Accelerated discovery of cell migration regulators using label-free deep learning–based automated tracking

Abstract

Cell migration underlies immune surveillance, tissue repair, embryogenesis, and—when dysregulated—tumor metastasis. Yet unlike proliferation, which can be profiled at scale, migration studies remain limited by labor-intensive imaging and analysis. Existing assays often forfeit single-cell resolution, require phototoxic fluorescent labeling, or depend on tedious manual tracking, restricting the range of molecular perturbations and microenvironmental contexts that can be examined. We present Deep learning Brightfield Imaging and cell Tracking (DeepBIT), a high-throughput platform that captures live-cell behavior in multiwell plates and uses a convolutional neural network to detect and track individual cells in brightfield videos—without labels or user bias. Brightfield images are paired with nuclear fluorescence images to generate diverse ground-truth datasets, enabling automated training and eliminating manual annotation. This scalability supports a data-driven approach to systematically dissect the regulation of cell migration. Using breast cancer cells as a testbed, we tracked ~1500 cells per well across 840 conditions—including 96 FDA-approved drugs at multiple doses, a range of extracellular matrix and growth factor combinations, and CRISPR knockouts of cytoskeletal genes—yielding ~1.3 million trajectories in 30 hours (~2 minutes per condition). This dataset revealed previously unrecognized motility modulators among FDA-approved compounds and uncovered strong context dependence; for example, TNF-α and RhoA could either suppress or promote migration in the same cells depending on extracellular cues. Together, DeepBIT provides an unbiased, label-free platform for single-cell motility profiling at a scale compatible with modern drug libraries and genomic perturbation tools, enabling systematic exploration and therapeutic targeting of cell migration.

Label-free deep learning tracks millions of cells to uncover how drugs, genes, and microenvironments shape migration.

INTRODUCTION

Cell migration is a fundamental and complex biological process that regulates various functions (1), including immune surveillance and response (2–5), embryonic development (6–8), and wound healing (9–12). Dysregulated cell motility contributes to multiple pathological conditions, such as chronic wounds, fibrosis, and hyper-inflammation (13), as well as aging (14). In oncology, motility plays a crucial role in tumor progression, allowing the spreading of cancer cells from the primary tumor and their dispersal to distant sites (15). Identifying the biophysical and molecular regulatory principles that drive cell migration in both healthy and disease contexts has led to the discovery of therapeutic targets (16–21)

Numerous cell migration assays have been developed to investigate the mechanisms and regulators of cell migration across diverse biological settings (22, 23). Among these, video-based cell tracking has emerged as a powerful method to study cell migration, as it provides single-cell resolution and temporal data to characterize and understand complex migration processes (24, 25). Tracking cell movements in videos remains a nontrivial task. A critical step in cell tracking is the accurate identification of nuclei/cells. Fluorescent labeling of cells is a common strategy to facilitate cell identification via image processing or machine vision algorithms. However, fluorescence imaging presents several challenges, including phototoxicity and alterations of cell physiology (26–29). In contrast, brightfield imaging perturbs cellular systems minimally, but detecting cell locations in these images is challenging due to low contrast and highly heterogeneous cell morphology. Moreover, while manual tracking can be used for brightfield videos (26–29), the extensive effort and subjectivity involved make it impractical.

Traditional cell tracking methods suffer from extremely low throughput, restricting the comprehensive study of cell migration at the molecular level (24, 25). Using conventional manual tracking software, an experienced user may track 50 cells in a video of 50 frames in ~1 hour. Hence, to track cells in 500 different conditions (~1500 cells per condition) would take a staggering 6 years. Moreover, emerging evidence highlights the context-dependent nature of motility regulation (30, 31), emphasizing the need for large-scale, network-based analyses to fully elucidate the regulatory roles of various molecules and biological context. Evaluating motility responses using existing molecular and drug compound libraries could be highly helpful for identifying motility regulators and potential therapeutic targets. However, the lack of high-throughput motility analysis workflows continues to pose a substantial challenge. Collectively, these limitations underscore the pressing need for high-throughput motility analysis of unlabeled cells to advance our understanding of cell migration and its intracellular and extracellular mechanisms of regulation.

Several computational methods have been developed for cell detection and tracking in time-lapse microscopy, including Cellpose, DeepAct, Usiigaci, and CellTraxx (32–35). While these deep learning (DL)–based approaches have advanced cell segmentation and automated tracking, they generally require well-curated training datasets for model fine-tuning to specific imaging conditions and cell types, which limits their generalizability and ease of use. For example, Cellpose provides broadly generalizable segmentation across diverse datasets but still depends on fluorescence labeling and benefits from retraining for specific imaging conditions. DeepAct and Usiigaci enable motion-based analysis but often require manual parameter optimization and retraining to adapt to different imaging modalities, whereas CellTraxx relies on phase-contrast images with well-separated cells for accurate tracking. Despite these advances, most existing tools remain constrained by their dependence on labeled training data, imaging modality, and limited scalability.

In this work, we developed a high-throughput, label-free cell motility analysis platform called Deep learning Brightfield Imaging and cell Tracking (DeepBIT). Using convolutional neural networks (CNNs), DeepBIT detects the nuclei of live cells in brightfield images, enabling automated and accurate cell tracking without the need for fluorescence labeling. A common challenge in training CNN models is the need for large, diverse, and representative datasets. We address this by integrating brightfield imaging, molecular labeling, and fluorescence microscopy to generate extensive and varied training datasets with ground-truth (GT) labels—eliminating the need for time-consuming manual annotations. DeepBIT can track thousands of cells across ~100 time frames in minutes, substantially enhancing the throughput of motility analysis. Its label-free imaging and automated tracking capabilities enable the large-scale deployment of motility studies for drug screening, system analysis, and high-dimensional interaction studies.

