Identification of reactive Borrelia burgdorferi peptides associated with Lyme disease

ABSTRACT

Borrelia burgdorferi, the agent of Lyme disease, is estimated to cause >400,000 annual infections in the United States. Serology is the primary laboratory method to support the diagnosis of Lyme disease, but current methods have intrinsic limitations that require alternative approaches or targets. We used a high-density peptide array that contains >90,000 short overlapping peptides to catalog immunoreactive linear epitopes from >60 primary antigens of B. burgdorferi. We then pursued a machine learning approach to identify immunoreactive peptide panels that provide optimal Lyme disease serodiagnosis and can differentiate antibody responses at various stages of disease. We examined 226 serum samples from the Lyme Biobank and the National Institutes of Health, which included sera from 110 individuals diagnosed with Lyme disease, 31 probable cases from symptomatic individuals, and 85 healthy controls. Cases were grouped based on disease stage and presentation and included individuals with early localized, early disseminated, and late Lyme disease. We identified a peptide panel originating from 14 different epitopes that differentiated cases versus controls, whereas another peptide panel built from 12 unique epitopes differentiated subjects with various disease manifestations. Our method demonstrated an improvement in B. burgdorferi antibody detection over the current two-tiered testing approach and confirmed the key diagnostic role of VlsE and FlaB antigens at all stages of Lyme disease. We also uncovered epitopes that triggered a temporal antibody response that was useful for differentiation of early and late disease. Our findings can be used to streamline serologic targets and improve antibody-based diagnosis of Lyme disease.

IMPORTANCE

Serology is the primary method of Lyme disease diagnosis, but this approach has limitations, particularly early in disease. Currently employed antibody detection assays can be improved by the identification of alternative immunodominant epitopes and the selection of optimal diagnostic targets. We employed high-density peptide arrays that enabled precise epitope mapping for a wide range of B. burgdorferi antigens. In combination with machine learning, this approach facilitated the selection of serologic targets early in disease and the identification of serological indicators associated with different manifestations of Lyme disease. This study provides insights into differential antibody responses during infection and outlines a new approach for improved serologic diagnosis of Lyme disease.

INTRODUCTION

Lyme disease, caused by infection with the spirochete Borrelia burgdorferi, is the most common tick-borne disease in the United States (1). An estimated >400,000 B. burgdorferi infections occur annually, with the severity ranging from mild to a systemic febrile illness (2). For clinical purposes, Lyme disease is divided into early localized, early disseminated, and late stages (3). The infection starts at the site of the tick bite, where the spirochetes are deposited in the dermis, multiply, and spread centrifugally through the dermis. The interaction with the host’s innate immune system results in an expanding erythema migrans rash, the typical primary sign of the infection, and is classified as early localized Lyme disease. If untreated, spirochetes can enter the bloodstream, disseminate, and establish infection at distant sites, causing diverse clinical manifestations (4). Early disseminated Lyme disease presentations include multiple erythema migrans lesions, early Lyme neuroborreliosis, and Lyme carditis. The hallmark of late Lyme disease in the United States is Lyme arthritis (5, 6).

Most laboratory tests used to support the diagnosis of Lyme disease are based on the detection of the antibody responses against B. burgdorferi in serum. Due to the time interval between infection and production of a detectable antibody response, patients with erythema migrans are usually negative at presentation. The United States Centers for Disease Control and Prevention (CDC)-recommended standard two-tier algorithm is positive in about 40% and the modified two-tier algorithms in about 50% of acute-phase samples from patients with erythema migrans (7). While patients with erythema migrans typically receive antibiotic therapy based on potential for exposure and the clinical presentation, improvements in laboratory testing that would shorten this window period would be helpful.

