NanTex enables computational multiplexing and phenotyping of organelles across super-resolution modalities

NanTex enables computational multiplexing and phenotyping of organelles across super-resolution modalities | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results NanTex enables computational multiplexing and phenotyping of organelles across super-resolution modalities View ORCID Profile Bela T.L. Vogler , View ORCID Profile Gregor J. Gentsch , View ORCID Profile Pablo Carravilla , View ORCID Profile Dominic A. Helmerich , View ORCID Profile Julia C. Heiby , Katharina Reglinski , View ORCID Profile William Durso , View ORCID Profile Teresa Klein , View ORCID Profile Felix Hildebrandt , Rohama Zahid , View ORCID Profile Katrin Spengler , View ORCID Profile Regine Heller , View ORCID Profile Christoph Kaether , View ORCID Profile Markus Sauer , View ORCID Profile Christian Eggeling , View ORCID Profile Christian Franke doi: https://doi.org/10.1101/2025.09.16.676574 Bela T.L. Vogler 1 Faculty of Physics and Astronomy, Institute of Applied Optics and Biophysics, Friedrich Schiller University Jena , Jena, Germany 2 Leibniz Institute of Photonic Technology e.V. , Jena, Germany , member of the Leibniz Centre for Photonics in Infection Research (LPI) , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Bela T.L. Vogler Gregor J. Gentsch 1 Faculty of Physics and Astronomy, Institute of Applied Optics and Biophysics, Friedrich Schiller University Jena , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Gregor J. Gentsch Pablo Carravilla 1 Faculty of Physics and Astronomy, Institute of Applied Optics and Biophysics, Friedrich Schiller University Jena , Jena, Germany 2 Leibniz Institute of Photonic Technology e.V. , Jena, Germany , member of the Leibniz Centre for Photonics in Infection Research (LPI) , Jena, Germany 3 Present address: Instituto Biofisika (UPV/EHU , CSIC), University of the Basque Country , Leioa, Spain and Ikerbasque Foundation for Science , Bilbao, Spain Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Pablo Carravilla Dominic A. Helmerich 4 Department of Biotechnology and Biophysics, Biocenter, University of Würzburg , Am Hubland, Würzburg, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Dominic A. Helmerich Julia C. Heiby 5 Leibniz Institute on Aging - Fritz Lipmann Institute (FLI) e.V. , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Julia C. Heiby Katharina Reglinski 1 Faculty of Physics and Astronomy, Institute of Applied Optics and Biophysics, Friedrich Schiller University Jena , Jena, Germany 2 Leibniz Institute of Photonic Technology e.V. , Jena, Germany , member of the Leibniz Centre for Photonics in Infection Research (LPI) , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site William Durso 5 Leibniz Institute on Aging - Fritz Lipmann Institute (FLI) e.V. , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for William Durso Teresa Klein 4 Department of Biotechnology and Biophysics, Biocenter, University of Würzburg , Am Hubland, Würzburg, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Teresa Klein Felix Hildebrandt 1 Faculty of Physics and Astronomy, Institute of Applied Optics and Biophysics, Friedrich Schiller University Jena , Jena, Germany 2 Leibniz Institute of Photonic Technology e.V. , Jena, Germany , member of the Leibniz Centre for Photonics in Infection Research (LPI) , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Felix Hildebrandt Rohama Zahid 6 Center for Molecular Biomedicine, Institute of Molecular Cell Biology, Jena University Hospital , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Katrin Spengler 6 Center for Molecular Biomedicine, Institute of Molecular Cell Biology, Jena University Hospital , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Katrin Spengler Regine Heller 6 Center for Molecular Biomedicine, Institute of Molecular Cell Biology, Jena University Hospital , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Regine Heller Christoph Kaether 5 Leibniz Institute on Aging - Fritz Lipmann Institute (FLI) e.V. , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Christoph Kaether Markus Sauer 4 Department of Biotechnology and Biophysics, Biocenter, University of Würzburg , Am Hubland, Würzburg, Germany 7 Rudolf Virchow Center, Research Center for Integrative and Translational Bioimaging, University of Würzburg , Würzburg, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Markus Sauer Christian Eggeling 1 Faculty of Physics and Astronomy, Institute of Applied Optics and Biophysics, Friedrich Schiller University Jena , Jena, Germany 2 Leibniz Institute of Photonic Technology e.V. , Jena, Germany , member of the Leibniz Centre for Photonics in Infection Research (LPI) , Jena, Germany 8 Jena Center for Soft Matter, Friedrich Schiller University Jena , Jena, Germany 9 Abbe Center of Photonics, Friedrich Schiller University Jena , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Christian Eggeling Christian Franke 1 Faculty of Physics and Astronomy, Institute of Applied Optics and Biophysics, Friedrich Schiller University Jena , Jena, Germany 8 Jena Center for Soft Matter, Friedrich Schiller University Jena , Jena, Germany 9 Abbe Center of Photonics, Friedrich Schiller University Jena , Jena, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Christian Franke For correspondence: christian.franke{at}uni-jena.de Abstract Full Text Info/History Metrics Supplementary material Data/Code Preview PDF Abstract Super-resolution microscopy (SRM) enables nanoscale visualization of cellular organelles but remains constrained by spectral overlap, labeling requirements, and temporal offsets in multicolor imaging. We present NanTex , a deep learning framework that introduces nanotexture as a universal descriptor of subcellular organization and achieves probabilistic demixing of multiple organelles from single-channel SRM images. Unlike segmentation methods that enforce exclusivity, NanTex preserves overlapping morphologies, enabling faithful reconstruction even in crowded regions. Trained on small curated datasets with augmentation, NanTex generalized across SMLM, MINFLUX, STED, SIM, and live-cell Airyscan, with modality-specific retraining where necessary. Demonstrated on cytoskeletal, endomembrane, and metabolic organelles, NanTex achieved high-fidelity reconstructions and extended uniquely to live-cell imaging, where it eliminated temporal misalignment and enabled quantitative tracking of vesicle-like ER subdomains. Notably, NanTex enables quantitative dynamic readouts directly from single-channel live-cell data, revealing and tracking hidden subdomains in real time without the need for additional labels. Beyond multiplexing, NanTex supported computational phenotyping by distinguishing structurally distinct yet molecularly identical populations, exemplified by nocodazole-induced microtubule depolymerization and its glyoxal-mediated modulation. Together, these results establish nanotexture as a paradigm-shifting descriptor for organelle identity and position NanTex as a modality-agnostic, label-efficient, and live-cell-ready strategy for quantitative nanoscale biology. Significance Statement NanTex reframes multiplexing as texture-based demixing, introducing nanotexture as a universal fingerprint of organelle identity. By enabling quantitative, label-efficient reconstructions across SRM modalities and live-cell phenotyping, including vesicle tracking from single-channel data, NanTex opens a path to studying organelle remodeling and disease progression with unprecedented fidelity across SRM modalities. Introduction Understanding how organelles organize themselves and interact at the nanoscale is central to cell biology. However, the diffraction limit long obscured these structures. Fluorescence-based super resolution microscopy (SRM) has transformed biological imaging by enabling the visualization and quantitative analysis of subcellular structures at nanometer resolution 1 , 2 . Techniques such as single-molecule localization microscopy (SMLM) 3 – 6 , 15 , stimulated emission depletion (STED) microscopy 7 , MINFLUX microscopy 8 , structured illumination microscopy (SIM) 9 , 10 , and Airyscan microscopy 11 have become indispensable tools for interrogating organelle architecture, dynamics, and interactions. Despite these advances, multiorganelle imaging remains fundamentally limited by the physics of fluorophores and acquisition. Spectral overlap, finite dye palettes, and the photophysical requirements of SRM 12 , 44 restrict simultaneous labeling, while live-cell imaging faces additional constraints from phototoxicity and temporal misalignment 2 . To address these bottlenecks, several multiplexing strategies have been proposed. Spectroscopic approaches such as fluorescence lifetime imaging microscopy (FLIM) 13 and polarisation-resolved imaging 14 extend color channels but remain restricted to a handful of species and are challenging to integrate across modalities. DNA-PAINT and Exchange-PAINT leverage iterative binding of labels to achieve large multiplexing capacity 15 , 16 , yet are inherently sequential and time-consuming, limiting their applicability to both dynamic and high-throughput imaging. Complementary computational segmentation methods including Cellpose 17 , 18 , Segment Anything 19 , and large-scale learning frameworks 20 all assume that structures can be discretely separated. In crowded regions where organelles overlap extensively these methods fail to preserve the overall nanoscale composition critical for biological interpretation. Here, we introduce NanTex (from “ nanotexture ”), a content-agnostic computational multiplexing framework that leverages nanoscale textural features as a generalizable structural fingerprint for organelles. Rather than segmenting discrete objects, NanTex performs probabilistic demixing through deep learning and thereby enabling the reconstruction of multiple identically labeled organelles from single-channel, monochromatic SRM images. This paradigm shift reframes multiplexing as a problem of organelle-specific demixing rather than spectral separation or structural delineation. Importantly, NanTex is data-efficient, requiring ∼10 images per organelle for training when combined with augmentation 26 , 37 , and broadly generalizes across SRM modalities, including SMLM, MINFLUX, STED, SIM, and live-cell Airyscan. We validate NanTex across fixed-cell and live-cell contexts, demonstrating accurate demixing of combinations of identically labeled microtubules, actin, clathrin, ER, mitochondria, peroxisomes, endosomes, lysosomes and vesicle-like ER-substructures even in crowded environments and challenging labeling conditions. Importantly, NanTex achieves live-cell monochromatic multiplexing, eliminating spectral cross-talk and temporal offsets that limit conventional multicolor imaging. For example, we applied NanTex to live-cell Airyscan data of ER–lysosome dynamics, building on our recent findings that implicate lysosome–ER contact sites of novel identified