RESULTS

DeepBIT system

Fluorescently tagged cells or nuclei enable easy and accurate detection of cell and nuclear locations, facilitating cell tracking using real-time imaging. However, fluorescent labeling can potentially alter the biological state of the observed cells and present issues such as phototoxicity and photobleaching, which limit the duration and temporal resolution of observations. In contrast, brightfield microscopy offers a label-free approach for live-cell imaging with minimal effects on the cells, allowing for extended observation time. However, accurately detecting cell locations in brightfield images is a challenging task (26–28).

Here, we established a pipeline that accurately detects the locations of label-free cells in brightfield images using CNNs, enabling high-throughput cell tracking and analysis (Fig. 1). Modern microscopy setups allow for the examination of motility patterns (i.e., trajectories) of individual cells from a vast number of different cell conditions in a 96-well layout, tracking cells over 4 mm² of area in each well with a sufficient temporal resolution of 10 min. Combined with high-throughput label-free analysis, our proposed workflow offers the opportunity to study a vast array of motility conditions and perturbations, such as compound screening, factorial extracellular designs, and molecular perturbations, thus enabling system-level analysis of cell motility.

Fig. 1. Workflow for plating/cell treatment, DL analysis, and DL model training.

Cells were seeded at 1000 to 2000 cells per well in 96-well plates and incubated at cell culture conditions (37°C and 5% CO₂). Following incubation, treatment conditions—such as alterations to the tissue microenvironment (TME), drug inhibition, or molecular perturbations—were added to each well accordingly. The plates were then placed on an inverted microscope with an on-stage incubator to maintain cell culture conditions. Time-lapse brightfield images were acquired at 10-min intervals over a minimum of 16 hours (4 positions per well; 384 positions per plate). The live nuclei in the brightfield images were detected using a custom DL model with DeepLabv3+ network architecture, and the trajectories of individual live nuclei were tracked using the previously established tracking methods.

We established an effective workflow for training a CNN model to detect nuclei locations in brightfield images, bypassing the need for time-consuming manual annotations of training datasets. Molecular labeling with Hoechst 33342 and propidium iodide (PI) was used to label live and dead cells, respectively. Both brightfield and fluorescent images of the cells were acquired simultaneously (Fig. 2A and fig. S1). The live and dead cell map (i.e., GT labels) corresponding to the brightfield images was then generated by processing the fluorescent images (Fig. 2A). To ensure that robust cell detection across various focal planes was achieved by the trained CNN, a z-stack of 11 brightfield images, representing both in-focus and slightly out-of-focus positions (±20 μm in z), was acquired at a given field of view (FOV), along with live and dead cell imaging (Fig. 2B).

Fig. 2. Validation of DL model to ensure high overall accuracy for nuclei prediction.

The DeepBIT model achieves high pixel-level accuracy and enables rapid tracking of thousands of cells compared with manual analysis. (A and B) Training datasets consisting of various cancer cell lines were used to allow for the model to learn to distinguish between live and dead nuclei. Brightfield images with corresponding fluorescent images (Hoechst 33342 for live cell detection and PI for dead cell detection) were acquired using the same microscopy setup as Fig. 1A. Corresponding brightfield and fluorescent images were combined to generate images annotated for background, live cell nuclei, and dead cell nuclei. These annotated images made up the data used to train the DL model, and the final model was used to annotate brightfield images for live and dead cell nuclei without and use of any additional markers/tags. (C) Nuclei prediction by the DL model. The brightfield images are at the top, and their respective nuclei predictions by the DL model are on the bottom. Brightfield image of MCF-7 and MDA-MB-231 and the corresponding DL predicted nuclei are present. Pixel-level overall accuracy reaches 98+% across cell types and densities in the test images. (D) Nuclei detection remains robust, and accuracy is retained even at different focal planes. (E) Nuclei detection remains robust, and accuracy is retained even at different cell densities. (F) Trajectory comparison between manual tracking (N = 30) and DL annotation/tracking (N = 2600). Results come from tracking done on the same sample video. (G) Comparison between individual cell speeds determined from manual tracking (N = 30) and DL annotation/tracking (N = 2600). Results come from tracking done on the same sample video. (H and I) Cell speed analysis of MDA-MB-231 cells in a 96-well plate using DeepBIT. Forty-eight wells were coated with collagen (50 μg/ml; bottom half), and 48 were uncoated controls. Brightfield images were acquired every 10 min for 16 hours. (H) Heatmap showing the average cell speed per well. (I) Dot plot showing cell migration speed of MDA-MB-231 cells cultured in collagen-coated versus uncoated wells. Each point represents the mean cell speed per well, and error bars indicate mean ± SD. Collagen coating significantly increased cell motility compared with the uncoated control.

