High-throughput methods to identify and classify modulators of nucleolar morphology and function

Abstract

Disclosed is a method coupling automated microscopy imaging with a deep neural network and computational modeling to classify distinct drug-induced nucleolar morphology changes with high accuracy and link them to nucleolar dysfunction. As such, this method can enable high-throughput characterization of nucleolar and other condensate phase behavior changes in chemical or genetic screens, directly linking condensate morphology changes to its physical mechanism of action for the first time.

Description

[0001] HIGH-THROUGHPUT METHODS TO IDENTIFY AND CLASSIFY MODULATORS OF NUCLEOLAR MORPHOLOGY AND FUNCTION

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS

[0003] The present application claims priority to U.S. Provisional Patent Application No. 63 / 544.506, filed October 17, 2023, the contents of which are incorporated by reference herein in its entirety.

[0004] TECHNICAL FIELD

[0005] The present disclosure is drawn to high-throughput analytical methods, and techniques for identifying and classifying modulators of nucleolar morphology and function in particular.

[0006] BACKGROUND

[0007] While intracellular condensates are increasingly recognized as regulators of important biological processes and as viable drug targets, methods that can robustly classify changes in mesoscale condensate morphology in vivo and link them to underlying functional perturbations within condensates are lacking. This is particularly relevant for the cell’s largest condensate, the nucleolus, which orchestrates the essential cellular process of ribosome biogenesis and is known to undergo distinct morphological changes in response to chemical and genetic perturbations, cell cycle changes, and in disease states.

[0008] BRIEF SUMMARY

[0009] In various aspects, a method for high throughput characterization and classification of nucleolar morphology changes upon treating cells with small molecules may be provided. The method may include providing a cell line. The cell line may be labelled with fluorescent protein markers, e.g., of multiple (preferably three, or at least three) nucleolar subcompartments.

[0010] The method may include treating the cell line with one or more small molecules (e.g., such as pharmaceutical agents) in each of a plurality of containers (such as wells, which may be defined by, e.g.. one or more 96- or 384-well plates).

[0011] The method may include allowing the one or small molecules to interact with cells in each well for a period of time (such as at least 1 hour).

[0012] The method may include fixing and staining cells with a nuclear dye in at least one well after treating with the one or more small molecules. The method may include capturing a plurality of images. Each image may include an image of one or more cells within one of the plurality of containers. The images may be captured automatically.

[0013] The method may include generating segmented images by segmenting the plurality of images. This may be done using, e.g., an automated segmentation algorithm. Each segmented image may include a single cell.

[0014] The method may include assigning a classification to each segmented image by subjecting the segmented image to a deep neural network trained to classify distinct nucleolar morphologies in a dose-dependent manner. Assigning a classification may include classifying nucleolar morphologies into classes where the nucleolar subphases display different phase behaviors that are linked to inhibiting distinct steps of nucleolar function in ribosome biogenesis. Assigning a classification may include assigning one of three possible morphology outcomes.

[0015] The method may include linking dose response (e.g., IC50 values) from nucleolar images to specific steps in ribosome biogenesis that are disrupted. The method may include identify ing hits based on the output from a trained neural network.

[0016] The method may include producing a dose-response curve from the plurality of images.

[0017] BRIEF DESCRIPTION OF FIGURES

[0018] Figure 1 is a schematic illustration of a mammalian cell nucleolus.

[0019] Figure 2 is a flowchart of a method.

[0020] Figures 3A-3C are images and schematic illustrations of unperturbed nucleolus (3A), nucleolar “caps'’ (3B), and nucleolar “necklaces” (3C) caused by inhibition of rRNA transcription with 10 pM CX-5461 or 6.25 pM flavopiridol (FVP) for 2 hours, respectively. Scale bar = 5pm.

[0021] Figure 4 is a schematic showing images presented to a ResNet50 neural network that classifies individual cells as having Untreated, Cap, or Necklace nucleolar morphology after calculating a pseudo-probability using Softmax function, and where, if the probability of the three classes does not exceed 0.8, cells are classified as “No Call”.

[0022] Figure 5 is a plot of network outputs in a two-dimensional “activation” space, based on the hidden layer X and Y coordinates. Averages of cells from both the training and test sets are placed in their expected regions as black-filled icons with white outlines.

[0023] Figure 6 is a confusion matrix of cells from the test set displays high accuracy for all three nucleolar morphology classes. Figure 7 is a series of images of cap and necklace phenotypes caused by additional drug perturbations. Act D = Actinomycin D (0.001 pM), BMH-21 (IpM). Oxalyplatin (25 pM), DRB = 5,6-dichlorobenzimidazole riboside (25 pM), alpha-amantin (6.25 pg / mL); THZ1 = (0.5 pM). Scale bar = 5pm.

