Predicting CMV status and HLA type

Intersect CMV ECOcluster TCRs with repertoires to predict CMV status and HLA type.

Published

August 10, 2026

1 Takeaways from this post

The CMV ECOcluster is a very good predictor of CMV status, even without accounting for donor HLA type
But donor HLA type does matter! Asian or Pacific Islander donors have fewer CMV ECOcluster TCRs.
Among CMV+ donors, the CMV ECOcluster can be used to predict donor HLA allele status for several alleles with high sensitivity and specificity
- Those HLA “models” could be used as a seed for stronger, generalized models

Code

# useful imports
import pandas as pd
from matplotlib import pyplot as plt

import itables
import seaborn as sns
from cmvividly.plots.style import set_style
set_style()

import itables
itables.init_notebook_mode()

#stats
from scipy.stats import mannwhitneyu

init_notebook_modetrusted

2 Load previously saved files related to the “Emerson” datasets

The TCR repertoires from the two “Emerson” cohorts are available through Adaptive’s ImmuneAccess portal. I downloaded the data, extracted the sample metadata and intersected the repertoire TCRs with the CMV ECOcluster, in this notebook. This intersection makes no attempt to account for donor HLA type.

Let’s load those processed files and prepare them for analysis.

Code

from cmvividly.data.access import load_emerson_metadata
pdf_sample_metadata = load_emerson_metadata()
print("'Emerson' cohort breakdown by CMV status:")
pd.crosstab(pdf_sample_metadata.cohort,
            pdf_sample_metadata.cmv_status)

'Emerson' cohort breakdown by CMV status:

^ⓘcmv_status	0.0	1.0
cohort
cohort_1	352	289
cohort_2	68	51

A sample of the TCR intersection of the CMV ECOcluster with the ‘Emerson’ repertoires:

Code

from cmvividly.data.access import load_cmv_ecocluster_intersect_emerson
pdf_tcr_matches_2026 = load_cmv_ecocluster_intersect_emerson()
pdf_tcr_matches_2026

