Ure 1) to synthesize 3D images for many values of the model parameters,Comparison of Microtubule Distributionsand then to pick the image that best matches the given real image (and thus to (-)-Calyculin A biological activity estimate the parameters that could have produced it). Our original method utilized 3D images, but 3D images of 1326631 intact whole cells are much less commonly available than 2D images. We therefore describe here a method of estimating 3D microtubule model parameters from 2D image fluorescence microscopy images of tubulin. We test our approach on the 3D images of HeLa cells previously used to develop the model, and then use it to compare microtubule distributions in different cell lines. Figure 2 provides an overview of the framework introduced in this paper. There are two sub-systems. One is for generating synthetic images of microtubules, and the other is for estimating the microtubule model parameters for real images through comparison with the synthetic images. We first obtained 2D fluorescence microscopy images for eleven cell lines. Each image contains two channels, one for microtubule staining and the other for nuclear staining. The images are segmented to find individual cell and nuclear boundaries. For each cell, we estimate a Point Spread Function (PSF), centrosome location and single microtu-bule intensity. On the basis of the segmented 2D cell and nuclear shapes, approximate 3D cell 1379592 and nuclear morphologies are generated. Given the model (Figure 1) and ranges of allowed values of its parameters (number of microtubules (N), mean of the length distribution (mu) and CP21 biological activity collinearity (a)), synthetic images of microtubule distributions are generated for each 3D morphology for each combination of allowed parameter values. Each raw synthetic image is then convolved with the estimated PSF and multiplied with the estimated single microtubule intensity to make it comparable to the real image. Numerical features are then calculated on every real cell image and the synthetic images for it. The matching method then selects one set of parameters for which the synthetic image is the closest to the real image in the feature space. Using this indirect method, we estimate the model parameters for 2D images from eleven human cell lines, and analyze the resulting parameters.Results 3D Cell and Nuclear Shape Generation from a 2D Slice of Microtubule Channel and Nucleus ChannelIn our earlier work, we described an indirect approach to estimate parameters of a generative model of microtubules that was conditioned on the shape of the cell and the nucleus [8]. These shapes were estimated from a 3D confocal stack of images of a total protein stain and a DNA stain respectively. Since the images we analyze in this paper are only 2D slices, we developed an approach to estimate an approximate 3D shape of a cell and nucleus from a 2D slice (purely for the purpose of being able to generate a synthetic microtubule distribution). The location of the centrosome was also estimated (see Methods). Figure 3 shows an example of microtubule and nucleus images and the resulting approximate 3D cell and nucleus shape models (see details in the section of “3D cell and nuclear morphology generation” in Methods). We also describe a method to detect the 3D coordinate of the centrosome from the microtubule image using a two step approach (see Methods). These models and centrosome location were then used to generate microtubules in the cytosolic space.Recovering 3D Microtubule Generative Model Para.Ure 1) to synthesize 3D images for many values of the model parameters,Comparison of Microtubule Distributionsand then to pick the image that best matches the given real image (and thus to estimate the parameters that could have produced it). Our original method utilized 3D images, but 3D images of 1326631 intact whole cells are much less commonly available than 2D images. We therefore describe here a method of estimating 3D microtubule model parameters from 2D image fluorescence microscopy images of tubulin. We test our approach on the 3D images of HeLa cells previously used to develop the model, and then use it to compare microtubule distributions in different cell lines. Figure 2 provides an overview of the framework introduced in this paper. There are two sub-systems. One is for generating synthetic images of microtubules, and the other is for estimating the microtubule model parameters for real images through comparison with the synthetic images. We first obtained 2D fluorescence microscopy images for eleven cell lines. Each image contains two channels, one for microtubule staining and the other for nuclear staining. The images are segmented to find individual cell and nuclear boundaries. For each cell, we estimate a Point Spread Function (PSF), centrosome location and single microtu-bule intensity. On the basis of the segmented 2D cell and nuclear shapes, approximate 3D cell 1379592 and nuclear morphologies are generated. Given the model (Figure 1) and ranges of allowed values of its parameters (number of microtubules (N), mean of the length distribution (mu) and collinearity (a)), synthetic images of microtubule distributions are generated for each 3D morphology for each combination of allowed parameter values. Each raw synthetic image is then convolved with the estimated PSF and multiplied with the estimated single microtubule intensity to make it comparable to the real image. Numerical features are then calculated on every real cell image and the synthetic images for it. The matching method then selects one set of parameters for which the synthetic image is the closest to the real image in the feature space. Using this indirect method, we estimate the model parameters for 2D images from eleven human cell lines, and analyze the resulting parameters.Results 3D Cell and Nuclear Shape Generation from a 2D Slice of Microtubule Channel and Nucleus ChannelIn our earlier work, we described an indirect approach to estimate parameters of a generative model of microtubules that was conditioned on the shape of the cell and the nucleus [8]. These shapes were estimated from a 3D confocal stack of images of a total protein stain and a DNA stain respectively. Since the images we analyze in this paper are only 2D slices, we developed an approach to estimate an approximate 3D shape of a cell and nucleus from a 2D slice (purely for the purpose of being able to generate a synthetic microtubule distribution). The location of the centrosome was also estimated (see Methods). Figure 3 shows an example of microtubule and nucleus images and the resulting approximate 3D cell and nucleus shape models (see details in the section of “3D cell and nuclear morphology generation” in Methods). We also describe a method to detect the 3D coordinate of the centrosome from the microtubule image using a two step approach (see Methods). These models and centrosome location were then used to generate microtubules in the cytosolic space.Recovering 3D Microtubule Generative Model Para.