[0024] Figure 8 is a bar chart showing fractions of cells from treatments in Figure 7 as classified by the neural network and calculated by the Softmax function.

[0025] Figures 9 and 10 are images showing a time course of nucleolar cap formation upon CX-5461 addition (10 pM) (9) and nucleolar necklace formation upon FVP addition (6.25 pM) (10).

[0026] Figure 11 is a plot showing neural network outputs of population averages for CX- 5461- or FVP-treated cells in 10 minute intervals plotted on activation space.

[0027] Figure 12 is a schematic showing vector algebra used to calculate a nucleolar morphology response.

[0028] Figure 13 is a plot showing nucleolar cap formation over time upon CX-5461 treatment as calculated per Figure 12.

[0029] Figure 14 is a plot showing nucleolar necklace formation over time upon FVP treatment as calculated per Figure 12.

[0030] Figures 15 and 16 are sets of images showing changes in nucleolar morphology' after 2 hours of CX-5461 treatment (15) or FVP treatment (16) at different concentrations. Scale bar = 5 pm.

[0031] Figure 17 is a plot showing neural network outputs for averages of cells treated with varying concentrations of CX-5461 and FVP in activation space.

[0032] Figure 18 is a schematic illustration showing functional readout assessing levels of RNA Pol I transcription for cap-forming drugs. During the last 15 minutes of drug treatment, cells are fed 5-ethynyl-uridine (5-EU) that can be visualized with a click-chemistry-containing fluorophore after fixation. 5-EU signal intensity7in the DFC is quantified as a proxy' of rRNA levels.

[0033] Figure 19 is a schematic illustration showing functional readout assessing levels of RNA Pol II transcription for necklace-forming drugs. Following drug treatment and fixation, cells are immunolabeled with an antibody detecting active Pol II transcription. Nuclear antibody signal is quantified as a proxy of Pol II CDK9-dependent Serine 2 phosphory lation.

[0034] Figure 20 is a graph showing dose-response curves for nucleolar morphology change (left axis) and 5-EU intensity change as described in Figure 18 (right axis) for cap-forming CX- 5461. ICso values represent 95% confidence intervals. Figure 21 is a graph showing dose-response curves for nucleolar morphology change (left axis) and Pol II antibody intensity change as described in Figure 19 (right axis) for necklace-forming FVP. IC50 values in parenthesis are in pM units and represent 95% confidence intervals.

[0035] Figure 22 is a graph showing dose-response curves for a variety of other cap- and necklace-forming drugs (see Figure 7).

[0036] Figure 23 is a graph showing a correlation between IC50 values from nucleolar morphology disruptions as measured by the neural network and literature-reported IC50 values from functional readouts.

[0037] Figure 24 is a schematic illustration showing rationale behind screening RNA-targeting ligands for condensate modulation.

[0038] Figure 25 is a schematic illustration of a drug screen workflow. Each plate contained training set drugs as controls in duplicate (DMSO, 10 pM CX-5461, and 6.25 pM FVP) as well as screening library compounds in duplicate at 3 different concentrations. Autofluorescent compounds were first removed by plate reader screening. Cells were then treated for 2 hours, fixed, stained, and subjected to the automated imaging and analysis pipeline.

[0039] Figure 26 is a plot showing the result of pilot screen showing population averages of each treatment’s probability of Untreated calls as calculated by the SoftMax function. Inlet shows positive controls and hit molecules with lowest probability of Untreated calls.

[0040] Figure 27 is a bar chart showing fractions of cells from control and hit compound treatments in Figure 26 as classified by the neural network and calculated by the Softmax function.

[0041] Figure 28 is a pair of example images of hits with highest “cap” calls. Scale bar = 5pm.

[0042] Figure 29 is an example image of a hit with the highest “No Call” calls. Scale bar = 5 pm.

[0043] Figure 30 is a bar chart showing Pol I transcription measurement via 5-EU metabolic labeling for controls and hit compounds. See Figure 18 for details.

[0044] Figure 31 is a bar chart showing Pol II transcription measurement via active Pol II antibody labeling for controls and hit compounds. See Figure 19 for details.

