Bio tcr bcr analysis repertoire visualization skills repertoire visualization

Version Compatibility

Reference examples tested with: matplotlib 3.8+, seaborn 0.13+, pandas 2.2+, numpy 1.26+; R circlize 0.4+, iNEXT 3.0+; VDJtools 1.2.1+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• R: packageVersion('<pkg>') then ?function_name to verify parameters
• CLI: <tool> --version then <tool> --help to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Note: matplotlib 3.9 removed plt.cm.get_cmap(name, N); use plt.get_cmap(name) and sample it, or matplotlib.colormaps[name].resampled(N). seaborn 0.13 deprecated bare palette= without hue=; assign hue= and legend=False.

Repertoire Visualization

"Plot my TCR/BCR repertoire" -> Render V-J usage, CDR3-length spectratype, clonal-space, clonal tracking, diversity/rarefaction, overlap, and similarity-network figures, and read each one correctly.

• CLI: vdjtools PlotFancyVJUsage, vdjtools RarefactionPlot (delegates plotting to R)
• R: circlize::chordDiagram (V-J), iNEXT::iNEXT + ggiNEXT (rarefaction/extrapolation)
• Python: matplotlib/seaborn for bespoke figures; numpy multinomial resampling for rarefaction; networkx + rapidfuzz for similarity networks

The governing principle: every figure inherits two upstream choices

A repertoire figure is only as valid as the numbers behind it, and those numbers depend on two decisions made before any plot is drawn:

1. The clonotype definition. A clonotype = CDR3 (nucleotide OR amino acid) + optionally V and J. Amino-acid CDR3 collapses convergent recombination (many nt rearrangements -> one aa clonotype), inflating apparent sharing and deflating richness; nt CDR3 is more conservative (Venturi et al. 2006 PNAS 103:18691). Every count on every axis shifts with this choice, so it must be stated in the figure and held constant across all samples in a comparison.
2. The sequencing depth. Richness, Shannon, clonality, and set-overlap (Jaccard, shared-clonotype counts) are strictly increasing functions of reads sequenced - the rare-clone tail never saturates. Diversity, rarefaction, and overlap figures are comparable across samples ONLY after depth normalization: downsample all samples to a common depth, or read rarefaction curves at a shared x. Comparing raw values across libraries of unequal depth measures sequencing effort, not biology (Chao et al. 2014 Ecol Monogr 84:45; Greiff et al. 2015 Genome Med 7:49). This caveat governs the diversity, rarefaction, and overlap recipes below.

Choosing the figure and reading it

Figure	What it reveals	How to compare correctly
V-J chord/circos	Combinatorial V-J pairing bias within one sample	Descriptive per sample; never compare raw usage across primer sets/platforms (primer bias masquerades as biology)
CDR3 spectratype	Clonal structure: Gaussian length distribution = polyclonal/naive, spikes = expansion	Weight by frequency to see expansions; by unique clonotypes to see underlying diversity
Clonal-space / proportion	Fraction of repertoire held by rare vs expanded clones	Bin by clone frequency (strata), not raw count; robust to depth if frequency-based
Clonal tracking	Expansion/contraction/persistence of clones over time	Downsample timepoints to common depth first; an "absent" clone is often a sampling zero
Rarefaction/extrapolation	Diversity comparison done right (interpolate to common depth)	Read all curves at a shared x; extrapolate at most 2-3x observed depth
Overlap heatmap	Pairwise repertoire similarity	Use Morisita-Horn (depth-robust) on depth-normalized samples; Jaccard is depth-biased
Similarity network	Clusters of related CDR3s (candidate specificity groups)	Structure depends entirely on the distance threshold; state it and test sensitivity

V-J usage: chord/circos and heatmap

The chord/circos shows which V and J segments pair, weighted by clonotype count, within one sample.

VDJtools route (delegates plotting to R; requires RInstall once):

vdjtools PlotFancyVJUsage -m metadata.txt output_dir/

R with circlize - a V-by-J count matrix becomes a chord diagram:

library(circlize)

plot_vj_chord <- function(clone_df) {
    vj_matrix <- table(clone_df$v_gene, clone_df$j_gene)
    chordDiagram(vj_matrix, transparency = 0.5, annotationTrack = c('grid', 'name'))
}

Python heatmap alternative (samples x V gene, or V x J for one sample) avoids a chord dependency and reads more quantitatively for many segments:

import seaborn as sns
import matplotlib.pyplot as plt

def plot_vj_heatmap(clone_df):
    vj = clone_df.pivot_table(index='v_gene', columns='j_gene', values='frequency', aggfunc='sum', fill_value=0)
    fig, ax = plt.subplots(figsize=(8, 6))
    sns.heatmap(vj, cmap='viridis', ax=ax)
    ax.set_title('V-J pairing frequency')
    return fig

