Figure 2. Landscape of human antibody class switching.
(A) State transition diagram of class switching. Classes are indicated as circles and possible switches as arrows. The radius of each circle indicates the relative abundance of the labeled class. The width of each arrow indicates the relative frequency of the switch (also reported in Table 3). Rare classes IgG4 and IgE have been omitted for clarity and are shown in Figure 2—figure supplement 8A. (B) Penetrance of direct switches from IgM/IgD. For each class, the fraction of sequences created by direct switching from IgM is shown (mean ± s.d. across n = 22 subjects for Sample and n = 14 subjects for Bio. Rep.). (C) Rates of CSR. The rate constant of each switch path was estimated by fitting an exponential probability distribution to the distribution of the number of somatic mutations accumulated prior to CSR (Figure 2—figure supplement 11). Distributions of rate constants for switch paths from IgM/IgD to activated classes (gray) and from an activated class to another activated class (white) having ≥500 examplesin both Sample and Bio. Rep. repertoires are shown.
DOI: http://dx.doi.org/10.7554/eLife.16578.013
Figure 2—source data 1. Counts of class switch events. .
Number of events observed for each possible switch from the class indicated by the row to the class indicated by the column. Data from D0 and D28 are provided separately. These data were used to calculate the switching rates depicted in Figure 2A.
elife-16578-fig2-data1.zip (581B, zip)
DOI: 10.7554/eLife.16578.014
Figure 2—figure supplement 1. Patterns of class switching measured using sequences with identical VDJ sequences but different constant regions are highly similar to those measured using the full lineage reconstruction approach.
(A) Origin of pairs of sequences having identical VDJ sequences but different constant region classes. PCR recombination artifacts (PCR chimeras) were detected by comparing the unique barcodes from each sequencing read. Specifically, a pair of sequences was identified as originating from PCR chimera if at least one V-region barcode was shared between the pair of sequences, accounting for ~5% of sequence pairs. (B) Landscape of antibody class switching measured using only pairs of sequences having identical VDJ sequences but different constant region classes, which did not originate from PCR chimeras. Top panels show the relative frequency of class switch events from the class indicated by the column to the class indicated by the row. Middle panels show the destination probability, which is the probability that a given sequence of class indicated by the column switches to the class indicated to the row. Bottom panels show the arrival probability, which is the probability that a given sequence of class indicated by the row arose via direct switch from the class indicated by the column. Sample (left) and biological replicate (right) are shown. (C) Comparison between the landscape of antibody class switching measured using only pairs of sequences having identical VDJ sequences but different constant region classes and the landscape measured using the full lineage reconstruction approach. The values that define the landscape (relative switch frequencies, destination probabilities, and arrival probabilities) are plotted against the values obtained using all parent-child sequence pairs. Squared Pearson correlation coefficient is shown.
Figure 2—figure supplement 2. Patterns of class switching measured using sequences inheriting all germline mutations from parent are highly similar to those measured using the full lineage reconstruction approach.
(A) Landscape of antibody class switching measured using only sequences inheriting all germline mutations from parent. Top panels show the relative frequency of class switch events from the class indicated by the column to the class indicated by the row. Middle panels show the destination probability, which is the probability that a given sequence of class indicated by the column switches to the class indicated to the row. Bottom panels show the arrival probability, which is the probability that a given sequence of class indicated by the row arose via direct switch from the class indicated by the column. Sample (left) and biological replicate (right) are shown. (B) Comparison between the landscape of antibody class switching measured using only sequences inheriting all germline mutations from parent and the landscape measured using the full lineage reconstruction approach. The values that define the landscape (relative switch frequencies, destination probabilities, and arrival probabilities) are plotted against the values obtained using all parent-child sequence pairs. Squared Pearson correlation coefficient is shown.
Figure 2—figure supplement 3. Patterns of class switching measured using sequences supported by consensus reads are highly similar to those measured using the full lineage reconstruction approach.
(A) Landscape of antibody class switching measured using only sequences supported by consensus reads formed from ≥3 sequencing reads. Top panels show the relative frequency of class switch events from the class indicated by the column to the class indicated by the row. Middle panels show the destination probability, which is the probability that a given sequence of class indicated by the column switches to the class indicated to the row. Bottom panels show the arrival probability, which is the probability that a given sequence of class indicated by the row arose via direct switch from the class indicated by the column. Sample (left) and biological replicate (right) are shown. (B) Comparison between the landscape of antibody class switching measured using only sequences supported by consensus reads and the landscape measured using the full lineage reconstruction approach. The values that define the landscape (relative switch frequencies, destination probabilities, and arrival probabilities) are plotted against the values obtained using all parent-child sequence pairs. Squared Pearson correlation coefficient is shown.
Figure 2—figure supplement 4. Landscape of class switching cannot be explained by random switching in proportion to the abundance of antibody classes.