lysosomal proteins 21 . Beyond multiplexing, NanTex uniquely supports computational phenotyping by disentangling structurally distinct but molecularly identical populations, exemplified by microtubule depolymerization and its glyoxal-mediated delay 33 , 34 . Together, our results establish nanotexture as a universal descriptor of organelle identity and NanTex as a paradigm-shifting tool for label-efficient and live-compatible structural biology that could enable quantitative nanomorphological analysis of cellular states, treatments, or disease perturbations. Results Haralick feature extraction identifies nanoscale organelle-specific textures Cellular organelles exhibit various morphologies across the micrometer to molecular scales, accessible by SRM 23 . We reasoned that sufficiently labelled and resolved organelles will possess an intrinsic, characteristic and self-similar nanometric fingerprint, i.e. a nanotexture . To demonstrate the extraction of quantifiable nanoscale textures from SRM images that could enable the differentiation of multiple organelles, we extracted classical Haralick textural descriptors 24 from single-molecule localization microscopy (SMLM) images of microtubules and endosomes, labelled by nanoparticles 47 , 48 ( see methods, Fig. 1 ). Using 26 features (mean and range values of 13 Haralick descriptors) computed in a sliding 7×7 pixel window, we observed partially separable distributions, with contrast-dependent features providing the clearest distinction ( Fig. 1b–c ). However, single-feature thresholding failed in overlapping regions, where textures intermix at the pixel scale ( Fig. 1d ). While this generally demonstrates the validity of an organelle-specific nanotexture, this classical segmentation approach failed to account for overlapping structures with different identities, inherent to most fluorescence imaging methods. Download figure Open in new tab Figure 1 | Haralick feature extraction identifies nanoscale organelle-specific textures. ( a ) Workflow for pixel-precise Haralick feature extraction. Single-molecule localization microscopy (SMLM) images (here, microtubules and endosomes) were subdivided into sliding-window regions of interest (ROIs), and 26 Haralick texture descriptors were computed for each ROI, yielding feature images encoding nanoscale structural information. ( b ) Example distribution of single Haralick descriptors, showing partial separation in mean contrast (yellow) and continuous distribution in mean correlation (blue). A simple threshold (red line) allows only limited discrimination. ( c ) Representative ground truth SMLM image (left) and the corresponding thresholded Haralick feature image (right), where endosomes (magenta) and microtubules (yellow) are partially separated but still exhibit misclassification in overlapping regions. ( d ) Quantitative performance of all 26 Haralick features expressed as F1-scores for organelle classification. Only a subset of descriptors exceeded an F1-score of 0.4, namely Angular Second Moment (F1 ≈ 0.50, marked 1 ), Contrast (F1 ≈ 0.48, marked 2 ), Correlation (F1 ≈ 0.41, marked 3 ), Sum of Squares Variance (F1 ≈ 0.45, marked 4 ), and Homogeneity (F1 ≈ 0.52, marked 5 ). These features showed moderate discriminative power, while the majority of descriptors performed substantially lower, highlighting that threshold-based approaches are insufficient for robust multiplexing. Scale bars: 5 µm. NanTex enables robust probabilistic demixing of monochromatic multi-organelle in silico SMLM data Addressing the limitations of classical segmentation approaches, we developed NanTex , a machine-learning-enabled computational multiplexing framework leveraging the previously described organelle-specific nanoscale textures. Unlike segmentation methods, NanTex utilizes probabilistic demixing enabled by a U-Net 25 trained to reconstruct multiple organelle output channels from monochromatic SRM input images ( see methods ). In short, the model employs a standard encoder-decoder architecture, with each convolutional block followed by batch normalization and ReLU activation. A 1×1 convolutional layer using linear regression outputs separate probabilistic predictions for each organelle. It is trained to produce the most probable component images contained within a superimposed multi-organelle grayscale input. Importantly, this generative approach preserves intensity and nanoscale nuances of overlapping organelles, fundamentally distinguishing NanTex from segmentation-based multiplexing methods. Being able to leverage heavy data augmentation ( see methods ) 26 , the network can be efficiently trained with ∼10 images per organelle and SRM modality, a requirement easily fulfilled by most SRM modalities and users. We first demonstrated NanTex ’s multiplexing capability using SMLM datasets with highest structural resolution, comprising three organelles often used in SRM benchmarking with distinct but overlapping structural profiles: microtubules, actin, and endosomes (i.e. vesicles) 27 , 28 . NanTex was trained on a stack of computationally superimposed and augmented sets of organelles (sourced partially from shareloc 42 , see methods ) with the task of re-generating the individual underlying images. This allowed for the successful multiplexing of different, in silico combined, multi-structural datasets and benchmarking of the network. Globally, NanTex reconstructed all three overlapping individual structures of the test set, i.e. microtubules, endosomes, and actin, with high fidelity, quantified by structural (multi-scale) self-similarity index measure 29 – 31 (SSIM = 0.80 - 0.87; MS-SSIM = 0.90 - 0.92, Fig. 3 , see Table 1 ). Importantly, multiplexing performance even rose within densely overlapping ROIs (SSIM = 0.84 - 0.90; MS-SSIM = 0.93 - 0.94, Fig. 3 ), where conventional segmentation completely failed to separate signals. While SSIM and MS-SSIM are full-reference, perception-informed metrics that assess similarity in luminance, contrast, and structure, and are therefore designed to correlate with human judgments of image quality (SSIM ranges from -1 to 1, with 1 indicating identical images), MS-SSIM in particular captures image fidelity across multiple spatial scales, providing greater alignment with perceptual quality than single-scale SSIM ( see Table 1 , Supplementary Discussion, Supplementary Fig. 1). Accordingly, the consistently high MS-SSIM values observed here (0.90–0.94) indicate near-perfect, perceptually indistinguishable reconstruction of the ground truth. These results highlight NanTex ’s unique ability to demix at least three spectrally identical and structurally overlapping organelles from a single monochromatic image, fundamentally surpassing segmentation-based approaches. Download figure Open in new tab Figure 2 | NanTex architecture and probabilistic demixing workflow. Schematic overview of the NanTex pipeline. Monochromatic SRM acquisitions (left) containing multiple organelles are subdivided into training patches and passed through a U-Net–style encoder–decoder network (middle). The network employs probabilistic demixing, directly reconstructing independent organelle channels rather than performing classical segmentation. Color-coded icons illustrate augmentation and network operations applied during training (rotation, flips, Gaussian noise/blur, gamma correction), which ensure robustness to variability in labeling density, photophysics, and imaging artifacts. The output (right) is a set of multiplexed organelle-specific channels that preserve both intensity and fine structural detail, here shown as ER (cyan), microtubules (yellow), and clathrin (magenta). This architecture enables accurate demixing even under high structural overlap, with training requirements of ∼10 input images per organelle and modality, augmented into several thousand training patches. Download figure Open in new tab Figure 3 | NanTex enables high-fidelity organelle specific multiplexing of triple-identity monochromatic SMLM data in silico. ( a ) Exemplary SRM image rendered from unrelated SMLM data of microtubules, actin and endosomes, combined in-silico in native greyscale ( left ) and its demixed multi-colour counterpart after NanTex analysis in false colours ( right , yellow: microtubules, cyan: actin, magenta: endosomes). ( b ) Magnified views of the region indicated in ( a ), exhibiting significant content in all three ground-truth channels ( from left to right: microtubules, endosomes, actin, sum), in comparison with ( c ) their respective counterparts after NanTex analysis. ( d ) Quantitative evaluation of NanTex reconstructions. Structural similarity (SSIM) and multi-scale SSIM (MS-SSIM) were computed globally across full fields of view ( a ) and locally in regions of dense overlap ( b-c ). NanTex achieved high global fidelity for all three organelles (microtubules: SSIM = 0.80, MS-SSIM = 0.91; endosomes: SSIM = 0.84, MS-SSIM = 0.90; actin: SSIM = 0.87, MS-SSIM = 0.92) and further improved similarity within overlapping ROIs (SSIM = 0.84–0.90, MS-SSIM = 0.93–0.94). These consistently high MS-SSIM values indicate near-perceptual indistinguishability from the ground truth, in stark contrast to the failure of intensity thresholding to recover overlapping structures. Scale bars: 5 µm (a), 1 µm (b-c). View this table: View inline View popup Download powerpoint Table 1 | Guidelines for interpretation of SSIM and MS-SSIM scores in NanTex benchmarking Encouraged by these in silico results, we next asked whether NanTex could achieve comparable multiplexing fidelity on experimental d STORM acquisitions of identically labeled organelles. Experimental validation of monochromatic SMLM multiplexing with NanTex To validate NanTex in realistic contexts, we acquired monochromatic d STORM datasets in which distinct organelles were deliberately labeled with the same fluorophore (AF647), resulting in composite images that appear structurally entangled. NanTex was then applied to demix these single-color overlays into separate channels corresponding to the original structures. In the first experiment, we multiplexed three organelles ( Fig. 4 , yellow: microtubules, magenta: clathrin, cyan: ER) recorded as independent AF647 d STORM acquisitions to ensure ground truth quantifyability. NanTex faithfully separated the three structures, maintaining high fidelity over 500 test images ( see methods ), with mean SSIM/MS-SSIM values of 0.86/0.94 for microtubules, 0.83/0.86 for clathrin, and 0.83/0.89 for ER. A representative ROI ( Fig. 4 ) shows slightly lower but still robust scores (microtubules: 0.82/0.94; clathrin: 0.67/0.77; ER: 0.75/0.85), consistent with the partially overlapping diffuse monomeric fractions of clathrin and ER, whereas the filamentous microtubules exhibited the highest fidelity. Together, these results confirm that NanTex preserves fine structural detail while tolerating biological complexity arising from shared molecular pools. Download figure Open in new tab Figure 4 | NanTex demixes experimental SMLM overlays of