{"paper_id":"121826df-df9e-4e76-bc4a-8e084c0f7b47","body_text":"An Interpretable 3D Bag-Of-Visual-Words Pipeline for \nVolumetric Microscopy Classification \nAnna E. Pittman*, Kirby R. Campbell*, Christophe Laumonnerie*, David J. Solecki* \n*Neuronal Cell Biology Division, Department of Developmental Neurobiology, St. Jude \nChildren’s Research Hospital, 262 Danny Thomas Place, Memphis, TN 38104, USA \n david.solecki@stjude.org \nCode repository: https://github.com/PittmanAEP/3DBoVW \nAbstract \nFluorescence microscopy increasingly produces complex volumetric datasets \nwhose biologically meaningful diYerences are diYicult to capture with hand-crafted \nmeasurements, especially when signal is distributed across three-dimensional space. \nHere, we present an interpretable 3D Bag-of-Visual-Words (BoVW) pipeline for \nclassification and analysis of volumetric microscopy data. The framework detects \nmultiscale local keypoints, computes rotationally robust 3D gradient-based descriptors, \nand aggregates them into image-level visual-word representations. These features are then \nused for low-dimensional visualization and logistic regression classification, while model \nweights are mapped back to the original volumes to generate attention maps that localize \ndiscriminative structures. We applied the pipeline to two cerebellar granule neuron \ndatasets spanning both ideal and non-ideal imaging conditions. In a near-isotropic lattice \nlight-sheet dataset of chromatin organization, the method separated control and NIPBL \nloss-of-function nuclei and supported accurate classification, with strongest performance \nin the facultative heterochromatin and H3.3 channels. Attention mapping and downstream \nconnected-component and Haralick analyses revealed that loss-of-function nuclei \ncontained more fragmented high-attention regions and smoother, more homogeneous \nchromatin-associated textures than controls. We then evaluated the same framework on \nan anisotropic confocal timelapse dataset of receptor clustering in dense neuronal \ncultures, where single-cell segmentation was impractical. Despite these challenges, the \nrepresentation captured the expected ligand-driven clustering response and resolved \nsubtler diYerences associated with a polarity protein overexpression. Together, these \nresults establish a simple, interpretable, and broadly applicable framework for extracting \nbiologically meaningful structure from volumetric microscopy datasets while preserving \nnative 3D context. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\n1. Introduction \nFluorescence microscopy has transformed the study of biological systems, enabling \ndirect visualization of subcellular structures and dynamics. However, the increasingly \nsophisticated imaging strategies that drive biological discovery also produce datasets that \nare large, heterogeneous, and diYicult to analyze1. 2D datasets, such as whole-slide \nhistology images, can be challenging due to their large individual sizes, while high-\nthroughput experiments generate tens of thousands of smaller fields of view that rapidly \noverwhelm manual and semi-manual workflows. These challenges are further \ncompounded in volumetric datasets and in modalities that probe intrinsically three-\ndimensional structures, for which open source, general-purpose analysis tools are \ncomparatively scarce. \nA common approach for analyzing 3D images is to treat them as stacks of 2D slices. \nThis is often done to reduce computational demands, simplify visualization, and mitigate \nanisotropy artifacts that are common along the z-dimension in microscopy data. For \nexample, open-source tools such as Cellpose typically segment volumetric data by running \n2D models on individual planes and then reconstructing 3D objects from the slice-wise \npredictions2. While this approach has been successful and eYicient for segmentation, it is \nless suitable for tasks that depend on genuine 3D structure, such as feature extraction and \nclassification of nuanced fluorescence signal. Down sampling or flattening the data to 2D \ncan discard important information about spatial relationships and topology. There is a need \nfor methods that operate directly on volumetric images, generating rich, interpretable \nfeature representations without reducing the data to lower dimensions prematurely. \nInterpreting volumetric fluorescence data is inherently diYicult for human observers. \nMeaningful visualization typically requires reducing 3D information to 2D slices or \nprojections, but this collapse discards spatial context and can obscure relationships that \nexist only in three dimensions. As a result, hand-crafted analyses often default to coarse \nsummary metrics (e.g., overall size or shape anisotropy) that capture only a small fraction \nof the available signal. There is therefore a need for methods that preserve native \ndimensionality while providing interpretable summaries that an expert can use to refine—\nor challenge—existing biological interpretations. \nHere, we present a 3D analysis framework based on a Bag-of-Visual-Words (BoVW) \npipeline with a custom 3D feature extractor. Local 3D fluorescence signal structure is \nsummarized using rotationally robust descriptors, and these descriptors are aggregated \ninto an image-level BoVW representation. The resulting feature vectors are used to train a \nlogistic regression classifier that distinguishes between relevant conditions and, via its \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nlearned weights, highlights intensity patterns and spatial motifs that are most informative \nfor classification. \nWe focus on cerebellar granule neurons (CGNs), an ideal system for method \ndevelopment because their maturation is tightly regulated and accompanied by \nstereotyped, well-characterized morphological transitions. In the developing cerebellum, \nCGN progenitors reside in the external granule layer (EGL) where they undergo clonal \nexpansion, then exit the cell cycle, diYerentiate, and migrate inward to form the mature \ngranule layer3. This progenitor-to-diYerentiated transition is marked by two hallmark events \ncentral to our study: extensive chromatin reorganization during diYerentiation4, and the \nproper direction of migratory behavior required for timely exit from the EGL5. Disrupting \neither process has major consequences—prolonged proliferation in the EGL is associated \nwith medulloblastoma6, while premature cell-cycle exit can reduce CGN number and \naYect proper lamination of the cerebellum, which has been linked to neurodevelopmental \ndisorders7,8. \nThese behaviors can be read out with interpretable fluorescence