This molecular imaging integration allowed us to collect a large dataset of images from both MDA-MB-231 and MCF-7 breast cancer cells—two commonly used cell lines in cancer research—at various cell densities, totaling 1584 brightfield images across 144 FOVs and 11 focal planes to train (N = 792) and test (N = 792) CNN models for cell detection (Fig. 2B). Live and dead nuclei images were acquired for each FOV to obtain the nuclei labels. The dataset is composed of cells at densities of 500 cells per well, 1000 cells per well, and 2000 cells per well in a 96-well plate. The molecular images were first converted to the label maps of live and dead cells through image processing. We trained the CNN model with the DeepLabV3+ framework (29), converting brightfield images into labeled live and dead cell images. The trained model labeled nuclei in the brightfield images of the testing dataset with an overall pixel-level accuracy of 98%+ for both cell types (Fig. 2C). The detection of live cells in labeled images showed a robust accuracy with an F1 score of 93.3 to 95.6% for MDA-MB-231 cells at different focal planes and cell densities. The cell detection had a better accuracy when images of cells were slightly out of the focal plane (Fig. 2D and fig. S2A). Similarly, the accuracy of the detection of MCF7 cells was robust across focal planes and cell densities, ranging from 90.7 to 96.3% (F1 score) (fig. S2C). Our trained model outperformed cell detection using the pretrained cyto3 Cellpose model (32), for which the F1 score was 76.83% on average for MDA-MB-231 cells and 70.41% for MCF-7 cells (fig. S2D).

Using our trained CNN, cell locations in each frame of live-cell brightfield movies could be effectively detected to reconstruct cell trajectories (Fig. 2F and movies S1 and S2). We further validated the tracked cell speed against manual tracking results (Fig. 2, F and G). Results showed that the average cell speed measured using DeepBIT (19.0 ± 6.6 μm/hour; N = 2600 cells) was consistent with the expected values from manual tracking (20.6 ± 5.0 μm/hour; N = 30 cells) at both the population and single-cell levels (fig. S3). For a single well of a 96-well plate, >2000 cells can be tracked in 100 frames (16 hours) within minutes using DeepBIT. The automated tracking throughput is substantially faster compared to manual tracking, which we estimated takes ~1 hour to track 50 cells in 50 consecutive frames. Using DeepBIT, we quantified the migration speed of MDA-MB-231 cells across a 96-well plate with and without collagen coating. As expected, cells cultured on collagen-coated wells exhibited significantly higher motility compared with uncoated controls. Cell speed measurements across replicate wells showed high consistency, demonstrating the reliability of DeepBIT-based tracking (Fig. 2, H and I). Overall, these results demonstrate that the established workflow can accurately analyze the motility of individual cells at high throughput.

DeepBIT enables the screening of motility regulation compounds

To demonstrate the utility and effectiveness of the DeepBIT workflow, we aimed to identify potential molecular modulators of cell motility by using compounds that inhibit cancer invasiveness and motility from Food and Drug Administration (FDA)–approved compound libraries. A total of 96 FDA-approved compounds targeting more than 16 distinct signaling pathways (table S1) were tested on MDA-MB-231 breast cancer cells at three different concentrations (0.01, 1, and 100 μM) (Fig. 3A). Of the 288 total conditions (96 unique compounds × 3 concentrations), our trained CNN model accurately determined nuclei locations and tracked cell motility under most conditions, successfully analyzing 280 conditions (Fig. 3, B and C). Eight compounds (table S2) at a concentration of 100 μM could not be accurately tracked due to the presence of a large number of particle-like objects in the brightfield images, resulting from the limited solubility of these compounds (fig. S4). In total, more than 600,000 cells were tracked across all analyzed conditions.

Fig. 3. The cell motility response of MDA-MB-231 to a panel of 96 FDA-approved compounds.

(A) Schematic of the compound screening workflow. Cells were incubated with three concentrations (0.01, 1, and 100 μM) of compounds for 24 hours before imaging. The panel of 96 FDA-approved compounds was categorized into 16 unique pathways. GPCR, G protein–coupled receptor; MAPK, mitogen-activated protein kinase; NF-κB, nuclear factor κB; mTOR, mammalian target of rapamycin. (B) Comparison between normalized cell speed and viability following treatment for each condition. Dotted lines represent cutoffs for “effective” conditions, and red numbers indicate notable conditions: (1) idebenone at 100 μM, (2) 1% DMSO; control, and (3) resminostat at 1 μM. (C) Representative cell trajectories for visualization of a (1) decrease, (2) control, and (3) increase in cell motility. Numbers correspond with highlighted conditions from (B). Scale bar, 50 μm. (D) Analysis of each pathway, showing the number of compounds which affect motility (dark gray), proliferation (gray), and viability (white) within each pathway category. (E) Venn diagram that summarizes which conditions substantially affect cell speed, proliferation, and viability out of the total 288 conditions tested (includes all compounds/concentrations). (F) Table listing the 13 conditions which caused a substantial increase (blue, +) or decrease (red, −) in cell speed without influencing proliferation or viability. For each condition, the compound name, pathway, target, and effective concentration is listed. BTK, Bruton’s tyrosine kinase; MEK, MAPK kinase; PDGFR, platelet-derived growth factor receptor; IL, interleukin. (G) Heatmaps which show the normalized effect of the 13 compounds from (F) on cell speed, proliferation, and viability. DMSO and resminostat were included as a control condition and substantial motility promoter, respectively. Red indicates inhibition, and blue indicated enhancement, relative to wild-type controls.

