Contrastive learning of dynamic processing body formation reveals undefined mechanisms of approved compounds

Summary

Membraneless organelles (MLOs) are liquid-like compartments that organize cellular functions through liquid-liquid phase separation of proteins and RNA. Their regulation is crucial for RNA metabolism, stress response, and signaling, yet leveraging their full spatial and quantitative diversity for phenotypic screening remains challenging. Here, we present processing body (PB)-scope, an unsupervised deep-learning framework for imaging-based screening of PBs, a representative MLO. The model was trained on >400,000 single-cell confocal images from a colon cancer cell line treated with 280 compounds. PB-scope enabled precise drug classification based on multiple PB features, including number, size, and spatial distribution. This approach uncovered phenotypic patterns that were previously obscured by the subtle and dynamic nature of PBs and highlighted a set of compounds that converge on Janus kinase (JAK) signaling as a regulator of PB dynamics. PB-scope is readily extensible to other MLOs, offering broad applicability in this emerging field of cell biology.

Graphical abstract

Highlights

• •PB-scope enables unsupervised phenotypic screening of P-bodies from live-cell images
• •Contrastive clustering analyzes >400 k single-cell images across 280 compounds and controls
• •PB-scope captures subtle spatial and quantitative P-body changes beyond size and number
• •JAK signaling is a regulator of P-body assembly, validated by JAK1/2 knockdown

Natural sciences; Biological sciences; Biochemistry; Cell biology; Computational Science; Image analysis; Drug Screening

Introduction

Phenotypic profiling has long been used in drug screening, primarily for compounds with known targets.^1,2,3,4,5 Recent advances in computational image recognition have enabled the development of novel methodologies for identifying previously unrecognized drug targets.⁶ In particular, phenotypic screening strategies that leverage deep learning models applied to large-scale fluorescent imaging data have shown promise in analyzing complex biological features. These models can uncover patterns and correlations that are often difficult to detect using traditional approaches, offering new insights into the mechanisms underlying disease progression and drug responses.^7,8,9,10,11 Recent studies have shown that a cellular morphology-based convolutional neural network (CNN) system can serve as a powerful tool for anti-senescence drug screening.¹⁰ It has also been shown that the nuclear morphology of human fibroblasts can predict senescence with an accuracy of up to 95% by utilizing a deep neural network (DNN).¹² However, these supervised approaches rely on well-characterized cell models or cells treated with drugs of known mechanisms of action (MOAs) as training labels.

In recent years, unsupervised learning, which does not rely on labeled data, has emerged as an alternative for exploratory analysis and clustering tasks.^13,14,15,16 Among the most influential frameworks in this domain is the Simple Framework for Contrastive Learning of Visual Representations (SimCLR).¹³ This contrastive learning approach revolutionizes self-supervised representation learning by maximizing agreement between augmented views of the same image while minimizing similarity with other instances in the batch, enabling the extraction of robust and discriminative features without requiring annotations. This makes it ideal for revealing hidden structures and patterns within the data. Such methods have been particularly effective in profiling phenotypic features of extensive cellular structure, such as mitochondria.¹⁴ As organelles central to cellular energy production, mitochondria exhibit intricate network structures in microscopic images, and changes in their morphology have been linked to various diseases and cellular stress responses. Such phenotypic changes provide a rich source of information for drug screening.^15,16,17 Recent studies have demonstrated the potential of self-supervised models that utilize re-identification networks to identify MOAs by capturing mitochondrial phenotypic alterations.¹⁸ Despite these promising developments, it remains uncertain whether a deep-learning framework can be extended to analyze high-content imaging screening based on the morphology of different subcellular compartments.

Membraneless organelles (MLOs) are dynamic subcellular compartments in eukaryotic cells formed through liquid-liquid phase separation of proteins and/or RNAs from the surrounding milieu.^19,20 Recent studies highlight their significance in diverse pathophysiological conditions.^21,22 Cytoplasmic processing body (P-body), as one representative MLO, is evolutionarily conserved from yeast to humans. It consists of translationally repressed mRNAs and numerous proteins associated with deadenylation and decapping, as well as mRNA translation repression.^23,24 P-bodies are highly dynamic, with posttranslational modifications of their core proteins, including phosphorylation and ubiquitination, that influence their turnover.^25,26,27 However, the mechanistic details of their assembly and disassembly remain largely elusive. While prior studies have focused on changes in P-body size and number, other features, such as subcellular positioning, may provide additional insights but pose technical challenges for accurate quantification.

To address this, we developed processing body (PB)-scope, an image-based screening platform that captures comprehensive P-body features at single-cell resolution, enabling the identification of small molecules that influence P-body assembly/disassembly. This unsupervised deep-learning framework employs contrastive clustering to analyze over 400,000 segmented single-cell images acquired using a spinning disk confocal microscope (320 nm resolution) from a colorectal cancer cell line treated with 280 compounds. Using in-house segmentation models, we validated that the PB-scope effectively integrates multiple P-body-related features, facilitating accurate compound grouping. This framework leverages the power of artificial intelligence to analyze and interpret complex biological data, offering insights that were previously difficult to obtain due to the dynamic and nuanced nature of MLOs. Notably, our analysis uncovered a class of drugs with a shared mechanism of action (MOA), inhibitors of the Janus kinase (JAK) signaling pathway, previously unknown in P-body regulation. Beyond identifying compounds that regulate PB assembly, PB-scope holds the potential for adapting to the study of other MLOs, offering broad applicability in this emerging field of cell biology.