Differential expression of outer surface proteins (Osp) enables B. burgdorferi to adapt to the diverse environments that the spirochete encounters in vertebrate and arthropod hosts and plays a key role in facilitating dissemination during vertebrate infection (8, 9). Along with a high degree of genetic heterogeneity among strains, variable antigenic expression plays a key role in the challenges of serological diagnosis of Lyme disease. A better understanding of temporal antigenic expression of B. burgdorferi could result in greater insights into pathogenesis as well as serological targets. Although comprehensive in vivo omics analyses of B. burgdorferi antigenic expression have been hampered by low spirochetemia, the examination of antibody responses could prove useful to identify stage-specific serologic indicators. In this study, we used the TBD-Serochip, a linear peptide microarray, to analyze IgG and IgM antibody responses to linear B. burgdorferi epitopes from patients diagnosed with different stages and manifestations of Lyme disease. We identified peptides that can be used to improve early diagnosis, as well as peptides that could be used to differentiate among disease manifestations and have the potential to improve antibody-based diagnosis of Lyme disease.

MATERIALS AND METHODS

TBD-Serochip

The Tick-Borne Disease Serochip (TBD-Serochip) is a slide-based peptide array used to catalog antibody responses to tick-borne pathogens (10). For each antigen selected for inclusion on the array, all protein sequences available as of October 2016 were downloaded from the NCBI protein database, aligned, and used to design 12-mer peptides that tile each protein with an 11-amino acid (aa) overlap to the preceding peptide in a sliding window pattern. Our prototype version of the TBD-Serochip included approximately 170,000 12-mer peptides per subarray and contained 12-mer peptides designed from antigenic sequences of eight tick-borne pathogens present in North America. For B. burgdorferi, this included 62 different antigens (including all paralogs) that are known to elicit an antibody response in humans (Fig. S1) (11, 12). For each antigen, we included the sequence of every genetic variant in the database for the 12-mer design. This included 12-mer peptides for 20 distinct OspC types and a wide range of recombinant sequences for VlsE. This approach enables the identification of all reactive portions for every examined antigen and demonstrates the impact of amino acid variation within a given epitope on antibody binding. Conversely, it can also inflate the number of significant reactive peptides due to cross-reactivity between different variants of the same 12-mer fragment (Fig. S2). The B. burgdorferi peptide component of the TBD-Serochip consisted of 91,338 peptides. The arrays were manufactured by Nimble Therapeutics.

Sample descriptions

The Lyme Disease Biobank

The Lyme Disease Biobank (LDB) sample repository includes well-characterized samples collected from patients with Lyme disease and healthy controls living in areas endemic for tick-borne disease (the Northeast and Upper Midwest) (13). The samples used in this study included sera from 38 confirmed acute Lyme disease cases, as determined by positive two-tiered serology, two positive ELISAs with erythema migrans > 5 cm, quantitative PCR (qPCR) and/or culture followed by PCR of the culture fluid for B. burgdorferi of whole blood from the acute-phase blood draw, or IgG seroconversion, with most being confirmed by two-tier serology (Table 1). The presence of an erythema migrans of >5 cm was noted in 25 patients (designated SEM-A), with six patients having >1 lesion. Four samples had evidence of B. burgdorferi infection by PCR and/or culture (Table 1). The cohort also included 31 probable Lyme disease cases, consisting of individuals with an erythema migrans rash of >5 cm but no confirmatory laboratory evidence (Table 2). We also included sera from 38 healthy controls living in endemic areas without a history of Lyme disease, all with negative serology. Samples were collected under IRB-approved protocols, and all participants provided written informed consent. Males represented the majority of enrolled cases (25 vs 13).

TABLE 1.

Description and testing data of the confirmed acute Lyme disease samples from the Lyme Disease Biobank cohort^a,b