identically labeled organelles. ( a ) Composite monochromatic d STORM image generated by combining individual AF647 acquisitions of microtubules, clathrin, and ER. ( b ) NanTex multiplexing results in false colours (yellow: microtubules, magenta: clathrin, cyan: ER). ( c ) Ground-truth channels of a randomly selected ROI. ( d ) Corresponding NanTex demixed channels and false-colour image of the same ROI. ( e ) Quantitative evaluation of the same ROI using SSIM and MS-SSIM, displayed in false colours. Global mean SSIM/MS-SSIM values were 0.86/0.94 for microtubules, 0.83/0.86 for clathrin, and 0.83/0.89 for ER; corresponding ROI values were 0.82/0.94, 0.67/0.77, and 0.75/0.85, respectively. The reduced scores for clathrin and ER reflect their partially overlapping diffuse monomeric fractions, while filamentous microtubules reached the highest fidelity. The representative example in ( a–e ), which was not cherry-picked, illustrates typical performance but yields slightly lower similarity scores than the overall dataset averages. Scale bars: 5 µm (a–b), 1 µm (c–d). We next validated NanTex on pairwise monochromatic d STORM data ( Fig. 5 ). For single-label microtubules and clathrin (AF647) that were imaged by d STORM simultaneously. NanTex successfully disentangled the cytoskeletal network from vesicular structures. Of note: clathrin-coated vesicles are consistently resolved despite spatial overlap with microtubules, without destroying their respective integrity. Importantly, the demixed microtubule channel shows virtually no unspecific background signal, consistent with the expected filamentous localization, while clathrin is robustly detected in both its vesicular and diffuse monomeric membrane fractions. For identically labeled microtubules and ER, NanTex also achieved a reliable separation, although the overall labeling efficiency for the combined structures was lower due to protocol mismatches. Despite this suboptimal staining, the remaining ER and microtubule network were accurately demixed, underlining the robustness of the model even in imperfect experimental conditions. Download figure Open in new tab Figure 5 | Experimental validation of monochromatic SMLM multiplexing with NanTex . ( a ) Representative d STORM dataset of identically labeled microtubules and clathrin (AF647). Top row: experimentally acquired monochromatic input ( left ) and NanTex multiplexing result ( right ; yellow: microtubules, magenta: clathrin). Bottom row: zoom-in of the indicated ROI showing ground truth (left), NanTex false-color overlay, and the corresponding demixed channels for microtubules and clathrin. An orange hexagon highlights a region where a clathrin-coated vesicle is cleanly separated from an overlapping microtubule. ( b ) Equivalent NanTex analysis of identically labeled microtubules and ER (cyan). Top row: experimentally acquired monochromatic input and NanTe x multiplexing result. Bottom row: zoom-in ROI with ground truth, false-color overlay, and individual NanTex channels. While labeling efficiency for ER was suboptimal due to protocol mismatches, the remaining structures are still cleanly multiplexed from microtubules. Scale bars: 5 μm (overview), 1 μm (zoom-ins). Together, these results establish that NanTex multiplexes monochromatic SMLM data across different organelles with high fidelity, preserves fine structural features, and tolerates experimental variability such as uneven labeling efficiency. NanTex demonstrates multiplexing capability across MINFLUX, STED, SIM, and Airyscan microscopy Having validated NanTex for SMLM, we next tested its generalizability across other fundamental SRM modalities ( Fig. 6 ), each with strengths and limitations: MINFLUX provides unmatched localization precision but is intrinsically low throughput 32 ; conversely, SIM, STED, and Airyscan are widely accessible but operate at lower resolution, potentially constraining nanotexture detection. Download figure Open in new tab Figure 6 | NanTex multiplexing capability across advanced SRM modalities . ( a ) Representative MINFLUX images of identically labeled clathrin and microtubules (AF647). Top: ground truth grayscale acquisitions; middle: NanTex multiplexing result (yellow: microtubules, magenta: clathrin) obtained with an SMLM-trained model without retraining; bottom: individual demixed channels. ( b ) Representative SIM analysis of identically labeled microtubules and clathrin. Top: monochromatic input; middle: NanTex multiplexing; bottom: single demixed channels. NanTex SIM reconstructions achieved high global MS-SSIM values (0.91 for microtubules, 0.87 for clathrin). ( c ) STED multiplexing of mitochondria (Tom20, Abberior STAR Red, cyan) and peroxisomes (PMP70, Abberior STAR Orange, magenta). Top: monochromatic acquisitions; middle: NanTex multiplexing; bottom: demixed channels. Quantitative evaluation showed robust demixing with MS-SSIM = 0.86 (mitochondria) and 0.90 (peroxisomes), despite lower single-scale SSIM values (∼0.61–0.67), consistent with STED’s reduced SNR. ( d ) Live-cell Airyscan recordings of ER (GFP-ATL3, green) and lysosomes (LysoTracker, magenta). Top: ground truth grayscale frames from a representative time course; middle: NanTex multiplexing; bottom: demixed channels. NanTex reconstructions achieved near-perfect fidelity for lysosomes (SSIM = 0.96, MS-SSIM = 0.94), while ER reconstructions scored lower at single-scale (SSIM = 0.53) but higher across scales (MS-SSIM = 0.75), reflecting robust recovery of global ER morphology. The entire time course is shown in Supplementary Video 1 and related ER-only Airyscan recordings (Supplementary Fig. 4, Supplementary Video 2) not only demonstrate vesicle demixing but also NanTex -enabled single-particle tracking of vesicular substructures, providing quantitative readouts of dynamic behavior. Scale bars: 1 μm (MINFLUX), 5 μm (SIM, STED, Airyscan). Applying our SMLM-trained model directly to MINFLUX images of identically labeled clathrin and microtubules yielded accurate multiplexing ( Fig. 6a ), establishing a proof-of-concept for cross-modality generalization without retraining - a particularly promising result given the low data yield of MINFLUX microscopy. However, performance degraded when the same model was applied to blurred d STORM or SIM datasets ( Supplementary Fig. 2), demonstrating that cross-modality transfer works when the modality’s effective resolution and/or contrast is close to the source domain. We next trained NanTex directly on lattice SIM datasets of microtubules and clathrin. Despite the modest lateral resolution (∼120 nm at 640 nm excitation), reconstructions achieved high MS-SSIM (0.91 for microtubules, 0.87 for clathrin), demonstrating that organelle-specific nanotextures can be recovered even under resolution-limited conditions ( Fig. 6b ). For STED, similar samples yielded poor performance ( Supplementary Fig. 3), consistent with the subpar STED utility of AF647. However, optimized dual-color STED imaging of mitochondria (Abberior STAR Red) and peroxisomes (Abberior STAR Orange) enabled robust multiplexing (MS-SSIM = 0.86 and 0.90; Fig. 6c ), despite modest single-scale SSIM values (∼0.61–0.67). This demonstrates that NanTex faithfully preserves global organelle structure under conditions of sufficient dye brightness and contrast, and further underscores the utility of multi-scale quality metrics in lower-SNR modalities. NanTex enables multiplexed live-cell imaging Encouraged by the strong SIM results, we next focused on live-cell applications. Compared to SIM, Airyscan microscopy offers similar resolution, but superior temporal performance and robustness making it the workhorse for dynamic multicolor recordings. Using U2OS cells expressing GFP-ATL3 and stained with LysoTracker, NanTex cleanly demixed ER and lysosomes from single-channel overlays ( Fig. 6d , Supplementary Video 1 ). Quantitatively, fidelity was near-perfect for lysosomes (SSIM = 0.96, MS-SSIM = 0.94) and robust for ER (SSIM = 0.53, MS-SSIM = 0.75), reflecting strong recovery of global morphology despite lower pixel-level contrast. Our data directly connects to our recent findings on newly discovered lysosomal proteins 21 and the emerging importance of ER–lysosome interactions in cellular physiology and pathology. In particular, ER–lysosome contact sites are increasingly recognized as key regulators in a wide range of disorders, where subtle perturbations of organelle communication underlie disease phenotypes. Importantly, NanTex also revealed previously hidden heterogeneity within single-labeled datasets. In Airyscan recordings of ER-only mRFP-KDEL U2OS cells, vesicle-like substructures became apparent that could not be captured by segmentation or orthogonal labeling, but were cleanly separated from the reticular ER by NanTex , consistent with ER-exit sites or sites of ER-phagy ( Supplementary Fig. 4, Supplementary Video 2 ). In addition to qualitative separation, NanTex enabled quantitative analysis of the demixed vesicle channel. Using TrackMate -based single-particle tracking 49 , we obtained mean-squared displacement (MSD) curves of vesicle-like structures across time-lapse recordings ( see methods , Supplementary Fig. 4e) 50 . The resulting diffusion coefficients extracted from MSD (time-average: 0.013±0.07 μm²/s, ensemble-average: 0.013±0.01 μm²/s) indicated a slow and homogeneous (ergodicity: 1.0) diffusion, while the anomalous diffusion alpha coefficients (time-average: 0.76±0.24, ensemble-average: 0.85) suggested confined movement, consistent with ER-bound compartments. This demonstrates that NanTex not only unmasks hidden structural subdomains but also renders them accessible to quantitative dynamic readouts directly from monochromatic live-cell data. This ability to uncover and track structurally distinct subdomains despite their identical label in real time, establishes live-cell imaging as one of the strongest potential application areas for NanTex . This could enable direct computational phenotyping of organelle heterogeneity and remodeling upon disease and treatment in real time in future studies. NanTex enables computational phenotyping of structurally distinct molecular populations Building on the cross-modality results, we next asked whether NanTex could be extended beyond separating different organelles to disentangling structural heterogeneity within a single molecular population. If successful, NanTex could also serve as a tool for computational phenotyping , i.e. detecting and quantifying structurally distinct subpopulations within otherwise indistinguishable molecular pools, inaccessible by orthogonal labelling strategies. To establish that NanTex recognizes structural nanotexture rather than spurious intensity cues, we conducted a controlled downsampling experiment on d STORM microtubule datasets, rendered with successively fewer localizations. Tubulin remained molecularly present but the filamentous nanotexture degraded into diffuse signal, leading to a sharp breakdown of SSIM and MS-SSIM values ( Supplementary