ANÁLISIS DE VALORACIONES Y DEMANDAS DE LA POBLACIÓN EN MATERIA DE SERVICIOS DE LOS
III.3. PERCEPCIÓN DE LOS USUARIOS SOBRE LOS SERVICIOS DE LOS ECOSISTEMAS DE LA
2.2.6.1 Tissues cluster by cell type and developmental stage
Quantifying the cell-type specific regulatory activity that influences
procreative fitness is of central interest to researchers in human epigenetics. However, quantification efforts have generally been limited to assessments of a small collection of properties such as transcription, chromatin accessibility, or enhancer activity. To
investigate the relationship between activity in differing cell types, we explored a clustering of 115 cell types based on differences in scores across all autosomal genomic positions in cell-type pairs (Figure 2.7). The combination of a broad-based quantification of activity based on selective pressure with cell-type sensitivity
provides a unique view into similarities and dissimilarities among cell types. Distance between cell types is measured by summing the absolute value of scores across all positions in the genome to form an L1 or “Manhattan” distance. As the score at each position represents a probability of being under selective pressure, the natural distance metric is in units of expected number of sites under selection (SUS) across the hg19 reference autosome. The minimum extent of any cell type in this space was
223,019,006 SUS and the maximum 264,070,590 SUS, with a median of 234,392,692 SUS. The minimum of the 6,555 distances between pairs of differing cell-types was 11,433,006 SUS and the maximum 61,602,121 SUS. The distance matrix was clustered using the default Ward-D2 clustering method on the R package V3.3.1. Projection from a 115-dimensional space to a 2-dimensional tree induces some distortion, making small distances appear smaller and large distances appear larger; however, some important relationships are clearly visible.
Figure 2.7: Tissue and developmental states cluster by FitCons2 scores. Clustering of 115
Roadmap cell types by variation in FitCons2 scores across all genomic positions shows strong tissue specificity. Immune and blood cell types cluster together (gold) and among them T-cells (outset, lower) group particularly tightly. Brain and neural cells (blue) show similar patterns of activity, and in particular fetal brain tissues (blue outset, bottom) cluster separately from adult tissues (blue outset, top). Similarly, fetal organ tissue (red outset) cluster within the broad cluster of internal organ tissues (red). Digestive tissue samples are shown in a grey outset, while the purple cluster contains skin and mesenchymal cell types. Tissue replicates of Fibroblasts, Keratinocytes and Melanocytes are immediately adjacent (purple outset, curly braces). Embryonic cells and induced pluripotent stem cells cluster tightly (green, outset) within the broader grouping of less differentiated progenitor cell types (green). While the projection to 2 dimension distorts distances, this figure has a natural distance scale in units of ∑𝑖∈ℎ𝑔19|ΔINSIGHT 𝜌|𝑖 , that is, hg19 position-summed differences in expected sites under selection. Broadly, this allows changes in epigenetic properties between differing tissue types within a single organism to be interpreted in a scale of selective pressure concordant with evolutionary turnover.
Cell types gather into clusters readily associated with tissue types,
corresponding to embryonic (green), blood (gold), neural (blue), connective tissue (purple) and internal organs (red & gray). As expected, replicate cell-types such as the pairs of fibroblasts, keratinocytes and melanocytes are immediately adjacent, showing highly similar patterns of FitCons2 scores (purple outset, braces). Strikingly, some cells cluster by differentiation stage while others cluster by related organ. Thus, induced pluripotent stem cells show similar patterns of genome wide activity as embryonic stem cells (green), while brain and neural cells cluster together regardless of developmental stage (blue). However, the induced H9 Neuronal Precursor cell line clusters with ESC’s, while within the neural tissue cluster embryonic brain tissues cluster together (blue outset, at bottom) and separately from adult brain tissues (blue outset, at top). The fetal neural tissue in the Brain & Neural cluster was sampled at 17- 20 weeks after gestation, suggesting a regulatory phase change between embryonic stem cells and differentiated fetal neural cells before 17 weeks of age. Similarly, fetal organ (red, outset) and fetal digestive tissues (grey outset, right) cluster together within their respective tissue types.
2.2.6.2 Identification of differentially active enhancers
We evaluated the ability of FitCons2 to distinguish between active and inactive states of specific regulatory elements by comparing a shared set of 1,026 FANTOM5 enhancers covering 375,480 genomic positions with differential activity among three cell types, GM12878, HUVEC, and H1 hESC. FANTOM571 enhancers are small loci (mean 366 bp, st. dev. 201 bp) identified using Cap Analysis Gene Expression88 (CAGE) to locate regions with divergent transcription that is associated with
enhancers. Differentially active enhancers are defined as those among the top 10% of CAGE read depths in at least one of the three cell types (active), and zero read depth in at least one of the three cell types (inactive). Enhancers that are neither active nor
inactive in a particular cell type are removed. For each cell type, the mean FitCons2 score for enhancers in each class is calculated using the relevant cell-type specific scoring. The mean score across positions for active elements in a cell type was consistently higher than the mean score across inactive elements (Figure 2.8), demonstrating sensitivity to the more highly active enhancers in each cell type.
To further characterize the predictive power in cell-type specific FitCons2 scorings, the distribution of scores for each combination of cell-type and activity state was generated and used to produce three ROC plots. These plots comparing the relative discriminative power of the same-cell scoring with off-cell scoring of the cell- type sensitive differential enhancers (Figure 2.9). In each plot, the true positives are positions in an active enhancer, while the true negatives are positions in an inactive
Figure 2.8: Tracking enhancer activity across cell-types. Mean FitCons2 scores of
differentially active enhancer across three cell types: GM12878 (gm), HUVEC (hu) and H1 hESC (h1). Mean cell-type specific scores of active enhancers are uniformly higher than mean scores of enhancers inactive in the same cell-type. The same set of enhancers is used in all three cell types, but the activity of each individual enhancer varies by cell-type.
enhancer. Each enhancer is active in at least one of the three cell types, and inactive in at least one of the others. In each case the same-cell score provided a higher AUC, indicating a better predictive accuracy for active enhancers. The mean same-cell AUC was 0.70, while the mean off-cell AUC was 0.46, no better than random assignment.
Figure 2.9: ROC plots showing FitCons2 scores tracking enhancer actitivy across cell- types. For a single collection of differentially active enhancers, positions in enhancers active
in a particular cell-type are the true positive, while positions in inactive enhancers are the true negatives. Each of the three panels represents FANTOM5 enhancer activity in one cell-type (GM12878, HUVEC, or H1 hESC), but is assessed using scores from each of the three cell types. Scores from cell-types matching the active enhancer cell-type show predictive power, with AUCs of 0.61 to 0.72. Scores from unmatched cell-types show little predictive power with AUCs of 0.37-0.55. The score variation used to discriminate active from inactive enhancers positions occurs at intermediate false positive rates because enhancers are relatively diffuse structures with a variable density of active positions and contain many low scoring positions even in the active state.