^ⓘsample_name	templates	productive_frequency	tcr_repertoire	vgene_repertoire	jgene_repertoire	cdr3	tcr_ecocluster	vgene_ecocluster	jgene_ecocluster	hla	hla_class	tcr_pgen	log10_tcr_pgen_eps	hla_cocluster	hla_cocluster_npos_hlamatch	hla_cocluster_nneg_hlamatch	hla_cocluster_auroc_hlaaware	hla_cocluster_auroc_hlaunaware	same_tcr
P00076	NaN	0.000052	CASSYGRVANTGELFF+TCRBV06-06+TCRBJ02-02	TCRBV06-06	TCRBJ02-02*01	CASSYGRVANTGELFF	CASSYGRVANTGELFF+TCRBV06-06+TCRBJ02-02	TCRBV06-06	TCRBJ02-02	DRB1*11:03	cii	1.656307e-10	-9.780859	h-DRB1_11_03-144	0	0	NaN	0.547883	True
P00076	NaN	0.000098	CASSLDTGQRNQPQHF+TCRBV05-05+TCRBJ01-05	TCRBV05-05*01	TCRBJ01-05*01	CASSLDTGQRNQPQHF	CASSLDTGQRNQPQHF+TCRBV05-05+TCRBJ01-05	TCRBV05-05	TCRBJ01-05	B*40:01	ci	2.116287e-11	-10.674425	h-B_40_01-223	5	11	1.000000	0.609832	True
P00076	NaN	0.000020	CASSSEWGTDTQYF+TCRBV05-05+TCRBJ02-03	TCRBV05-05*01	TCRBJ02-03*01	CASSSEWGTDTQYF	CASSSEWGTDTQYF+TCRBV05-05+TCRBJ02-03	TCRBV05-05	TCRBJ02-03	DRB3*02:02	cii	7.981617e-11	-10.097909	h-DRB3_02_02-99	20	25	0.966000	0.706735	True
P00076	NaN	0.000089	CASSEGQGDSPLHF+TCRBV05-01+TCRBJ01-06	TCRBV05-01*01	TCRBJ01-06*01	CASSEGQGDSPLHF	CASSEGQGDSPLHF+TCRBV05-01+TCRBJ01-06	TCRBV05-01	TCRBJ01-06	DQA101:02+DQB106:02	cii	1.077448e-09	-8.967604	h-DQA1_01_02_DQB1_06_02-7	4	13	1.000000	0.591077	True
P00076	NaN	0.000109	CASSWGNTGELFF+TCRBV05-01+TCRBJ02-02	TCRBV05-01*01	TCRBJ02-02*01	CASSWGNTGELFF	CASSWGNTGELFF+TCRBV05-01+TCRBJ02-02	TCRBV05-01	TCRBJ02-02	B*40:01	ci	4.219798e-08	-7.374708	h-B_40_01-223	5	11	1.000000	0.609832	True
P00076	NaN	0.000052	CASSRGTDVNTEAFF+TCRBV05-01+TCRBJ01-01	TCRBV05-01*01	TCRBJ01-01*01	CASSRGTDVNTEAFF	CASSRGTDVNTEAFF+TCRBV05-01+TCRBJ01-01	TCRBV05-01	TCRBJ01-01	DRB1*11:03	cii	2.956767e-10	-9.529183	h-DRB1_11_03-144	0	0	NaN	0.547883	True
P00076	NaN	0.000068	CASSLSRGGLNGYTF+TCRBV27-01+TCRBJ01-02	TCRBV27-01*01	TCRBJ01-02*01	CASSLSRGGLNGYTF	CASSLSRGGLNGYTF+TCRBV27-01+TCRBJ01-02	TCRBV27-01	TCRBJ01-02	B*40:01	ci	1.645806e-10	-9.783621	h-B_40_01-223	5	11	1.000000	0.609832	True
P00076	NaN	0.000041	CASSLEQARYGYTF+TCRBV05-01+TCRBJ01-02	TCRBV05-01*01	TCRBJ01-02*01	CASSLEQARYGYTF	CASSLEQARYGYTF+TCRBV05-01+TCRBJ01-02	TCRBV05-01	TCRBJ01-02	DRB1*11:03	cii	4.028607e-10	-9.394845	h-DRB1_11_03-144	0	0	NaN	0.547883	True
P00076	NaN	0.000033	CASSLSRGGLNGYTF+TCRBV27-01+TCRBJ01-02	TCRBV27-01*01	TCRBJ01-02*01	CASSLSRGGLNGYTF	CASSLSRGGLNGYTF+TCRBV27-01+TCRBJ01-02	TCRBV27-01	TCRBJ01-02	B*40:01	ci	1.645806e-10	-9.783621	h-B_40_01-223	5	11	1.000000	0.609832	True
P00076	NaN	0.000031	CASSRETSYEQYF+TCRBV18-01+TCRBJ02-07	TCRBV18-01*01	TCRBJ02-07*01	CASSRETSYEQYF	CASSRETSYEQYF+TCRBV18-01+TCRBJ02-07	TCRBV18-01	TCRBJ02-07	DRB3*02:02	cii	2.847252e-09	-8.545574	h-DRB3_02_02-99	20	25	0.966000	0.706735	True
(148,098 more rows not shown)

3 Summarize the intersection per sample

Let’s evaluate three metrics on their ability to separate CMV+ from CMV- donors:

match count (n_matches): the number of TCRs in the sample that match the CMV ECOcluster
breadth: the fraction of TCRs in the sample that match the CMV ECOcluster
depth: the sum of productive frequencies of TCRs in the sample that match the CMV ECOcluster

Code

# count matches per row
import numpy as np
pdf_tcrmatch_rowcounts = pdf_tcr_matches_2026.groupby("sample_name").agg({
    "tcr_repertoire": "count",
    "productive_frequency": "sum"
}).reset_index()
pdf_tcrmatch_rowcounts.columns = ["sample_name", "n_matches", "depth"]
pdf_exactmatch_rowcounts_withmeta = pdf_sample_metadata.merge(pdf_tcrmatch_rowcounts, how="left", on="sample_name").fillna(0)
pdf_exactmatch_rowcounts_withmeta["breadth"] = pdf_exactmatch_rowcounts_withmeta["n_matches"] / pdf_exactmatch_rowcounts_withmeta["productive_rearrangements"]