[0045] DETAILED DESCRIPTION

[0046] Quantitative cell imaging has provided fundamental insights into how subcellular organization affects cellular activities and, ultimately, cell fates. Furthermore, imaging and measuring these changes in disease states established the area of high-content screening (HCS) for drug discovery. At the same time, high-content images are complex, and deciphering the biological meaning behind changes in pixel-based phenotypes is a formidable challenge. While a priori knowledge of metrics of interest (e.g. organelle size, shape, marker intensity) helps in streamlining analysis and phenotype interpretation, it carries a risk of overlooking other relevant or even novel changes that could provide important biological insights. Advances in deep learning algorithms for images, ‘"computer vision”, are proving helpful in addressing this issue. As a result, deep learning methods have enabled simultaneous monitoring of many subcellular structures and reinvigorated interest in HCS of both genetic and chemical libraries, leading to discoveries of novel fundamental biology' and of novel therapeutic targets.

[0047] The present disclosure seeks to introduce deep learning to imaging of biomolecular condensates, a newly discovered form of subcellular structures that caused a paradigm shift in our view of cellular organization. Condensates locally concentrate biomolecules such as nucleic acids and proteins to execute specific functions. Because they lack a surrounding membrane, condensates can form and dissolve on short timescales; they also dynamically respond to changes in their environment, including drug treatments. Therefore, microscopically visible condensates with well-characterized functions and disease relevance represent attractive readouts for phenoty pic screening. Indeed, HCS of stress granules, cytoplasmic condensates implicated in frontotemporal dementia (FD) and amyotrophic lateral sclerosis (ALS), yielded insights into small molecule features that prevent stress granule formation. However, the mechanism behind the drug-induced changes in the measured features such as stress granule number and area has yet to be established. In general, linking changes in condensate structure to the underlying functional alterations is important because condensate mesoscale structures, biophysics, and molecular composition are known to change in disease states, leading to the term “condensatopathies” as well as enthusiasm for condensate-modifying drugs as a novel therapeutic strategy. Hence, an understanding of the molecular underpinnings behind these differences without extensive secondary screens to understand drug mechanisms of action would greatly assist in expediting the discovery of fundamental principles and modulators of condensates in disease.

[0048] The nucleolus is a good model system to study structure-function relationships of condensates. Its well-characterized function in ribosome biogenesis is compartmentalized into 3 layers (FIG. 1). Specifically, for the nucleolus (10), the fibrillar center (FC) (12) contains ribosomal RNA (rRNA) transcriptional machinery, the dense fibrillar center (DFC) (14) harbors processing enzymes necessary for rRNA processing, and the granular component (GC) (16) houses ribosomal and non-ribosomal proteins that facilitate rRNA folding and pre- ribosome formation. Importantly, previous studies have shown that inhibition of rRNA transcription and processing coincides with distinct changes in nucleolar integrity, suggesting that the maintenance of nucleolar structure depends on its function in ribosome biogenesis. Furthermore, these phenotypes are also therapeutically relevant because ribosome biogenesis is a prime cancer target, as cancer cells need to upregulate this process to cope with the rise in protein synthesis necessary for unrestricted cell growth and proliferation. The reliance of cancer cells on ribosome biogenesis makes them more sensitive to treatments that inhibit this process compared to healthy cells. Changes in nucleolar morphology have thus generated interest for their use as high-throughput microscopy readouts to identify regulators of ribosome biogenesis. For example, genetic screens counting nucleoli numbers in human breast cancer cells led to the identification of proteins important for ribosome biogenesis at various levels, including findings of cell cycle regulating proteins as important contributors to this process. Additionally, chemical screens that quantified changes in nucleolar protein partitioning revealed a role of protein phosphorylation by transcriptional kinases as co-regulators of RNA polymerase I and II. Finally, the development of the “iNo score", a measurement that combines several pre-determined metrics of nucleolar morphology, provided insights into ribosomal and nucleolar proteins to the integrity of this condensate. While these seminal studies established the nucleolus as a reliable indicator of defects in ribosome biogenesis, they rely on metrics that do not classify or directly link distinct disruptions to underlying functional changes. Further, because they all require nucleolar segmentation and measure pre-determined features, they can potentially overlook previously unseen changes that could point to novel biology.

[0049] As detailed below, the approach disclosed herein addresses conventional issues by coupling automated microscopy imaging with a deep neural network to classify distinct drug- induced nucleolar morphology changes with high accuracy and link them to nucleolar functional changes. As such, the disclosed method can enable high-throughput characterization of nucleolar and other condensate morphology changes in chemical or genetic screens, directly linking condensate morphology changes to its underlying biological dysfunction for the first time.

[0050] The disclosed method can be used for high throughput characterization and classification of nucleolar morphology changes, which may be useful for, e g. screening cells treated with small molecules, genetic perturbations, or exposure to oxidative stress.

[0051] Referring to FIG. 2, a disclosed method (200) may include providing (210) a cell line. The cell line may be labelled with fluorescent protein markers, e.g., of multiple (preferably three, or at least three) nucleolar subcompartments. Cells exhibiting nucleoli with a tripartite structure may include, e.g., mammalian cells (such as human cells). Other cells that may be appropriate for use here may include, e.g.. yeast cells (such as S', cerevisiae cells).