markers, including \nchromatin markers that report nuclear organization and cell-surface guidance receptors \nthat report migration-related signaling necessary for proper migratory direction and \nprogression out of the EGL. Our 3D nuclear dataset targets chromatin organization using \nfour diYerent chromatin markers imaged with lattice light-sheet microscopy, whose near-\nisotropic resolution, a rarity in the microscopy world, makes the volumetric structure well \nsuited for true 3D analysis. The variety of chromatin markers allows us to investigate which \nmarkers contain condition-specific changes and gives us a deep understanding of the \nunderlying biology. \nOur 3D timelapse dataset captures receptor clustering dynamics before and after \nligand addition, in both control cells and cells with perturbation of a polarity pathway. This \ndesign provides two complementary scales of phenotypic change: a strong, visually \nstriking ligand-driven clustering response and a subtler modulation associated with the \ngenetic perturbation. The dataset also reflects common constraints of dense neuronal \ncultures—extensive overlap and intermingling processes make single-cell segmentation \nand cropping impractical. Additionally, it also reflects common microscopy constraints: \nconfocal acquisition yields pronounced axial anisotropy and sampling steps chosen for \nspeed lead to sparse axial sampling, compromising the resolution in the axial dimension. \nTogether, these factors create a stringent test of algorithmic robustness under realistic, \nnon-ideal imaging conditions that deviate from clean, isotropic volumes. \nA central design principle throughout this work is minimal complexity. We aim to build \nthe simplest pipeline that adequately addresses a given dataset, adding additional layers \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nof complexity only when the data requires them. The BoVW backbone was chosen in order \nto pursue the most transparent simple pipeline; we avoid layering on neural networks and \nopt instead for a simple logistic regression model to dig deeper into the data. \nThe full 3D image analysis pipeline—rotationally robust local descriptors, BoVW \naggregation, and a linear classifier with mappable weights—enables condition separation \nwhile retaining 3D context and interpretability in both the ideal (isotropic single cell crops) \nand more realistic (many overlapping cells and anisotropic acquisition) use cases. \nUltimately, it’s a broadly applicable framework designed to generalize across a variety of \nfluorescent datasets. \n2. Related Works \nBag-of-Visual-Words (BoVW), inspired by bag-of-words models in text9, represents an \nimage by quantizing local intensity descriptors into a dictionary of “visual words” and \nsummarizing each image as a histogram of those words10. These histograms form compact \nfeature vectors that support tasks such as image retrieval, clustering, and classification. \nThey are often visualized using dimensionality reduction methods such as principal \ncomponent analysis (PCA) or uniform manifold approximation and projection (UMAP). Our \nmethod adopts the BoVW paradigm for its interpretability and modularity, allowing us to \ndesign the local descriptor and sampling strategy to properly reflect genuinely 3D \nstructures from biological fluorescent volumetric imaging whereas standard BoVW \npipelines often rely on 2D inputs, or 3D images that have been broken down into 2D slices. \nBoVW has been applied to biomedical imaging in multiple contexts, particularly in 2D \npathology where local tissue morphologies recur across large cohorts. For instance, Cruz-\nRoa et al. use BoVW-style representations for histology analysis, demonstrating that \ndictionaries of local patterns can capture meaningful visual content in H&E imagery11. \nRelated unsupervised “visual phenotype” approaches use dictionaries to discover disease-\nassociated morphologies and link them to clinical or molecular endpoints. Powell et al. \nconstruct image-derived visual words from TCGA glioma slides to predict survival and \nassociate predictive phenotypes with signaling activity, illustrating how dictionary-based \nrepresentations can bridge image patterns to biological hypotheses12. Similarly, Lee et al. \npropose an unsupervised bag-of-words framework on renal biopsy whole-slide images to \npredict functional outcomes13, while noting that missing spatial information can contribute \nto misclassifications—highlighting a recurring tradeoY between compactness and spatial \nspecificity. In contrast to these primarily 2D, tissue-scale settings, our focus is cellular or \nsubcellular organization in volumetric fluorescence microscopy, where discriminative cues \narise from 3D structures within the images. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nBoVW has also been applied to CT volumetric data; Feulner et al. apply a bag-of-words \nrepresentation to CT volumes for body-part estimation, demonstrating that local \ndescriptors can compactly represent 3D scans14. However, that analysis broke down the \n3D image into 2D slices before feature extraction. While that was suYicient for the analysis \nof CT volumes, volumetric fluorescence microscopy generates complex 3D structures that \nmust be analyzed in their native dimensionality. Our approach addresses these \nmicroscopy-specific constraints by extracting local 3D descriptors within each image and \naggregating them into an image-level BoVW vector used for classification and \ninterpretation. \n A key challenge in adapting BoVW to fluorescence volumes is handling arbitrary \norientation of subcellular structures. Prior work on rotation robustness typically relies on \ncanonical orientation assignment, orientation pooling, or expressing gradients in a relative \nreference frame. Our descriptor follows the latter strategy: we adapt a rotation-invariant \nhistogram of oriented gradients (HOG) variant that replaces absolute gradient orientation \nwith locally relative measurements15 and extend the formulation to 3D microscopy \nvolumes (Methods). \nTaken together, prior work shows that dictionary-based representations can summarize \ncomplex imagery and support downstream prediction, including in biomedical and even \nvolumetric settings. At the same time, standard BoVW pipelines rely on 2D inputs either \nfrom 2D images or 2D slices of 3D volumes. This motivates our framework: a BoVW pipeline \ndesigned for volumetric fluorescence signal, using rotationally robust 3D local descriptors \nand a linear classifier whose weights can be mapped back to informative intensity patterns \nand spatial motifs. \n3. Datasets and Problem Setup \nProblem definition \nTraditional analyses rely on an expert observer to recognize global trends and then \ndesign concrete measurements that capture those trends. In some cases, this is \nstraightforward—for example, measuring cell velocity for a motility phenotype. However, as \nthe underlying signal becomes more nuanced with less obvious trends, it is increasingly \ndiYicult to identify a single, hand-crafted quantity that adequately summarizes the \nphenotype. This is especially true for structural properties such as chromatin organization, \nwhere the phenotype might be visually apparent but not easily reducible to a small set of \nintuitive descriptors. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nA further challenge arises when the signal is inherently three-dimensional. For a human \nobserver, it is extremely diYicult to meaningfully interpret volumetric structure without \ncollapsing the data into 2D slices or projections. Although these projections can be \ninformative—and may even be suYicient in some settings—they capture only a fraction of \nthe information present in the volume, leaving much of the signal unexamined. \nComputational methods do not face the same constraint. Models can operate directly \non the data in its native dimensionality, enabling them to interrogate the full 3D structure \nrather than a compressed projection. This makes it possible to leverage volumetric \ninformation that is diYicult to access by eye and to reveal trends that might otherwise \nremain hidden. \nOur goal is to complement careful expert inspection with a flexible, data-driven \nrepresentation that can capture phenotypic diYerences without requiring an a priori choice \nof a relevant measurement. By operating directly on volumetric data in its native \ndimensionality, this approach preserves 3D spatial context and enables the analysis to \nleverage structural information that is diYicult to articulate or quantify by hand. To this end, \nwe build on a Bag of Visual Words (BoVW) backbone and develop a pipeline that (i) isolates \nimage structures that diYer between conditions and (ii) provides interpretable summaries \nof those structures. \n \nFigure 1 Dataset Acquisitions: Cartoon illustrations show microscope setups and outputs. Spinning disk confocal \nmicroscopes image fast but su median nuclear volume) and smaller control nuclei (< \nmedian nuclear volume), with volumes derived from the 3D nuclear segmentations (for \nanalysis on non-stratified datasets see the Supplementary Information). NIPBL knockdown \nled to increased nuclear volume; enriching for larger nuclei therefore enriched for cells \nmost aYected by the perturbation. For the timelapse dataset the dictionary was learned \nand applied to the entire dataset. \n4.3 Logistic Regression Classifier \nWe trained a logistic regression classifier on the normalized BoVW vectors to quantify \nthe extent to which the representation captured the phenotype. The input features were the \n200-dimensional normalized frequency vectors, and the target labels were the known \nexperimental conditions (i.e. control vs. LOF). Logistic regression was used with standard \nL2 regularization. For the nuclear images a two-class model was trained (control vs LOF) \nand for the membrane receptor images a four-class model was trained to account for \ncontrol and overexpression pre and post addition of ligand. Model performance for the two-\nclass model was summarized by the area under the receiver-operating characteristic curve \n(AUC-ROC), computed from the predicted probabilities across cells. \nThe fitted model provided a coeYicient for each visual word, indicating the degree and \ndirection with which that word contributed to predicting the correct class. Words that \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\npushed towards the correct class were given positive weights, words which pushed \ntowards the incorrect class (confounding information) were given negative weights. \n4.4 Attention Map Construction \nTo localize discriminative information within each image, we decomposed the linear \nclassifier’s image-level decision value into patch-level contributions. For each image, we \nfirst reconstructed the TF–IDF and L2-normalized visual word vector used during logistic \nregression training and computed the contribution of each word to the decision value as \n𝑤!𝑥!, where 𝑤!is the learned logistic regression coeYicient and 𝑥!is the normalized TF–IDF \nweight of word 𝑘. We then distributed each word’s contribution back to individual patches \nin proportion to the absolute usage of that word in each patch, such that patches \ncontaining more of a given word received a larger share of that word’s contribution. \nSumming these contributions over all words yielded a scalar attention score for each \npatch. By construction, the sum of patch scores equals the image-level classifier score (up \nto numerical precision), providing a faithful spatial decomposition of the linear decision \nfunction. \nFor intuition, consider a visual word that contributes 10 arbitrary units to the classifier \nscore for a given nucleus. If this word is distributed equally across 10 patches (each patch \ncarrying the same absolute usage of that word), then each of those patches receives 1 unit \nof that contribution, and their patch-level contributions sum to the original 10 units. \nPatch-level attention scores were then mapped back onto the 3D image volume using \nthe original patch locations, producing a volumetric attention map for each nucleus. \nAttention maps were rescaled on a per-nucleus basis (e.g., min–max normalization) to \nfacilitate visualization and comparison across images. \n \n4.5 Nuclear Blob and Texture Analysis \nTo quantify the spatial organization of high-attention regions, the attention maps were \nthresholded to define a binary mask of “high-attention” voxels. A single global threshold \nwas chosen empirically based on the distribution of attention scores, and this threshold \nwas applied uniformly across nuclei. \nWithin each binary attention map, connected-component labeling in 3D was used to \nidentify isolated high-attention “blobs. ” For each nucleus, we recorded (i) the number of \nblobs and (ii) the volume (voxel count) of each blob. These measurements were compared \nbetween conditions to assess diYerences in the fragmentation and spatial extent of the \nmost informative regions. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nTo characterize the local intensity patterns within high-attention regions, we computed \nHaralick texture features on 3D patches. For each image, voxel intensities within the \nnuclear mask were clipped to the 1st–99th percentiles, normalized to [0,1], and quantized \nto 32 gray levels, with background voxels set to zero. High-attention patches (as defined by \nthe attention threshold) were extracted from these quantized volumes, and Haralick \nfeatures were computed using the Mahotas Python package (mh.features.haralick, \ndistance = 1, ignore_zeros = True). This function constructs gray-level co-occurrence \nstatistics over multiple directions and, with return_mean = True, returns the standard set \nof 13 Haralick features averaged over directions for each patch. \nEach high-attention patch was summarized by a 13-dimensional Haralick feature \nvector. We analyzed these features at two levels. First, for each nucleus we computed the \nmean value of each Haralick feature across its high-attention patches, yielding one 13-\ndimensional summary vector per image. To examine the full distribution of texture values \nacross high-attention regions, we pooled patch-level feature values across all images \nwithin a condition and estimated kernel density distributions (KDEs) for individual Haralick \nfeatures. \n5. Experiments and Results \n5.1 Experimental Setup \nWe evaluated the proposed BoVW framework on two microscopy datasets (i) a 3D \nchromatin architecture dataset used to test whether the 3D BoVW representation supports \nboth condition discrimination and localization of discriminative structure within nuclei and \n(ii) a 3D timelapse receptor clustering dataset, which tests robustness on a realistic \nacquisition where images are anisotropic and cannot be reduced to single-cell crops. \n3D chromatin dataset (segmented single-nucleus volumes) \nIn the chromatin dataset, each sample corresponded to a segmented and cropped \nnucleus, enabling direct cell-level comparison across conditions (control vs. NIPBL \nknockdown). The primary analysis focused on a custom designed H3K27me3 H2AK119ub \nbiosensor for facultative heterochromatin (manuscript in preparation), and the same \nworkflow was also applied to additional markers (Hoechst, H3.3, and CTCF). Each probe \nwas chosen to focus on diYerent types of chromatin. The facultative heterochromatin \nprobe accumulates in facultative heterochromatin, which is heterochromatin whose \naccessibility can be developmentally regulated. H3.3 is associated with transcriptionally \nactive chromatin, CTCF is enriched at topologically associated domain boundaries, and \nHoechst binds the minor groove preferentially in AT-rich regions. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nBecause biological variability in knockdown extent introduced heterogeneity, we \nperformed downstream analyses on size-stratified subsets (larger LOF nuclei and smaller \ncontrol nuclei, split at the median nuclear volume) to enrich for cells most aYected by the \nperturbation while reducing overlap driven by size-related variability. \nTo quantify the spatial organization of discriminative regions, attention maps were \nthresholded to define high-attention voxels, followed by 3D connected-component \nanalysis to measure the number and size distribution of high-attention “blobs” per nucleus. \nTo further characterize local structure within these regions, we computed 3D Haralick \ntexture features on high-attention patches and compared feature distributions between \nconditions. \n3D timelapse receptor dataset (frame-wise field-of-view volumes) \nFor the receptor clustering timelapse dataset, sequences were split into individual \ntimepoints, yielding a series of 3D frames treated as independent samples for \nrepresentation learning and classification. Unlike the chromatin dataset, individual cell \nsegmentation and cropping were infeasible due to high cell density and overlap, and the \nconfocal acquisition produced pronounced axial anisotropy, creating a challenging “non-\nideal” test case. Labels captured both a strong ligand-driven clustering response (pre- vs \npost-ligand) and a subtler modulation associated with pathway perturbation. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\n \n5.2 3D Chromatin Architecture Results \nPrevious work in the lab attempted to distinguish control and NIPBL loss-of-function \n(LOF) conditions using hand-crafted features derived from expert visual inspection. Here, \nwe sought to identify similar or additional trends using an unsupervised Bag of Visual \nWords approach adapted to 3D data. The normalized frequency vectors exhibited \nclustering behavior, (Figure 3B), that separated control and NIPBL LOF cells. Some overlap \nFigure 3 Facultative Heterochromatin biosensor results. A: 3D renderings of the biosensor channel overlaid with the \npositive attention maps to highlight areas of high attention. B: UMAP results from the normalized image embedding \nvectors C: A logistic regression model was trained on the image labels and normalized embedding vectors. The model \ntrained with an AUC-ROC of 0.979. D: Selected kernel density estimate graphs from various Haralick Features. The KDEs \nwere generated by analyzing patch-level Haralick feature values and summing over the two conditions. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nbetween conditions as seen, consistent with biological variability in the degree of \nknockdown, diYerences in diYerentiation status, and other sources of heterogeneity. \nHowever, when size-based stratification was applied, the separation between control and \nLOF clusters became more pronounced (SI Figure 4). \nThis clustering, the separation between control and loss of function seen in the UMAP , \nindicated that the algorithm had captured biologically relevant intensity patterns—i.e., \nsome combination of visual words was suYicient to discriminate between control and LOF \nnuclei. However, the representation itself does not, a priori, reveal which features drive the \nseparation. Although one can inspect normalized frequency vectors to identify words that \nare more prevalent in one condition than the other, the soft assignment (linear \ncombinations of many words per patch) makes it diYicult to intuitively understand what \ncommon words “look like” in terms of raw image structure. In principle, two patches might \ndiYer because one has a high coeYicient for word 1 but low for word 2, whereas another \nhas high values for both words, and so on across many words. \nTo understand which words carried the most discriminative information, we trained a \nlogistic regression (LR) model on the normalized frequency vectors using the known \ncondition labels (control vs. NIPBL LOF). The LR model achieved an area under the \nreceiver-operating characteristic curve (AUC-ROC) of 0.979, confirming that the