# add log10-transformed columns for plotting
# add an epsilon to breadth/depth for log10 transformation
EPS = 1e-4
pdf_exactmatch_rowcounts_withmeta["log10_breadth"] = np.log10(pdf_exactmatch_rowcounts_withmeta["breadth"] + EPS)
pdf_exactmatch_rowcounts_withmeta["log10_depth"] = np.log10(pdf_exactmatch_rowcounts_withmeta["depth"] + EPS)
pdf_exactmatch_rowcounts_withmeta

^ⓘsample_name	cmv_status	cohort	productive_rearrangements	sex	ethnicity	race	Tissue Source	age	hla_class_i	n_matches	depth	breadth	log10_breadth	log10_depth
P00076	1.0	cohort_1	59552	Female	Unknown Ethnicity	Unknown racial group	PBMC,Peripheral blood lymphocytes (PBL),T cells,gDNA	48.0	0	264.0	0.011752	0.004433	-2.343605	-1.926224
Keck0076_MC1	1.0	cohort_2	287910	Male	Non-Hispanic or Latino	Asian or Pacific Islander	PBMC,Peripheral blood lymphocytes (PBL),T cells,gDNA	27.0	0	387.0	0.012810	0.001344	-2.840382	-1.889078
P00402	1.0	cohort_1	471962	Male	Non-Hispanic or Latino	Caucasian	PBMC,Peripheral blood lymphocytes (PBL),T cells,gDNA	42.0	HLA-A03,HLA-A26,HLA-B38,HLA-B44	599.0	0.003363	0.001269	-2.863543	-2.460584
P00655	0.0	cohort_1	194142	0	Unknown Ethnicity	Unknown racial group	PBMC,Peripheral blood lymphocytes (PBL),T cells,gDNA	0.0	HLA-A02,HLA-B40,HLA-B*51	131.0	0.000745	0.000675	-3.110831	-3.073262
Keck0067_MC1	1.0	cohort_2	153019	Male	Non-Hispanic or Latino	Asian or Pacific Islander	PBMC,Peripheral blood lymphocytes (PBL),T cells,gDNA	28.0	0	135.0	0.001694	0.000882	-3.007781	-2.746235
P00181	0.0	cohort_1	233425	Male	Non-Hispanic or Latino	Caucasian	PBMC,Peripheral blood lymphocytes (PBL),T cells,gDNA	37.0	HLA-A01,HLA-A03,HLA-B08,HLA-B49	143.0	0.000736	0.000613	-3.147144	-3.077621
P00334	0.0	cohort_1	126607	Male	Unknown Ethnicity	Unknown racial group	PBMC,Peripheral blood lymphocytes (PBL),T cells,gDNA	47.0	HLA-A02,HLA-A03,HLA-B07,HLA-B14	56.0	0.000508	0.000442	-3.265749	-3.215779
P00043	0.0	cohort_1	181775	Female	Unknown Ethnicity	Unknown racial group	PBMC,Peripheral blood lymphocytes (PBL),T cells,gDNA	0.0	HLA-A24,HLA-A30,HLA-B13,HLA-B51	98.0	0.000510	0.000539	-3.194412	-3.214335
P00580	0.0	cohort_1	149611	Male	Unknown Ethnicity	Unknown racial group	PBMC,Peripheral blood lymphocytes (PBL),T cells,gDNA	54.0	HLA-A01,HLA-A03,HLA-B08,HLA-B15	88.0	0.001305	0.000588	-3.162290	-2.852307
P00636	0.0	cohort_1	202287	Male	Non-Hispanic or Latino	Caucasian	PBMC,Peripheral blood lymphocytes (PBL),T cells,gDNA	42.0	HLA-A11,HLA-A32,HLA-B07,HLA-B15	136.0	0.000827	0.000672	-3.112207	-3.033045
(776 more rows not shown)

3.1 Visualize the intersection by cohort and CMV status

Visualize the difference in this intersection between CMV+ and CMV- donors. Three metrics: * match count (n_matches): the number of TCRs in the sample that match the CMV ECOcluster * breadth: the fraction of TCRs in the sample that match the CMV ECOcluster * depth: the sum of productive frequencies of TCRs in the sample that match the CMV ECOcluster