[0052] Fluorescent protein markers are well understood in the art. Such markers may include, e.g., green fluorescent protein (GFP), enhanced green fluorescent protein (eGFP), yellow fluorescent protein (YFP), red fluorescent protein (RFP) (including mCherry. DsRed, etc.), cyan fluorescent protein (CFP), blue fluorescent protein (BFP) (including mTagBFP, mTagBFP2, etc.), and / or a near-infrared fluorescent protein.

[0053] The markers may be fused to a protein that is known to localize to a target region. For example, polymerase I is predominantly localized in the FC, so an appropriate subunit may be fused to the fluorescent marker to incorporate the marker into the FC. Similar approaches can be used for each compartment.

[0054] Example 1

[0055] Human osteosarcoma cells (U2-OS) were chosen because of their clinical relevance since they are patient-derived, as well as amenability to imaging due to their epithelial tissue origin. Fluorescent protein markers of all three nucleolar compartments were incorporated by stably expressing proteins that localize to these regions. Specifically, the FC was visualized with Polymerase I subunit RPA16 fused to enhanced green fluorescent protein (EGFP), the DFC was visualized with Nop56, a core component of the ribosomal small subunit (SSU) processome fused to mCherry, and the GC was labeled with monomeric blue fluorescent protein (mTagBFP)-fused nucleophosmin (NPM1), an abundant GC protein that assists in the maturation and flux of ribosomal particles by directly interacting with large ribosomal subunit precursors.

[0056] All lentiviral DNA plasmids were generated using the FM5 lentiviral vector containing Ubiquitin C promoter. DNA encoding mTagBFP2-NPMl and Nop56-mCherry were PCR- amplified using the Q5 Hot Start High-Fidelity Master Mix (NEB) with primers synthesized by IDT. PCR products w ere inserted into the linearized vector using the In-Fusion HD cloning kit (Takara) per manufacturer’s instructions. Plasmid inserts were confirmed by Sanger sequencing. The RPA16-GFP plasmid was previously reported.

[0057] To create stably expressing U2OS lines containing constructs of interest, lentivirus was produced in Lenti-X 293T cells. Specifically, cells were grown to -50% confluency in 6-well plates and transfected with the FM5 aforementioned plasmids alongside helper plasmids PSP and VSVG using FUGENE HD® transfection reagent (Promega). After three days, supernatant was harvested and used immediately for transduction of U2OS cells plated at -40% confluency. Three days after transduction, cells were passaged for stable maintenance and subjected to fluorescence-activated cell sorting (FACS) where only cells containing BFP, GFP, and mCherry fluorescence were used in subsequent experiments.

[0058] All cell lines were grown in a humidified incubator at 37 °C with 5% CO2 in Dulbecco’s modified Eagle’s medium (Gibco) supplemented with 10% FBS (Atlanta Biological) and 100 U / mL penicillin-streptomycin (Thermo Fisher).

[0059] This cell line, hereon referred to as the three-color nucleolar line, was confirmed to display visible canonical phenotypes of nucleolar "caps” and nucleolar "necklaces” caused by small molecule drugs that inhibit rRNA transcription (CX-5461) and processing (flavopiridol, FVP), respectively. See FIGS. 3A-3C.

[0060] CX-5461 is the first known selective Pol I inhibitor that directly targets Pol I, inducing cell death via autophagy in a variety of solid tumor cells but not in healthy somatic cells. FVP is an inhibitor of cyclin-dependent kinases (CDKs), including CDK9 that is a component of positive transcription elongation factor b (P-TEFb). This kinase is necessary for phosphorylation and subsequent elongation of RNA polymerase II. The discovery that Pol II transcriptional kinase inhibitors disrupt early rRNA processing by Burger and co-workers led to the hypothesis that Pol II transcripts can regulate nucleolar ribosome biogenesis. Indeed, studies have shown that CDKs govern nucleolar structure maintenance and that certain Pol II- transcribed non-coding RNA species localize to nucleoli and regulate nucleolar integrity7.

[0061] Referring to FIG 2, the method (200) may include treating (212) the cell line with one or more small molecules (e.g.. such as pharmaceutical agents) in each of a plurality of containers (such as wells, which may be defined by, e.g., one or more 96- or 384-well plates). The method may include allowing (214) the one or small molecules to interact with cells in each well for a period of time (such as at least 1 hour).