frequency \nvectors alone were suYicient for accurate classification (Figure 3C). The LR model also \nprovided a weight (coeYicient) for each visual word, quantifying the extent to which that \nword contributed to predicting the LOF versus control class. Words that predicted the \ncorrect class were assigned positive weights, words that predicted the incorrect class were \nassigned negative weights. \nWe next asked which regions within each nucleus contributed most strongly to this \nclassification. Using the logistic regression coeYicients as weights for the visual words, we \ncomputed an “attention” score for each patch and mapped these scores back into the 3D \nvolume to generate attention maps (Figure 3A). To describe these high-attention regions \nquantitatively, we first thresholded the attention maps to define “high-attention” voxels and \nperformed 3D particle analysis. These maps revealed distinct spatial patterns in control \nversus LOF nuclei. Control nuclei tended to contain a small number of relatively large, \ncontiguous high-attention regions that spanned substantial portions of the nuclear volume. \nIn contrast, LOF nuclei displayed multiple smaller, more fragmented high-attention regions \nthat were spatially restricted, indicating a shift from broad, extended domains to more \nfragmented patterns of informative signal (SI Figure 6). \nFinally, we examined the local intensity patterns within these high-attention regions \nusing 3D Haralick texture features. Each high-attention patch was summarized by the \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nstandard set of 13 Haralick features derived from gray-level co-occurrence statistics. \nComparing feature distributions between conditions revealed consistent trends: LOF \nnuclei exhibited high-attention regions with more smoothly varying textures, greater \nhomogeneity, increased structural regularity, and stronger statistical coupling of \nneighboring intensities (Figure 3D). Collectively, these metrics indicate that, in the regions \nmost informative for classification, LOF nuclei display smoother, less punctate chromatin-\nassociated signal than control nuclei. Thus, the 3D BoVW framework, coupled with a \nsimple linear classifier and texture analysis, not only separates control and NIPBL LOF \nnuclei but also points to specific, biologically interpretable diYerences in chromatin texture \nunderlying this separation. \nThe primary analysis above focused on the facultative heterochromatin probe, which is \nenriched in facultative heterochromatin, but the same pipeline was applied independently \nto the other chromatin markers in the dataset. Hoechst—a commonly used DNA dye that \nbinds the minor groove with preference for AT-rich regions—produces the familiar pattern \nof bright, compact heterochromatin and dimmer euchromatin. In this dataset, it was the \nleast informative channel: UMAP embeddings of the BoVW frequency vectors showed no \nseparation between control and LOF cells, and the logistic regression model performed \npoorly (AUC–ROC = 0.716, SI Figure 1). \nCTCF labeling, which marks genomic sites involved in higher-order chromatin \norganization and is enriched at loop anchors and topologically associating domain (TAD) \nboundaries, showed only weak structure in the embedding. UMAP again revealed minimal \nseparation, although classification improved modestly relative to Hoechst (AUC–ROC = \n0.817), suggesting subtle but detectable condition-dependent diYerences (SI Figure 2). \nFinally, H3.3—associated with transcriptionally active chromatin—carried a strong \ncondition-sensitive signal. This channel produced the second-best separation in UMAP \n(after the facultative heterochromatin probe) and supported robust classifier performance \n(AUC–ROC = 0.950), consistent with substantial reorganization of features linked to active \nchromatin under the perturbation (SI Figure 3). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\n \n5.3 3D Timelapse Results \nThe receptor clustering timelapse dataset provides a stringent test case because it \ndeparts from the “clean” single-cell setting used for the nuclear volumes. Each 3D frame \ncorresponds to an entire field of view containing densely packed neurons with substantial \noverlap. In addition, confocal acquisition produces pronounced axial anisotropy, and each \nfield of view contains unavoidable biological heterogeneity (cells at diYerent diYerentiation \nstates and variable expression levels). \nFigure 4 DCC Timelapse dataset results. A: representative images from the data showing CGN cultured with and without \nPard3 overexpression [scale bar = 10 µm] before and after addition of the ligand netrin. B: UMAP graph of image \nembedding vectors. C: Representative images overlaid with the positive attention maps. D: Confusion matrix after training \na 4-class logistic regression model. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nDespite these constraints, prior conventional analyses based on manual/algorithmic \nsegmentation of receptor clusters followed by size and morphology measurements \nshowed that ligand addition (netrin) increases the number of guidance receptor Deleted in \nColorectal Cancer (DCC) clusters and Partitioning defective 3 (Pard3) overexpression (OE) \npromotes larger and more persistent membrane-associated DCC clusters5. These \nestablished eYects make the dataset a useful benchmark for representation learning: the \ndistinction between pre- and post-netrin frames is visually apparent and serves as an \ninternal positive control, whereas diYerences between control and Pard3-overexpressing \nsamples are subtler and provide a more stringent test of sensitivity to perturbation-induced \nphenotypes. Because the dataset has been previously analyzed with handcrafted features \nand segmentation, we could also evaluate whether the 3D BoVW representation \nrecapitulates known results without explicit segmentation. \nAs expected, the most prominent separation identified by the algorithm, visualized with \nUMAP in Figure 4B, was between pre- and post-netrin images. Beyond this, the method \nalso distinguished control cells from Pard3-overexpressing cells, indicating that SIFT \nkeypoints encoded as visual words captured suYicient information to replicate, in an \nunsupervised manner, results that previously required manual segmentation and feature \nextraction. \nThe clustering results also reflected underlying biological complexity. The \noverexpression of Pard3 was achieved by the nucleofection of the Pard3 plasmid. As a \nresult, although all Pard3-overexpressing cells expressed more Pard3 than controls, the \nmagnitude of overexpression