dNEMO: a tool for quantification of mRNA and punctate structures in time-lapse images of single cells

Many biological processes are regulated by single molecules and molecular assemblies within cells that are visible by microscopy as punctate features, often diffraction limited. Here we present detecting-NEMO (dNEMO), a computational tool optimized for accurate and rapid measurement of fluorescent puncta in fixed-cell and time-lapse images. The spot detection algorithm uses the à trous wavelet transform, a computationally inexpensive method that is robust to imaging noise. By combining automated with manual spot curation in the user interface, fluorescent puncta can be carefully selected and measured against their local background to extract high quality single-cell data. Integrated into the workflow are segmentation and spot-inspection tools that enable almost real-time interaction with images without time consuming pre-processing steps. Although the software is agnostic to the type of puncta imaged, we demonstrate dNEMO using smFISH to measure transcript numbers in single cells in addition to the transient formation of IKK/NEMO puncta from time-lapse images of cells exposed to inflammatory stimuli.

activated through induced proximity with other signaling mediators (12)(13)(14)(15)(16). In cells that express EGFP fused to NEMO exposed to inflammatory cytokines, EGFP-NEMO transiently localizes to punctate fluorescent structures near the plasma membrane (16,17). The number and timescales of EGFP-NEMO-enriched puncta reveal quantitative properties of receptor-associated protein complexes that transmit information from the inflammatory milieu into the NF-κB transcriptional system.
Here we present detecting NEMO (dNEMO), a free application that uses waveletbased spot detection and supervised segmentation to detect and measure properties of fluorescent puncta in fixed-cell and time-lapse images. We show that the wavelet-based approach is significantly faster than traditional Gaussian fitting methods and allows for almost real-time interaction with single cells in quantitative imaging data. Intuitive tools for cell segmentation, spot inspection, and background correction, in addition to manual and automated selection of puncta based on quantifiable features (e.g. size, location, fluorescence) ensure that single cell data are of the highest quality. Results are formatted for easy coordination with other software packages, such as single-particle tracking applications and other analyses of structural dynamics (18). Using smFISH and live-cell data for EGFP-NEMO as demonstrations, we show that dNEMO is a versatile workspace for rapid, precise, and robust measurement of fluorescent puncta in digital images.

RESULTS: dNEMO identifies near diffraction-limited fluorescent puncta in 2D and 3D images
Wavelet-based approaches are used in image analysis for de-noising, compression, and feature extraction with low computational cost (19,20). In wavelet-based feature extraction applications, the source image is decomposed into wavelet maps, a series of images where contrast is enhanced for particular spatial features. Since the wavelet transform sequentially applies a different convolution matrix at successive levels of the algorithm, the size and qualities of spatial features that are enhanced in each wavelet map can be modulated.
The à trous wavelet transform accurately detects and localizes isotropic diffractionlimited spots such as fluorescent mRNA puncta in single molecule FISH images (21,22).  (22,23), the L2 wavelet map (Figure 1B, 3 rd column; see also Movie S1) enhanced contrast for puncta at or near the diffraction limit. At higher levels, larger puncta were more resolved (L3 wavelet map; Figure 1B, 4 th column) at the expense of reduced clarity for smaller puncta. Although users of dNEMO can select a wavelet map appropriate for their application, the L2 wavelet map was used for all subsequent experiments to detect small molecular assemblies.
To identify fluorescent puncta near the diffraction limit, the L2 wavelet map was segmented using thresholding and watershed algorithms in dNEMO. Briefly, the background is removed from the L2 wavelet map by setting pixel values below a userdefined threshold to zero and multiplying other pixel values by -1 to generate an inverted wavelet map. A watershed is then applied to define local basins in the inverted wavelet map. The dNEMO interface updates the source image in real-time to assist users with a first-pass visual estimation for puncta identified under the chosen threshold value ( Figure   2, top panel). The spatial coordinates for each basin and the surrounding path are used to define the centroid and perimeter respectively for each of the fluorescent punctum.
Puncta are then evaluated to prevent over-segmentation, where a single punctum with a noisy spatial distribution of fluorescence is erroneously segmented by the watershed into two or more puncta. To resolve over-segmentation, the paths that connect all pairs of centroids separated by less than 10 pixels are scanned for a minimum intensity value. If a minimum is not found, the puncta are consolidated, and properties of the merged punctum are recalculated. Similarly, centroids falling within 3 pixels of each other are combined because this is within the resolution limit of our optical imaging system.
For analysis of puncta in 3D images, wavelet maps are produced for each 2D slice of the image stack. Centroids identified in each slice are referenced against centroids in adjacent slices and puncta are merged between slices of the image stack if the X-Y coordinates of their centroids fall within a Euclidean distance of 2 pixels. Using this approach in parallel with fluorescence information for the same punctum in adjacent image slices, an axial component of each centroid can be calculated and overlapping spots can be resolved (Supplemental Figure 2). Fluorescence properties for each 3D punctum can be either aggregated across image slices or measured at the image slice that corresponds to the axial centroid. Once user-defined settings are established, dNEMO is typically run in batch processing mode so that the methods for puncta detection are identical across all images in an experiment.