[0062] In some embodiments, the period of time is at least 1 minute, 5 minutes, 10 minutes, 20 minutes, 30 minutes, 1 hour, 2 hours, 3 hours, 4 hours, 5 hours, or 6 hours up to 1 hour, 2 hours, 3 hours, 6 hours, 12 hours, 24 hours, 36 hours, or 48 hours, including all logical ranges and subcombinations thereof. For example, in some embodiments, the period of time may be 1 minute to 24 hours, 10 minutes to 2 hours. 1 hour to 48 hours, or 1 hour to 6 hours.

[0063] Example 2

[0064] Cells from Example 1 were plated for at least 12 hours before chemical treatment. Each cell w as exposed to a predetermined concentration of one or more compounds. CX-5461 , FVP, BMH-21, Oxalyplatin, DRB, and THZ1 were purchased from MedChemExpress. Actinomycin D and alpha-amanitin were purchased from Sigma. Pladienolide B was purchased from Tocris Bioscience. All compounds were purchased as powders and resuspended to 10 mM stock concentration, except for Actinomycin D (1 mM), and alpha-amanitin (1 mg / mL). CX- 5461 was resuspended in 50 mM sodium phosphate (pH 3.5), oxalyplatin and alpha-amanitin were resuspended in water, and the remaining compounds were resuspended in DMSO.

[0065] Referring to FIG. 2, the method may include fixing (216) and staining cells with a nuclear dye in at least one well after treating with the one or more small molecules.

[0066] Example 3

[0067] Following the small molecule treatment in Example 2. some cells were fixed in 4% paraformaldehyde (diluted in DPBS from a 16% paraformaldehyde stock, Electron Microscopy Science) for 15 minutes at room temperature and washed three times for 5 minutes each in DPBS. Nuclei were stained with the SiR-DNA kit (Spirochrome, CY-SC007) at a 1 :5000 dilution in DPBS for at least 3 hours.

[0068] Referring to FIG. 2, the method may include capturing (218) a plurality of images. Each image may include an image of one or more cells within one of the plurality of containers. The images may be captured automatically.

[0069] It should be understood that the images may be captured at any time before, during, and / or after the cells have been exposed to the small molecules. For example, images may be captured before exposure, at approximately the time of exposure, and periodically (such as every' 5 minutes, every' 10 minutes, every' 20 minutes, etc.) after the exposure, until it is desired to stop capturing images (which may be a predetermined amount of time, such as the previously mentioned period of time for exposing the cells to the small molecules).

[0070] Example 4

[0071] After fixing, the three-color nucleolar U2OS line was imaged on a Yokogawa CSU-W1 spinning-disc confocal microscope with a Yokogawa CSU-W1 spinning-disc confocal microscope with a Nikon CFI Plan Apo VX 60X water immersion objective (numerical aperture, 1.20) with a Nikon Water Immersion Dispenser (WID) on a Nikon Eclipse Ti2 body using an ORC A-Fusion BT back-thinned high quantum efficiency sCMOS camera. mTagBFP- NPM1, RPA194, Nop56-mCherry. For the three-color nucleolar line, mTagBFP2-NPMl, RPA194-GFP, Nop56-mCherry, and SiR-DNA were imaged using 405, 488, 561 and 647 nm laser lines, respectively. Using Nikoms JOBS software module, an automated image acquisition protocol was developed where five 3x3 images per well taken (608 x 608 pm each). Images were batch-denoised (4 channels) using Nikon Elements Denoise.ai module and batch- deconvolved (20 iterations) with the Richardson-Lucy algorithm.

[0072] Referring to FIG. 2, the method may include generating (220) segmented images by segmenting the plurality of images. This may be done using, e.g., an automated segmentation algorithm. Broadly, the segmentation aspect involves identifying cells from a larger image, and the automated aspect involves cropping each cell to be compatible with the neural network. For example, for segmenting cells, the IdentifyPrimary Objects module in CellProfiler software may be utilized, with Minimum Cross Entropy and filtering based on FormFactor (circularity7); other related software packages may be used for these automated steps. The software finds the cells, which segments the image into background and a unique identifier for each cell in the image. That provides x-y coordinates of each cell, which is used to then crop out images containing a cell at the center, which are then fed into the network. Each segmented image may include a single cell. Such techniques are well-understood in the art.

[0073] The method may include assigning (230) a classification to each segmented image by subjecting the segmented image to a deep neural network trained to classify distinct nucleolar morphologies in a dose-dependent manner. Assigning a classification may include classifying (232) nucleolar morphologies into classes where the nucleolar subphases display different phase behaviors that are linked to inhibiting distinct steps of nucleolar function in ribosome biogenesis. Assigning a classification may include assigning (234) one of three possible morphology outcomes.

