plechaty, A., McDonnell, M. T., Borisevich, A. Y., Davern, S. M., Sanders, B. C., & Williams, A. N. (2026). CryoBlob – a JAX-based, reference-free nanoparticle detector for low-dose Cryo-EM. SoftwareX, 36, Article 103242. https://doi.org/10.1016/j.softx.2026.103242

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103242

Keywords: cryogenic electron microscopy, nanoparticle detection, CryoBlob, JAX, low-dose imaging, Laplacian of Gaussian, differentiable programming, nanomedicine, drug delivery, PLGA nanocarriers, watershed transform, image processing, scale-space analysis, automated particle picking, computer vision, materials science, Wiener deconvolution

News Source: Denise Maddox. (November 19, 2026). CryoBlob Detects Nanoparticles in Low-Dose Cryogenic Electron Microscopy Without Prior Training. Scienmag.

Tags: automated particle pickingComputer VisionCryoBlobcryogenic electron microscopydifferentiable programmingDrug deliveryimage processingJAXLaplacian of Gaussianlow-dose imagingmaterials scienceNanomedicinenanoparticle detectionPLGA nanocarriersscale-space analysiswatershed transformWiener deconvolution
Share12Tweet7Share2ShareShareShare1

Related Posts

Digital Transformation Bolsters Enterprise Resilience Among Listed Chinese Companies

Digital Transformation Bolsters Enterprise Resilience Among Listed Chinese Companies

November 19, 2026
Fly Ash Filler Strengthens Solid Masonry Blocks Made With Recycled Concrete Aggregates

Fly Ash Filler Strengthens Solid Masonry Blocks Made With Recycled Concrete Aggregates

November 19, 2026

Analytical Model Reveals Fractional Thermal Memory During Geothermal Shut-In Periods

November 19, 2026

Digital Maturity Enables IT Investment Gains in Chinese Manufacturing Firms

November 19, 2026

POPULAR NEWS

  • Anlotinib Combined with Cranial Radiotherapy Shows Promise for Lung Cancer Brain Metastases

    Anlotinib Combined with Cranial Radiotherapy Shows Promise for Lung Cancer Brain Metastases

    29 shares
    Share 12 Tweet 7
  • Curcumin Reduces Liver Steatosis in Rats by Altering Gut Microbiota and Epigenetics

    29 shares
    Share 12 Tweet 7
  • Air Sampling in Bat Caves Captures Mammalian Viruses Across Optimized Volumes

    29 shares
    Share 12 Tweet 7
  • Rimegepant Provides Meaningful Relief as Second- or Third-Line Acute Migraine Therapy

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Jiangsu Sports Facility Study Identifies Economic Spillovers and Regional Siphon Effects

Adjunctive Breathing Exercises Modestly Improve Certain Childhood Asthma Lung Measures, Review Finds

Quran Recitation Linked to Reduced Labor Pain and Blood Pressure in Study

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.