To ensure that these inhibitors only affected cell motility and not cell viability and proliferation, our live-dead assay was performed for all compounds at each dose. A condition was defined as having an effect on motility, proliferation, or viability if it caused a 25% or greater change compared to the dimethyl sulfoxide (DMSO) controls (Fig. 3, B and C). Among all the tested compounds, we found that 32 affected motility, 42 affected proliferation, and 21 affected viability (Fig. 3D and table S3). Among the 22 neuronal signaling inhibitors tested, 10 were found to affect motility, representing the largest group of motility inhibitors identified in this study. (Fig. 3D). In particular, we found that serotonin [5-hydroxytryptamine (5-HT)] receptor antagonists substantially affected cell motility (fig. S5).

Since motility inhibition could result from cell killing (and dead cells cannot actively move), we examined the association between cell motility, proliferation, and viability among motility inhibitors. Out of the 46 total conditions that influenced motility, we identified 13 conditions that specifically affected motility without substantially affecting proliferation or viability (Fig. 3E). Thus, most of the conditions that affected motility also affected either proliferation or viability. For the 5-HT receptor antagonists, we found only one of the five compounds affected only motility (Fig. 3F). As expected, we found that all the compounds that affected viability (N = 21) also affected motility and proliferation. We also identified 12 conditions that affected both proliferation and motility.

Among the 13 compounds that specifically affected cell motility, four—mafenide (carbonic anhydrase inhibitor), dexmedetomidine (adrenergic receptor agonist), glucosamine, and moguisteine—enhanced cell motility, while nine—ibrutinib, fluocinonide, beclomethasone, trametinib, ramosetron, omipalisib, crenolanib, levobupivacaine, and dexamethasone acetate—inhibited motility (Fig. 3, F and G). Most motility-inhibiting compounds have previously been reported to suppress cell migration in various contexts, with the exception of ramosetron, a 5-HT₃ receptor antagonist for which no link to cell motility has been previously reported. In contrast, the effects of the identified motility-promoting compounds are not well established, apart from dexmedetomidine, which has been shown to enhance migration in certain cancer cell types. We found that resminostat, an epigenetic drug, could induce an ~80% increase in MDA-MB-231 cell motility at a dosage of 1 μM compared to DMSO controls without affecting viability but slightly lowering proliferation. However, resminostat caused complete cell death at 100 μM and had no effect at 0.01 μM (Fig. 3, C and G). Overall, our results demonstrate that our DeepBIT workflow can efficiently screen libraries of compounds and identify potential regulators for cell motility.

DeepBIT enables combinatorial analysis of extracellular regulators of motility

We next used the DeepBIT platform for a systematic analysis of potential extracellular regulators of motility by exploring cell motility responses to combinatorial conditions. A subset of microenvironmental factors known to individually influence breast cancer cell motility was investigated, including cytokine stimulation [epidermal growth factor (EGF) and tumor necrosis factor–α (TNF-α)] (31), extracellular matrix (collagen) (36), and variations in serum [fetal bovine serum (FBS)] (Fig. 4A) (37). In addition, to determine cell-dependent responses, five breast cancer cell lines of varying invasive potential were examined (Fig. 4, A and B) including MDA-MB-231, MCF-7, SUM149, SUM159, and HCC1954 cancer cells. These cells were chosen because they are routinely used for cancer modeling in vitro and in animal models.

Fig. 4. The cell motility response of a breast cancer cell panel to combinations of serum concentration, extracellular matrix coating, and cytokines.

(A) Workflow diagram outlining the total unique conditions tested within this study. Cells were serum-starved for 16 to 24 hours and then incubated with their respective cytokines (100 ng/ml) for 4 hours before imaging. A total of 500,000+ total cells were tracked over 120 combinations with three replicates. (B) Representative cell trajectories for visualization of the fastest and slowest condition for each breast cancer cell line. Scale bars, 100 μm. (C) Heatmap summarizing the motility (top) and persistence (bottom) effects from the various combinations of microenvironmental factors on breast cancer cells. (D) Comparison between the persistence and displacement values across all conditions and cell lines. Each point represents a unique combination of microenvironmental factors. (E) Plots showing the percentage change in speed and persistence induced by EGF (circles) and TNF-α (squares) in MDA-MB-231 and MCF-7 cells. Each point represents a change caused by a soluble factor compared to a baseline combination of factors without the factor of interest. (F) Comparison of MDA-MB-231 percentage change in speed in response to EGF and TNF-α across all conditions. Shows the difference in magnitude of response depending on the context in which a soluble factor is introduced. Statistical analysis was performed by one-way analysis of variance (ANOVA): ***P < 0.001. (G) Plots showing the percentage change in speed and persistence induced by EGF and TNF-α across all cell lines. Each point represents a change caused by a soluble factor compared to a baseline combination of factors without the factor of interest. (H) Hierarchical clustering analysis of speed and persistence changes among the five breast cancer cell lines. Illustrated are the similarities and differences in response to microenvironmental factors between cell lines.

Our trained DeepBIT platform was able to accurately determine nuclei locations and track cell motility for all the above cell lines and conditions, corresponding to a total of 120 unique conditions (Fig. 4C and fig. S6). More than 500,000 cells were tracked across the analyzed conditions in three biological replicates. Examination of the heatmaps summarizing speed and persistence values (i.e., how consistently a cell maintains its velocity; see the definition in Materials and Methods) (38) reveals the important effect of extracellular cues in the modulation of cell motility. Collagen and EGF generally increased motility, while the effect of TNF-α varied across cell lines. There was a direct correlation between speed and persistence observed across all cell lines, which suggests that cells that move faster tend to exhibit more directional movement (Fig. 4D). Overall, MDA-MB-231 cells had the highest motility; however, their persistence was more susceptible to modulation by microenvironmental factors compared to other tested cancer cells.

