Figure 1.
Schematic illustration of experimental setup for dataset construction for morphology-based prediction model construction.
Figure S1 shows the illustration of usage of the objective morphology-based prediction model, and its major technological achievements using this dataset. The initial sample (P1) was divided into three separate culture samples (SEED, PRE, and DIFF) at each passage. SEED samples were mainly used for the continuous-passage culture until termination of growth (P9). From the cell yield at each passage of the SEED samples, population doubling time (PDT) was calculated, and taken as the experimentally determined potential. DIFF samples derived from each passage were divided into three differentiation cultures (samples O, A, and C for osteogenic, adipogenic, and chondrogenic differentiation, respectively) and grown under the indicated conditions for 3–4 weeks. The differentiation values of samples O, A, and C were experimentally quantified by individual staining protocols. The staining results were then converted by image-processing analysis to obtain the experimentally determined differentiation potentials. The three types of differentiation potentials together with the population doubling potential (population doubling time: PDT) were designated as “multiple differentiation potentials” of the hBMSCs. PRE samples consisted of sample I (for imaging) and sample R (for RNA extraction). From sample I in each passages, phase-contrast image were acquired at 24 h intervals over 4 days. Acquired images were then converted by image processing to obtain morphological features from every cell in all images (see also Fig. S2 and S3 for the details of image processing). Morphological features were statistically processed to yield transformed morphological features through data cleansing and statistical calculations, and the results were used as the input features. Sample R were subjected to total RNA extraction for gene-expression analysis. Either or both morphological features or/and gene-expression data were combined (input parameters), and arranged with the experimentally determined potentials of the hBMSCs (output parameters) to constitute training data for construction of prediction models.
Figure 2.
Representative morphological images of continuously passaged hBMSCs.
Columns indicate passage numbers, indicated as P-number. Rows indicate hBMSC lot names. (A) Phase-contrast microscopic images (10×) prior to differentiation culture (sample I). Scale bar, 50 μm. (low-resolution cellular images shown in Figure S4) (B) Alizarin red staining after 2 weeks of osteogenic differentiation culture (sample O). Scale bar, 200 μm. (C) Oil red staining after 3 weeks of adipogenic differentiation culture (sample A). Scale bar, 200 μm. (D) Alcian blue staining after 4 weeks of chondrogenic differentiation culture (sample C). Scale bar, 200 μm. From P7–P9, near the termination of growth, differentiation samples could not be prepared for (B) and (C) because of the lack of cell numbers. In (D), when the pellet sizes were smaller than 200 μm, we declined to produce specimens from the sample on the grounds that the differentiation culture had not been successful.
Figure 3.
Quantified experimentally determined differentiation values and population doubling times of hBMSCs.
Green bar, Lot A; blue bar, Lot B; pink bar, Lot C. Passage numbers are indicated as P2–P9. (A) Bar plots of average stained areas of Alizarin red–stained samples (n = 6). (B) Bar plots of average stained areas of Oil red–stained samples (n = 6). (C) Bar plots of stained areas in Alcian blue–stained samples (n = 1), normalized by their pellet size. (D) Line plots of PDT. Green diamonds, Lot A; blue squares, Lot B; pink triangles, Lot C. Error bars indicate standard deviation (s.d.).
Figure 4.
Heat map of gene-expression transitions and passage numbers.
Genes were clustered by hierarchical clustering for indicating clusters that correlate to the passage number increases. The red boxed cluster is the cluster of genes that correlate to passage number within all cell lots, indicating non–patient-specific passage-related genes. The relationship between colors and normalized values of gene expression is illustrated in the explanatory heat map at lower right.
Figure 5.
Schematic illustration matrix of prediction feature data profile and usage concepts of prediction models.
Six types of morphological feature conversion methods are proposed as M-patterns. Briefly, M-patterns are numbered in order of the amount of efforts required to prepare for model construction. M-patterns 1–4 require four images at 24-hour intervals; M-pattern 5 requires two images each on days 1 and 4; and M-pattern 6 requires only one image on the first day. For parameters described as “linked”, each morphological feature is not only used as the data for each time point, but this information is also converted into the changing ratio between time points. For “non-linked” parameters, morphological features are used as they are. Averages, quintile points, and groups of distribution representatives were compared to find the best statistical parameter to represent the morphological features measured in all individual cells in an image. Therefore, M-patterns 1–4 were designed to increase the amount of information about cellular distribution for incrementing the heterogeneity of cells.
Figure 6.
Comparisons of prediction models.
For the detailed definition of “M-pattern”, see Material and Methods. Conceptual illustration of M-pattern appears in Figure 4. Potential I, osteogenic differentiation rate; potential II, adipogenic differentiation rate; potential III, chondrogenic differentiation rate; potential IV, PDT. Each matrix involving line plots consists of three columns, separated by dotted lines, representing differences among lots (Lots A, B, and C). In each column, horizontal axis represents passage numbers, from P2 on the left to P9 on the right. Upper number at the shoulder of each matrix indicates scaled error rate, i.e., the median value of prediction errors among all the samples, normalized by the experimental values. The lower number at the shoulder of each matrix indicates the correlation coefficient. Blue line plot represents the value of experimentally determined values. Red line plot represents the prediction values from the prediction models. Greater overlap between blue and red line plots and minimum differences across passages and lot differences corresponds to higher predictive performance, represented by lower scaled error rate and higher correlation coefficient.