Fig. 5), validating that NanTex faithfully recognizes the self-similar tubular structure itself, not auxiliary metadata. Guided by these controls, we applied NanTex to microtubule depolymerization phenotypes in human umbilical vein endothelial cells (HUVECs) treated with nocodazole 33 , with or without glyoxal supplementation at non-fixative concentrations 22 , 34 . Using the SMLM-trained model (microtubules, ER, clathrin), NanTex demixed intact filamentous microtubules (yellow) from diffuse tubulin pools (magenta) despite their identical antibody labeling ( Fig. 7 a-h ). The resulting per-cell polymerization index (PolIndex, Fig. 7 i, see methods ) distributions revealed significant and quantifiable phenotypes, i.e. drastic loss of filament integrity under nocodazole treatment (PolIndex = −0.61 ± 0.25) and a significant trend toward delayed depolymerization in the presence of both, nocodazole and glyoxal (PolIndex = 0.16 ± 0.23), while glyoxal alone resulted in a modest reduction (PolIndex = 0.42 ± 0.30) relative to untreated cells (PolIndex = 0.65 ± 0.22). Download figure Open in new tab Figure 7 | NanTex enables computational phenotyping of microtubule depolymerization in HUVECs. ( a–d ) Representative d STORM images of α-tubulin–labeled microtubules (AF647) in primary HUVECs under four conditions: untreated, glyoxal-stabilized (1 mM, 48 h), nocodazole-treated (10 µM, 5 min), and combined nocodazole + glyoxal treatment. ( e–h ) Corresponding NanTex multiplexing results, separating intact filamentous microtubules (yellow) from diffuse depolymerized tubulin (magenta). Insets ( g,h ) highlight representative regions where NanTex resolved diffuse signals and filaments despite their identical label. ( i ) Boxplots with jittered single-cell points of the polymerization index (PolIndex) for Untreated (n = 15), Glyoxal (n = 14), Nocodazole (n = 13), and Nocodazole + Glyoxal (n = 13) groups of cells. Two-sided Mann– Whitney U tests (unadjusted): Untreated vs. Nocodazole, p = 7.9×10⁻⁶ (****); Nocodazole vs. Nocodazole + Glyoxal, p = 2.61×10⁻⁵ (**** ) ; Untreated vs. Glyoxal, p = 0.0275 (*). Additional contrasts (not indicated): Glyoxal vs. Nocodazole, p = 1.75×10⁻⁵; Nocodazole + Glyoxal vs. Untreated, p = 7.44×10⁻⁵; Glyoxal vs. Nocodazole + Glyoxal, p = 0.0186. Star notation: ns (p ≥ 0.05), * (p < 0.05), ** (p < 0.01), *** (p < 0.001), **** (p < 0.0001). Boxplots show the median (orange center line); the interquartile range Q1–Q3 (box); whiskers extend to the most extreme data points within 1.5× the interquartile range from the quartiles; all individual data points are overlaid. Scale bars: 10 μm. Interestingly, even untreated cells in this dataset exhibited a detectable diffuse fraction (PolIndex = 0.65 ± 0.22), in contrast to our high-purity training datasets where filamentous integrity was near complete (PolIndex = 0.93 ± 0.04; Supplementary Fig. 6). This discrepancy traced back to antibody choice: an in-house conjugate with degree of labeling ∼1 yielded low background, whereas the commercial antibody used here introduced higher unspecific signal. Thus, beyond biological phenotyping, NanTex also functions as a quantitative quality-control tool, benchmarking labeling reagents and preparation protocols against gold-standard datasets. Together, these data establish NanTex as a computational phenotyping tool that detects and quantifies structurally distinct subpopulations within a single molecular pool and reveals a glyoxal-mediated delay of nocodazole-induced depolymerization. This opens new opportunities for quantitative structural biology, from ultrastructural remodeling of the ER to mitochondrial response subtype characterization, and provides a generalizable framework for systematic benchmarking of sample preparation strategies across laboratories and users. Discussion Multiplexed nanoscale imaging has long been constrained by the physics of fluorophores and detection, with spectral overlap and limited dye palettes restricting simultaneous labeling, particularly in live cells. NanTex reframes this problem as one of texture-based reconstruction rather than spectral separation or structural segmentation and defines a new paradigm for computational multiplexing. By exploiting nanoscale “nanotextures” as organelle fingerprints, NanTex achieves probabilistic demixing of single-channel images, preserving overlapping morphologies even under dense overlap conditions that segmentation-based approaches cannot resolve. Unlike recent segmentation frameworks such as Cellpose, Segment Anything, or µSAM that discretize objects, NanTex reconstructs overlapping organelles by leveraging nanotexture, positioning it as a complementary but fundamentally distinct class of computational multiplexing. A central outcome of this work is the demonstration that nanotexture serves as a universal descriptor of organelle identity. Because NanTex requires only ∼10 representative images per organelle and modality for effective training, it can be readily adapted across systems, making it a plug-and-play solution for users who wish to extend multiplexing beyond their immediate labeling and modality constraints. Across SMLM, MINFLUX, STED, SIM, and live-cell Airyscan, NanTex consistently demixed identically labeled organelles ranging from cytoskeletal filaments (microtubules, actin) to vesicular systems (clathrin, endosomes, lysosomes, peroxisomes) and reticular networks (ER, mitochondria). Importantly, NanTex not only recovers expected structures but also uncovers hidden subdomains within single-labeled datasets, such as vesicle-like ER substructures in live cells. This ability to reveal, quantify, and track structural heterogeneity in real time makes live-cell imaging the strongest application space for NanTex , directly enabling computational phenotyping of organelle remodeling during cellular dynamics. Moreover, the ability to track demixed vesicular structures in ER-only datasets ( e.g. Supplementary Fig. 4) demonstrates that NanTex can extend beyond visualization to provide quantitative, dynamic readouts in live-cell contexts. In nocodazole-treated HUVECs, NanTex separated intact microtubule filaments from diffuse depolymerized tubulin, uncovering a glyoxal-mediated phenotype not accessible with labeling-based strategies. We note that benchmarking NanTex requires metrics that capture perceptual and structural fidelity. We therefore focus on SSIM and MS-SSIM, which reflect human-perceived image similarity and multi-scale preservation of structure. Segmentation-based metrics (e.g., F1/Jaccard on binarized structure masks, Manders-type assignment for channel specificity) are ill-suited, as NanTex does not classify discrete objects but reconstructs overlapping features. This is a critical conceptual distinction and positions NanTex adjacent to segmentation pipelines. While we demonstrate the broad applicability of NanTex , there are caveats to be aware of (see also Supplementary Discussion ). Out-of-focus regions and organelle heterogeneity represent two central limitations. Our blurring experiments and SMLM microtubule datasets illustrate that in areas of pronounced defocus, NanTex no longer assigns organelle identity, a behavior that is both expected and conceptually reassuring, as it reflects the loss of true texture information rather than spurious assignments. Similarly, organelles such as ER and mitochondria display considerable morphological diversity, ranging from fragmented to elongated networks. Importantly, this study emphasized NanTex to be data-efficient, with ∼10 training images per organelle, when combined with augmentation, in part to lower its adoption barrier with novice users. However, we anticipate significant performance gains with larger datasets, as well as a more robust capture of the full spectrum of organelle heterogeneity. We envision a community-wide effort to share organelle-specific SRM data that would enable training of holistic models capable of demixing ten or more organelles simultaneously. Looking forward, NanTex provides a framework for significant methodological expansion. Integration with transformer-based architectures 35 could better capture long-range structural dependencies, while adversarial (GAN) 36 , 46 layers may further enhance the realism of demixed reconstructions. Beyond fluorescence SRM, the concept of texture-guided reconstruction could also extend to volumetric SRM data, correlative EM, or even (annotated) label-free modalities 45 , where structural nanotextures provide discriminative cues. In summary, NanTex establishes nanotexture as a paradigm-shifting descriptor for organelle identification. By enabling label-efficient, cross-modality, and live-compatible multiplexing, it bridges computational image analysis with quantitative cell biology. Its potential for real-time phenotyping and for tracking disease-relevant remodeling positions NanTex as a foundation for the next generation of high-content, clinically relevant structural biology. Materials Availability All data needed to evaluate the conclusions of this study are provided in the main text, supplementary information, and/or the Key Resources Table ( Supplementary Table 1 ). Plasmids generated for this study have been deposited at Addgene (accession numbers listed in the Key Resources Table). No new unique reagents were generated beyond those listed. Data and Code Availability All raw microscopy image files and processed source data supporting the findings of this study have been deposited at Zenodo 51 – 56 and are freely accessible. A comprehensive overview is provided in Supplementary Table 2 . All code used for Haralick feature extraction, NanTex model training, inference, and evaluation (including trained network weights for all SRM modalities) is made available via GitHub and will be archived on Zenodo upon publication. Custom analysis scripts for quantitative benchmarking (SSIM, MS-SSIM, F1-score, etc.) are provided alongside training/inference pipelines in the GitHub repository. Prior to publication: NanTex weights are being made available to reviewers upon request. Post publication: Pretrained NanTex weights for SMLM, SIM, STED, and Airyscan used in this study are made available (see Zenodo/GitHub), but we emphasize that retraining may be required if imaging parameters or resolution differ substantially. Competing Interests B.T.L.V., G.J.G., C.F. have submitted a patent application related to the technology presented in this study. All other authors declare no competing interests. Declaration of generative AI and AI-assisted technologies in the writing process The authors used DeepL and ChatGPT to assist with style, grammar, and vocabulary during the preparation of this work to ensure that it is grammatically and linguistically correct. After using these tools, the authors reviewed and edited the content as needed. We, the authors, declare full responsibility for the content of the published article. Methods Sample Preparation for ER live-cell imaging Plasmid construction The expression construct