Code

# boxplot of matches by CMV status
f, (ax1, ax2, ax3) = plt.subplots(1, 3, figsize=(12, 6))

sns.boxplot(data=pdf_exactmatch_rowcounts_withmeta, x="cohort", y="n_matches", hue="cmv_status", ax=ax1)
ax1.set_title("Matched TCRs")
sns.boxplot(data=pdf_exactmatch_rowcounts_withmeta, x="cohort", y="log10_breadth", hue="cmv_status", ax=ax2)
ax2.set_title("Breadth (log10)")
# hide ax2's legend
ax2.get_legend().remove()
sns.boxplot(data=pdf_exactmatch_rowcounts_withmeta, x="cohort", y="log10_depth", hue="cmv_status", ax=ax3)
ax3.set_title("Depth (log10)")
ax3.get_legend().remove()

f.suptitle("ECOcluster matched TCRs / breadth / depth\nby cohort and CMV status")
f.tight_layout()

All three metrics show separation, but breadth and depth, which account for sequencing depth, show far better separation than match count.

4 Assess diagnostic performance

Build ROC curves using match count, breadth and depth.

Code

from cmvividly.plots.roc_utils import plot_roc, calc_auroc
f, ax = plt.subplots(figsize=(5, 5))

auroc_matches = calc_auroc(pdf_exactmatch_rowcounts_withmeta, "n_matches", "cmv_status")
auroc_breadth = calc_auroc(pdf_exactmatch_rowcounts_withmeta, "breadth", "cmv_status")
auroc_depth = calc_auroc(pdf_exactmatch_rowcounts_withmeta, "depth", "cmv_status")

plot_roc(pdf_exactmatch_rowcounts_withmeta, "n_matches", label_col="cmv_status", ax=ax, label=f"matches (AUC={auroc_matches:.2f})")
plot_roc(pdf_exactmatch_rowcounts_withmeta, "breadth", label_col="cmv_status", ax=ax, label=f"breadth (AUC={auroc_breadth:.2f})")
plot_roc(pdf_exactmatch_rowcounts_withmeta, "depth", label_col="cmv_status", ax=ax, label=f"depth (AUC={auroc_depth:.2f})")
ax.set_title("ROC curves predicting CMV status\nusing raw match count, breadth and depth")
pass

So, after accounting for repertoire depth in one of two ways, the ECOcluster is a very good predictor of CMV status, even though we haven’t accounted for donor HLA type at all.

4.1 Break down performance by race

But that doesn’t mean HLA doesn’t matter, even for CMV diagnosis! Race is a loose but useful proxy for HLA type. The CMV ECOcluster was built on repertoires from a wide variety of donors, some explicitly sourced to increase donor diversity. Nevertheless, well over half were Caucasian. Let’s see how these donors break down by the race reported in the metadata:

Code

pdf_exactmatch_rowcounts_withmeta.race.value_counts()

^ⓘ	count
race
Caucasian	460
Unknown racial group	258
Asian or Pacific Islander	46
African Race	11
Native American or Alaska Native	10
0	1

Those values are straight from the ImmuneAccess file. I’m not sure what race “0” means, so I’ll leave that out. I’ll also leave out the groups with very few donors.

We might expect to see the classifier do better on Caucasians than on other groups, because they share more alleles with the repertoires ECOclusters were built on. Let’s take a look.

Code

pdf_withrace_withenough = pdf_exactmatch_rowcounts_withmeta[
    (pdf_exactmatch_rowcounts_withmeta.race != 0)
    & (~pdf_exactmatch_rowcounts_withmeta.race.isin(["African Race", "Native American or Alaska Native"]))
]
   
f, (ax1, ax2, ax3) = plt.subplots(1, 3, figsize=(12, 4))

sns.boxplot(y="race", x="n_matches", hue="cmv_status", data=pdf_withrace_withenough, ax=ax1)
ax1.set_title("Matched TCRs")
sns.boxplot(y="race", x="log10_breadth", hue="cmv_status", data=pdf_withrace_withenough, ax=ax2)
ax2.set_yticks([])
ax2.set_title("Breadth (log10)")
ax2.get_legend().remove()
sns.boxplot(y="race", x="log10_depth", hue="cmv_status", data=pdf_withrace_withenough, ax=ax3)
ax3.set_yticks([])
ax3.set_title("Depth (log10)")
ax3.get_legend().remove()

f.tight_layout()

Visually, it looks like, among CMV+ donors, breadth, depth and # matches are all lower in the Asian or Pacific Islander group than in the other groups. Let’s test for significance with a two-sided Mann-Whitney U test.