[0074] The method may include linking (240) a dose response (D-R) (e.g., IC50 values) determined from nucleolar images to specific steps in ribosome biogenesis that are disrupted. The method may include identifying (250) hits based on the output from the computational model.

[0075] The disclosed techniques may utilize an automated microscopy and machine learning- powered image analysis pipeline. The pipeline may include light source(s) to illuminate and activate the fluorescent protein tags in cells in containers, which may have been exposed to small molecules (such as wells in the arrays). The pipeline may include one or more cameras to capture images of the fluorescent protein tags. The pipeline may include one or more processors (which may be coupled to memory and / or non-transitory computer-readable storage devices as appropriate). The processor(s) may be operably coupled to the cameras and light sources, and may be configured to, collectively, control the cameras and / or light sources. The processors may be configured to receive images from the cameras. The processors may be configured to pre-process the images. Pre-processing may include, e.g., de-speckling or other image adjustments to ensure appropriate images are fed into a neural network. Pre-processing may include segmentation of the images as disclosed herein. Pre-processing may include feature extraction; such is well-understood in the art. The features may then be fed to a trained neural network for, e.g., identifying hits, etc. Example 5

[0076] The three-color nucleolar line and the "‘training set" small molecules from Examples 1- 4 were utilized as controls to establish an automated microscopy and machine learning- powered image analysis pipeline and to classify unperturbed, nucleolar cap, and nucleolar necklace phenotypes. The pipeline may include the containers into which the cells and small molecules may be added (e.g., wells in an array). The pipeline may include light source(s) to illuminate and activate the fluorescent protein tags. The pipeline may include one or more cameras to capture images of the fluorescent protein tags. The pipeline may include one or more processors (which may be coupled to memory and / or non-transitory computer-readable storage devices as appropriate). The processor(s) may be operably coupled to the cameras and light sources, and may be configured to. collectively, control the cameras and / or light sources. The processors may be configured to receive images from the cameras. The processors may be configured to pre-process the images. Pre-processing may include, e.g., de-speckling or other image adjustments to ensure appropriate images are fed into a neural network. Preprocessing may include segmentation of the images as disclosed herein. Pre-processing may include feature extraction; such is well-understood in the art. The features may then be fed to a trained neural network for, e.g., classification.

[0077] Specifically, in this example, the three-color nucleolar line was seeded in 96 well-plates and treated with cap-forming CX-5461, necklace-forming FVP, or DMSO for 2 hours. Following cell fixation and staining with a nuclear dye, plates were subjected to automated imaging and nuclear segmentation. To improve robustness, each cell is cropped and augmented with random flips and noise. This approach yielded tens of thousands of cells from each treatment category that were then used to train several convolutional neural networks (CNNs). By assessing the loss function, it was found that the ResNet50 CNN had the highest performance for both the test and training sets. To facilitate visualization of individual cells spread from each treatment category, an encoder-decoder hidden layer was added after ResNet50 that collapses all differentiating features from neurons into two values (X and Y) and is linearly connected to the 3 class outcomes - P(Untreated), P(Cap), and P(Necklace), see FIG. 4.

[0078] This enabled each cell to be assigned with an X and Y coordinate and plot it in a 2- dimensional “activation space”, where the colors are determined by network output categories as probabilities calculated by the SoftMax function. See FIG. 5. It was found that this trained network yielded higher than 98% accuracy; most of the predicted network “calls” for different conditions matched the true condition label as visualized by the confusion matrix. See FIG. 6. It was then confirmed that the disclosed network was not limited to only detecting phenotypes specific to CX-5461 and FVP treatment. Cells that had been treated with other small molecules known to inhibit rRNA transcription (such as Actinomycin D, BMH-21, and oxaliplatin) as well as those known to inhibit rRNA processing via Pol II transcriptional modulation (such as DRB, THZ-1, and alpha-amanitin) w ere examined. All transcription and processing inhibitors also induced cap or necklace phenotypes, respectively (see FIG. 7) and were classified as expected by population proportions from network outputs (see FIG. 8).

[0079] The method may include producing (260) a dose-response curve from the plurality of images.

[0080] Example 6

[0081] The magnitude of a small molecule's effect on its cellular target and associated functional outcome is time- and concentration-dependent. It is therefore hypothesized that such effects can be visually captured as gradual alterations in nucleolar morphology7and that those progressions can be quantified by the disclosed network. Indeed, upon imaging hundreds of cells in 10-minute interv als over 2 hours, a gradual change from an unperturbed interphase nucleolus to the cap morphology was observed upon treatment with CX-5461. See FIG. 9. An equivalent transition to a necklace morphology was observed over 2 hours of FVP treatment. See FIG. 10.