Sample ID	Patient origin	EM > 5 cm at enrollment	Antibiotic therapy at enrollment (days)	B. burgdorferi culture	B. burgdorferi culture fluid PCR	B. burgdorferi qPCR	Whole-cell lysate ELISA	C6 peptide ELISA	VlsE/ PepC10	Western blot IgM	Western blot IgG	Two-tier testing Result	Initial discriminatory model prediction	Tester Set Prediction
LYM-997	NY	YES	YES (1)	NA	NA	NEG	RE	POS	NA	POS	IT	POS	POS	NA
LYM-1237	NY	YES	NO	NA	NA	NEG	RE	POS	NA	POS	IT	POS	POS	NA
LYM-1232	WI	YES	YES (1)	NA	NA	NEG	NA	POS	POS	NEG	NEG	NEG	POS	NA
LYM-1227	NY	YES	NO	NO	POS	NEG	NR	NEG	NA	IT	NEG	NEG	POS	NA
LYM-1214	WI	YES	NO	NA	NA	NEG	NA	POS	NA	NEG	POS	POS	POS	NA
LYM-1206	NY	YES	NO	NA	NA	NEG	BL	POS	NA	POS	IT	POS	POS	NA
LYM-1203	NY	YES	NO	NO	POS	NEG	RE	POS	NA	IT	IT	NEG	POS	NA
LYM-1200	NY	YES	YES (1)	NO	POS	NEG	RE	NEG	NA	POS	IT	POS	NEG	NA
LYM-1199	NY	YES	NO	NO	NEG	NEG	NR	POS	NA	POS	IT	POS	NEG	NA
LYM-1191	NY	YES	YES (1)	NA	NA	NEG	RE	POS	NA	POS	IT	POS	POS	NA
LYM-1160	MA	YES	NO	NA	NA	NEG	RE	POS	NA	IT	IT	NEG	POS	POS
LYM-1131	MA	YES	NO	POS	POS	NEG	NR	NEG	NA	IT	IT	NEG	NEG	NA
LYM-1114	NY	YES	NO	NA	NA	POS	NR	POS	NA	IT	IT	NEG	NEG	NEG
LYM-1110	NY	YES	NO	NA	NA	POS	NR	NEG	NA	NEG	IT	NEG	POS	NA
LYM-1107	NY	YES	NO	NA	NA	NEG	RE	POS	NA	IT	POS	POS	POS	NA
LYM-1099	WI	YES	NO	NA	NA	POS	NA	NEG	NEG	NEG	NEG	NEG	NEG	NEG
LYM-1097	NY	YES	YES (10)	NA	NA	NEG	RE	POS	NA	POS	IT	POS	POS	POS
LYM-1093	WI	YES	YES (1)	NA	NA	NEG	NA	POS	POS	POS	NEG	POS	POS	NA
LYM-1034	WI	YES	YES (1)	NA	NA	NEG	NA	POS	POS	POS	NEG	POS	POS	NA
LYM-1031	NY	YES	NO	NA	NA	NEG	BL	POS	NA	IT	IT	NEG	POS	NA
LYM-1016	MA	YES	NO	NEG	NEG	NEG	RE	POS	NA	POS	POS	POS	POS	NA
LYM-1015	MA	YES	YES (1)	NEG	NEG	NEG	NR	POS	NA	POS	IT	POS	POS	NA
LYM-1005	NY	YES	NO	NA	NA	NEG	NR	POS	NA	POS	IT	POS	POS	POS
LYM-1002	NY	YES	NO	NA	NA	NEG	NR	POS	NA	POS	IT	POS	POS	NA
LYM-1001	NY	YES	NO	NA	NA	NEG	RE	POS	NA	IT	IT	NEG	POS	NA
LYM-998	NY	NO	NO	NA	NA	NEG	RE	POS	NA	POS	IT	POS	POS	POS
LYM-1207	WI	NO	NO	NA	NA	NEG	NA	POS	NA	NEG	NEG	NEG	POS	NA
LYM-1181	NY	NO	NO	NA	NA	NEG	NR	POS	NA	POS	IT	POS	POS	NA
LYM-1127	NY	NO	YES (1)	NEG	NEG	NEG	RE	POS	NA	IT	POS	POS	POS	NA
LYM-1055	NY	NO	NO	NA	NA	NEG	BL	POS	NA	POS	IT	POS	POS	NA
LYM-1204	NY	NO	NO	NA	NA	NEG	NR	POS	NA	POS	IT	POS	NEG	NA
LYM-1137	MA	NO	NO	NEG	NEG	NEG	RE	POS	NA	POS	POS	POS	POS	NA
LYM-1123	NY	NO	NO	NEG	NEG	NA	NR	POS	NA	POS	NEG	POS	POS	NA
LYM-1086	WI	NO	YES (2)	NA	NA	NEG	NA	POS	POS	POS	NEG	POS	POS	NA
LYM-1080	WI	NO	NO	NA	NA	NEG	NA	POS	POS	POS	POS	POS	POS	NA
LYM-1022	WI	NO	YES (2)	NA	NA	NEG	NA	NEG	POS	POS	NEG	POS	POS	POS
LYM-1013	NY	NO	NO	NA	NA	POS	BL	NEG	NA	NEG	IT	NEG	NEG	NA
LYM-1010	NY	NO	NO	NA	NA	NEG	RE	NEG	NA	POS	IT	POS	POS	NA