Code

pdf_api = pdf_withrace_withenough[pdf_withrace_withenough.race == "Asian or Pacific Islander"]
pdf_notapi = pdf_withrace_withenough[pdf_withrace_withenough.race != "Asian or Pacific Islander"]
pdf_api_cmvpos = pdf_api[pdf_api.cmv_status == 1]
pdf_notapi_cmvpos = pdf_notapi[pdf_notapi.cmv_status == 1]

mwu_p_matches = mannwhitneyu(pdf_api_cmvpos.n_matches, pdf_notapi_cmvpos.n_matches).pvalue
mwu_p_breadth = mannwhitneyu(pdf_api_cmvpos.breadth, pdf_notapi_cmvpos.breadth).pvalue
mwu_p_depth = mannwhitneyu(pdf_api_cmvpos.depth, pdf_notapi_cmvpos.depth).pvalue

print(f"Two-sided MWU p-values:")
print(f"# matches: {mwu_p_matches:.1e}")
print(f"breadth: {mwu_p_breadth:.1e}")
print(f"depth: {mwu_p_depth:.1e}")

Two-sided MWU p-values:
# matches: 6.2e-05
breadth: 1.7e-04
depth: 4.9e-03

Yes, these differences are highly significant, particularly for match count and breadth. In the preprint, we took great pains to account for those differences, using inferred HLA types on all the donors and comparing breadths against reference distributions for donors with the associated allele.

Let’s take a look at ROC curves within each group:

Code

f, ax = plt.subplots(figsize=(5, 5))
for race in sorted(pdf_withrace_withenough.race.unique().tolist()):
    pdf_race = pdf_withrace_withenough[pdf_withrace_withenough.race == race]
    auroc_matches = calc_auroc(pdf_race, "breadth", "cmv_status")
    plot_roc(pdf_race, "breadth", label_col="cmv_status", label=f"{race} (AUC={auroc_matches:.2f})", ax=ax)
ax.legend(bbox_to_anchor=(0.2, 0.5), loc='upper left')
ax.set_title("ROC curves predicting CMV status\nusing breadth, by race")

We do markedly worse for the Unknown racial group due to a number of CMV- donors with very high breadth. There are a lot of people in that group, so the difference isn’t likely due to chance. More likely, those donors have HLA alleles whose CMV-COclusters aren’t as well differentiated w.r.t. CMV status and lack other alleles that are better differentiated.

Also, from the boxplots and MWU tests above, we also know that, if we throw all those donors together, API CMV+ donors will be less well-separated from CMV- donors of other races.

5 Infer donor CMV+ donor HLA type from the TCR repertoire

Can we get something useful out of the HLA-bound nature of the CMV TCRs? Can we use it to infer the status of CMV+ donors with respect to the HLA alleles present in the CMV ECOcluster?

If we consider two CMV+ donors, one with HLA allele A*02:01 and one without, only the first donor should have A*02:01-associated TCRs responding to CMV. Since the CMV ECOcluster records the HLA allele associated with each TCR, we can use the A*02:01-associated CMV ECOcluster TCRs in a CMV+ donor’s repertoire to predict whether that donor has A*02:01.

Limitations:

All we’ve got is Class I HLA type, for a subset of donors, so we’re entirely blind to Class II
- (I think the full HLA types for these donors might be floating around somewhere? If you know how to get them, let me know and I’ll redo all this!)
The HLA metadata has values like HLA-A*01, while the HLA-COclusters in the CMV ECOcluster are associated with alleles like A*01:01. My workaround here will be to add :01 to the end of the HLA metadata values and see if any HLA-COclusters match. That’s a bit of a hack, and I might be missing some modelable alleles this way, but it demonstrates the idea.