(A) Landscape of antibody class switching measured after shuffling parent-child pairs of sequences. Top panel shows the relative frequency of class switch events from the class indicated by the column to the class indicated by the row. Middle panel shows the destination probability, which is the probability that a given sequence of class indicated by the column switches to the class indicated to the row. Bottom panel shows the arrival probability, which is the probability that a given sequence of class indicated by the row arose via direct switch from the class indicated by the column. (B) Comparison of the landscapes of antibody class switching before and after shuffling parent-child pairs of sequences. The values that define the landscapes (relative switch frequencies, destination probabilities, and arrival probabilities) are plotted against each other. Squared Pearson correlation coefficient is shown.
Figure 2—figure supplement 5. Landscape of class switching saturates with respect to sequencing depth.
Rarefaction analysis of class switching landscapes of five subjects.Sequencing reads were sampled to varying depth, and the class switching landscape was measured in each case, the values of the relative switch frequency are plotted. The relative switch frequency is obtained by dividing the number of switches for a given transition by the maximum number of switches observed for any transition. For each subject, 5 replicate subsamples were performed at each depth, and the values obtained in these replicates are indicated by points, while the line connects medians of the replicates.
Figure 2—figure supplement 6. Rarefaction analysis indicates that switch intermediates are robustly detected.
Sequencing reads were subsampled to varying depth for the five subjects shown in Figure 2—figure supplement 5 with five replicate subsamplings at each depth. Data from all five subjects was pooled and used to calculate the fraction of switches from A to C indicated by the title of each panel that were direct (A -> C) and indirect (A -> B -> C). Median across replicates is indicated by the red line.
Figure 2—figure supplement 7. Class switching landscape is not sensitive to the lineage clustering cutoff parameter.
Clustering to identify clonal lineages of antibodies was performed on all repertoires from D0 with varying values of the clustering cutoff parameter ranging from 0.80 to 0.95. The class switching landscape was then calculated. In this calculation, we included only lineages having ≤2500 sequences in every parameter setting to ensure computational tractability. The landscape in each case is plotted against the landscape measured when the cutoff is 0.95. Squared Pearson correlation is shown.
Figure 2—figure supplement 8. Landscape of class switching in humans.
(A) Class switch state transition diagram including the rare classes IgG4 and IgE. Classes are indicated as pies and possible switches are indicated as arrows. Radius of each pie indicates the relative abundance of the class. The width of each arrow indicates the relative frequency of the switch (also reported in Table 2). (B) Heatmaps showing the class switch landscape as an average across subjects. Top panel shows the relative frequency of class switch events from the class indicated by the column to the class indicated by the row. Middle panel shows the destination probability, which is the probability that a given sequence of class indicated by the column switches to the class indicated to the row. Bottom panel shows the arrival probability, which is the probability that a given sequence of class indicated by the row arose via direct switch from the class indicated by the column. Sample (left) and biological replicate (right) are shown.
Figure 2—figure supplement 9. Comparisons of class switch landscapes across individuals.
(A and B) Differences between class switch landscapes (measured by Jensen-Shannon distance). Distances were calculated between the vectors representing (A) relative switch frequency or (B) destination probability. Lane 1 compares the two biological replicates for each subject. Lanes 2 and 3 compare pairs of subjects (identical twins or unrelated individuals). Median is indicated by red line. (C) Comparison of class switching landscapes of identical twins and unrelated pairs of subjects. Destination probabilities of identical twin pairs (red) and all possible pairs of unrelated subjects (blue) are plotted against one another. Intraclass correlation coefficient (ICC) for twins and unrelated pairs was calculated using bootstrap resampling of pairs of subjects (1000 replicates) and reported in the legend (5th to 95th percentile range). (D) Comparison of the relative switch frequency of all possible class switches across subjects. Each point indicates the relative frequency of the switch indicated on the x-axis for an individual subject. Median is indicated by red line. Relative frequency of all switches is similar across all subjects.
Figure 2—figure supplement 10. Class switching landscapes of individual subjects.
(A) Relative switch frequency, (B) destination probability, and (C) arrival probability are shown for each subject. Twins are shown on the same row and zygosity is indicated by row label.
Figure 2—figure supplement 11. Measurement of rates of class switching.
(A) Motifs analyzed to characterize the rate of CSR. (B) Distributions of mutations accumulated prior to class switching. Switches from IgM/IgD to activated classes (IgG, IgA, IgE) are plotted separately from switches between activated classes, as indicated by color. The p value of Kolmogorov-Smirnov test, two-sample comparing these two distributions is shown. (C) Cumulative probability of class switching as mutations accumulate. Origin and destination of class switch are indicated by color. (D) Rate constants of class switching along all switch paths where we observed >250 direct switches. Exponential distributions were fitted to the distributions of the number of mutations accumulated prior to class switching (see examples in panel E) and the rate constant was extracted. (E) Examples of exponential distributions (modified to have an additional parameter for non-zero y-intercept [CDF(x) = 1 – exp(-ax) + b]) fitted to the empirical distributions of the number of mutations accumulated prior to class switching. Fit was performed using the curve_fit function in the scipy.optimize module in Python, which implements the Levenberg-Marquardt nonlinear least squares algorithm. Rate constants of the fitted exponential distributions are shown in panel D.