CHAPTER 3. HMX reveals tertiary interactios within multiple stable substructures of the
4.5. GUIDELINES FOR OF RING-MAP DATA ANALYSIS 93
4.5.3. Spectral clustering Analysis 99
Spectral clustering analysis allows for the identification of multiple native structures in a single pool of RNA. The ability to accurately identify and separate distinct
conformations by spectral clustering analysis can be determined by three criteria; the eigengap values, fraction of the minor cluster, and the differences between the cluster profiles.
4.5.3.1.Eigengap value analysis
Determining the number of clusters that can be definitively separated is an important first step in evaluating clustering data. The quantity and quality of cluster separation can be determined by evaluating the eigengap plot. If a sample has multiple, well-defined clusters the sample will have an eigengap of 0.03 or greater. The number of possible clusters can be determined by the position of the eigengap. A sample with two clusters will have an eigengap greater than 0.03 in the second position, three clusters will have an eigengap greater than 0.03 in the third position, and so on. Figure 4.14 outlines the expected eigengap profiles for samples with multiple clusters.
4.5.3.2.Fraction of the minor cluster
The clustering algorithm partitions all sequences into a derived cluster and so the fraction of each cluster in an RNA sample can be determined. However, reactivity profiles for small fractions cannot be reliably determined. In this study, fractions of less then 10% were not considered well-resolved clusters. If a sample has a low population cluster it should not be considered when producing the reactivity profiles for different clusters.
4.5.3.3.Difference in cluster profiles
It is also important to compare the resulting profiles from the separated clusters, especially when eigengap values are close to 0.03. Clusters, which represent distinct
structures, can be identified when reactivity at certain positions are preferentially represented in a single cluster. It is important to note that a sample with poorly defined clusters will still produce reactivity profiles for each cluster. While poorly defined cluster profiles may show small differences in reactivity, no position will be exclusively represented in either cluster.
4.5.3.4.Analyzing the TPP riboswitch.
The analysis of the TPP riboswitch with decreasing levels of structure serves as a good example of how clustering data should be analyzed. The eigengap values indicate that the saturating ligand sample has potentially thee clusters, the no-ligand sample has two, and the no Mg2+ sample has one cluster (Fig. 4.18A). The smallest cluster of the saturating ligand sample is only 5%, so only two reactivities profiles can be generated for this sample (Table 4.1). For the no-ligand and no Mg2+ samples, the eigengap values are just above and below
the eigengap threshold, respectively. As a result, the differences between the resulting profiles should be compared to determine if each cluster represents a distinct conformation (Fig. 4.18B). The no ligand sample has a few nucleotides, indicated by blue circles, which are exclusively represented in a single cluster. This verifies that each profile represents a distinct conformation. The no Mg2+ sample produces profiles in which position both increase
and decrease in reactivity between cluster profiles but no positions are exclusively
Figure 4.18 Determining number of clusters in TPP samples using eigengap values and cluster profiles. (A) Eigengaps of saturating ligand, no ligand and no Mg2+ indicate that samples have 2,1 and 0 clusters respectivly. How well samples are seprated can be
determined by how above or below the eigengap threshold egengap values are. Samples with multiple positons above the egiengap threshold determine number of clusters. (B) reactivity profiles for culsters derived from TPP samples. The major cluster (Red) is compaired to minor cluser (Blue). The more difinitivly separated clusters have positions in which are preferentially represented in a single cluster (blue circles). The poorley separated, no Mg2+ sample produces two clusters with few major differences between reactive positions.
0 0.02 0.06 0.1 2 4 6 8 10 Results_TPP_noLig_L#3F9515A No*ligand 0 0.02 0.06 0.1 2 4 6 8 10 Results_TPP_noMg_lo#3F950FE No*ligand*+*no*Mg2+ 0 0.02 0.06 0.1 2 4 6 8 10 Results_TPP_+Lig_lo#3F951D9 Saturating*ligand 0 0.1 0.3 20 40 60 80 Frequency 0 0.1 0.3 20 40 60 80 Major*cluster*(81%) Major*cluster*(83%) Minor*cluster*(19%) Minor*cluster*(17%) 0 0.1 0.2 20 40 60 80 Major*cluster*(71%) Minor*cluster*(29%) Eigengap Eigengap*number Sequence*position A B Eigengap*threshold
4.5.3.5.Additional criteria for evaluating data for spectral clustering.
In addition to the general criteria set for all RING-MaP data, additional criteria must be considered to determine if data is suitable for clustering analysis. A mutation frequency threshold was used to set the number of positions used for clustering analysis. Only positions with a sufficiently high mutation rate were considered. In this work, only positions with a mutation frequency greater than 0.02 were used. The numbers of positions used in spectral clustering calculations for each RNA are outlined in the final column of Table 4.2. The number of positions used should reflect the number of single-stranded A and C nucleotides in an RNA. If the number of nucleotides analyzed does not fit this criteria for spectral
clustering, it may reflect a larger problem with the sample quality and one or more of the criteria used to determine the quality of RING-MaP data should be addressed.
RING-MaP provides data that can be analyzed to identity structurally correlated nucleotides and the presence of multiple structures within a pool of RNA. These analyses provide a unique view of RNA structure that has not been possible in a single experiment before. The criteria outlined here provide a simple set of guidelines to assist data analysis of new RNAs and an efficient way to evaluate data used for RING-MaP calculations.