varied, producing overlap between conditions in both pre- \nand post-netrin images. \nTemporal heterogeneity was likewise evident in the low-dimensional embedding \n(UMAP). Post-netrin data points that clustered near pre-netrin points corresponded to the \nearliest time frames after netrin addition. Because timelapse imaging involved multiple \nfields of view, some cells were exposed to netrin slightly longer than others before the first \npost-stimulation frame was acquired, leading to a continuum of response states rather \nthan a sharp pre/post boundary. \nFeeding the normalized histograms into a four-class logistic regression model revealed \nwhich conditions were easier to classify and which conditions were more confusing. The \nmodel made no errors when classifying the control pre-netrin images and the Pard3 OE \npost-netrin images but had more diYiculty with the other two classes (Figure 4D). \nTogether, these results show that the 3D BoVW framework (i) captured the expected \nlarge-scale ligand-driven change (pre- vs post-netrin), (ii) resolved more subtle diYerences \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nassociated with polarity pathway perturbation (control vs Pard3 overexpression), and (iii) \nencoded both biological heterogeneity (variable expression) and temporal heterogeneity \n(variable time after stimulation). A remaining limitation is that analysis is performed at the \nlevel of whole fields of view rather than individual cells; given dense cultures and mixed \ndevelopmental states, field-level aggregation likely contributes to the partial overlap \nbetween conditions and sets an upper bound on separability in this dataset. \n6. Discussion \nThe design of this pipeline was guided by the goal of creating a broadly applicable tool \nfor analyzing complex microscopy datasets, rather than a method tailored exclusively to \none specific dataset. By operating directly on 3D image structure and learning data-driven \nvisual vocabularies, the framework is well suited to problems in which diYerences between \nconditions are visually apparent to an expert but resist simple, hand-crafted quantification. \nOne natural application is screening. In experiments with many genetic perturbations, \nthe pipeline can be used to generate image-level embeddings and quantify how strongly \neach condition deviates from a control distribution. Ranking conditions by their divergence \nfrom control can help prioritize perturbations for follow-up studies. A similar strategy could \nbe applied across imaging channels, for example to identify which marker exhibits the most \ndiscriminative phenotype in a multi-channel experiment. \nBeyond screening, the method can assist in interpreting which aspects of image \nstructure distinguish conditions. Because the downstream classifier operates on \ninterpretable Bag-of-Visual-Words features, its weights can be mapped back to image \nregions and intensity textures that drive classification. In this sense, the algorithm is \nintended to complement, not replace, expert evaluation: it provides quantitative \nsummaries and spatial attributions that can sharpen and guide biological interpretation. \nThe approach also has limitations. First, feature extraction is driven by SIFT-like \nkeypoints in the intensity landscape, making the method inherently intensity-based. In \ndatasets where the signal is extremely homogeneous, with few local maxima or minima, \nkeypoint detection—and therefore the BoVW representation—may fail to capture \nmeaningful variation. Second, in its current form the pipeline operates on a single channel \nat a time. It does not directly model relationships between channels, and therefore cannot \nexploit information that is present only in the joint spatial organization or colocalization of \nmultiple markers. \nDespite these constraints, applying the pipeline independently to multiple chromatin \nmarkers proved informative, because each channel emphasized a diYerent aspect of \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nnuclear organization. Hoechst provided little discriminative signal: UMAP showed minimal \nseparation and the logistic regression model performed poorly, suggesting that overall DNA \ndistribution (heterochromatin vs euchromatin) was largely unchanged between conditions \n(or that any changes were not well captured by an intensity-keypoint representation). In \ncontrast, H3.3—associated with transcriptionally active chromatin—contained a richer \nand more condition-sensitive structure, yielding clear separation in UMAP and supporting \nclassifier training. The primary probe, for facultative heterochromatin, similarly carried \nstrong discriminatory information, producing robust separation and model performance, \nwhereas CTCF fell in between, with weaker but detectable separation and modest \nclassification accuracy. \nViewed biologically, this channel-by-channel behavior is itself a useful readout. NIPBL \nknockdown is expected to disrupt cohesin-mediated loop extrusion and thereby alter gene \nregulation and chromatin architecture. The pronounced changes observed in H3.3 are \nconsistent with a strong impact on transcriptionally active regions. Interestingly, facultative \nheterochromatin also showed marked diYerences, even though it is enriched in facultative \nheterochromatin. This suggests that the perturbation aYects nuclear architecture broadly, \nproducing measurable reorganization even in regions not classically considered \ntranscriptionally active. \nMore generally, running the same analysis across multiple markers transforms a \n“single-score” classification into a comparative, mechanistic probe: channels that \nseparate well likely report on aspects of organization most aYected by the perturbation, \nwhile channels that do not may indicate relative stability (or limited sensitivity of the \nrepresentation for that signal type). This multi-channel perspective therefore increases the \nbiological insight gained from the same set of nuclei, enabling a more nuanced \ninterpretation of how diYerent chromatin compartments respond to genetic perturbation. \nThe chromatin architecture dataset presented here illustrates an ideal use case. An \nexpert observer can readily see that NIPBL knockdown alters nuclear morphology and \nchromatin organization, but an a priori feature set to quantify these diYerences is not \nobvious. The pipeline identifies image regions and texture patterns that are most \ninformative for distinguishing control and knockdown nuclei. In NIPBL-depleted cells, \ncohesin dysfunction produces enlarged nuclei with large chromatin-poor “voids, ” \neYectively compressing chromatin into a