Results

PB-scope: An unsupervised deep learning framework for large-scale phenotypic screening of P-body regulators

The framework of the PB-scope is illustrated in Figure 1. To investigate the in vivo dynamics of endogenous P-bodies, we generated a DDX6-GFP knock-in HCT116 cell line for fluorescent labeling of P-bodies (Figure S1). These cells were then treated with 278 FDA-approved kinase inhibitors and two control compounds previously reported to affect P-body assembly in mammalian cells, MG132,²⁸ and thapsigargin²⁹ (Table S1). Although other perturbations such as heat shock are also known to affect P-bodies, we focused on compounds with clearer relevance to potential therapeutic pathways. Time-lapse high-resolution image data were collected using the CQ1 Benchtop High-Content Analysis System (Yokogawa) at 1 h intervals over 8 h post-treatment (Figure 1A). We simultaneously monitored P-bodies, mitochondria, nuclei, and cell membrane in living cells (Figure 1B). The mitochondrial channel was used for cell segmentation via Cellpose,^30,31 while the remaining three channels contributed to the construction of a high-quality dataset comprising over 400,000 single-cell images (Figure 1C). Subsequently, we trained a contrastive clustering model³² to extract features and cluster phenotypes associated with various drug treatments (Figures 1D and S2).

Figure 1.

Overview of PB-scope: an unsupervised deep learning-based framework for large-scale phenotypic screening on P-bodies

(A) HCT116 cells stably expressing DDX6-GFP were plated in 96-well plates, treated with 280 compounds at 10 μM concentrations, and subjected to high-content imaging using the CQ1 confocal quantitative imaging system.

(B) The analyzed images consist of four channels: (1) bright-field image for cellular morphology, (2) mitochondrial network, (3) processing body, and (4) nucleus. Merged composite demonstrates spatial relationships between these subcellular compartments. Scale bar, 10 μm.

(C) Mitochondrial channels were processed through Cellpose 3.0 to generate a curated dataset containing over 400,000 high-quality single-cell images.

(D) A contrastive clustering framework was implemented for unsupervised feature extraction, followed by UMAP dimensionality reduction to identify compounds with analogous mechanism-of-action (MOA) profiles through cluster localization analysis.

(E) Quantitative analysis of P-body formation followed by drug treatment.

(F) Mechanistic evaluation of lead compounds via imaging analysis.

Contrastive clustering adopts a two-branch architecture that performs data augmentation separately and then utilizes the contrastive learning process to enhance feature representation and clustering accuracy. The model decomposes the clustering task into two distinct yet interconnected levels: cluster-level contrastive learning and instance-level learning. Each row of the feature matrix can be interpreted as a soft label for an instance, representing the probability distribution of the instance belonging to each cluster. Conversely, each matrix column embodies the cluster representation, illustrating how different instances are distributed across clusters and effectively indicating which instances should be grouped together within each cluster. This dual-level approach allows the model to simultaneously refine the representation of data points and their cluster assignments. As a result, the model achieves a more comprehensive understanding of the data, facilitating more accurate and meaningful clustering outcomes.

The choice of data augmentation is critical in contrastive learning, as it significantly influences model performance.¹³ In this study, we applied a suite of augmentation techniques optimized for fluorescent microscope data, including standard transformations such as cropping and resizing, horizontal flipping, grayscale conversion, and standard color jittering. In addition, we implemented a color space transformation³³ called multi-channel color jitter (MCjitter) that independently takes the intensity of each fluorescent channel into our augmentation strategy. MCjitter works by independently augmenting the intensity of each fluorescent channel, thereby mitigating the effects of channel-specific noise and preserving the integrity of the underlying biological signals. This augmentation technique proved instrumental in enhancing the model’s ability to learn robust and discriminative features from the fluorescent images. Furthermore, to tackle the data imbalance problem, we introduced focal contrastive loss³⁴ into the PB-scope. This loss function is designed to dynamically adjust the weights assigned to different samples based on their difficulty, giving more emphasis to hard-to-classify instances and minority classes. By incorporating MCjitter and the focal contrastive loss function, our contrastive clustering framework achieves a notable accuracy of 63.27% in unsupervised learning, outperforming baseline comparisons across key evaluation metrics (Figure S3). Specifically, it demonstrated superior performance in adjusted Rand index (ARI) and normalized mutual information (NMI), underscoring its effectiveness in capturing meaningful cluster structures and aligning with cluster labels.

Phenotypic features extracted by the model were projected using Uniform Manifold Approximation and Projection (UMAP),³⁵ allowing for the visualization of compound clustering based on the shared MOAs and functional effects on P-body. Our approach was further validated through rigorous quantitative evaluations (Figures 1E and 1F). Beyond P-body regulation, this approach establishes a generalizable platform for investigating drug actions through unsupervised phenotypic analysis of MLOs.

Identify drugs inducing P-body phenotypes via PB-scope

Using our contrastive clustering model, we analyzed 282 conditions (280 drug treatments and dimethyl sulfoxide [DMSO]/NA controls) following a 6 h treatment regimen. For each drug, 30 cell images were randomly selected from the test set to generate UMAP embeddings (Figure S4). From this initial cohort, 41 drugs exhibited distinct spatial clustering in UMAP projection, which were subsequently categorized into five groups based on their distribution in the UMAP embedding space (Figure 2A). Two independent replicate experiments validated the reproducibility of our approach: five drug groups displayed consistent distribution patterns across both replicate experiments (Figures 2B–2G). Notably, structural analogs, such as T0374/T0374L in group 1 and T1791/T1791L in group 5, exhibited spatial proximity in UMAP (Figure S5), aligning well with our expectations that structurally similar compounds often induce the same phenotypic change.

Figure 2.

