Fig 1.
(a) Summary of model learning and classification pipeline. (b) Illustration of coordinate system for probability density function. For each pixel in an image, distance between it and the nearest point on the nuclear membrane (L1) and between it and the nearest point on the cell membrane (L2) are calculated and used to calculate the radial position (r) as L1/(L1+L2). In addition, the distance to the nearest point on a segmented microtubule (d) and the angle between the pixel and the major axis of the cell (α) are calculated. (c) A two-color image of a vesicular protein (TFRC, transferrin receptor, green) and microtubules (red) in a U-2OS cell. (d) Segmented image of microtubules (red) and puncta (green). (e) Remaining background intensity.
Table 1.
Proteins used to define punctate subpatterns in this study.
Fig 2.
Distribution of cells of the combinations of proteins and cell lines in the first three principal components learned from the whole feature space by PCA.
The number for each point indicates the protein index and the color indicates the cell line (red for A-431, green for U-2OS and blue for U-251MG). The gray ellipses represent the scope of 1.5 standard deviations, which contain about 50% to 80% of cells. The arrows summarize the composition of each principal component by showing the direction in which each feature increases (see S1 Table for the list of features). The left panel shows the first and second principal components while the right panel shows the second and third principal components. Feature projections with a magnitude less than 0.1 were removed for visualization purposes.
Table 2.
Ability to distinguish 11 punctate classes.
Classifiers were trained using 5-fold cross validation, and the class of the held out image was predicted. Results are shown for classifiers constructed using all features, and values in parentheses are for training without the microtubule features.
Table 3.
Top-ranked proteins assigned to one of the ten high-confidence subpatterns.
The top protein for each cell type for each subpattern (except Coated Pits) is included if its separability is less than 0.70 (which is more selective than the threshold determined in S2 Fig). The separability measures for all proteins are included in S1 Dataset.
Fig 3.
Comparison of model features for different patterns.
The values for each feature were z-scored to put them on the same scale across features, and the average value for each feature is shown as a function of the feature number (see S1 Table for feature definitions).
Fig 4.
Graphical model representation for the Bayesian hierarchical framework of generative model of puncta conditioned on cell geometry and microtubules.
A nuclear shape is drawn from dn, a cell shape is drawn from dc, dependent on the nuclear shape [8]. A microtubule pattern is synthesized from dm dependent on the generated cell and nuclear shape [9]. The distribution of shape and positions of puncta, dp, is modeled with components pp, which models the position of puncta dependent on the cell, nucleus and microtubule pattern, and np, sp and ip, which independently model the number, size and intensity of puncta. The background pattern is similarly generated dependent on the cell, nucleus and microtubule pattern with pb, and its intensity is determined with ib.
Fig 5.
Representative images from four patterns and corresponding synthesized images in U-2OS cells.
The left column shows cell images closest to the median of parameter space for cells of that pattern, and the right column shows synthesized cells from the generative model of protein pattern conditional on cell geometry and microtubules of the left panel. The green, red and blue channels represent puncta, microtubules and nuclei, respectively.
Fig 6.
Synthetic cell image containing eleven punctate patterns.
Synthetic distributions for all patterns were independently created in the same cell; this assumes that positions of puncta do not affect each other (e.g., that peroxisomes are not more or less likely to be near RNP bodies). The nucleus is shown in dark grey and microtubules in light gray. Colors for patterns are the same as in Fig 3.