pInd20_GFP-ATL3 was generated using the Gateway cloning system (Thermo Fisher Scientific). The entry clone pDONR221_GFP-ATL3 (kindly provided by Laura Behrendt) was recombined with the destination vector pInd20 (Addgene plasmid #44012) via an LR recombination reaction according to the manufacturer’s instructions. The resulting expression vector pInd20_GFP-ATL3 was verified by restriction digestion and sequencing. Cell culture mRFP-KDEL U2OS cells were generated as described in 21 and were cultured in DMEM high glucose media (Sigma) supplemented with 10% heat inactivated FBS (Sigma). Cells were maintained at 37°C, 5% CO 2 and 95% humidity in a CO 2 incubator. Cells were seeded in 8 well chambered cover glass systems with high-precision glass coverslips (No. 1.5H, thickness 170 ± 5 µm, Thermo Fisher, cat. #155409) with 20x10 3 cells per well. Prior to live-cell imaging of mRFP-KDEL U2OS the medium was changed to phenol-free, complete DMEM media (Thermo Fisher). Generation of U2OS cell line expressing inducible GFP-ATL3 A Tet-inducible lentiviral vector (pInducer20) carrying GFP-ATL3 (pInd20_GFP-ATL3) was used to generate U2OS stable cell lines. Lentiviral particles were produced in HEK293T/LentiX cells using a three-plasmid system. Briefly, cells were plated at a density of 6.0 × 10⁶ per 10-cm dish in 10 mL growth medium (DMEM supplemented with 10% FBS and 1% Penicillin/Streptomycin). The following day, cells were co-transfected with 10 µg pInd20_GFP-ATL3, 10 µg pPAX2 (Addgene plasmid #12260), and 2.5 µg pMD2.G (Addgene plasmid #12259) using polyethyleneimine (PEI-Max, 1 mg/mL stock) in Opti-MEM (Thermo Fisher Scientific). After a 20 min complexation period, the transfection mixture was added to cells in the transfection medium (DMEM supplemented with 2% FBS, without Penicillin/Streptomycin). Culture medium was replaced with growth medium 16–18 h post-transfection. Viral supernatants were collected at 48 h, 60 h, and 72 h post-transfection, filtered through a 0.45 µm syringe filter, aliquoted, and stored at −80 °C. For infection, U2OS wild-type cells were plated to reach 30–40% confluency the following day and exposed to viral supernatant supplemented with 8 µg/mL protamine sulfate in a final volume of 5 mL infection medium (DMEM). In parallel, a control condition was processed identically but without viral supernatant. Twenty-four hours post-infection, medium was replaced with complete DMEM (Thermo Fisher, cat. no. 61965-026). To select stable transductants, G418 (Geneticin) was added at a final concentration of 400 µg/mL. Medium containing G418 was renewed every 2–3 days until non-transduced control cells were fully eliminated. Lysosome staining U2OS_pInd20_GFP-ATL3-WT were plated at a density of 5.0 × 10⁴ cells per dish on 35-mm glass-bottom dishes (MatTek) in 2 mL of complete DMEM without phenol red (Thermo Fisher Scientific, cat. no. 21063-029). To induce expression of GFP-ATL3, doxycycline was added at a final concentration of 1 µg/mL. Cells were maintained under induction conditions for 48 h prior to imaging. For lysosomal labeling, cells were incubated with LysoTracker™ Red DND-99 (Thermo Fisher Scientific) at a final concentration of 60 nM in complete DMEM. Cells were stained for 30 min at 37 °C under standard culture conditions. Immunofluorescence labeling of microtubules, clathrin, and ER Cos-7 cells were grown in 8 well chambered cover glass systems with high-precision glass coverslips (No. 1.5H, thickness 170 ± 5 µm) and rinsed briefly in pre-warmed PBS (37 °C) to remove medium. Cells were permeabilized for 1–2 min using 0.3% glutaraldehyde and 0.25% Triton X-100 in cytoskeleton buffer (10 mM MES pH 6.1, 150 mM NaCl, 5 mM EGTA, 5 mM glucose, 5 mM MgCl₂), followed by fixation for 10 min in 2% glutaraldehyde in the same buffer. After two washes in PBS, residual autofluorescence was quenched with 0.1% NaBH₄ in PBS for 7 min. Cells were washed once in PBS, then blocked in 5% BSA/PBS for 30 min. For microtubule staining, cells were incubated with rabbit anti-α-tubulin primary antibodies (2 ng/µl in blocking solution, 60 min at RT), followed by three washes (1× PBS, 2× PBS/0.1% Tween-20, 5 min each). Secondary labeling was performed with Alexa Fluor 647-conjugated goat anti-rabbit IgG (4 ng/µl, 45 min in blocking solution). Clathrin-coated pits and vesicles were stained with rabbit anti-clathrin heavy chain antibodies (2 ng/µl) and corresponding Alexa Fluor 647-conjugated goat anti-rabbit secondary antibodies (4 ng/µl). ER structures were labeled with rabbit anti-Calnexin primary antibodies in the same workflow. After secondary labeling, cells were washed (1× PBS, 2× PBS/0.1% Tween-20) and post-fixed with 4% formaldehyde in PBS for 5 min. Final washes were performed in PBS, and samples were stored in PBS containing 0.05% sodium azide at 4 °C until imaging. For antibody conjugation, Alexa Fluor 647-NHS was coupled to secondary goat anti-rabbit antibodies in sodium tetraborate buffer (100 mM, pH 9.5) using spin-desalting columns (40 kDa MWCO) for purification. A 2x molar dye excess was used to achieve a degree of labeling (DOL) of ∼1.1. The DOL was determined by UV–vis absorption spectroscopy. HUVEC cell culture and treatment Primary human umbilical vein endothelial cells (HUVECs) were isolated from anonymously acquired human umbilical cords according to the “Ethical principles for Medical Research Involving Human Subjects” (Declaration of Helsinki 1964) as previously described 22 , 43 . The protocol was approved by the Jena University Hospital Ethics Committee. The donors were informed and gave written consent. Briefly, after rinsing the cord veins with 0.9% NaCl, endothelial cells were detached with collagenase (0.01%, 3 min at 37 °C), suspended in M199/10% FCS, washed once (500 × g, 6 min) and seeded on a cell culture flask coated with 0.2% gelatin. Full growth medium (M199, 17.5% FCS, 2.5% human serum, 7.5 µg/ml ECGS, 7.5 U/ml heparin, 680 µM glutamine, 100 µM vitamin C, 100 U/ml penicillin, 100 µg/ml streptomycin) was added 24 h later. Cells were cultured until confluence and experiments usually carried out with HUVEC from the second passage. Nocodazole-treated microtubule samples: For nocodazole experiments, HUVECs were seeded on double-coated high-precision glass coverslips (No. 1.5H, 24 mm, gelatin 1%) at a density of ∼2.2 × 10⁵ cells per coverslip in 6-well plates. Cells were cultured under standard endothelial growth conditions for 24 h before addition of glyoxal (1 mM) for 48 h. To induce acute microtubule depolymerization, cells were treated with 10 µM nocodazole (Sigma M1404-2MG) for 5 min prior to fixation. Experimental groups included (i) untreated control cells, (ii) glyoxal-treated cells, (iii) nocodazole-treated cells, and (iv) combined GO + nocodazole treatment. Additional controls consisted of primary-only and secondary-only antibody samples. Cells were fixed in glutaraldehyde fixation buffer for 15 min, washed with PBS, and quenched with 0.1% NaBH₄ (3 × 10 min). After permeabilization in 0.1% Triton X-100 and blocking with 3% BSA, microtubules were stained with rabbit monoclonal anti-α-tubulin antibody (ab18251, Abcam; 2 µg/ml, overnight at 4 °C) followed by Alexa Fluor 647-conjugated goat anti-rabbit secondary antibody (4 µg/ml, 1 h at room temperature). Nuclei were counterstained with Hoechst 33342 (1 µg/ml, 10 min). Mitochondria and peroxisomes for STED HEK 293 (ATCC, VA, USA) and cells were maintained in a culture medium consisting of DMEM with 4500 mg glucose/L, 110 mg sodium pyruvate/L supplemented with 10% foetal calf serum, glutamine (2 mM) and penicillin–streptomycin (1%). The cells were cultured at 37 °C/8.5% CO2. Cells were grown on a #1.5 μ-Dish 35 mm. For immunolabeling, the cells were fixed with 3% formaldehyde in PBS, permeabilized for 5 min with 0,1 % Triton X-100 and blocked with 2% BSA + 5% FCS in PBS for 1 h at room temperature. Samples were incubated with primary antibodies in blocking buffer for 1 h at room temperature. To stain peroxisomes, antibodies recognizing PMP70 (Anti-PMP70 antibody (ab3421), rabbit polyclonal, Abcam, Cambridge, UK) were used. Mitochondria were labelled using an antibody against TOM20 (Tom20 Antibody (F-10): sc-17764, mouse monoclonal, Santa Cruz Biotechnology, Dallas, TX). After several washing steps, the cells were incubated for 30 min with secondary antibodies conjugated to Abberior STAR orange (rabbit) and Abberior STAR Red (mouse) (Abberior Instruments, Goettingen, Germany) diluted 1:250 in 1% BSA in PBS. After several washing steps, the slides were mounted on a drop of Mowiol (Sigma-Aldrich). Data Acquisition and Imaging Modalities SMLM imaging Clathrin, ER, and microtubule SMLM imaging was performed on a customized d STORM platform based on an inverted Olympus IX-71 microscope equipped with a 60× / 1.45 NA oil immersion objective, described in detail elsewhere 6 , 27 . Alexa Fluor 647 was excited with a 641 nm laser at ∼1.5-3.5 kW cm⁻² in a switching buffer consisting of 100 mM β-mercaptoethylamine (MEA), 10% (w/v) glucose, 0.5 mg ml⁻¹ glucose oxidase, and 40 µg ml⁻¹ catalase in PBS, adjusted to pH 7.4 with KOH. After a brief epi-illumination step to suppress out-of-focus fluorescence, 30,000– 50,000 frames were recorded in HiLo mode at 33 ms per frame on an Andor iXon DU-897 EMCCD camera with 128 nm optical pixel size. Nocodazole-related datasets were acquired on a ZEISS Elyra 7 operated in d STORM mode, using a 63× / 1.45 NA TIRF objective and identical dye chemistry and buffer conditions as above. Frame sequences (∼20,000–50,000 frames, 25–33 ms per frame) were recorded in HiLo mode on a pco.edge sCMOS camera with 100 nm optical pixel size and processed identically in rapidSTORM 41 . MINFLUX imaging MINFLUX imaging was performed on a commercial MINFLUX system (Abberior Instruments) using identical biological samples and labeling conditions as described for SMLM, with Alexa Fluor 647 as the fluorophore. A freshly prepared imaging buffer was used containing 30 mM cysteamine (MEA), 0.5 g/L glucose oxidase (stored at –20 °C), 40 µg/mL catalase (stored at 4 °C), and 10% TRIS/Cl buffer (pH 8.0, 4 °C). For drift correction, 100 nm gold beads (fiducials) were introduced by adding 200 µL suspension into ibidi 8-well chambers for 5 min, followed by three PBS wash steps. Imaging was performed in MBM stabilization mode, optimized for long-term acquisition. Samples were pre-switched for 60–120 s at 640 nm laser wavelength and 50% laser power to reduce emitter density to the optimal range for MINFLUX localization. Subsequent imaging was carried out overnight at 10% laser power. Both 2D and 3D default imaging sequences provided by Abberior Instruments were employed without further modification. SMLM Rendering and Preprocessing Raw SMLM image stacks were processed in rapidSTORM 41 with free FWHM fitting enabled and 500 photons lower intensity threshold. SMLM and MINFLUX localization data were rendered into grayscale 8-bit TIFF images using rapidSTORM with 10 nm/pixel resolution, intensity scaling based on photon count, and bilinear pixel-interpolation. Localization files were converted from proprietary formats (.smlm, .msr, etc.) into plain-text files readable by rapidSTORM using custom scripts if necessary. Rendered images were cropped and converted to uniform resolution and scale. For MINFLUX data, additional histogram equalization and morphological opening (2-pixel kernel) were applied to suppress depth-projected artifacts. SIM imaging SIM imaging was performed on a ZEISS Elyra 7 system (laser module: 000000-0239-500) equipped with four laser lines (405, 488, 561, 641 nm) and a mercury vapor lamp. A Plan-Apochromat 63×/1.4 oil immersion objective was used in combination with the built-in Optovar (1.6×), yielding an effective pixel size of 62 nm at the detector. For comparative d STORM measurements, a 63×/1.45 NA oil immersion TIRF objective was employed. A quad-band dichroic beamsplitter and emission filter (LBF 405/488/561/642) was used for all acquisitions. Each SIM dataset was acquired in lattice SIM mode as a z-stack with a physical step size of 110 nm, where each plane consisted of 13 phase-shifted raw images. 