To further assess how individual cytokines influence cell motility under different microenvironmental factors, we measured changes in speed and persistence in response to EGF and TNF-α for both MDA-MB-231 and MCF-7 cells across the baseline conditions tested in this study (Fig. 4, E and F). For example, to calculate the effect of EGF under baseline conditions that included TNF-α, collagen, and 0% serum, we identified paired conditions with and without EGF while keeping the other factors constant. The percentage change in cell speed and persistence between these two conditions was then quantified to represent the specific effect of EGF. The same approach was applied to determine the effect of TNF-α across all microenvironmental contexts. We found that MDA-MB-231 and MCF-7 cells displayed distinct cytokine-dependent motility patterns, with MDA-MB-231 showing enhanced responsiveness to EGF and MCF-7 demonstrating greater sensitivity to TNF-α across conditions. In MDA-MB-231 cells, EGF generally increased both speed and persistence, but the magnitude varied widely (from +7.7 to +48.6% in cell speed) depending on the microenvironment. For instance, EGF enhanced motility by 48.6% in the presence of TNF-α, 0.1% FBS, and collagen on a flat substrate, but this effect dropped to 11.6% when FBS concentration was increased to 10% (Fig. 4F). Similarly, TNF-α—typically linked to increased cancer cell motility (30, 31, 39)—showed both positive and negative effects on MDA-MB-231 motility, depending on extracellular context. TNF-α generally promoted motility, except in environments containing only collagen, where it had a negative effect on migration (Fig. 4F). Together, these results highlight that microenvironmental factors critically shape the regulatory impact of motility signals.

We further explored the motility regulatory effects across different breast cancer cell lines. EGF generally enhanced cell motility in all five cell lines under most conditions. However, a notable negative effect was observed in HCC1954 cells, where EGF reduced cell speed by 18.9% under 10% FBS without collagen or TNF-α (fig. S7). Consistent with earlier findings, TNF-α exhibited a dual role in motility regulation across all cell lines (Fig. 4G). Notably, the negative regulation by TNF-α varied by cell line and context. For example, under no collagen, no EGF, and 10% FBS, TNF-α reduced motility in MCF-7 and HCC1954 cells, while under the same conditions, TNF-α strongly promoted motility in SUM149 cells (fig. S8). Hierarchical clustering of normalized persistence and speed values revealed which cell lines had similar motility responses to extracellular factors (Fig. 4H and fig. S8). SUM149 and SUM159 cell lines clustered closely particularly in speed and relatively closely in persistence, which suggests that they share motility characteristics in response to stimuli. MDA-MB-231 and MCF-7 formed distinct clusters of regulated migration, which suggests that they have unique migration strategies in response to microenvironmental factors compared to other breast cancer cells.

Overall, our results demonstrate that breast cancer cell motility is highly context-dependent—including EGF acting primarily as a promigratory factor and TNF-α having pro- and antimigratory effects. These findings, made possible thanks to our high-throughput assay, provide insight into how microenvironmental factors regulate cancer cell migration and could help identify potential targets for therapeutic strategies.

Combining DeepBIT and CRISPR enables deep profiling of motility regulators under diverse microenvironmental conditions

Last, we demonstrated that DeepBIT enables deep phenotypic profiling of molecules that regulate cell migration by incorporating molecular perturbation methods, such as CRISPR. We knocked out RHOA, ARPC2, and CTTN in MDA-MB-231 cells; molecules that regulate the assembly and architecture of actin filament network. We measured their motility responses across 16 distinct microenvironmental conditions. These conditions were defined by a combinatorial framework incorporating cytokine stimulation (EGF and TNF-α), extracellular matrix components (collagen), and serum (FBS) (Fig. 5A). In total, >250,000 cells were tracked across the 64 analyzed conditions in three biological replicates.

Fig. 5. The impact of CRISPR KOs on cancer cell motility across diverse microenvironmental conditions.

(A) Workflow diagram outlining the total unique MDA-MB-231 knockouts (KOs) and conditions tested within this study. KO/wild-type cells were serum-starved for 16 to 24 hours and then incubated with their respective cytokines (100 ng/ml) for 4 hours before imaging. (B) Representative cell trajectories for visualization of the fastest and slowest condition for each unique KO. Scale bars, 50 μm. (C) Comparison between the persistence and displacement values across all conditions and cell lines. Each point represents a unique combination of microenvironmental factors. (D) Plots showing the change in speed induced by the KO in MDA-MB-231 cells compared to the negative control speed. Each point represents a unique combination of environmental conditions. (E) Comparison of percentage change in speed in response to RHOA (circle), ARPC2 (square), and CTTN (triangle) KOs across all microenvironmental conditions. Shows the difference in magnitude of response depending on the context in which KO is introduced. Statistical analysis was performed by one-way ANOVA: *P < 0.05, **P < 0.01, and ***P < 0.001.

Molecular knockouts (KOs) can influence cell morphology, and we found that RHOA KO induced a notable morphological transformation (fig. S9). Yet, our trained CNN model continued to successfully detect cell nuclei and track motility across all KO experiments, including RHOA KO. Our results revealed that, once more, cell motility under KO conditions exhibited a wide range of behaviors, depending on extrinsic factors (fig. S10). The slowest motility across all tested KOs and controls consistently occurred in the absence of external stimuli (i.e., no FBS, collagen, EGF, or TNF-α). In contrast, the highest motility responses for both KOs and controls were observed under conditions with different combinations of three or more external stimuli (Fig. 5, B to E).