smaller accessible volume. The learned features \nemphasize smoother, more homogeneous chromatin signal within these restricted regions, \nconsistent with a model in which chromatin is packed more uniformly into the remaining \nterritory. In contrast, control nuclei exhibit a more punctate and heterogeneous chromatin \ntexture distributed over a larger volume, which the classifier associates with the control \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\ncondition. Thus, the quantitative output of the algorithm aligns with and refines the \nqualitative biological interpretation. \nAcross datasets, the pipeline was built following a “minimal complexity” principle: we \nbegan with the simplest variant capable of separating conditions and introduced additional \ncomponents only as needed. For the timelapse DCC clustering dataset, SIFT keypoint \ndetection combined with a 3D BoVW representation was suYicient to distinguish not only \nthe presence or absence of ligand (netrin) but also the more subtle eYect of Pard3 \noverexpression on receptor organization. \nTo assess whether a more complex deep-learning representation oYered an advantage, \nwe trained a masked autoencoder (MAE) on the same nuclear dataset and used the \nlearned embeddings for downstream classification. The MAE-based embeddings \nunderperformed our 3D BoVW representation (SI Fig. 8; AUC-ROC 0.837 for 3D MAE and \n0.666 for 2D projection MAE, versus 0.979 for the facultative heterochromatin BoVW \nanalysis) reinforcing our choice of a simpler pipeline that not only performed better in this \nsetting but also retained direct interpretability through mappable visual words and \nattention maps. \nAll code for the 3D pipeline is made publicly available on the 3D BoVW repository on \nGitHub [https://github.com/PittmanAEP/3DBoVW], enabling researchers to apply, adapt, \nand extend these tools to diverse imaging modalities and biological questions. \nAcknowledgements \nWe are indebted to Sharon King and Rebecca Petersen of the Department of \nDevelopmental Neurobiology Neuroimaging Laboratory at St. Jude Children’s Research \nHospital for maintaining and aligning the instruments used in this study’s lattice light-sheet \nimaging sessions. The Solecki lab is funded by the American Lebanese Syrian Associated \nCharities and by grants R01 NS066936, R01 NS104029, and R01 NS139519 from the \nNational Institute of Neurological Disorders. The content of this manuscript is solely the \nresponsibility of the authors and does not necessarily represent the oYicial views of the \nNIH. \n \n \n \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nReferences: \n \n1. Campbell, K. R., Hallada, L. P ., Huang, Y .-S. & Solecki, D. J. From Blur to Brilliance: The \nAscendance of Advanced Microscopy in Neuronal Cell Biology. Annu. Rev. Neurosci. 47, \n235–253 (2024). \n2. Stringer, C., Wang, T., Michaelos, M. & Pachitariu, M. Cellpose: a generalist algorithm for \ncellular segmentation. Nat. Methods 18, 100–106 (2021). \n3. Alder, J., Lee, K. J., Jessell, T. M. & Hatten, M. E. Generation of cerebellar granule neurons \nin vivo by transplantation of BMP-treated neural progenitor cells. Nat. Neurosci. 2, 535–\n540 (1999). \n4. Correlative three-dimensional super-resolution and block-face electron microscopy of \nwhole vitreously frozen cells | Science. \nhttps://www.science.org/doi/full/10.1126/science.aaz5357. \n5. Laumonnerie, C. et al. Siah2 antagonism of Pard3/JamC modulates Ntn1-Dcc signaling \nto regulate cerebellar granule neuron germinal zone exit. Nat. Commun. 16, 355 (2025). \n6. Yang, Z.-J. et al. Medulloblastoma can be Initiated by Deletion of patched in Lineage-\nRestricted Progenitors or Stem Cells. Cancer Cell 14, 135–145 (2008). \n7. Hatten, M. E. CENTRAL NERVOUS SYSTEM NEURONAL MIGRATION. Annu. Rev. \nNeurosci. 22, 511–539 (1999). \n8. Reiner, O. & Sapir, T. LIS1 functions in normal development and disease. Curr. Opin. \nNeurobiol. 23, 951–956 (2013). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\n9. Dept. of Computer Science, Jigawa State Colledge of Education, Gumel, Nigeria, \nAbubakar, H. D., Umar, M., & Dept. of Computer Sceince, Faculty of Science, Sokoto \nState University, Sokoto, Nigeria. Sentiment Classification: Review of Text Vectorization \nMethods: Bag of Words, Tf-Idf, Word2vec and Doc2vec. SLU J. Sci. Technol. 4, 27–33 \n(2022). \n10. Csurka, G., Dance, C. R., Fan, L., Willamowski, J. & Bray, C. Visual Categorization \nwith Bags of Keypoints. \n11. Cruz-Roa, A., Caicedo, J. C. & González, F . A. Visual pattern mining in histology \nimage collections using bag of features. Artif. Intell. Med. 52, 91–106 (2011). \n12. Powell, R. T. et al. Identification of Histological Correlates of Overall Survival in \nLower Grade Gliomas Using a Bag-of-words Paradigm: A Preliminary Analysis Based on \nHematoxylin & Eosin Stained Slides from the Lower Grade Glioma Cohort of The Cancer \nGenome Atlas. J. Pathol. Inform. 8, 9 (2017). \n13. Lee, J. et al. Unsupervised machine learning for identifying important visual features \nthrough bag-of-words using histopathology data from chronic kidney disease. Sci. Rep. \n12, 4832 (2022). \n14. Feulner, J. et al. Estimating the body portion of CT volumes by matching histograms \nof visual words. in Medical Imaging 2009: Image Processing vol. 7259 625–632 (SPIE, \n2009). \n15. Cheon, M., Lee, W., Hyun, C.-H. & Park, M. Rotation Invariant Histogram of Oriented \nGradients. Int. J. FUZZY Log. Intell. Syst. 11, 293–298 (2011). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\n16. Famulski, J. K. et al. Siah Regulation of Pard3A Controls Neuronal Cell Adhesion \nDuring Germinal Zone Exit. Science 330, 1834–1838 (2010). \n17. Singh, S. et al. Zeb1 controls neuron diYerentiation and germinal zone exit by a \nmesenchymal-epithelial-like transition. eLife 5, e12717 (2016). \n18. Trivedi, N. et al. Drebrin-mediated microtubule–actomyosin coupling steers \ncerebellar granule neuron nucleokinesis and migration pathway selection. Nat. \nCommun. 8, 14484 (2017). \n19. Kullmann, J. A. et al. Oxygen Tension and the VHL-Hif1α Pathway Determine Onset \nof Neuronal Polarization and Cerebellar Germinal Zone Exit. Neuron 106, 607-623.e5 \n(2020). \n20. Pachitariu, M., Rariden, M. & Stringer, C. Cellpose-SAM: superhuman generalization \nfor cellular segmentation. 2025.04.28.651001 Preprint at \nhttps://doi.org/10.1101/2025.04.28.651001 (2025). \n21. Lowe, D. G. Distinctive Image Features from Scale-Invariant Keypoints. Int. J. \nComput. Vis. 60, 91–110 (2004). \n22. Dupre, R., Argyriou, V ., Greenhill, D. & Tzimiropoulos, G. A 3D Scene Analysis \nFramework and Descriptors for Risk Evaluation. in 2015 International Conference on 3D \nVision 100–108 (2015). doi:10.1109/3DV .2015.19. \n23. Salton, G. & Buckley, C. Term-weighting approaches in automatic text retrieval. Inf. \nProcess. Manag. 