Identification of drugs with analogous P-body phenotypes via PB-scope

(A) UMAP cluster analysis of 41 compounds demonstrates a distinct spatial distribution (negative control: DMSO), based on which five drug clusters were identified.

(B–F) Experimental validation across two independent biological replicates reveals consistent cluster distribution patterns. Time-lapse montages (0–8 h, 1 h intervals) display characteristic phenotypic evolution for each compound group. Scale bars, 5 μm.

(G) The negative control group exhibited a uniform distribution in the UMAP, which serves as a baseline for comparing the distribution patterns of other assembled groups.

Given the distinctive clustering observed at the 6-h time point, we further analyzed the data after shorter treatment regimens to see the sensitivity of our method. As shown in Figure S6, PB-scope was capable of distinguishing treatment groups as early as 3 h post-treatment, with clustering resolution progressively improving over time. As for the influence on P-body dynamics, we observed that drugs in groups 1 and 3 led to a reduction in the number of P-bodies, accompanied by an increase in their size (Figures 2B and 2D). In contrast, groups 2 and 4 showed no visible changes in P-body size or numbers compared to the DMSO control (Figures 2C, 2E, and 2G). Additionally, group 5 induced cellular death around 3 h post-treatment (Figure 2F). These results demonstrate PB-scope’s ability to resolve the full spectrum of drug responses—from subtle P-body modulation to acute cytotoxicity.

Quantitative analysis of P-body phenotypes

To further investigate the differences among the five drug groups identified through clustering, particularly to explore the latent phenotypic landscape associated with P-bodies in groups 2 and 4, we performed a quantitative analysis of P-body number, the DDX6-GFP abundance and the localization of P-bodies. Given the inherent challenges in manually annotating P-body boundaries with high accuracy, we adopted a simulation-based and deep learning-assisted detection approach to approximate P-body counts and fluorescence intensity (see STAR Methods for details, Figures S7 and S8). We first generated a synthetic dataset by simulating cells and phase-separated particles (Figures 3A and S7). Cells were modeled as ellipsoids of varying sizes, with background noise following a Gaussian distribution. This simulation dataset was then used to train a YOLOv7³⁶ network for particle detection (Figures 3B and 3C). Before detection, the experimental images were converted to 8-bit format, and their intensity values were normalized to the range of 450–1,000 using Fiji.³⁷ Using this pipeline, we quantified the average number of P-bodies and fluorescent intensity of the DDX6-GFP across compound groups (Figures 3D–3G). (Figures 3D and 3E are examples from each group. The detailed results are summarized in Table S2.) We observed a distinct effect in P-body formation for MG132 and thapsigargin, similar to previously reported: MG132 continuously decreases the number over 8 h,²⁸ while thapsigargin temporarily reduces the number of P-bodies within 6 h^29,38 (Figure S9).

Figure 3.

Quantification and mechanisms of action analysis of selected drugs

(A) A simulation model of intracellular P-body was constructed to generate synthetic P-body distributions with ground truth annotations.

(B) A YOLO-v7 architecture trained on synthetic datasets was implemented for automated identification and quantitative analysis of P-body formation.

(C) Example of P-body detection, achieving >95% agreement with manual annotations (Figure S8).

(D) P-body numbers per cell in the time course under different drug treatment groups.

(E) DDX6-GFP intensity (a.u.) per cell under different drug treatment groups. Error bars represent the STD of three independent analyses for (D) and (E).

(F) Quantitative analysis of P-body numbers at 6 h post-treatment across different drug groups.

(G) Quantitative analysis of DDX6-GFP intensity (a.u.) at 6 h post-treatment across different drug groups. The p-values were determined using the two-tailed Mann-Whitney U test for (F) and (G).

The statistical significance compared with DMSO was indicated as ∗∗∗p < 0.001; ∗p < 0.05; ns, no significant difference. Data points that lay outside the 15%–85% range were deemed outliers and excluded from the statistical analysis.

(H and I) Mechanism of action (MOA) profiling for drugs in Groups 1 and 3.

The quantification of P-body numbers reveals a particularly pronounced decline in groups 1, 3, and 5, with group 5 exhibiting the most significant decrease in 6-h post-treatment (Figure 3F). In contrast to other groups, the quantitative analysis of DDX6 fluorescent intensity shows that group 1 exhibited a notable increase in fluorescence intensity (Figure 3G). We then calculated the ratio of total fluorescence intensity to P-body count and found that individual P-body fluorescence increased in group 1 (Figure S10). This trend may be due to the fusion of small P-bodies. Conversely, groups 3 and 5 displayed a significant decrease in cellular fluorescence intensity. We performed quantitative analyses of DDX6 expression levels and found that treatment with group 3 inhibitors, such as T1829 (ruxolitinib) and T7503 (upadacitinib), did not reduce DDX6 protein levels (Figure S11). This result suggested that the decrease of fluorescence in group 3 may be due to the dispersion of protein accumulation. For group 5, we observed dramatic morphological changes associated with drug-induced cell death (Figures 2F, 3F, and 3G). These analyses also revealed nuanced effects of specific compounds on P-body dynamics, particularly in drug groups that did not exhibit substantial morphological changes under visual inspection (groups 2 and 4). While group 2 does not show changes in the number of P-bodies, the variation in P-body number is much higher compared to other groups (Figure 3F). This may be caused by the heterogeneous P-body formation, which is difficult to quantify using traditional methodologies. Additionally, spatial analysis of P-bodies in group 2 indicates that the distances from P-bodies to the nucleus decreased (Figure S12). Furthermore, group 4 exhibits a slight decrease in the number of P-bodies. These findings highlight PB-scope’s unique sensitivity in identifying compounds that induce nuanced phenotypic changes, which would be overlooked in standard screening paradigms.