Code

pdf_sample_metadata_withhla = pdf_sample_metadata[pdf_sample_metadata.hla_class_i.notnull()]
pdf_sample_metadata_withhla_cmvpos = pdf_sample_metadata_withhla[pdf_sample_metadata_withhla.cmv_status == 1]
print(f"Number of CMV+ donors with HLA Class I data: {len(pdf_sample_metadata_withhla_cmvpos)}")

all_alleles = set()
for hla_str in pdf_sample_metadata_withhla_cmvpos.hla_class_i:
    alleles = hla_str.split(",")
    all_alleles.update([x for x in alleles if x != "HLA MHC Class I"])  # there's some malformed metadata in there.
print(f"HLA Class I alleles present among CMV+ donors: {len(all_alleles)}")

alleles = sorted(all_alleles)
pos_counts = []
neg_counts = []
for allele in sorted(all_alleles):
    pdf_sample_metadata_withhla[allele] = pdf_sample_metadata_withhla.hla_class_i.apply(lambda x: 1 if allele in str(x).split(",") else 0)
    pos_counts.append(pdf_sample_metadata_withhla[(pdf_sample_metadata_withhla[allele] == 1) & (pdf_sample_metadata_withhla.cmv_status == 1)].shape[0])
    neg_counts.append(pdf_sample_metadata_withhla[(pdf_sample_metadata_withhla[allele] == 1) & (pdf_sample_metadata_withhla.cmv_status == 0)].shape[0])

pdf_pos_neg_counts = pd.DataFrame({
    "allele": sorted(all_alleles),
    "pos_count": pos_counts,
    "neg_count": neg_counts
})

f, ax = plt.subplots()    
sns.histplot(data=pdf_pos_neg_counts, x="pos_count", ax=ax, bins=20)
ax.set_title("CMV+ donor counts per Class I allele")
pass

Number of CMV+ donors with HLA Class I data: 274
HLA Class I alleles present among CMV+ donors: 59

A lot of those alleles have very few allele+ donors among the CMV+ donors. Luckily, we don’t really need to train a model! We just need enough donors to assess breadth as a predictor of allele status.

Let’s require at least 10 CMV+ donors with that allele to assess breadth as a predictor of allele status.

Code

alleles_enough_cmvpos = pdf_pos_neg_counts[pdf_pos_neg_counts.pos_count >= 10].allele.tolist()
print(f"Class I alleles with at least 10 CMV+ donors: {len(alleles_enough_cmvpos)}")
# eliminate alleles with no associated HLA-COcluster
alleles_enough_cmvpos_with_hlacocluster = [
    allele for allele in alleles_enough_cmvpos if allele[len("HLA-"):] + ":01" in pdf_tcr_matches_2026.hla.unique().tolist()
]
print(f"...and an associated HLA-COcluster: {len(alleles_enough_cmvpos_with_hlacocluster)}")

Class I alleles with at least 10 CMV+ donors: 28
...and an associated HLA-COcluster: 12

So, there are 12 or so alleles we can reasonably evaluate.

Let’s evaluate breadth on the associated HLA-COcluster as a predictor for having each allele.

Code

from cmvividly.plots.roc_utils import calc_roc_summary

pdf_tcr_matches_2026["hla_for_join"] = pdf_tcr_matches_2026.hla.apply(
    lambda x: x.split("*")[0])
# Join the HLA metadata with the breadth metric and assess breath as a predictor
pdf_sample_metadata_withhla_cmvpos = pdf_sample_metadata_withhla[pdf_sample_metadata_withhla.cmv_status == 1]


def calc_rocs_all_models(pdf_metadata, title_suffix):
    rows = []
    f, ax = plt.subplots()
    
    for allele in alleles_enough_cmvpos_with_hlacocluster:    
        allele_for_join = allele[len("HLA-"):] + ":01"
        pdf_matches_this_allele = pdf_tcr_matches_2026[pdf_tcr_matches_2026.hla == allele_for_join]
        pdf_rowcounts_this_allele = pdf_matches_this_allele.groupby("sample_name").agg({
            "tcr_repertoire": "count",
            "productive_frequency": "sum"
        }).reset_index()
        pdf_rowcounts_this_allele_withmeta = pdf_metadata.merge(
            pdf_rowcounts_this_allele, how="left", on="sample_name").fillna(0)
        pdf_rowcounts_this_allele_withmeta["breadth"] = pdf_rowcounts_this_allele_withmeta["tcr_repertoire"] / pdf_rowcounts_this_allele_withmeta["productive_rearrangements"]
        pdf_rowcounts_this_allele_withmeta["depth"] = pdf_rowcounts_this_allele_withmeta["productive_frequency"]
        auroc_breadth = calc_auroc(pdf_rowcounts_this_allele_withmeta, "breadth", label_col=allele)
        sens_98_spec_breadth = calc_roc_summary(pdf_rowcounts_this_allele_withmeta, "breadth", label_col=allele,
                                                specificities=[0.98])["senses_at_specs"][0]
        auroc_depth = calc_auroc(pdf_rowcounts_this_allele_withmeta, "depth", label_col=allele)
        rows.append({
            "allele": allele,
            "n_cmvpos": sum(pdf_rowcounts_this_allele_withmeta[allele] == 1),
            "auroc_breadth": auroc_breadth,
            "auroc_depth": auroc_depth,
            "sens_98_spec_breadth": sens_98_spec_breadth
        })
        plot_roc(pdf_rowcounts_this_allele_withmeta, 
                 "breadth", label_col=allele, ax=ax, label=allele)
    ax.set_title(f"ROC curves predicting HLA allele status\nusing breadth, by allele, {title_suffix}")
    ax.legend(bbox_to_anchor=(1.01, 1.05), loc='upper left')
    f.set_size_inches(5,5)
    