CDR3 spectratype: length distribution reveals clonal structure

The spectratype is a histogram of CDR3 length. A roughly Gaussian, bell-shaped distribution indicates a diverse, polyclonal (naive-like) repertoire; skew or discrete spikes at particular lengths indicate clonal expansion(s). The classic immunoscope spectratype (and VDJtools CalcSpectratype) bins nucleotide length, where in-frame clones sit 3 nt apart and the periodicity is part of the readout; amino-acid length is a common simplification - state which is plotted. Weighting by read/UMI frequency shows expansions; weighting by unique clonotypes shows the underlying diversity - the two views can look opposite, so label which is plotted.

def plot_spectratype(clone_df, length_col='cdr3_length'):
    fig, ax = plt.subplots(figsize=(9, 5))
    bins = range(clone_df[length_col].min(), clone_df[length_col].max() + 2)
    ax.hist(clone_df[length_col], bins=bins, weights=clone_df['frequency'], color='steelblue')
    ax.set_xlabel('CDR3 length (aa)')
    ax.set_ylabel('Frequency')  # Gaussian = polyclonal; spikes = clonal expansion
    return fig

Clonal-space / proportion: how much repertoire the big clones hold

Bin clones into frequency strata (Rare / Small / Medium / Large / Hyperexpanded) and stack their summed frequency. Because strata are defined on frequency, this view is comparatively depth-robust and shows clonal-space homeostasis at a glance. A treemap of top clones is an alternative when individual dominant clones matter.

import numpy as np

def plot_clonal_space(clone_df, sample_col='sample'):
    # Frequency-based strata (immunarch homeo convention); frequency makes this depth-robust
    edges = [0, 1e-4, 1e-3, 1e-2, 1e-1, 1.0]
    labels = ['Rare', 'Small', 'Medium', 'Large', 'Hyperexpanded']
    clone_df = clone_df.copy()
    clone_df['stratum'] = pd.cut(clone_df['frequency'], bins=edges, labels=labels)
    space = clone_df.groupby([sample_col, 'stratum'], observed=True)['frequency'].sum().unstack(fill_value=0)
    fig, ax = plt.subplots(figsize=(8, 5))
    space[labels].plot(kind='bar', stacked=True, ax=ax, colormap='viridis')
    ax.set_ylabel('Fraction of repertoire')
    return fig

Clonal tracking across timepoints

Track individual clone frequencies over an ordered sample set (vaccination, infection, therapy). A line plot of the top clones, or an alluvial/stream for the same data, shows expansion, contraction, and persistence. Downsample timepoints to a common depth before declaring contraction - a clone scored absent at one timepoint is frequently below the sampling floor, not truly gone.

def plot_clone_tracking(clone_df, top_n=10, clone_col='cdr3_aa', time_col='timepoint'):
    top = clone_df.groupby(clone_col)['frequency'].sum().nlargest(top_n).index
    fig, ax = plt.subplots(figsize=(9, 5))
    for clone in top:
        d = clone_df[clone_df[clone_col] == clone].sort_values(time_col)
        ax.plot(d[time_col], d['frequency'], marker='o', label=clone[:12])
    ax.set_xlabel('Timepoint'); ax.set_ylabel('Clone frequency')
    ax.legend(bbox_to_anchor=(1.02, 1), loc='upper left', fontsize=7)
    return fig

Rarefaction/extrapolation: the correct way to compare diversity

A bare bar of Shannon (or observed richness) across samples of unequal depth is misleading - it plots depth as much as biology. The defensible comparison is a rarefaction/extrapolation curve: interpolate each sample down to (and extrapolate modestly above) a shared depth, then read diversity at a common x. iNEXT computes this for Hill numbers q=0/1/2 with confidence intervals (Hsieh et al. 2016 Methods Ecol Evol 7:1451; Chao et al. 2014 Ecol Monogr 84:45).

R route (gold standard, gives CIs):

library(iNEXT)

# count_list: named list of per-sample integer clonotype-count vectors
out <- iNEXT(count_list, q = c(0, 1, 2), datatype = 'abundance')
ggiNEXT(out, type = 1)  # diversity vs sample size, read at a common x

VDJtools route: vdjtools RarefactionPlot -m metadata.txt output_dir/.

