Using different random seeds. The predefined parameters for the validation bed

Using different random seeds. The predefined parameters for the validation bed were get Gracillin estimated from each testing library separately. The resulting errors are shown in Table 2 and indicate that the system accurately recovers model parameters from 2D slices of synthetic images.Estimating Microtubule Parameters for Images of Eleven Cell Lines3D microtubule model parameters were estimated from 2D fluorescence microscopy images of eleven cell lines collected as described previously [1], with the application of the whole framework including library generation, feature calculation and matching (see Methods). This dataset consisted of 112 A-431 cells, 114 from U-2OS cells, 94 U-251MG cells, 38 RT-4 cells, 110 PC3 cells, 51 Hep-G2 cells, 35 HeLa cells, 77 CaCo2 cells, 66 A-549 cells, 70 Hek-293 cells and 54 MCF-7 cells. Figure 4 showsexamples of query images and the corresponding images synthesized using the parameters estimated from them. Note that the synthetic images are not “exactly” the same as the corresponding real ones in every single microtubule, because the goal of the generative models is to learn the underlying MedChemExpress SIS-3 distribution of microtubules from which the real images were drawn. Hence from Figure 4, we can see that synthetic images are similar to real ones in terms of the distribution of microtubules. There is an underlying assumption that the cells from the same cell line tend to have some level of consistency in the distribution of microtubules. Therefore, we measured the coefficient of variation (the ratio of the standard deviation to the mean) for the estimated parameters of real cells. The resulting values for the number of microtubules ranged from 0.28 to 0.60 and from 0.21 to 0.43 for the mean of the length distribution. We show the frequency distribution of each of the three parameters for every cell line in Figure 5. It shows that most of the cell lines have quite close to a normal distribution for both the number of microtubules and mean length. Some may deviate a little to have a Gamma distribution-like shape. For the collinearity, due to the computational efficiency, we only used three candidate values in the library of synthetic images, so we cannot draw any significant conclusions. The scatter plot of the two dimensional parameter space (number of microtubules and mean length) estimated from those cell lines is shown in Figure 6. The plot shows the variation in number of microtubules, in mean length and in joint correlation of the two. We will compare them in the next section.Comparison of Microtubule DistributionsFigure 3. Generation of 3D cell geometry (cell shape and nuclear shape) from real 2D slices of the microtubule and nucleus channels. (A) Example of a real 2D cell image (tubulin channel) and its approximate bottom shape. (B) Cartoon of an X-Z projection of a cell on a substrate. (C) Example of a generated 3D cell shape containing 8 stacks (1.6 microns). (D) Illustration of inputs and outputs for the procedure. doi:10.1371/journal.pone.0050292.gComparing Microtubule Distributions Across Eleven Cell LinesComparing bivariate distributions of the number of microtubules and the mean of length. We compared thebivariate distribution of the estimated number of microtubules and the mean of length across different cell lines. We first compared the covariances using Box’s M test. The p-value for this comparison was<0 which indicates that we can readily reject the null hypothesis of homogeneity of covariances. Next, we u.Using different random seeds. The predefined parameters for the validation bed were estimated from each testing library separately. The resulting errors are shown in Table 2 and indicate that the system accurately recovers model parameters from 2D slices of synthetic images.Estimating Microtubule Parameters for Images of Eleven Cell Lines3D microtubule model parameters were estimated from 2D fluorescence microscopy images of eleven cell lines collected as described previously [1], with the application of the whole framework including library generation, feature calculation and matching (see Methods). This dataset consisted of 112 A-431 cells, 114 from U-2OS cells, 94 U-251MG cells, 38 RT-4 cells, 110 PC3 cells, 51 Hep-G2 cells, 35 HeLa cells, 77 CaCo2 cells, 66 A-549 cells, 70 Hek-293 cells and 54 MCF-7 cells. Figure 4 showsexamples of query images and the corresponding images synthesized using the parameters estimated from them. Note that the synthetic images are not ``exactly'' the same as the corresponding real ones in every single microtubule, because the goal of the generative models is to learn the underlying distribution of microtubules from which the real images were drawn. Hence from Figure 4, we can see that synthetic images are similar to real ones in terms of the distribution of microtubules. There is an underlying assumption that the cells from the same cell line tend to have some level of consistency in the distribution of microtubules. Therefore, we measured the coefficient of variation (the ratio of the standard deviation to the mean) for the estimated parameters of real cells. The resulting values for the number of microtubules ranged from 0.28 to 0.60 and from 0.21 to 0.43 for the mean of the length distribution. We show the frequency distribution of each of the three parameters for every cell line in Figure 5. It shows that most of the cell lines have quite close to a normal distribution for both the number of microtubules and mean length. Some may deviate a little to have a Gamma distribution-like shape. For the collinearity, due to the computational efficiency, we only used three candidate values in the library of synthetic images, so we cannot draw any significant conclusions. The scatter plot of the two dimensional parameter space (number of microtubules and mean length) estimated from those cell lines is shown in Figure 6. The plot shows the variation in number of microtubules, in mean length and in joint correlation of the two. We will compare them in the next section.Comparison of Microtubule DistributionsFigure 3. Generation of 3D cell geometry (cell shape and nuclear shape) from real 2D slices of the microtubule and nucleus channels. (A) Example of a real 2D cell image (tubulin channel) and its approximate bottom shape. (B) Cartoon of an X-Z projection of a cell on a substrate. (C) Example of a generated 3D cell shape containing 8 stacks (1.6 microns). (D) Illustration of inputs and outputs for the procedure. doi:10.1371/journal.pone.0050292.gComparing Microtubule Distributions Across Eleven Cell LinesComparing bivariate distributions of the number of microtubules and the mean of length. We compared thebivariate distribution of the estimated number of microtubules and the mean of length across different cell lines. We first compared the covariances using Box's M test. The p-value for this comparison was<0 which indicates that we can readily reject the null hypothesis of homogeneity of covariances. Next, we u.