3D datasets were recorded to optimize optical sectioning and suppress out-of-focus contributions, but only 2D SIM reconstructions were used for downstream analysis. Exposure time per phase was 250 ms, and optical gratings were selected according to the system’s default settings. SIM reconstruction was carried out in ZEN Black (Zeiss) with the internal reconstruction tool using the "precise" quality setting, no baseline cut, and default Wiener filtering. STED imaging STED imaging was performed on an Abberior STEDYCON system (Abberior Instruments GmbH) mounted on an Olympus IX83 inverted microscope equipped with a UPlanXApo 100×/1.45 NA oil immersion objective. Images were acquired with an effective pixel size of 20 nm, a pixel dwell time of 10 µs, and 3–10 line accumulations depending on the size of the region of interest. Mitochondria were labeled with Abberior STAR ORANGE and excited with a pulsed 561 nm laser (0.5–3.2 µW, adjusted to labeling density), while peroxisomes were labeled with Abberior STAR RED and excited with a pulsed 640 nm laser (16–33 µW). In both cases, fluorescence depletion was achieved with a pulsed 775 nm fiber laser operated at ∼ 225 mW in the sample plane. Reported laser powers correspond to mean values at the specimen plane; values were approximated based on calibrated power measurements taken on different experimental days and adjusted with internal correction factors for optical losses. Emission signals were split between two avalanche photodiode detectors (SPCM-AQRH, Excelitas Technologies): the STAR ORANGE channel was collected between 580–630 nm, and the STAR RED channel between 650–700 nm. Time-gated detection was applied between 1–7 ns after excitation. All acquisitions were controlled using STEDYCON Smart Control software (version 7.1.53, Abberior Instruments). Airyscan live-cell imaging Airyscan live-cell imaging was performed on a Zeiss LSM 980 confocal microscope equipped with an Airyscan2 detector in Fast 2D mode. Cells were maintained at 37 °C and 5% CO₂ in a stage-top incubation chamber (PeCon). Images were acquired using a Plan-Apochromat 63×/1.4 oil immersion objective. GFP-ATL3 was excited at 488 nm and detected at 523 nm, while LysoTracker Red was excited at 561 nm and detected at 579 nm; mRFP was excited at 590 nm the emission was captured from 612 nm. Sequential image series were collected with a pixel size of 43 nm selecting a region of interest of 26.48 × 26.48 µm, resulting in a frame time interval of ∼ 260 ms. Images were optionally deconvolved with the Airyscan processing tool of the ZEN Blue software (v3.10). Haralick Feature Extraction and Classical Clustering Feature computation Classical textural descriptors were generated using the Mahotas Python library (v1.4.13) 38 . For each pixel, a set of 13 Haralick features were extracted from gray-level co-occurrence matrices (GLCMs), including angular second moment, contrast, correlation, variance, inverse difference moment, sum average, sum variance, sum entropy, entropy, difference variance, difference entropy, information measures of correlation, and maximal correlation coefficient. To capture local heterogeneity, both the mean and peak-to-peak range of each Haralick feature were calculated within a 7×7 sliding window. This resulted in a 26-dimensional feature vector per pixel. Normalization All feature values were scaled to the range [0,1] by min–max normalization, applied independently per feature. Scaling ensured comparability across different samples and avoided dominance of high-variance features. Feature maps were border-trimmed during sliding-window calculation to prevent artifacts at image edges. F1-Score Haralick feature maps were generated from computational superimpositions of two mono-structural images. These input images were then masked to isolate the regions of interest. To evaluate segmentation performance, we compared the nonzero pixels in the overlay of the original images with the nonzero pixels in the thresholded feature maps. This comparison was performed across multiple threshold levels. The performance of each threshold was quantified using the F1-score as implemented in scikit-learn (v1.2.1), with the highest score being returned for a comparison. As semantic segmentation is a binary decision on the pixel level (either Microtubules or LNP), regions of overlap were considered to be belonging to the more present structure, in this case Microtubules. Clustering Unsupervised clustering was performed on the 26-dimensional feature stack using k-means (k = 5) as implemented in scikit-learn (v1.1.3). Each pixel was assigned to one of five clusters by minimizing squared Euclidean distance in feature space. This procedure yielded cluster maps representing regions of locally similar nanotexture. Evaluation While Haralick clustering proved useful for feature exploration and highlighted major differences in nanotexture between organelles, it was insufficient for precise multiplexing in regions of structural overlap. Specifically, cluster boundaries frequently fragmented continuous structures (e.g., filamentous microtubules) and failed to preserve co-localized signals. These observations motivated the development of NanTex , which extends beyond classical feature-driven segmentation to probabilistic demixing of overlapping structures. NanTex Computational superimposition and first level data augmentation To generate train, test and validation data, mono-structural images were rescaled [0, 255], resized (padded or cut), rotated against each other (0°, 90°, 180°, 270°), and overlaid summatively. Permuting over all combinations of samples per structure and rotations, the amount of available data had been increased substantially while creating different structural aggregates. Data preprocessing and second level augmentation All input fluorescence images were preprocessed by standardizing each patch to zero mean and unit variance. Ground-truth channels corresponding to individual organelles were min–max normalized to the range [0,1]. Basic augmentation was applied using the Albumentations library 37 , 38 to increase model robustness to photophysical variability and imaging artifacts. Augmentations included: randomized patch cutouts (256x256), randomly applied horizontal and vertical flips, and Median blur (kernel size in [3,5]. All data processing involved has been performed on an AMD Ryzen Threadripper 2920X 12-Core CPU leveraging a custom implementation of the ray library to realize multiprocessing. Network architecture NanTex employs a UNet–style convolutional encoder–decoder with four resolution levels. Each level consists of two 3×3 convolutions followed by batch normalization and ReLU activation. Downsampling is implemented via 2×2 max pooling, while upsampling uses transposed convolutions. Skip connections ensure retention of fine spatial information. The network processes 32 feature maps per patch. The demixing head is a 1×1 convolutional layer with linear activation, producing a grayscale output for each channel (256×256 each), corresponding to the predicted organelle structures. Training protocol Training was performed in PyTorch (v1.12) 39 on an NVIDIA RTX A5000 GPU. The Adam optimizer was used with learning rate 1×10−5, β1=0.9, β2=0.99, and ε=1×10−8. The loss function was defined as mean squared error (MSE) between predicted and ground-truth images. Variable batch sizes (e.g. 16, 32 patches) were used. Training typically converged after ∼80–120 epochs of varying sizes (e.g. 16, 32, 64 cycles), monitored by validation MSE and SSIM. Prediction and post-processing Multi-structural images were subsequently loaded, resized (to prevent shape mismatches), standardized to zero-mean-unit-variance, and divided into patches (usually 256x256 px), before being passed through the model sequentially, skipping empty patches to increase execution speed and prevent hallucinations. Predicted patches were reassembled into full-size images using a sliding window with stride = 256 px (no overlap). Background correction was applied by histogram-based binning, where the five most frequent intensity values per channel were zeroed to suppress background contributions. Quantitative evaluation with SSIM and MS-SSIM To benchmark NanTex reconstructions against ground-truth (GT) channels, we used the pytorch- msssim implementation of the Structural Similarity Index ( SSIM ) and its multi-scale extension ( MS-SSIM ), two perception-aligned image quality metrics that compare local luminance, contrast, and structure rather than raw pixel errors. SSIM is computed on small windows and combines three terms, luminance, contrast, and structure, into a single score in [−1, 1] (practically [0, 1] after intensity normalization). MS-SSIM extends SSIM across image pyramids, weighting coarse and fine scales to better reflect human/perceptual judgment and robustness to small misregistrations 29 – 31 . We report both because SSIM is highly sensitive to high-frequency losses (e.g., blur), while MS-SSIM more faithfully summarizes fidelity across spatial scales (useful when nanotexture is preserved but slightly smoothed). For completeness, PSNR was also computed as an internal quality control metric but not used for decision-making, as it correlates weakly with perceived quality compared to SSIM-type metrics. Computation All images were 8-bit or normalized to [0, 255]. SSIM used a Gaussian window (11×11, σ = 1.5), constants C1 = (0.01L)², C2 = (0.03L)² with dynamic range L = 1, and “reflect” boundary handling (scikit-image, v0.22; parameters per default) 40 . MS-SSIM used 5 scales with down-sampling by 2 and the original weighting vector [0.0448, 0.2856, 0.3001, 0.2363, 0.1333] 30 . Metrics were computed (i) globally per full field of view and (ii) locally on 256×256 patches/ROIs (reported as mean ± s.d. across patches). All scores are reported per GT/predicted organelle channel pair. Controls and caveats