[0082] After analyzing the results in the disclosed network and plotting population average coordinates for each time frame on the activation space, it was gratifyingly observed that CX- 5461 - and FVP-treated cells moved from the “Untreated” to “Cap” and “Necklace” space, respectively. See FIG. 11.

[0083] It w as then reasoned that one could utilize standard vector algebra to quantify these effects. See FIG. 12. For each treatment, the standard vector was defined as the vector drawn between the trained untreated to the trained treated point, and the vectors for each time frame were projected onto the standard vector. The magnitude of each time frame vector projected onto the standard vector is then defined as the “nucleolar morphology response”. With this approach, one can plot the rate of nucleolar cap or necklace formation and observe concentration-dependent effects. See FIGS. 13-14.

[0084] The concentration dependence prompts the question of w ether one could measure dose-response curves by incubating cells with vary ing concentrations of each drug for the same amount of time. Indeed, when the three-color nucleolar line was treated with increasing concentrations of drugs for 2 hours, similarly gradual changes in cap- and necklace-formation rates were observed. See FIGS. 15-16. Specifically, with CX-5461 treatment, it was observed that while the nucleolar GC layer appears to shrink and round at intermediate concentrations (0. 125 - 1 pM), FC fusion and positioning on the GC surface (i.e. cap formation) only occurs at higher concentrations (5 and 10 pM). See FIG. 15. Similarly, treating cells with increasing concentrations of necklace-forming FVP showed that lower concentrations (0.05 - 0.1 pM) cause GC layer fragmentation, but GC rounding and further separation from FC and DFC layers that are characteristic of necklaces occur at higher concentrations (0. 125 pM and higher). See FIG. 16.

[0085] These cells were then subjected to the disclosed automated imaging and analysis pipeline to visualize the gradual change in activation space with increasing drug concentrations. See FIG. 17. By measuring the nucleolar morphology response with vector projections for each concentration, we can plot dose-response curves and measure corresponding ICso values. CX-5461 treatment provided an ICso range from 0.28-0.38 pM (95% confidence interval, see FIG. 20). FVP treatment resulted in an order of magnitude lower range (IC50 = 0.074-0.104 pM, see FIG. 21) relative to CX-5461, which corresponds to nucleolar morphology changes being visible at lower FVP concentrations compared to CX-5461 (see FIGS. 15-16). Combined, these results illustrate the remarkable capability of the disclosed network to accurately quantify subtle, incremental changes in nucleolar morphology for both time and concentration ranges.

[0086] It was next tested if the degree of nucleolar morphology7disruption quantified by the disclosed network matches the extent of cellular target engagement and consequent functional inhibition by these drugs. For cap-forming drugs such as CX-5461, one can measured nucleolar rRNA transcriptional output (see FIG. 18). Specifically, the pulse-chase metabolic labeling paradigm was adapted by feeding cells with a uridine analog that can be visualized with a fluorophore using click chemistry. This enabled the quantification of RNA intensity7in the nucleolus as a proxy of transcriptional activity after treating cells with increasing drug concentrations as described. This measurement provided dose-response curves for CX-5461 similar to those obtained w ith the nucleolar morphology measurement; IC50 values ranged from 0.15-0.24 pM (95% confidence interval, see FIG. 20) and were also in agreement with previously reported IC50 values in which rRNA levels were measured by qRT-PCR upon CX- 5461 treatment. For FVP-treated cells, active RNA Polymerase II levels were measured via immunofluorescence (see FIG. 19), as a proxy of Pol II phosphorylation by CDK9 that necklace-forming drugs inhibit. Again, IC50 values similar to those measured by the nucleolar morphology metric were obtained (see FIG. 21) and it was found that they are in close agreement with FVP-induced rRNA processing defects as quantified via Northern blots. Further, it was shown that the disclosed method can quantify a range of responses from other cap- and necklace-forming drugs (see FIG. 7, 8, and 22), measuring low-nanomolar to high micromolar IC50 values. All measurements were in close agreement with previously reported functional measurements for these drugs (see FIG. 23).

[0087] In addition to nucleoli, the disclosed approach was able to be applied in a network trained on nuclear speckles, and dose-response curves for speckle modulators were obtained. The quantified IC50 values matched function-related measurements reported previously.

[0088] Thus, the extent of condensate morphology disruption quantified by the disclosed network is predictive of a drug’s mechanism of action (MO A), and as such has the potential to be used as a phenotypic screening platform for discovering novel small molecule modulators of condensate mesoscale structure and function.