To see whether the observed phenotype fits with the known MOA of the drugs, we checked the known targets of different compounds groups (Figures 3H, 3I, and S13). Interestingly, apoptosis-related pathways such as Fms-like tyrosine kinase (FLT) and Vascular Endothelial Growth Factor Receptor (VEGFR) were frequently identified as targets in group 1 (Figure 3H). Indeed, group 5 exhibited an apparent apoptotic phenotype, and group 1 displayed close geometric proximity to group 5 in the UMAP embedding (Figure 2). Furthermore, JAK inhibitors stand out in groups 1 and 3, which showed opposite effects in intensity, but similar trend on number (Figures 3F–3I). Together, these observations underscore the biological phenotypic coherence and robustness of our clustering framework. These findings validate that the drugs identified through clustering methods indeed induce changes in cellular phenotypes, thereby confirming the reliability of the screening approach. While visually observed changes in cellular phenotypes provide an intuitive reflection of the overall physiological state of cells under drug influence, quantitative analysis could gain a more detailed profile of even nuanced phenotypic changes.

The JAK signaling pathway is involved in P-body regulation

To further understand the influence of inhibitors on P-body, we focused on inhibitors targeting JAK as they show different UMAP embedding among groups 1 and 3 (Figures 3H and 3I). When we explored the functional targets of the drugs in groups 1 and 3, we found that two out of six and four out of six inhibit JAK1 and JAK2, respectively. To validate the results, we further conducted immunofluorescence experiments on cells treated with small interfering RNA (siRNA) targeting either JAK1 or JAK2 to clarify the effects of this signaling cascade on the quantitative characteristics of P-bodies (Figures 4A, 4B, and S14). We observed an increased number of P-bodies in HCT116 cells when JAK1 or JAK2 was knocked down. These observations suggest a potential association between JAK signaling and the inhibition of P-body formation. As the JAK pathway is a general signal transduction pathway related to stress response and transcription (Figure 4C), the protein composition or overall structural integrity of P-bodies must be altered in the absence of functional JAK1 and JAK2. Thus, our screen revealed a previously unrecognized regulatory role for the JAK signaling pathway in maintaining and regulating P-body assembly.

Figure 4.

Perturbation of JAK leads to enhanced P-bodies

(A) HCT116 cells were knocked down using JAK1 and JAK2 siRNA, and immunostained for P-body components DDX6 (magenta) and EDC4 (green). The nuclei were visualized with DAPI (blue). Scale bar, 10 μm.

(B) Quantification of P-body number per cell across three experimental groups. Statistical significance determined by an unpaired t test was indicated as ∗∗∗p < 0.001.

(C) Model of JAK-STAT signaling pathway-mediated P-body regulation. JAK is activated when cytokines or growth factors bind to their respective receptors, leading to receptor dimerization, JAK and STAT phosphorylation, and subsequent transcriptional regulation. Inhibition of the pathway by knockdown of JAK1/2 leads induction of P-body formation.

(D) Summary of JAK inhibitors identified in this work that modulate P-body formation.

Discussion

In this work, we established a robust computational framework for large-scale drug screening based on P-body phenotypes through unsupervised deep learning. We processed and analyzed a dataset comprising over 400,000 microscopic images from HCT116 cells treated with 280 compounds (Figure 1). The PB-scope is built based on a contrastive clustering framework with two distinct augmentation branches and MCjitter model that provides better UMAP clustering with similar MOA and pathways (Figures S2 and S4), enhancing the potential in identifying promising drug candidates. Additionally, we developed a deep learning-based detection model for quantitative analysis of P-bodies and used it to confirm the efficacy and robustness of our initial screening results (Figures 3 and S12). Finally, the identification of JAK pathways in P-bodies regulation was validated with independent perturbation assays (Figure 4).

Compared to existing phenotypic screening methods,^39,40,41 the PB-scope presents numerous benefits. First, our approach introduces deep learning for feature extraction and clustering of large-scale image data, enabling unbiased discovery of novel cellular functions. Second, it overcomes the limitations of supervised methods by eliminating the need for labeled data. Moreover, the self-supervised contrastive learning architecture circumvents the constraints of traditional supervised approaches, allows the discovery of compounds with previously uncharacterized MOA. These technical advances allowed the identification of novel compound clusters showing unexpected P-body modulation effects that warrant further investigation (Figure 2).

A particularly striking finding was the association between JAK inhibitors and P-body regulation (Figures 3H, 3I, and 4). Our analysis identified JAK1/2 as recurrent targets in groups 1 and 3, and subsequent validation via siRNA knockdown and immunofluorescence confirmed that JAK inhibition enhances P-body formation. Interestingly, group 1 includes JAK2-selective inhibitors, while groups 3 and 4 include inhibitors targeting JAK1 or JAK1/2 (Figures 4D and S15). Although JAK1 and JAK2 have similar functions, the effectiveness of their inhibitors may differ. These suggest that JAK inhibitors may have different cellular physiological consequences and require attention to their molecular MOA when used. Notably, prior studies have demonstrated that P-bodies can be affected by oxidative stress⁴² and viral infection,⁴³ and the JAK-STAT pathway is one of the core pathways involved in these stress responses. The JAK-STAT pathway may regulate the phosphorylation of certain core PB proteins, thereby influencing protein-protein interactions or impacting cellular organization to alter P-body dynamics.⁴⁴ These observations suggest an underexplored link between JAK-STAT signaling and P-body dynamics, providing new insights into how kinase pathways influence P-body formation.