SSIM can be deteriorated by large uniform backgrounds. Small lateral misregistrations depress SSIM; MS-SSIM mitigates this but does not replace proper alignment. Two unrelated images with similar low variance can yield mediocre SSIM; our negative controls (unrelated organelles, natural images, voids) produced low SSIM/MS-SSIM, validating specificity (Supplementary Fig. 2). Because SSIM-type metrics remain intensity-aware, we also report intensity-agnostic metrics elsewhere when the biological question concerns presence/absence rather than brightness (e.g., computational phenotyping). Modality-specific considerations Notably, in STED and Airyscan datasets, single-scale SSIM values were reduced due to the inherently lower SNR and contrast of these modalities, yet MS-SSIM remained high (0.86–0.90 for STED; > 0.74 for Airyscan), accurately reflecting preservation of large-scale organelle morphology. This highlights that multi-scale metrics are more appropriate for benchmarking NanTex in lower-SNR imaging regimes where local contrast is degraded but global structures are faithfully demixed. Reporting For each modality and organelle type we report global and patch-wise SSIM/MS-SSIM (main text and figure panels), and we visualize per-pixel SSIM heatmaps to localize discrepancies where meaningful. When summarizing across many fields, we present distributions (box plots) and the median with interquartile range. Interpretation (rule-of-thumb for NanTex SRM demixing) Absolute thresholds depend on modality/SNR. In our experience, MS-SSIM is typically slightly higher than SSIM when large-scale morphology is correct but fine detail is modestly smoothed; the converse (MS-SSIM ≤ SSIM) flags scale-dependent artifacts (e.g., high-frequency hallucinations). We therefore use both metrics jointly: high MS-SSIM with high SSIM indicates faithful demixing; divergence between the two prompts visual inspection or complementary, intensity-agnostic measures if applicable. Table 1 provides an overview consistent with qualitative meaning in our datasets and controls (also see Supplementary Fig. 2) Single Particle Tracking of Demixed Vesicles Demixed time-lapse image stacks were analyzed using TrackMate (v7.14.0) 49 . Vesicles were detected with the LoG (Laplacian-of-Gaussian, 15 pixel object diameter, quality threshold = 0.5) detector with sub-pixel localization enabled and the median filter option active to suppress noise. Tracks were built with the advanced Kalmann Tracker (search radius 20 px, max frame gap 2 frames). Mean-squared displacement (MSD) curves were computed per track from the exported coordinates and processed with custom Python scripts as described in detail in 50 . Computational Phenotyping of Microtubule Depolymerization Computational phenotyping of microtubule morphology subtypes was performed on d STORM images of untreated, glyoxal-treated, nocodazole-treated, and combined glyoxal/nocodazole-treated HUVEC cells. The NanTex model trained on untreated microtubule SMLM, clathrin and ER data was applied to demix each image into two channels: (i) intact microtubules (microtubule channel) and (ii) diffuse tubulin (combined other channels). For each cell, the number of positive pixels per channel was determined after thresholding at the 95th percentile of background intensity distribution to exclude noise. The polymerization index was then defined straight forwardly as where N mic denotes the number of pixels assigned to the intact microtubule channel and N rest the number of pixels assigned to the diffuse tubulin channel. This metric provides a normalized, cell-wise readout ranging from –1 (entirely depolymerized) to +1 (entirely filamentous). Values closer to zero correspond to mixed phenotypes or unspecific backgrounds. Single-cell polymerization index values were compiled across conditions to assess treatment-dependent phenotypes. Glyoxal treatment alone yielded intermediate values, whereas nocodazole exposure produced a strong shift toward negative indices, consistent with microtubule depolymerization. The combined treatment condition enabled assessment of potential additive or synergistic effects. Statistical analysis was performed on the per-cell polymerization index distributions across conditions. Group comparisons were evaluated using Welch’s t -tests alongside non-parametric Mann–Whitney U tests to account for small sample sizes and potential variance heterogeneity. Effect sizes were additionally reported (Cohen’s d , rank-biserial correlation) to capture the magnitude of differences independent of sample size. P-values < 0.05 were considered significant. Given the limited sample sizes for these proof-of-concept experiments (n = 3–5 cells per condition), statistical interpretation relied not only on p -values but also on effect size measures, which provide a more robust estimate of biological relevance under small- n conditions. Welch’s t -tests were used for group comparisons allowing unequal variances, and Mann–Whitney U tests served as a non-parametric validation. Effect sizes were reported as Cohen’s d (parametric) and rank-biserial correlation (non-parametric). This combined approach ensures that both statistical significance and the magnitude of observed effects are transparently communicated, with effect sizes serving as the primary criterion where statistical power was limited. Supplementary Figures Supplementary Figure 1. Illustration of SSIM and MS-SSIM specificity as controls for NanTex evaluation. Supplementary Figure 2. Resolution-dependence of NanTex performance. Supplementary Figure 3. Suboptimal NanTex performance with AF647 STED data. Supplementary Figure 4. NanTex demixing and single particle tracking of live-cell Airyscan datasets. Supplementary Figure 5. Validation of NanTex recognition specificity using reduced localization density. Supplementary Figure 6. High-purity microtubule d STORM datasets used as reference for quality control. Supplementary Videos Supplementary Video 1. Live-cell Airyscan: ER–lysosome dynamics demixed by NanTex . Supplementary Video 2. NanTex demixing of ER-only super-resolution Airyscan live-cell data. Acknowledgements C.F. is grateful for the support of all Franke Lab members. C.F. and C.E. especially thank Silvia Ruthardt for administrative assistance. We thank Michael Habeck and Parsa Khaleghi for fruitful discussions regarding model transferability and depolarization quantification. C.F. is supported by the Thüringer Aufbaubank (TAB, FGR 0060 KI-supER) and the DFG via the Collaborative Research Center PolyTarget (CRC 1278, project number 316213987, project B07). P.C. has received funding from the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement no. 892232. J.C.H. is supported by the Fritz-Thyssen Foundation (award number: 10.20.1.022MN), and the NCL Foundation (NCL-Stiftung). C.K. is supported by the DFG (KA 1751/9-1). B.T.L.V. is supported by the Thüringer Landesgraduiertenstipendien 2024 and has received funding through the Honors Programme of the Friedrich-Schiller-Universät Jena. B.T.L.V is especially grateful for all the discussions and exchange made possible by the DL@MBL 2022 initiative as well as for the funding provided therefore by the Marine Biological Laboratory and the Howard Hughes Medical Institute. We thank the Microverse Imaging Center for providing microscope facility support. The Microverse Imaging Center is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – EXC 2051 – Project-ID 39071386 and project number 316213987 (CRC 1278, project Z01). The ELYRA 7 was funded by the Free State of Thuringia with grant number 2019 FGI 0003. The LSM 980 was funded by the Free State of Thuringia with grant number 2019 FGI 0001. We further acknowledge the DFG (Instrument funding MINFLUX Jena INST 275_405_1; Instrument funding modular STED INST 1757/25-1 FUGG; GRK M-M-M: GRK 2723/1 – 2023 – ID 44711651), the State of Thuringia (TMWWDG), the Leibniz Association (Leibniz Collaborative Excellence Programme, project AMPel – project number K548/2023), and the Free State of Thuringia (TAB; AdvancedSTED / FGZ: 2018 FGI 0022; Advanced Flu-Spec / 2020 FGZ: FGI 0031) for the work on the STED and MINFLUX microscopes. The authors also gratefully acknowledge support from the FLI Core Facilities Proteomics, and Imaging. Footnotes https://github.com/Zatyrus/NanTex https://doi.org/10.5281/zenodo.17120603 https://doi.org/10.5281/zenodo.17120791 https://doi.org/10.5281/zenodo.17121085 https://doi.org/10.5281/zenodo.17121094 https://doi.org/10.5281/zenodo.17121098 https://doi.org/10.5281/zenodo.17121104 References 1. ↵ Schermelleh , L. , et al. Super-resolution microscopy demystified . Nat. Cell Biol. 21 , 72 – 84 ( 2019 ). doi: 10.1038/s41556-019-0333-6 OpenUrl CrossRef PubMed 2. ↵ Jacquemet , G. , Carisey , A. F. , Hamidi , H. , Henriques , R. & Leterrier , C . The cell biologist’s guide to super-resolution microscopy . J. Cell Sci . 133 , jcs240713 ( 2020 ). doi: 10.1242/jcs.240713 OpenUrl Abstract / FREE Full Text 3. ↵ Hess , S. T. , Girirajan , T. P. & Mason , M. D . Ultra-high resolution imaging by fluorescence photoactivation localization microscopy . Biophys. J . 91 , 4258 – 4272 ( 2006 ). doi: 10.1529/biophysj.106.091116 OpenUrl CrossRef PubMed Web of Science 4. Rust , M. J. , Bates , M. & Zhuang , X . Sub-diffraction-limit imaging by stochastic optical reconstruction microscopy (STORM) . Nat. Methods 3 , 793 – 795 ( 2006 ). doi: 10.1038/nmeth929 OpenUrl CrossRef PubMed Web of Science 5. Heilemann , M. et al. Subdiffraction-resolution fluorescence imaging with conventional fluorescent probes . Angew. Chem. Int. Ed . 47 , 6172 – 6176 ( 2008 ). doi: 10.1002/anie.200802376 OpenUrl CrossRef PubMed Web of Science 6. ↵ van de Linde , S. et al. Direct stochastic optical reconstruction microscopy with standard fluorescent probes . Nat. Protoc . 6 , 991 – 1009 ( 2011 ). doi: 10.1038/nprot.2011.336 OpenUrl CrossRef PubMed 7. ↵ Hell , S. W. & Wichmann , J . Breaking the diffraction resolution limit by stimulated emission . Opt. Lett . 19 , 780 – 782 ( 1994 ). doi: 10.1364/OL.19.000780 OpenUrl CrossRef PubMed Web of Science 8. ↵ Balzarotti , F. et al. Nanometer resolution imaging and tracking of fluorescent molecules with minimal photon fluxes . Science 355 , 606 – 612 ( 2017 ). doi: 10.1126/science.aak9913 OpenUrl Abstract / FREE Full Text 9. ↵ Gustafsson , M. G . Surpassing the lateral resolution limit by a factor of two using structured illumination microscopy . J. Microsc . 198 , 82 – 87 ( 2000 ). doi: 10.1046/j.1365-2818.2000.00710.x OpenUrl CrossRef PubMed Web of Science 10. ↵ Shao , L. , Kner , P. , Rego , E. H. & Gustafsson , M. G . Super-resolution 3D microscopy of live whole cells using structured illumination . Nat. Methods 8 , 1044 – 1046 ( 2011 ). doi: 10.1038/nmeth.1734 OpenUrl CrossRef PubMed Web of Science 11. ↵ Huff , J . The Airyscan detector from ZEISS: confocal imaging with improved signal-to-noise and resolution . Nat. Methods 12 , 1205 – 1210 ( 2015 ). doi: 10.1038/nmeth.3441 OpenUrl CrossRef 12. ↵ Grimm , J. B. & Lavis , L. D . Caveat fluorophore: an insiders’ guide to small-molecule fluorescent labels . Nat. Methods 19 , 149 – 158 ( 2022 ). doi: 10.1038/s41592-021-01338-6 OpenUrl CrossRef PubMed 13. ↵ Niehörster , T. et al. Multi-target spectrally resolved fluorescence lifetime imaging microscopy . Nat. Methods 13 , 257 – 262 ( 2016 ). doi: 10.1038/nmeth.3817 OpenUrl CrossRef PubMed 14. ↵ Chen , M. et al. Spectrally resolved super-resolution microscopy with wide-field spectral single-molecule localization microscopy . Nat. Commun . 