Local background correction and cell segmentation for accurate quantification of puncta in single cells
Slow-varying background from non-specific dye accumulation and free fluorescent proteins that are not part of molecular assemblies, among other sources, will contribute to the measured intensity of a punctum. To correct for these effects, dNEMO collects local

Spot detection in dNEMO is rapid and accurate
In comparison with spot-fitting methods, such as 3D Gaussian or maximum likelihood estimation (26), the wavelet-based approach in dNEMO does not require iterative estimation of parameters or image pre-processing steps. To compare against our application, we selected the software package FISH-quant (27) primarily because it implements a 3D Gaussian fitting method for detection of transcripts and it's actively used in the research community. For example, we had previously used FISH-quant to detect single mRNA molecules in the context of stochastic transcription events (28). To demonstrate keyframing we analyzed a time-lapse image of CRISPR/Cas9modified U2OS cells that express EGFP-tagged NEMO from its endogenous gene locus in response to IL-1 (16). Formation of NEMO puncta in single cells were tracked by making keyframe adjustments to cell segmentation polygons ( Figure 4A; see also Movie S2).
Fluorescent properties of NEMO puncta were followed over the time-lapse to monitor time-courses for adaptive changes in NEMO puncta numbers in addition to distributions for fluorescent properties for NEMO puncta over time in each cell ( Figure 4B). By selecting appropriate parameters, such as the wavelet map level and limiting boundaries for puncta intensity or size, punctate structures can be accurately measured and curated in digital images to produce high-quality single-cell datasets.

DISCUSSION:
In this work we have shown that dNEMO is an effective tool for quantification of fluorescent puncta in fixed-cell and time-lapse images. The à trous wavelet algorithm progressively removes high frequency noise within fluorescent images and can be used to enhance puncta near the diffraction limit and larger. When compared with established methods, dNEMO performs with comparable or better localization accuracy, depending on the amount of noise in the source image, and is significantly faster. Although dNEMO is suitable for detection of relatively bright structures in epifluorescence images, model fitting may still be the preferred method for detection of structures with lower signal-tonoise (29,30). Keyframing in dNEMO provides an effective interface to curate single cell data and correct for systematic effects in imaging data. We demonstrate dNEMO using fixed-cell smFISH and live-cell enrichment of EGFP-NEMO to puncta and expect dNEMO will also excel at quantifying other fluorescence reporters, including components of the central dogma, protein assemblies, and bright vesicular structures.
Updates to dNEMO are expected to further reduce its runtime and enhance its capabilities. One notable limitation of the current dNEMO implementation is the disproportionate amount of overhead dedicated to the over-segmentation algorithm. We expect that these can be mitigated through updates for parallelization of the oversegmentation process or by modifications to the watershed algorithm that reduce the computational expense while maintaining accuracy. Beyond runtime improvements, one of the largest bottlenecks is the manual segmentation of cells. We are actively considering experimental methods for labeling and incorporating an automated cell segmentation approach into dNEMO, either directly or through a plug-in system where users can choose their own cell segmentation method. Finally, we are also streamlining dNEMO data structures for compatibility with existing single-particle tracking packages (31), so that time-varying properties of single puncta can be tracked and associated with singlecell responses.
Tools dedicated to the processing of biological images have enabled many studies of single cell variability and dynamics, and contributed to the discovery of emergent cellular properties. dNEMO fills a gap in the scientific community by providing a simple workspace for users to interact with biological puncta in fluorescence microscopy images that are central to fundamental cellular processes. The software is controlled with a MATLAB user interface or as a stand-alone executable, and is available as Supplementary Software or at https://github.com/recleelab along with a user manual and test data used to generate the figures in this article.