In summary, PB-scope has constructed a comprehensive conceptual framework that combines advanced imaging techniques, cell segmentation algorithms, unsupervised comparative learning, and in vitro biochemical experiments. This framework could be easily adapted to analyze other MLOs for phenotypic screening, which could accelerate the process of new drug discovery and development by capturing complex stress responses to various compounds. Future work could expand the dataset and refine deep learning models to enhance accuracy and robustness. Ultimately, our framework has the potential to revolutionize the drug screening process and accelerate the discovery of new and effective therapies.

Limitations of the study

While the current framework demonstrates significant advantages over existing methods, three technical constraints require future improvements. First, unsupervised drug screening cannot ensure consistent drug effects across all cells due to intercellular heterogeneity. Future research may need to incorporate more refined cell sorting techniques or develop algorithms that adjust for cellular heterogeneity. Second, unsupervised drug screening lacks precise quantitative criteria. P-bodies show relatively simple morphological features that can be described by size and number. As with mitochondrial phenotype, it is not a typical case that the results of contrastive clustering can be quantifiable. New quantitative methods need to be explored to evaluate the impact of drugs on cells more accurately. Additionally, current methods often require z-projection processing of images, which may result in the loss of valuable information about cell structure and spatial distribution. By utilizing point cloud models such as Pointnet^45,46 to capture the phenotypic features of cellular structures such as P-bodies and mitochondria, we could preserve the three-dimensional spatial structural characteristics while minimizing the consumption of computational resources.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Tatsuhisa Tsuboi (ttsuboi@sz.tsinghua.edu.cn).

Materials availability

The cell line and oligonucleotides used in the study are listed in the key resources table. All materials and reagents in this study are available from the lead contact with a completed materials transfer agreement.

Data and code availability

• •The single cell dataset utilized for model training, along with all model weights, is publicly accessible via Zenodo at https://doi.org/10.5281/zenodo.14591158. The simulation dataset utilized for P-body detection can be found via Zenodo at https://doi.org/10.5281/zenodo.15202103. The annotations of FDA-approved drugs were gathered from Drugbank.
• •The Pytorch codes for P-body drug screening can be found at https://github.com/todhc22skjicea/PB-scope. P-body simulations, detection, and quantitative analysis are available at https://github.com/todhc22skjicea/PB-scope-detection and https://github.com/catastrxphic/pb_analysis.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

We thank the members of Tsuboi and Chen laboratory for their helpful discussions and feedback on the paper. This work was supported in part by the Key Research and Development Program of the Ministry of Science and Technology (grant nos. 2024YFE0102700 and 2023YFA0914303), Shenzhen Science and Technology Innovation Commission WDZC20220811144737001, the Jilin Fuyuan Guan Food Group Co., Ltd, startup fund OD2021031C, Interdisciplinary Research and Innovation Fund JC2022008, Overseas Research Cooperation Fund HW2024009 from Tsinghua SIGS (to T.T.), National Natural Science Foundation of China (32470590 and 32100613), Shenzhen Science and Technology Innovation Commission (JCYJ20210324104605014 and KQTD20180411143432337), Shenzhen Key Laboratory of Gene Regulation and Systems Biology ZDSYS20200811144002008, and Shenzhen-Hong Kong Institute of Brain Science-Shenzhen Fundamental Research Institutions 2024SHIBS0002 (to W.C.). We thank the authors of https://github.com/ManiadisG/DivClust⁴² for making their code public.

Author contributions

Q.Z., W.C., and T.T. conceived and designed the project; Q.Z., X.P., D.P., M.Z., Y,L., Z.S., Z.C., and L.F. performed wet experiments and analyzed image data; D.S., M.C.C.A, X.Y., and T.T. performed image quantification and computational analysis; D.S., Q.Z., M.C.C.A, X.Y., L.F., W.C., and T.T. wrote the manuscript.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit polyclonal anti-DDX6	Proteintech	Cat#14632-1-AP; RRID: AB_2091264
Mouse monoclonal anti-EDC4	Santa Cruz Biotechnology	Cat#sc-376382; RRID: AB_10988077
Chemicals, peptides, and recombinant proteins
The library of 278 FDA-approved kinase inhibitor compounds	TargetMol	L1610
MG132	TargetMol	T2154; Cat#133407-82-6
Thapsigargin	TargetMol	TQ0302; Cat#67526-95-8
Hoechst 33342	Beyotime	Cat# C1022
MitoTracker Deep Red FM	Yeasen	Cat# 40743ES50
DAPI	MCE	Cat# HY-D0814
Critical commercial assays
Hieff Clone® Universal One Step Cloning Kit	Yeasen	10922ES50; Cat#10922
Deposited data
Single cell dataset utilized for model training, along with all model weights	This paper	https://doi.org/10.5281/zenodo.14591158
Simulation dataset utilized for P-body detection	This paper	https://doi.org/10.5281/zenodo.15202103.
Experimental models: cell lines
HCT116 cells	ATCC	CCL-247
Oligonucleotides
siRNA targeting sequence: JAK1: CCGTATCTCTCCTCTTTGT	This paper	N/A
siRNA targeting sequence: JAK2: GGAAATCTGAGGCAGATTA	This paper	N/A
Primers for JAK1: for 5′-GGTCAGCATTAACAAGCAGGACAA-3′, rev 5′-AGCCATCTACCAGGGACACAAAG-3′	This paper	N/A
Primers for JAK2: for 5′-GGTGCTGAAGCTCCTCTTCT -3′, rev 5′-CCGTGCACAAAATCATGCCG-3′	This paper	N/A
Primers for β-actin: for 5′-AGACTTCGAGCAGGAGATGG-3′, rev 5′-CAGGCAGCTCATAGCTCTTCT-3′	This paper	N/A
Recombinant DNA
Plasmid: pSpCas9(BB)-T2A-Puro	Ran et al.45	Addgene, Cat#48139
Software and algorithms
Yolo v7	Wang et al.36	https://github.com/WongKinYiu/yolov7?spm=a2ty_o01.29997173.0.0.10f4c921fMaPPr
Cellpose 3.0	Stringer et al.31	https://cellpose.readthedocs.io/en/latest/index.html
UMAP	McInnes et al.46	https://joss.theoj.org/papers/10.21105/joss.00861
Other
CellVoyager™ CQ1 Benchtop High-Content Analysis System	Yokogawa	https://www.yokogawa.com/solutions/products-and-services/life-science/high-content-analysis/cellvoyager-cq1/
Carl Zeiss microscope (Axio Observer 7)	Carl Zeiss	https://www.zeiss.com/microscopy/en/home.html