[0089] Example 7

[0090] The nucleolus harbors RNA molecules transcribed by all three RNA polymerases as well as an abundance of RNA-binding proteins. Indeed, recent work has shown that RNA transcripts sene as both a scaffold for a multitude of dynamic and nucleolar layer-specific RNA-DNA, RNA-RNA. and RNA-protein interactions to facilitate correct ribosome assembly. Many of these species are dysregulated in disease states and have been posed as therapeutic targets, but there is currently a lack of high-throughput methodologies to expedite the discovery' of small molecule modulators of these interactions in cells. It is envisioned that applying the disclosed method in a screen of bioactive RNA-binding small molecules can fill this gap. See FIG. 24.

[0091] 20 commercially available small molecules were purchased from the R-BIND 2.0 database, which contains validated RNA binders with demonstrated activity in cell culture or animal models as well as three known bacterial ribosome inhibitors. Each molecule was screened at three different concentrations, in duplicate, and alongside training set controls in a 96 well plate. See FIG. 25.

[0092] To prevent false positives caused by intrinsically fluorescent compounds, their fluorescence was first quantified in a plate reader, which led to the discarding of 3 compounds prior to treating the cells and subjecting them to the disclosed imaging and analysis pipeline. See FIG. 25. The average probability of "Untreated" calls for each replicate were plotted, observing a cluster of CX-5461 and FVP controls as well as three hit molecules (Harmine, hVEGF-IN-1, and Topotecan) in the low er left cluster of this graph. See FIG. 26. While the control treatments were correctly classified into their population averages, harmine and hVEGF-IN-1 were classified as Caps, and the Topotecan (TPT) hit had a majority of “No Call” classifications. See FIG. 27.

[0093] It was first confirmed that hit compounds Harmine and hVEGF-IN-1 indeed caused cap morphologies as visualized by high-resolution microscopy. See FIG. 28. As expected, this morphology change was accompanied by a drop in 5-EU nucleolar signal similar to the CX- 5461 positive control (see FIG. 30). Notably, however, these hit compounds also led to a decrease in active Polymerase II signal similar to FVP (see FIG. 31). It is reasoned that these compounds are non-specific intercalating transcription inhibitors, similar to Actinomycin D. Lastly, the hit compound TPT had the highest categorization of “No call” cells, which the disclosed network labels as a morphology cannot be classified as untreated, cap, or necklace with high confidence (see FIGS. 4, 27). In line with this categorization, high-resolution imaging of TPT-treated cells revealed inverted nucleolar morphologies relative to untreated nucleoli, but without significant FC enlargement characteristic of caps or separation of FC and DFC layers from the GC as is characteristic of necklaces (see FIG. 29). As such, TPT does not cause the same functional effects as cap- and necklace-forming compounds do. See FIGS. 30-31.

[0094] Collectively, these results show the power of the disclosed method as a reliable screening platform that can accurately identify novel small molecules that induce known nucleolar structure and function perturbations as well as its ability to discover novel nucleolar morphology changes that hint at uncharacterized nucleolar biology.

Claims

What is claimed:1 . A method for high throughput characterization and classification of nucleolar morphology changes upon treating cells with small molecules, comprising: providing a cell line labelled with fluorescent protein markers of one or more nucleolar subcompartments; treating the cell line with one or more small molecules in each of a plurality of wells; capturing a plurality of images, each image including an image of one or more cells within one of the plurality of wells; generating segmented images by segmenting the plurality of images; and assigning a classification to each segmented image by subjecting the segmented image to a deep neural network trained to classify distinct nucleolar morphologies in a dosedependent manner.

2. The method of claim 1, wherein the small molecules are pharmaceutical agents.

3. The method of claim 1, further comprising linking a dose response from nucleolar images to specific steps in ribosome biogenesis that are disrupted.

4. The method of claim 3, further comprising identifying hits based on output from a neural network.

5. The method of claim 1, wherein assigning a classification includes classifying nucleolar morphologies into classes where nucleolar subphases display different phase behaviors that are linked to inhibiting distinct steps of nucleolar function in ribosome biogenesis.

6. The method of claim 1, further comprising producing a dose-response curve from the plurality of images.

7. The method of claim 1, wherein the plurality of wells is defined by one or more 96- or 384- well plates.

8. The method of claim 1, further comprising allowing the one or small molecules to interact with cells in each well for a period of time.

9. The method of claim 8, wherein the period of time is at least 1 hour.

10. The method of claim 1, further comprising fixing and staining cells with a nuclear dye in at least well after treating with the one or more small molecules.

11. The method of claim 1, wherein the plurality of images are captured automatically.

12. The method of claim 11, wherein an automated segmentation algorithm is used to segment the plurality of images.

13. The method of claim 12, wherein each segmented image includes a single cell.

14. The method of claim 1, wherein assigning a classification includes assigning one of three possible morphology outcomes.

Images