We found a strong association between cell speed and movement persistence across all KO groups and conditions (Fig. 5C). At lower speeds, persistence was highly correlated with speed, whereas at higher speeds, persistence plateaued. These results suggest a “universal” motility response characterized by the nonlinear relationship between cell speed and persistence (40).

We further examined the effects of these molecular manipulations on cell motility under different conditions by evaluating changes in cell speed relative to the scramble control across 16 baseline conditions, including the presence or absence of EGF, TNF-α, collagen, and serum. The loss of RHOA induced distinct effects on cell motility that depended on extrinsic conditions. Under baseline conditions where cells readily migrated at high speed, RHOA KO further enhanced cell migration. In contrast, when cells were already in baseline conditions with lower migration speed, RHOA KO led to a decrease in motility (Fig. 5D and fig. S10). KO of ARPC2, on the other hand, generally led to a reduction in cell speed. The extent of this reduction was strongly associated with the cell’s migration capability, with an estimated ~35% decrease in speed observed across all tested conditions (Fig. 5D). The mixed effects of CTTN KO in MDA-MB-231 cells were also observed, with most conditions showing minimal decreases in speed. We further examined how extrinsic conditions are associated with different molecular KOs. In the presence of collagen and serum, RHOA KO significantly increased cell speed (Fig. 5E). Under other conditions, RHOA KO caused a slight decrease in cell motility or had minimal effects. The ARPC2 KO consistently reduced cell motility across most conditions, with the largest effect observed in the presence of collagen (Fig. 5E). Minor reductions in speed were observed for CTTN KO under certain conditions (Fig. 5E).

Our findings demonstrate that motility regulation via gene targeting is also context-dependent. ARPC2 is clearly essential for maintaining migration speed, whereas RHOA and CTTN appear to have a more context-dependent role. These results, rendered possible by our high-throughput assay, provide insight into how cytoskeletal regulators influence cancer cell motility and suggest targets that could be used to modulate migration. Last, we demonstrated that the workflow can be adapted to target cell types with distinct morphologies—specifically lymphocytes (Jurkat T cells)—through model retraining. The retrained model accurately detected lymphocytes, achieving an F1 score of ~0.94 across focal planes (fig. S11). Together, these results further highlight the general utility of the DeepBIT workflow for studying cell motility in diverse cell types.

DISCUSSION

We demonstrate that the proposed DeepBIT framework enables accurate and efficient tracking of cell motility in label-free brightfield videos using a trained CNN. Our workflow addresses the challenges of training accurate CNN models for nuclei detection by integrating fluorescent cell labeling with brightfield microscopy. This approach generates an extensive training dataset without the need for potentially time-consuming and subjective manual annotations. The automated labeling of nuclei in brightfield imaging provides a foundation for high-throughput analysis of cell migration behaviors at a single-cell resolution in a time-resolved manner. In addition, it facilitates large-scale screening of motility inhibitors and enables the exploration of motility responses at complex molecular intersections.

The CNN model developed in this study for live-nuclei detection performs effectively when applied to epithelial or mesenchymal cell types such as MCF-7 and MDA-MB-231. However, its performance may be limited when applied to cells with markedly different morphologies. For example, although the model was trained across a range of cell densities, extremely high-density cultures can produce brightfield patterns that differ substantially from those represented in the training dataset, potentially reducing detection accuracy. Likewise, cell types with distinctive morphologies are likely be more challenging to detect reliably (32, 41, 42). Furthermore, the current DeepBIT model was trained using images acquired at an effective 5× resolution (pixel size, ~1.3 μm); images obtained under different optical configurations or contrast modalities (e.g., differential interference contrast or alternative brightfield settings) may not generalize well to the pretrained model (43, 44).

Despite these limitations, the molecular labeling–based workflow we introduce enables efficient retraining of the model for specific imaging conditions, optical modalities, and cell types. This flexibility allows DeepBIT to be readily adapted for applications such as lymphocyte tracking or other specialized cell populations. Although retraining CNN models traditionally requires large, high-quality annotated datasets, our workflow automates label generation, substantially reducing manual effort and enabling scalable, robust retraining across diverse experimental settings. While our analyses of cell migration under different pharmacological and microenvironmental conditions serve primarily as proof-of-concept demonstrations of DeepBIT’s utility, future studies that combine DeepBIT with complementary molecular assays will be valuable for uncovering deeper biological insights. The central contribution of this work is the establishment of a generalizable, accessible computational framework for large-scale, label-free cell tracking that can be adapted across a broad range of imaging systems, cell types, and biological contexts.

The role of Ras homolog family member A (RhoA) in cancer invasion remains controversial, as findings from individual studies have reported both proinvasive and inhibitory effects when RhoA is inhibited (45, 46). Our study also demonstrates that RhoA KO induces a wide range of motility responses, from inhibition to minimal or even enhanced motility, depending on the presence of collagen in the environment. These results underscore the complex interplay between extrinsic and intrinsic cellular factors in regulating motility. A single-axis (piecemeal) analysis of molecular function provides a limited context and may not accurately reflect its role in the complex in vivo environment, emphasizing the need for high-throughput approaches to capturing a more comprehensive picture and precision of biology.