11 , 4352 ( 2020 ). doi: 10.1038/s41467-020-18065-6 OpenUrl CrossRef PubMed 15. ↵ Jungmann , R. , Avendaño , M. S. , Woehrstein , J. B. et al. Multiplexed 3D cellular super-resolution imaging with DNA-PAINT and Exchange-PAINT . Nat. Methods 11 , 313 – 318 ( 2014 ). doi: 10.1038/nmeth.2835 OpenUrl CrossRef PubMed 16. ↵ Gwosch , K. C. et al. MINFLUX nanoscopy delivers 3D multicolor nanometer resolution in cells . Nano Lett . 19 , 2641 – 2646 ( 2019 ). doi: 10.1021/acs.nanolett.9b00508 OpenUrl CrossRef 17. ↵ Stringer , C. et al. Cellpose: a generalist algorithm for cellular segmentation . Nat. Methods 18 , 100 – 106 ( 2021 ). doi: 10.1038/s41592-020-01018-x OpenUrl CrossRef PubMed 18. ↵ Pachitariu , M. et al. Cellpose 2.0: how to train your own cell segmenter . Nat. Methods 20 , 852 – 860 ( 2023 ). doi: 10.1038/s41592-023-01897-0 OpenUrl CrossRef 19. ↵ Kirillov , A. , et al. Segment Anything. arXiv 2304.02643 ( 2023 ). 20. ↵ Greenwald , N. F. et al. Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep learning . Nat. Biotechnol . 40 , 555 – 565 ( 2022 ). doi: 10.1038/s41587-021-01094-0 OpenUrl CrossRef 21. ↵ Ghoochani , A. et al. Proteomic insights into lysosomal biology in health and disease . bioRxiv 2024.10.14.618295 ( 2024 ). doi: 10.1101/2024.10.14.618295 OpenUrl Abstract / FREE Full Text 22. ↵ Di Sanzo , S. , Spengler , K. , Leheis , A. et al. Mapping protein carboxymethylation sites provides insights into their role in proteostasis and cell proliferation . Nat. Commun . 12 , 6743 ( 2021 ). doi: 10.1038/s41467-021-26982-6 OpenUrl CrossRef PubMed 23. ↵ Sigal , Y. M. et al. Visualizing and discovering cellular structures with super-resolution microscopy . Science 361 , 880 – 887 ( 2018 ). doi: 10.1126/science.aau1044 OpenUrl Abstract / FREE Full Text 24. ↵ Haralick , R. M. , Shanmugam , K. & Dinstein , I . Textural features for image classification . IEEE Trans. Syst. Man Cybern. SMC - 3 , 610 – 621 ( 1973 ). doi: 10.1109/TSMC.1973.4309314 OpenUrl CrossRef PubMed 25. ↵ Ronneberger , O. , Fischer , P. & Brox , T . U-Net: convolutional networks for biomedical image segmentation . Med. Image Comput. Comput. Assist. Interv . 9351 , 234 – 241 ( 2015 ). doi: 10.1007/978-3-319-24574-4 \_28 OpenUrl CrossRef 26. ↵ Shorten , C. & Khoshgoftaar , T. M . A survey on image data augmentation . J. Big Data 6 , 60 ( 2019 ). doi: 10.1186/s40537-019-0197-0 OpenUrl CrossRef 27. ↵ Franke , C. , Sauer , M. & van de Linde , S . Photometry unlocks 3D information from 2D localization microscopy data . Nat. Methods 14 , 41 – 44 ( 2017 ). doi: 10.1038/nmeth.4073 OpenUrl CrossRef PubMed 28. ↵ Franke , C. , Repnik , U. , Segeletz , S. et al. Correlative single-molecule localization microscopy and electron tomography reveals endosome nanoscale domains . Traffic 20 , 601 – 617 ( 2019 ). doi: 10.1111/tra.12671 OpenUrl CrossRef PubMed 29. ↵ Wang , Z. , Bovik , A. C. , Sheikh , H. R. & Simoncelli , E. P . Image quality assessment: from error visibility to structural similarity . IEEE Trans. Image Process . 13 , 600 – 612 ( 2004 ). doi: 10.1109/TIP.2003.819861 OpenUrl CrossRef PubMed Web of Science 30. ↵ Wang , Z. , Simoncelli , E. P. & Bovik , A. C . Multiscale structural similarity for image quality assessment . Proc. Asilomar Conf. Signals 2 , 1398 – 1402 ( 2003 ). doi: 10.1109/ACSSC.2003.1292216 OpenUrl CrossRef 31. ↵ Hore , A. & Ziou , D . Image quality metrics: PSNR vs . SSIM. Proc. Int. Conf. Pattern Recognit . 2366–2369 ( 2010 ). doi: 10.1109/ICPR.2010.579 OpenUrl CrossRef 32. ↵ Ostersehlt , L. M. , Jans , D. C. , Wittek , A. , et al. DNA-PAINT MINFLUX nanoscopy . Nat. Methods 19 , 1072 – 1075 ( 2022 ). doi: 10.1038/s41592-022-01577-1 OpenUrl CrossRef PubMed 33. ↵ Vasquez , R. J. , Howell , B. , Yvon , A. M. , Wadsworth , P. & Cassimeris , L . Nocodazole-induced microtubule disassembly in mammalian cells . J. Cell Sci . 110 , 1603 – 1614 ( 1997 ). PMID: 9224780 OpenUrl Abstract / FREE Full Text 34. ↵ Schnell , U. et al. Glyoxal as an alternative fixative to formaldehyde in immunostaining and super-resolution microscopy . Nat. Commun . 12 , 4844 ( 2021 ). doi: 10.1038/s41467-021-25122-6 OpenUrl CrossRef PubMed 35. ↵ Ghoochani , A. et al. Lysosomal proteomics uncovers new regulators of cellular homeostasis . Front. Cell Dev. Biol . 12 , 1484329 ( 2024 ). doi: 10.3389/fcell.2024.1484329 OpenUrl CrossRef 36. ↵ Rivenson , Y. , Wang , H. , Wei , Z. et al. Virtual histological staining of unlabelled tissue-autofluorescence images via deep learning. Nat . Biomed. Eng . 3 , 466 – 477 ( 2019 ). doi: 10.1038/s41551-019-0362-y OpenUrl CrossRef PubMed 37. ↵ Buslaev , A. et al. Albumentations: fast and flexible image augmentations . Information 11 , 125 ( 2020 ). doi: 10.3390/info11020125 OpenUrl CrossRef 38. ↵ Coelho , L. P . Mahotas: open source software for scriptable computer vision . J. Open Res. Softw . 1 , e3 ( 2013 ). doi: 10.5334/jors.ac OpenUrl CrossRef 39. ↵ Paszke , A. , et al. PyTorch: an imperative style, high-performance deep learning library . Adv. Neural Inf. Process. Syst. 32 , 8024 – 8035 ( 2019 ). doi: 10.48550/arXiv.1912.01703 OpenUrl CrossRef 40. ↵ van der Walt , S. et al. scikit-image: image processing in Python . PeerJ 2 , e453 ( 2014 ). doi: 10.7717/peerj.453 OpenUrl CrossRef PubMed 41. ↵ Wolter , S. et al. Real-time computation of subdiffraction-resolution fluorescence images . Nat. Methods 9 , 1040 – 1041 ( 2012 ). doi: 10.1038/nmeth.2224 OpenUrl CrossRef PubMed Web of Science 42. ↵ Ouyang , W. , Bai , J. , Singh , M. K. et al. ShareLoc — an open platform for sharing localization microscopy data . Nat. Methods 19 , 1331 – 1333 ( 2022 ). doi: 10.1038/s41592-022-01659-0 OpenUrl CrossRef PubMed 43. ↵ Spengler , K. , Kryeziu , N. , Große , S. , Mosig , A. S. & Heller , R . VEGF triggers transient induction of autophagy in endothelial cells via AMPKα1 . Cells 9 , 687 ( 2020 ). doi: 10.3390/cells9030687 OpenUrl CrossRef 44. ↵ Franke , C. et al. Approach to map nanotopography of cell surface receptors. Commun . Biol . 5 , 218 ( 2022 ). doi: 10.1038/s42003-022-03152-y OpenUrl CrossRef PubMed 45. ↵ Ounkomol , C. , Seshamani , S. , Maleckar , M. M. , Collman , F. & Johnson , G. R . Label-free prediction of fluorescence images from transmitted-light microscopy . Cell 173 , 1552 – 1565 ( 2018 ). doi: 10.1016/j.cell.2018.04.053 OpenUrl CrossRef PubMed 46. ↵ Dosovitskiy , A. , et al. An image is worth 16×16 words: transformers for image recognition at scale . arXiv 2010.11929 ( 2020 ). doi: 10.48550/arXiv.2010.11929 OpenUrl CrossRef 47. ↵ Paramasivam , P. et al. Endosomal escape of delivered mRNA from endosomal compartments limits efficacy . J. Cell Biol . 221 , e202110137 ( 2022 ). doi: 10.1083/jcb.202110137 OpenUrl CrossRef 48. ↵ Holick , C. T. et al. Poly(2-ethyl-2-oxazoline) (POx) as poly(ethylene glycol) (PEG)-lipid substitute for lipid nanoparticle formulations . Small 21 , 2411354 ( 2025 ). doi: 10.1002/smll.202411354 OpenUrl CrossRef 49. ↵ Tinevez , J.-Y. et al. TrackMate: An open and extensible platform for single-particle tracking . Methods 115 , 80 – 90 ( 2017 ). doi: 10.1016/j.ymeth.2016.09.016 OpenUrl CrossRef PubMed 50. ↵ Vogler , B.T.L. , et al. , Parameter Optimization for Iterative MINFLUX Microscopy enabled Single Particle Tracking . bioRxiv 2025.06.10.658837; doi: 10.1101/2025.06.10.658837 OpenUrl Abstract / FREE Full Text 51. ↵ Gentsch , G. J. , Vogler , B. T. L. , & Franke , C . ( 2025 ). NanTex SRM Dataset — SMLM I (ShareLoc) [Data set] . Zenodo . doi: 10.5281/zenodo.17120603 OpenUrl CrossRef 52. Franke , C. , Gentsch , G. J. , & Vogler , B. T. L . ( 2025 ). NanTex SRM Dataset — SMLM II (Olympus IX-71 dSTORM) [Data set] . Zenodo . doi: 10.5281/zenodo.17120791 OpenUrl CrossRef 53. Gentsch , G. J. , Vogler , B. T. L. , & Franke , C . ( 2025 ). NanTex SRM Dataset — SIM (ZEISS Elyra 7 lattice SIM) [Data set] . Zenodo . doi: 10.5281/zenodo.17121085 OpenUrl CrossRef 54. Gentsch , G. J. , Vogler , B. T. L. , & Franke , C . ( 2025 ). NanTex SRM Dataset — STED 1 (Leica Infinity Line) [Data set] . Zenodo . doi: 10.5281/zenodo.17121094 OpenUrl CrossRef 55. Hildebrandt , F. , Gentsch , G. J. , Vogler , B. T. L. , & Franke , C . ( 2025 ). NanTex SRM Dataset — STED 2 (Abberior STEDYCON) [Data set] . Zenodo . doi: 10.5281/zenodo.17121098 OpenUrl CrossRef 56. ↵ Heiby , J. , Durso , W. , Gentsch , G. J. , Vogler , B. T. L. , & Franke , C . ( 2025 ). NanTex SRM Dataset — Airyscan (Zeiss LSM 880) [Data set] . Zenodo . doi: 10.5281/zenodo.17121104 OpenUrl CrossRef View the discussion thread. Back to top Previous Next Posted September 18, 2025. Download PDF Supplementary Material Data/Code Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following NanTex enables computational multiplexing and phenotyping of organelles across super-resolution modalities Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share NanTex enables computational multiplexing and phenotyping of organelles across super-resolution modalities Bela T.L. Vogler , Gregor J. Gentsch , Pablo Carravilla , Dominic A. Helmerich , Julia C. Heiby , Katharina Reglinski , William Durso , Teresa Klein , Felix Hildebrandt , Rohama Zahid , Katrin Spengler , Regine Heller , Christoph Kaether , Markus Sauer , Christian Eggeling , Christian Franke bioRxiv 2025.09.16.676574; doi: https://doi.org/10.1101/2025.09.16.676574 Share This Article: Copy Citation Tools NanTex enables computational multiplexing and phenotyping of organelles across super-resolution modalities Bela T.L. Vogler , Gregor J. Gentsch , Pablo Carravilla , Dominic A. Helmerich , Julia C. Heiby , Katharina Reglinski , William Durso , Teresa Klein , Felix Hildebrandt , Rohama Zahid , Katrin Spengler , Regine Heller , Christoph Kaether , Markus Sauer , Christian Eggeling , Christian Franke bioRxiv 2025.09.16.676574; doi: https://doi.org/10.1101/2025.09.16.676574 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Biophysics Subject Areas All Articles Animal Behavior and Cognition (7640) Biochemistry (17706) Bioengineering (13902) Bioinformatics (41978) Biophysics (21465) Cancer Biology (18611) Cell Biology (25528) Clinical Trials (138) Developmental Biology (13387) Ecology (19920) Epidemiology (2067) Evolutionary Biology (24332) Genetics (15615) Genomics (22519) Immunology (17747) Microbiology (40424) Molecular Biology (17194) Neuroscience (88662) Paleontology (667) Pathology (2839) Pharmacology and Toxicology (4827) Physiology (7650) Plant Biology (15160) Scientific Communication and Education (2046) Synthetic Biology (4302) Systems Biology (9826) Zoology (2271)