Experimental model and study participant details

Construction of cell lines and cell culture

Knock-in (KI) cells were generated from wild-type HCT116 cells. A single-guideRNA (sgRNA) (5′-ttatcaatgttgctcggaat-3′) with overhangs for BbsI (NEB, R3539) restriction was inserted into the pSpCas9(BB)-T2A-Puro (Addgene, 48139). The homologous repair donor template was inserted into the linearized Tet-On 3G vector backbone (courtesy of Dr. Ruijun Tian, Southern University of Science and Technology) using Gibson assembly (YEASEN, 10922ES50). The transfection was carried out using the bicistronic nuclease plasmid with the corresponding donor plasmid at the ratio of 2 to 1, and a total of 9 μg plasmid were transfected in HCT116 cells as described above. The fluorescent-positive single cells were sorted into 96-well plates by FACS to expand single-cell clones. The single-cell clones of homozygous KI cells were selected in which DDX6-GFP replaced WT DDX6. Primers used to genotype single clones are as follows: aggtggtatgttctgtgactgt (forward), aggactctgaatttaagtactgcta (reverse).

HCT116 DDX6-GFP KI cells were maintained in Dulbecco’s Modified Eagle Medium (DMEM; Gibco, #10566016), supplemented with 10% fetal bovine serum (FBS; GenClone, #25-550) and 1% penicillin/streptomycin (P/S; Gibco, #15140122), in a humidified atmosphere containing 5% CO2 at 37°C. Prior to passaging the cells onto imaging plates, they were gently washed with DPBS (Gibco, #14190144) to remove any residual medium and detached using 0.05% Trypsin-EDTA (Gibco, #25300054) to facilitate cell dissociation. A total of 12,000 cells per well were seeded into 96-well plates in 200 μL of medium.

Method details

Drug treatments

The library of 278 FDA-approved kinase inhibitor compounds and small molecular compounds MG132 and thapsigargin were purchased from TargetMol (Table S1). Cells were cultured in 96-well plates for 48 hours before drug treatment. All wells were treated with a 10μM concentration of the drugs. The live cell imaging data were collected over 8 hours. Dimethyl sulfoxide (DMSO) was set as a negative control to assess the cell phenotype at the basal level.

Imaging

Live-cell imaging was performed using the CellVoyager™ CQ1 Benchtop High-Content Analysis System (Yokogawa) equipped with a 40× objective lens with NA = 0.95. Four fields of view were captured per well. Z-stack imaging was conducted with 0.5 μm step sizes across a total depth of 10 μm, utilizing an autofocus routine on the 405 nm channel with 2 μm steps over a 10 μm range. For fluorescent staining, nuclei were labeled with 5 μg/mL Hoechst 33342 (Beyotime, #C1022), and mitochondria were labeled with 50 nM MitoTracker Deep Red FM (Yeasen, #40743ES50). After a 30-minute incubation with the dyes, cells were washed twice with PBS, and fresh medium (Gibco, #21063029) supplemented with 10% FBS (GenClone, #25-550) and 1% Penicillin-Streptomycin (Gibco, #15140122) was added. All imaging procedures were conducted under standard cell culture conditions (37 °C, 5% CO₂).

Data preprocessing

The original image had dimensions of 10 × 2000 × 2000. We carried out maximum projection processing along the Z-axis. Subsequently, cell segmentation was accomplished via the mitochondrial channel utilizing Cellpose 3.0 software.^30,31 We used mitochondria as a more reliable marker to segment cells than cell membranes, since they sometimes overlap. Following this, cells whose centers of the region of interest (ROI) were located more than 70 pixels away from the image boundary were excluded. With the center of each retained ROI as a reference point, squares measuring 150 × 150 pixels were generated and employed to extract individual cells precisely. Across 280 compounds and 8-time points, over 400,000 images were ultimately selected for analysis from 282 drug treatment conditions (280 compounds and DMSO, NA). These images were then split into a training and test dataset at a ratio of 7:3.

Structure of unsupervised clustering model

The neural network structure of the PB-scope is built based on a contrastive clustering framework.³² We divide the model into two branches, each subject to distinct data augmentation operations (Figure 1D). Specifically, we elaborate on two data enhancement methods for each instance $X_i$, through which we generate a set of 2N enhanced data samples, denoted as:

$$\begin{array}{c}\left\{X_1^a,X_2^a,\dots \dots X_n^a,X_1^b,X_2^b,\dots \dots X_n^b\right\}\end{array}$$ (Equation 1)

We utilize two independent MLPs to project features into the row and column space, where instance-level and cluster-level contrastive learning are conducted separately. We adopt ResNet34⁴⁷ as the backbone network for feature extraction, which can effectively capture similar features of the samples with its powerful deep residual learning capability. In the training process, we employed the Adam optimizer,⁴⁸ with an initial learning rate of 0.0001. No weight decay or learning rate scheduler is applied. Due to time and memory limitations, the batch size is set to 128, and the model is trained for 200 epochs before testing (Figure S16).