24, 513–523 (1988). \n \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nSI Figures \n \n \nSI Figure 1 Hoechst results A: 3D renderings of the Hoechst channel overlaid with the \npositive attention maps to highlight areas of high attention. B: UMAP results from the \nnormalized image embedding vectors C: A logistic regression model was trained on the image \nlabels and normalized embedding vectors. The model trained with an AUC-ROC of 0.716. D: \nSelected kernel density estimate graphs from various Haralick Features. The KDEs were \ngenerated by analyzing patch-level Haralick feature values and summing over the two \nconditions. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\n \n \n \nSI Figure 2 CTCF Results A: 3D renderings of the CTCF channel overlaid with the positive attention \nmaps to highlight areas of high attention. B: UMAP results from the normalized image embedding \nvectors C: A logistic regression model was trained on the image labels and normalized embedding \nvectors. The model trained with an AUC-ROC of 0.817. D: Selected kernel density estimate graphs \nfrom various Haralick Features. The KDEs were generated by analyzing patch-level Haralick feature \nvalues and summing over the two conditions. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\n \n \nSI Figure 3 H3.3 Results A: 3D renderings of the H3.3 channel overlaid with the positive attention \nmaps to highlight areas of high attention. B: UMAP results from the normalized image embedding \nvectors C: A logistic regression model was trained on the image labels and normalized embedding \nvectors. The model trained with an AUC-ROC of 0.950. D: Selected kernel density estimate graphs \nfrom various Haralick Features. The KDEs were generated by analyzing patch-level Haralick feature \nvalues and summing over the two conditions. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\nSI Figure 4 EAects of volume splits on classification. Analyses were performed either on the full dataset without volume \nsplitting (top row, A–E) or after size stratification (bottom row, F–J), using smaller control nuclei and larger NIPBL LOF \nnuclei split at the median nuclear volume, as described for the chromatin dataset analyses. A and F , kernel density \nestimates of nuclear volume distributions for control and LOF nuclei; F shows the mean volume for each condition and \nthe applied split. B and G, out-of-fold ROC curves for logistic regression models trained on normalized BoVW vectors from \nthe facultative heterochromatin channel, with AUC increasing from 0.863 to 0.979 after stratification. C and H, \ncorresponding UMAP embeddings for facultative heterochromatin. D and I, out-of-fold ROC curves for logistic regression \nmodels trained on H3.3 BoVW vectors, with AUC increasing from 0.756 to 0.950 after stratification. E and J, corresponding \nUMAP embeddings for H3.3. These results show that volume-based stratification reduces overlap between control and \nNIPBL LOF nuclei and improves condition separation in both channels. \n \n \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\n \nSI Figure 5 Rotational Invariance. A minimal test dataset was generated from two individual facultative heterochromatin \nbiosensor nucleus crops, one control and one NIPBL LOF . For each volume, eight rotated variants were created and \nanalyzed together with the original image using the full pipeline. The resulting UMAP embedding shows that rotated \nversions cluster with their corresponding source image rather than separating by orientation, consistent with the \nrotationally robust descriptor design described in the manuscript \n \n \nSI Figure 6 Connected-component (“blob”) analysis of positive attention regions in facultative heterochromatin biosensor \nexpressing CGNs. A, representative nucleus overlaid with the attention map, with positive and negative patch contributions shown \non a signed scale. B, positive attention regions retained after thresholding the attention map (> 0.05). C, binary mask after connected-\ncomponent labeling, with individual high-attention components identified for downstream quantification. D, distribution of average \nconnected-component volume per image for control and NIPBL LOF nuclei. E, distribution of the number of connected components \nper image for each condition. Consistent with the attention-map trends described in the manuscript, control nuclei tended to contain \nfewer, larger connected high-attention regions, whereas NIPBL LOF nuclei showed a greater number of smaller, more fragmented \nhigh-attention regions \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\n \nSI Figure 7 EAects of normalization on image intensity. A-C, intensity histograms of the facultative heterochromatin \nchannel from a selected control image under various normalization conditions. D-F, i n t e n s i t y h i s t o g ra m s of the facultative \nheterochromatin channel from a selected NIPBL LOF image under various normalization conditions. G-I, ROC curves from \ntraining the LR model on the entire dataset under various normalization conditions. \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint \n\n \n \nSI Figure 8 Comparison with embedding vectors from custom trained masked autoencoder. Full facultative \nheterochromatin biosensor channel nuclear volumes, or 2D maximum intensity projections, were used to train a custom \nmasked autoencoder (MAE) to generate embedding vectors to compare to the embedding vectors from the Bag-of-visual-\nwords pipeline. The MAE was implemented as a 3D convolutional masked autoencoder (ConvMAE3D; base width 64) \ntrained with 80% blockwise masking (4 × 16 × 16 voxels) for 150 epochs using AdamW (learning rate = 1 × 10⁻⁴) and \nmasked L1 reconstruction loss; image-level embeddings were obtained by mean-pooling bottleneck features for \ndownstream logistic regression. A,C: original images (volumetric or 2D max projection). B,F: images after 80% masking \nwas done. Only the nucleus was masked to avoid having the MAE learn background textures. C,G: reconstruction results \nafter 150 epochs of training. D,H: the embedding vectors from the trained MAE were used to train a 2-class logistic \nregression model in the same way as with the BoVW vectors in the paper. Both cases (volumetric and 2D) did not perform \nas well as the BoVW vectors. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted April 22, 2026. ; https://doi.org/10.64898/2026.04.21.719969doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}