Automated cell tracking in brightfield images substantially reduces manual effort and enables high-throughput analysis of cell motility, facilitating advances in systems and precision biology. In this study, we analyzed a total of 840 experiments (including all repeats and conditions) with 100+ frames per experiment, requiring ~30 hours of processing time and tracking ~1.2 million cells. By comparison, we estimate that manual tracking by a trained researcher achieves a throughput of 50 cells across 50 frames per hour, meaning that the same task would take ~48,000 hours (~5.5 years) to complete manually. This stark contrast highlights the potential of automated tracking for large-scale motility studies. Our assay reveals that 5-HT antagonists can substantially affect cell motility, implicating neurotransmitter signaling pathways in the regulation of breast cancer migration. While the role of 5-HT receptors, or serotonin receptors, is well established in neurotransmitter regulation, recent studies have also highlighted their involvement in tumorigenesis and tumor progression across various tumor types (47, 48). In breast cancer, overexpression of 5-HT has been observed in patients with triple-negative breast cancer (49) and is associated with increased protumor activity. In addition, treatment with 5-HT agonist antidepressants in patients with cancer has been linked to a higher risk of cancer recurrence (50–52). Our findings further demonstrate that a subset of 5-HT antagonists induces antimigratory effects in cancer cells, independent of their antiproliferative activities. These results provide further evidence of the role of the 5-HT–associated pathway in regulating cancer invasion.

MATERIALS AND METHODS

Cell lines and culture

MDA-MB-231 (HTB-26), MCF-7 (HTB-22), and HCC1954 (CRL-2338) cells were purchased from the American Type Culture Collection, and SUM149 (HUMANSUM-0003004) and SUM159 (HUMANSUM-0003006) were obtained from BioIVT. The breast cancer cell lines were cultured in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% FBS (Corning, 35-010-CV) and 1% penicillin-streptomycin (P/S; Sigma-Aldrich, P0781-100ML). Cells were maintained at 37°C and 5% CO₂ in an incubator for passage numbers less than 15. Cell were trypsinized (trypsin EDTA, Sigma-Aldrich, T4049-500ML) and passaged at 70 to 80% confluency, every 2 to 3 days.

Live and dead cell assay and analysis

Cells were plated at a density of 1500 to 2000 cells per well depending on the cell line. Following an overnight incubation and/or cell tracking experiment, well plates were stained using Hoechst 33342 (Thermo Fisher Scientific, H21492) and PI (Thermo Fisher Scientific, P1304MP) and imaged in two channels, 395 and 555 nm. Cells were identified in the fluorescent images and were classified as live or dead using filters based on Hoechst 33342 and PI intensity.

Deep learning model training

Cells were plated at varying densities of 500, 1000, and 2000 cells per well in 96-well plastic-bottom plates (Corning, 3603). After plating, cells were incubated overnight to allow attachment to the surface. The following day, cells were stained with Hoechst 33342 (3 μg/ml) and PI (3 μg/ml) to label live and dead nuclei, respectively, for 30 min. For imaging, brightfield images were first captured at 11 different focal planes with a 2-μm z-step, offset above and below the in-focus plane, using a Nikon TI-E microscope with 10× objective. Immediately after brightfield imaging, fluorescence images of Hoechst 33342 and PI were acquired for each FOV to minimize potential differences caused by cell motility. Four FOVs were imaged per well.

Live and dead cell–labeled images for each FOV were generated from the fluorescent images. A bandpass filter was applied to reduce background noise and improve the signal-to-noise ratio (53, 54). Intensity thresholds were manually determined to identify positively stained regions. Since Hoechst stains both live and dead nuclei, regions positive for both Hoechst and PI were classified as PI-positive (dead) cells. The brightfield images were also normalized before training. First, the background field was estimated and subtracted from each brightfield image. The background field was calculated using a two-dimensional median filter with a window size of 300 × 300 pixels. After background subtraction, the images were rescaled to have a mean intensity of 100 arbitrary units (a.u.) and an SD of 30 a.u. This was achieved by dividing the pixel intensities of the background-corrected images by their SD, multiplying by 30, and then adding 100.

The image set is then split into a training and testing dataset to train a DeeplabV3+ network (29) to detect live and dead cell label from brightfield images. The image split is performed at FOV levels such that the whole z-stack images are assigned either to training or testing. The images are first down-sampled at twofold, and the image size for training models is set to 1024 × 1024 pixels. The networks are trained with 30 epochs. The model performance is evaluated on the basis of the testing image. All computational procedures including image processing and model training were performed using MATLAB (The MathWorks, Natick, MA).

Evaluation model performance in cell tracking

To evaluate the effectiveness of the trained model in cell tracking, an independent image set containing brightfield images with corresponding 4′,6-diamidino-2-phenylindole (DAPI) and PI nuclear staining was used. The trained model was first applied to the brightfield images to generate live-cell nuclei predictions. Marker-controlled watershed segmentation was then implemented to separate clustered nuclei in the predicted label images. GT live nuclei locations were determined from the corresponding DAPI and PI images acquired simultaneously for each brightfield image. GT live cells were defined as DAPI⁺/PI⁻ nuclei after segmentation of the DAPI channel using a previously established nuclei segmentation pipeline (53, 54). Detected cells and GT live cells were considered a match [true positive (TP)] if their centroids were within 10 μm (~the size of a nucleus). Unpaired GT nuclei were counted as false negatives (FN), and unpaired detected live cells were counted as false positives (FP). Precision, sensitivity, and F1 score were then computed as follows: Precision = TP/(TP + FP), sensitivity = TP/(TP + FN), and F1 = 2TP/(2TP + FP + FN).