Python resampling route (interpolation by multinomial subsampling) when iNEXT is unavailable:

def rarefaction_curve(counts, depths, reps=20, rng=None):
    # Interpolate observed richness by drawing 'm' reads without-replacement-like via multinomial
    rng = rng or np.random.default_rng(0)
    counts = np.asarray(counts, dtype=float)
    p = counts / counts.sum()
    total = int(counts.sum())
    richness = []
    for m in depths:
        if m > total:  # interpolation only; do not extrapolate past observed depth here
            richness.append(np.nan); continue
        obs = [np.count_nonzero(rng.multinomial(m, p)) for _ in range(reps)]
        richness.append(np.mean(obs))
    return richness

Overlap heatmap: pick a depth-robust metric

An N-by-N similarity heatmap summarizes pairwise repertoire overlap. The metric choice is the decision: Morisita-Horn is abundance-weighted and near-invariant to sample size, so it is the default across unequal depths; Jaccard (presence/absence) is dominated by the shallower sample's depth and should not be compared across unequal-depth pairs. Compute the matrix in vdjtools-analysis on depth-normalized samples, then render it here. Label the metric in the title.

def plot_overlap_heatmap(overlap_matrix, metric='Morisita-Horn'):
    fig, ax = plt.subplots(figsize=(7, 6))
    sns.heatmap(overlap_matrix, annot=True, fmt='.2f', cmap='YlOrRd', vmin=0, vmax=1, square=True, ax=ax)
    ax.set_title(f'Repertoire overlap ({metric})')  # state the metric; Jaccard is depth-biased
    return fig

Clonotype-similarity network

Nodes are CDR3s, edges connect sequences within a chosen distance, and clusters are candidate specificity groups. The network structure depends entirely on the threshold: too loose chains distinct clones together, too tight fragments real groups. State the threshold and metric, and test sensitivity before interpreting clusters. Same-length CDR3s with Hamming distance is the conservative default; Levenshtein allows indels.

import networkx as nx
from rapidfuzz.distance import Levenshtein

def build_similarity_network(clone_df, max_norm_dist=0.15, clone_col='cdr3_aa'):
    # normalized_similarity in [0,1]; edge when 1 - similarity <= threshold. Structure is threshold-dependent.
    clones = clone_df[clone_col].unique()
    g = nx.Graph()
    g.add_nodes_from(clones)
    for i, a in enumerate(clones):
        for b in clones[i + 1:]:
            if 1 - Levenshtein.normalized_similarity(a, b) <= max_norm_dist:
                g.add_edge(a, b)
    return g

Common Errors

Symptom	Cause	Fix
Diversity "differs" between groups but tracks library size	Bare Shannon/richness bar across unequal depth	Downsample to common depth or plot rarefaction curves read at a shared x (iNEXT/RarefactionPlot)
Overlap heatmap shows huge differences driven by one shallow sample	Jaccard/shared-count on raw, unequal-depth counts	Use Morisita-Horn on depth-normalized samples; label the metric
Network clusters change completely on re-run with a new cutoff	Arbitrary distance threshold; single-linkage chaining	Fix and report the metric + threshold; run a sensitivity sweep; prefer Hamming on same-length CDR3s
Two figures disagree on sharing/diversity	Built on different clonotype definitions (nt vs aa, +/- V/J)	Hold one clonotype definition constant across all figures in a comparison and state it
Spectratype "expansion" vanishes when re-plotted	Switched between frequency-weighted and clonotype-weighted histogram	Choose one weighting per figure and label it (frequency shows expansions)
V-usage differences between batches look biological	Multiplex-PCR primer bias	Compare usage only within one protocol, or use UMI/5'RACE data
plt.cm.get_cmap(name, N) raises AttributeError	Removed in matplotlib 3.9+	Use plt.get_cmap(name) or matplotlib.colormaps[name].resampled(N)

Related Skills

• vdjtools-analysis - Compute the diversity/overlap inputs (depth-normalized)
• mixcr-analysis - Produce clonotype tables
• scirpy-analysis - Single-cell clonal-expansion overlays
• data-visualization/ggplot2-fundamentals - General ggplot2 grammar
• data-visualization/heatmaps-clustering - Overlap/usage heatmap techniques

References

• Shugay M, et al. VDJtools: unifying post-analysis of T cell receptor repertoires. PLoS Comput Biol 2015; 11(11):e1004503. (PlotFancyVJUsage, RarefactionPlot, overlap metrics.)
• Chao A, Gotelli NJ, Hsieh TC, Sander EL, Ma KH, Colwell RK, Ellison AM. Rarefaction and extrapolation with Hill numbers. Ecol Monogr 2014; 84(1):45-67. (Interpolation/extrapolation framework.)
• Hsieh TC, Ma KH, Chao A. iNEXT: an R package for rarefaction and extrapolation of species diversity (Hill numbers). Methods Ecol Evol 2016; 7(12):1451-1456. (Rarefaction/extrapolation curves with CIs.)
• Greiff V, et al. A bioinformatic framework for immune repertoire diversity profiling. Genome Med 2015; 7:49. (Hill-number diversity profiling of repertoires.)
• Venturi V, et al. Sharing of T cell receptors in antigen-specific responses is driven by convergent recombination. PNAS 2006; 103(49):18691-18696. (aa-clonotype sharing inflated by convergence.)
• Chao A. Nonparametric estimation of the number of classes in a population. Scand J Stat 1984; 11:265-270. (Chao1 richness estimator.)