Multichannel color jitter

The data augmentation strategies²³ of PB-scope include resizing, cropping, converting to grayscale, horizontal flipping, Gaussian Blur, and Multi-channel Color Jitter, a data enhancement strategy for fluorescent images. In traditional contrast learning, the RGB channel intensities of natural images are usually correlated, as they often reflect the properties of the same object. However, when it comes to microscope images, each channel usually represents multiple cellular or tissue components tagged with distinguishable fluorescent markers.⁴⁹ Therefore, we incorporate an additional color perturbation model into the augmentation method to adjust the intensity of each channel differently to accommodate the noise introduced by individual laser sources.

For an image $I$ with dimensions $h\times w$ and color channels $c\in \left\{R,G,B\right\}$, the transformed image $I^{\prime }$ is defined as:

$$\begin{array}{c}{I}^{\prime }\left(h,w,c\right)=\{\begin{array}{c}T\left(I\left(h,w,c\right)\right)\text{if}r<p\\ I\left(h,w,c\right),\text{otherwise}\end{array})\end{array}$$ (Equation 2)

T(x) encompasses transformations of brightness, contrast, and saturation.

Loss function of contrast clustering

The objective function for our model comprises two components: the instance-level and cluster-level contrastive losses as:

$$\begin{array}{c}Loss=L_{ins}+L_{clu}\end{array}$$ (Equation 3)

The instance loss $l_i$ and cluster loss ${\hat{l}}_i$ for a given sample $X_i$ is defined as:

$$\begin{array}{c}{l}_i=-{\left(1-s_i\right)}^{\gamma }\log \left(s_i\right)\end{array}$$ (Equation 4)

$$\begin{array}{c}{\hat{l}}_i=-{\left(1-{\hat{s}}_i\right)}^{\gamma }\log \left({\hat{s}}_i\right)\end{array}$$ (Equation 5)

Where $s_i$ represents the similarity between sample pairs are defined as:

$$\begin{array}{c}{s}_i^a=\frac{\exp \left(\text{sim}\left(z_i,z_j\right)/\tau \right)}{\sum _{j=1}^N\left[\exp \left(\text{sim}\left(z_i,z_j\right)/\tau \right)+\exp \left(\text{sim}\left(z_i,z_j\right)/\tau \right)\right]}\end{array}$$ (Equation 6)

$$\begin{array}{c}{\hat{s}}_i=\frac{\exp \left(\text{sim}\left({\hat{z}}_i,{\hat{z}}_j\right)/\tau \right)}{\sum _{j=1}^M\left[\exp \left(\text{sim}\left({\hat{z}}_i,{\hat{z}}_j\right)/\tau \right)+\exp \left(\text{sim}\left({\hat{z}}_i,{\hat{z}}_j\right)/\tau \right)\right]}\end{array}$$ (Equation 7)

Where $\text{sim}\left(z_i,z_j\right)$ represents the cosine distance between the features of two different samples and $\tau$ serves as temperature parameters to regulate the degree of softness in each loss component.

Overall, the instance-level and cluster-level contrastive losses are computed as follows:

$$\begin{array}{c}{L}_{ins}=\frac{1}{2N}\sum _{i=1}^N(l_i^a+l_i^b)\end{array}$$ (Equation 8)

$$\begin{array}{c}{L}_{clu}=\frac{1}{2M}\sum _{i=1}^M\left(\widehat{l_i^a}+\widehat{l_i^b}\right)-H\left(Y\right)\end{array}$$ (Equation 9)

where $H\left(Y\right)$ is the entropy of the cluster assignment probabilities.

Clustering analysis

30 cell images were randomly selected for each drug from the test set to conduct clustering tests, and UMAP plots were generated. Forty-two drugs exhibiting noticeable clustering phenomena were identified and subsequently divided into five groups based on their UMAP distributions. Repeated experiments were carried out on two independent sets of images treated with drugs for 6 hours.

P-body detection and quantification

For the detection of P-body, we generated a dataset by simulating cells and phase-separated particles (Figure S7). The cells were modeled as ellipsoids of varying sizes, and the background noise within them was assumed to follow a Gaussian distribution. For the phase-separated particles, their intensity is formulated as:

$$\begin{array}{c}{B}_i=A_i\times \exp \left(-\frac{2dis}{{\sigma }_g^2}\right)+N\end{array}$$ (Equation 10)

Where $B_i$ is the brightness, $Ai$ is the parameter for the amplitude of the body’s intensity, and $N$ represents the background noise.

We then took z-slices and generated corresponding labels for each. Subsequently, the generated dataset was used to train the YOLOv7 network. The images were first converted to 8-bit format. Subsequently, they were normalized to a range of 450 - 1000. Detection was performed to identify visible P-bodies, resulting in detection boxes. The confidence level for detection was set at 0.25. The number of P-bodies was then counted, and this count was divided by the number of cells to calculate the average number of P-bodies per cell. The protein expression level was determined by the average fluorescence intensity of the cell ROI subtracting the intensity of the background as:

$$\begin{array}{c}{I}_i=\frac{\sum _{\left(x,y\right)\in R_i}I\left(x,y\right)}{R_i}-B_g\end{array}$$ (Equation 11)

where $R_i$ represents the area of the cell ROI and $B_g$ is the average intensity of the background. Since the quantification does not account for intensity adjustments due to photobleaching, the intensity shows overall a downward trend. These analyses were performed on three independent images, and their standard deviation was calculated.