Cell motility assay

Black 96-well plates (Corning, 3603) were coated with of collagen type I, rat tail (50 μg/ml; Corning, 354249) for 30 min. Breast cancer cells were seeded at so that cells would reach ~2000 cells per well before imaging. To ensure even cell distribution, the 96-well plate was placed on a shaker at 500 rpm for 2 min following seeding and allowed to settle for 20 min before being placed in an incubator. Cells underwent desired treatments following an overnight incubation. Following treatment, 96-well plates were placed onto a microscope (Nikon Eclipse Ti-E) with an on-stage incubator (Tokai Hit, INU-TIZW) to maintain cell culture conditions of 37°C and 5% CO₂. Time-lapse brightfield images were acquired at 10-min intervals over a minimum of 16 hours (4 positions per well; 384 positions per plate). Images were obtained at ×10 magnification.

FDA-approved compound treatment

MDA-MB-231 cells were seeded at 1000 cells per well in a 96-well plate. Cells were treated with 0.01, 1, and 100 μM FDA-approved compounds randomly selected from a single plate of an FDA-approved screening library (Selleck Chem, L4300-03) for 24 hours. Treated cells underwent time-lapse imaging with subsequent live/dead cell analysis. “Control” samples were supplemented with DMSO vehicle in place of FDA-approved compounds.

CRISPR lipofection

MDA-MB-231 cells were seeded in 12-well plates so that they would reach 60 to 80% confluency on the day of transfection. Cells were transfected with TrueCut Cas9 Protein (2500 ng per well, Thermo Fisher Scientific, A36498) and RHOA (CRISPR1031551_SGM), ARPC2 (CRISPR995699_SGM), and CTTN (CRISPR925694_SGM) True Guide sgRNA (single guide RNA; 480 ng per well, Thermo Fisher Scientific, A35533) using Lipofectamine CRISPRMAX (Thermo Fisher Scientific, CMAX00003) for 48 hours. Following transfection, cells were trypsinized and seeded into 96-well plates for high-throughput cell motility assay conditioning and analysis. Nontargeting sgRNA-treated (Thermo Fisher Scientific, A35526) cells and cells treated with Lipofectamine CRISPRMAX only were used as controls.

Serum starvation and cytokine treatment

MDA-MB-231 and HCC1954 were seeded at 1500 cells per well, and MCF-7, SUM159, and SUM149 were seeded at 1250 cells per well in a 96-well plate. Cell culture medium was replaced with starvation medium (DMEM supplemented with 0, 0.1, 10% FBS, respectively, and 1% P/S) following an overnight incubation after plating. Cells were serum-starved for 16 to 24 hours. EGF (100 ng/ml; PeproTech, AF-100-15), TNF-α (PeproTech, 300-01A), and EGF + TNF-α was added to each well respectively and incubated with cells for 4 hours before imaging. Control samples underwent serum starvation and were supplemented with additional starvation medium in place of cytokines.

Deep learning cell detection, tracking, and analysis

To track the cell motility, the live cell nuclei locations in the brightfield images are first detected using the trained deep learning (DL) model after twofold downsampling. The marker-controlled watershed was then implemented to segmentation, and the detected live nuclei images and locations of the segmented nuclei object are then measured (54). Objects with area less than 15 μm² are excluded from further analysis. Once the nuclei locations are obtained from all timeframes, cell trajectories are tracked using previously established methods (54, 55). Cell instantaneous speed and persistence are calculated for each tracked object. Cell instantaneous speed was calculated at a time lag (τ) of 1 hour using the following equation

$$\text{Speed}\left(\mathrm{\tau }\right)=<\frac{\sqrt{{\left[x(t+\mathrm{\tau })-x\left(t\right)\right]}^2+{\left[y(t+\mathrm{\tau })-y\left(t\right)\right]}^2}}{\mathrm{\tau }}$$

where τ represents the time lag, <… > indicates time averaging, t represents the instantaneous time, and (x,y) are the coordinates for cell location at a given point in time.

Persistence is defined as the ratio of net distance traveled (D_net) to integrated distances traveled (D_int), calculated using the following equations

$$D_{\text{net}}=\sqrt{{\left[x\left(t_{\text{end}}\right)-x\left(t_0\right)\right]}^2+{\left[y\left(t_{\text{end}}\right)-y\left(t_0\right)\right]}^2}$$

$$D_{\text{int}}=\sum _{t=t_0+N\mathrm{\tau }, N=0,1\dots }\sqrt{{\left[x(t+\mathrm{\tau })-x\left(t\right)\right]}^2+{\left[y(t+\mathrm{\tau })-y\left(t\right)\right]}^2}$$

In these equations, t₀ represents the initial time point, t_end represents the final time point collected, and N represents the total number of time steps, taking into account the time lag. The time lag for calculating D_int is a single frame.

Statistical analysis

One-way analysis of variance (ANOVA) was performed using GraphPad Prism and was used to determine statistical significance. Results were considered significant at *P < 0.05, **P < 0.01, and ***P < 0.001.

Supplementary Materials

The PDF file includes:

Figs. S1 to S11

Legends for tables S1 to S3

Legends for movies S1 and S2

References

Other Supplementary Material for this manuscript includes the following:

Tables S1 to S3

Movies S1 and S2