The dNEMO user interface
The user interface is a MATLAB-based application which provides several means of interaction with single-channel images and movies. Users load a given image or movie into the application. The two overarching processes of the application, cellular

The à trous wavelet transform
Implementation of the à trous wavelet algorithm is largely adapted from Izeddin and colleagues (22). Briefly, the raw image is convolved with a matrix to create a wavelet map of the initial image. The L1 kernel is initially populated with values supplied by the third  properties of interest are measured as defined by the user, including centroid location, size, and intensity of the punctum, among others (see guide packaged with software).

Local background correction
Background correction is performed locally for individual puncta in the source image. A binary mask is created representing the regions identified by the à trous wavelet transform and dilated by a user-defined number of pixels (Supplemental Figure 3). Pixel intensity values are collected from the annular ring around each punctum and used as the punctum's local background. An additional user-defined parameter can be assigned to offset the inner diameter of the annular ring. The buffer region excludes the background pixels that immediately surround the punctum and may contain out-of-focus light from the fluorescent source. The background pixel values measured in the annular ring for a punctum are averaged and subtracted from each pixel identified within the punctum in the source image.

Cell segmentation
Segmentation of individual cells is performed by the user using an interactive polygon tool. This polygon can be further adjusted by the user in subsequent frames to account for morphology changes and cell movements over time. Cell segmentation uses the keyframing approach described above.

Simulated images with diffraction-limited objects
Three-dimensional matrices of size 512 x 512 x 64 (X, Y, Z) were populated with zeros.

Comparison with FISH-quant results
Simulated images containing theoretical point spread functions (PSFs) approximating diffraction-limited objects of known coordinates and intensities were analyzed separately using dNEMO and FISH-quant (27). Puncta were identified in both dNEMO and FISH-quant using the tools available in the "Spot Filter" or "Spot Detection" components, respectively. Additional post-processing of the puncta identified using FISH-quant was performed using the "Thresholding" tool supplied in FISH-quant. Puncta were considered successfully identified if the measured centroid was within a Euclidean distance of 2 of the true centroids. Error rates for puncta localization accuracy were determined as the simulated image was subjected to Gaussian noise of increasing standards of deviation (Supplemental Figure 4).

Comparing smFISH transcripts identified in raw and deconvolved images
NFKBIA transcripts detected by smFISH in HeLa cells were obtained from a previous study (25) and were deconvolved with SoftWoRx using hardware specifications for the DeltaVision microscope (Applied Precision, GE Healthcare Life Science) used for the original image acquisition. Both images were analyzed with dNEMO and mean intensities for the same puncta were compared between images. We show both the uncorrected mean intensity and mean intensity corrected for local background (Supplemental Figure   4e). The R 2 value for identified puncta is improved significantly (Supplemental Figure 4f), demonstrating the accuracy for measurement of background-corrected puncta.

SUPPLEMENTAL INFORMATION:
Supplemental information includes four figures and two movies can be found with this article online.

ACKNOWLEDGMENTS
We thank Sanjana Gupta, Chaitanya Mokashi, and David Schipper for helpful discussions. We also thank Suzanne Gaudet for the use of smFISH images. This work was funded by NIH grant (R35-GM119462) to R.E.C.L.