Spatial distance analysis by object segmentation

For P-body detection, the fluorescence channel corresponding to the P-body marker was extracted and preprocessed by intensity smoothing to enhance contrast. Images were then filtered using a Gaussian kernel to reduce high-frequency noise, followed by adaptive Otsu thresholding to generate binary masks in which P-body structures were defined as foreground objects. Small spurious objects were removed through morphological opening and a minimum size criterion to prevent the segmentation of merged structures or debris. For each segmented object, centroid coordinates, contours, and binary masks were recorded for subsequent spatial analysis.

Nuclei were segmented using an analogous workflow with channel-specific adaptations. The nuclear channels were processed independently, applying Gaussian blurring, adaptive Otsu thresholding, and morphological opening with a minimum area of 3 pixels to exclude noise and very small objects. Centroid coordinates and binary masks were obtained for all nuclei in each field of view.

To quantify spatial relationships between P-bodies and Nuclei, pairwise Euclidean distances were calculated between the centroids of P-bodies and those of nuclei. For each P-body (i), the nearest-neighbor distance to nuclei (j) was defined as:

$$d_i=\underset{j}{\min }\sqrt{\left({\left(x_i-x_j\right)}^2+{\left(y_i-y_j\right)}^2\right)}$$

where ($x_i$, $y_i$) denotes the centroid of P-body (i) and ($x_j,y_j$) denotes the closest point to nuclear (j) in pixel units. Pixel-based distances were converted to micrometers using the microscope calibration factor:

$$d_{\mathrm{\mu }m}=d_{px}\times 0.1625$$

For each image, the mean and standard deviation of the nearest-nucleus distances were calculated across all P-bodies and used for statistical comparisons between conditions.

Immunofluorescence

HCT116 cells were seeded in 24-well plates covered with circular microscope cover glass (Nest, 801010) and cultured in DMEM with solvent or compounds. Then, cells were washed twice with cold 1xDPBS (Gibco, 14190144) and were fixed using 10% neutral buffered formalin (Bioss, C2034) at room temperature for 30 minutes, following permeabilization with methanol at −20 °C for 30 minutes, and blocking at room temperature for another 30 minutes using antibody dilution buffer (1xDPBS, 3% BSA). After the primary antibody was incubated overnight at 4°C, cells were then incubated with the diluted secondary antibody for 1 hour at room temperature and 1xDAPI (MCE, HY-D0814) to stain nuclei. Imaging was performed on an inverted fluorescence microscope (BioTeK). As primary antibodies, we used DDX6 rabbit polyclonal antibody (Proteintech, 14632-1-AP) and EDC4 mouse monoclonal antibody (Santa Cruz Biotechnology, sc-376382).

Quantification of P-body number in immunofluorescence images

Randomly selected immunofluorescence images from control, JAK1 siRNA, and JAK2 siRNA knockdown groups in HCT116 cells were analyzed to quantify P-body numbers. Individual cells with clearly defined boundaries and no overlap with neighboring cells were manually selected for analysis, while cells with ambiguous borders or overlapping regions were excluded. DDX6 channel was converted to 8-bit grayscale format. A consistent threshold range of 30 to 255 was applied across all images to detect and count P-bodies.

RNAi knockdown

HCT116 cells were transfected with negative control (NC) or gene-specific siRNA oligos using Lipofectamine™ RNAiMAX Transfection Reagent (Cas No.13778150) and the cells were analyzed 60 hours after transfection. The siRNA sequences are CCGTATCTCTCCTCTTTGT for JAK1 and GGAAATCTGAGGCAGATTA for JAK2.

RNA extraction and expression analysis

The total RNA of cultured cells was extracted by RNA isolator Total RNA Extraction Reagent (Vazyme, R401) followed by isopropanol precipitation according to the standard protocols. Reverse transcription was performed with cDNA Synthesis Master Mix (Vazyme, R223). Quantitative real-time PCR was carried out using 2xSYBR Master Mix (Vazyme, Q711). Gene expression was normalized to β-Actin RNA. Primer sequences used for gene expression analysis are as following: JAK1_for 5′-GGTCAGCATTAACAAGCAGGACAA-3′, JAK1_rev 5′-AGCCATCTACCAGGGACACAAAG-3′; JAK2_for 5′-GGTGCTGAAGCTCCTCTTCT -3′, JAK2_rev 5′-CCGTGCACAAAATCATGCCG-3′; β-actin_for 5′-AGACTTCGAGCAGGAGATGG-3′, β-Actin _rev 5′-CAGGCAGCTCATAGCTCTTCT-3′.

Quantification and statistical analysis

The image-based drug screening was conducted in two replicates over 8 hours. The images from both experiments were combined for training and testing. During testing, 30 single-cell images were randomly selected from the test set for each drug to perform clustering analysis. No data were excluded from the study. The quantitative model was validated through three experimental trials, with the standard images verified by two experimental experts.

Published: January 31, 2026

Supplemental information

Table S1. The compounds used in this study

The drug library comprises a collection of 278 FDA-approved kinase inhibitors, MG132, and thapsigargin.

Table S2. Quantification of P-bodies in cells treated with compounds

Quantification of average P-body numbers and DDX6 expression levels per cell treated with different compounds. DDX6 expression levels were quantified by calculating fluorescence intensity within cells across three independent images.

Table S3. Distance between P-body and nucleus in cells treated with compounds

The distance from the P-body to the nucleus in cells treated with different compounds was measured.