    pdf_allele_aurocs = pd.DataFrame(rows).sort_values("auroc_breadth", ascending=False)
    return pdf_allele_aurocs

pdf_allele_aurocs_cmvpos = calc_rocs_all_models(pdf_sample_metadata_withhla_cmvpos, "CMV+ donors")

Let’s take a look at the AUROCs and sensitivities at 98% specificity for those models.

Code

f, ax = plt.subplots()
sns.scatterplot(data=pdf_allele_aurocs_cmvpos, y="auroc_depth", x="auroc_breadth")
ax.set_title("AUROC for breadth vs depth predicting\nHLA allele status among CMV+ donors")
sns.lineplot(x=[0, 1], y=[0, 1], color="gray", linestyle="--", ax=ax)
f.set_size_inches(5, 5)
pdf_allele_aurocs_cmvpos

^ⓘ	allele	n_cmvpos	auroc_breadth	auroc_depth	sens_98_spec_breadth
0	HLA-A*01	71	0.985985	0.983071	0.887324
8	HLA-B*18	26	0.949132	0.944789	0.807692
6	HLA-B*08	55	0.945870	0.942715	0.800000
4	HLA-A*31	15	0.930116	0.900772	0.733333
2	HLA-A*03	62	0.917985	0.913421	0.677419
1	HLA-A*02	130	0.909348	0.908280	0.707692
10	HLA-B*38	15	0.882111	0.883655	0.733333
3	HLA-A*11	27	0.846304	0.859049	0.259259
11	HLA-B*51	31	0.823311	0.823975	0.580645
5	HLA-A*68	30	0.822199	0.820287	0.533333
(2 more rows not shown)

AUROCs based on breadth and on depth are strikingly similar, with a slight advantage to breadth-based models.

Just to be sure, let’s take a look at those same “models” in CMV- donors. There should be no signal, because those donors shouldn’t have CMV-responding TCRs.

Code

pdf_sample_metadata_withhla_cmvneg = pdf_sample_metadata_withhla[pdf_sample_metadata_withhla.cmv_status == 0]

pdf_allele_aurocs_cmvneg = calc_rocs_all_models(pdf_sample_metadata_withhla_cmvneg, "CMV- donors")

Huh! Most of the alleles are total garbage, as expected. But there’s some signal in A*01. There’s only one A*01:01 HLA-COcluster in the CMV ECOcluster, but perhaps it’s not totally “pure”… maybe it’s dominated by CMV-specific TCRs, but some TCRs also respond to other exposures.

6 Applications of HLA allele models usable on CMV+ donors

So, all but one of those “models” (HLA-B*35) has notable sensitivity at essentially perfect specificity, and they all have AUROC > 0.77. The best 6 models have >70% sensitivity at 98% specificity.

Quite likely, a number of the other alleles represented in the CMV ECOcluster have similar performance, but we don’t have enough CMV+ donors with those alleles in this dataset to assess performance. And of course we haven’t even been able to look at Class II alleles, since we didn’t have any data there.

I think the best HLA allele models, combined with the CMV ECOcluster’s ability to predict CMV status and enough public repertoires, could allow us to “pseudolabel” inferred CMV+ donors with status on those HLAs, learn signal from exposures other than CMV, and build HLA allele models that could be used on both CMV+ and CMV- donors.