^a NA: data not available; RE: reactive; NR: not reactive; IT: indeterminate; BL: borderline; POS: positive; NEG: negative; NY; New York; WI: Wisconsin; MA: Massachussetts.

^b Samples with an infection identified by qPCR and/or culture are shown in bold. B. burgdorferi qPCR was performed in whole blood samples.

TABLE 2.

Description and testing data of probable Lyme samples from the Lyme Disease Biobank cohort^a,b

Sample ID	Patient origin	EM > 5 cm at enrollment	Antibiotic therapy at enrollment (days)	B. burgdorferi qPCR	Whole-cell lysate ELISA	C6 Peptide ELISA	VlsE/ PepC10	Western blot IgM	Western blot IgG	Two-tier Testing result	Discriminatory model prediction
LYM-1008	NY	YES	NO	NEG	BL	NEG	NA	IT	IT	NEG	POS
LYM-1011	NY	YES	NO	NEG	NR	NEG	NA	NEG	IT	NEG	POS
LYM-1048	NY	YES	NO	NEG	NA	NEG	NEG	NEG	NEG	NEG	POS
LYM-1081	NY	YES	NO	NEG	NR	POS	NA	IT	NEG	NEG	POS
LYM-1105	NY	YES	NO	NEG	NA	NEG	POS	NEG	NEG	NEG	POS
LYM-1178	NY	YES	NO	NEG	NR	NEG	NA	IT	IT	NEG	POS
LYM-1186	NY	YES	YES (2)	NEG	NR	NEG	NA	IT	IT	NEG	POS
LYM-1210	NY	YES	NO	NEG	NR	NEG	NA	IT	IT	NEG	POS
LYM-991	NY	YES	NO	NEG	NR	POS	NA	IT	NEG	NEG	POS
LYM-1006	WI	YES	NO	NEG	NR	NEG	NA	NEG	IT	NEG	NEG
LYM-1014	NY	YES	NO	NEG	NR	NEG	NA	NEG	IT	NEG	NEG
LYM-1032	NY	YES	NO	NEG	NR	NEG	NA	IT	IT	NEG	NEG
LYM-1039	WI	YES	YES (1)	NEG	NA	IND	NA	NEG	NEG	NEG	NEG
LYM-1054	WI	YES	NO	NEG	NA	NEG	NA	NEG	NEG	NEG	NEG
LYM-1082	WI	YES	NO	NEG	NA	NEG	EQUIV	NEG	NEG	NEG	NEG
LYM-1083	WI	YES	NO	NEG	NA	NEG	NEG	POS	NEG	NEG	NEG
LYM-1094	NY	YES	NO	NEG	RE	NEG	NA	IT	IT	NEG	NEG
LYM-1096	WI	YES	YES (1)	NEG	NA	NEG	NEG	NEG	NEG	NEG	NEG
LYM-1098	WI	YES	YES (1)	NEG	NA	POS	NEG	NEG	NEG	NEG	NEG
LYM-1100	WI	YES	YES (2)	NEG	NA	NEG	NEG	NEG	NEG	NEG	NEG
LYM-1109	CA	YES	YES (1)	NEG	NR	NEG	NA	IT	NEG	NEG	NEG
LYM-1133	UT	YES	NO	NEG	NR	NEG	NA	NEG	IT	NEG	NEG
LYM-1153	WI	YES	NO	NEG	NA	NEG	NEG	NEG	NEG	NEG	NEG
LYM-1154	NY	YES	YES (1)	NEG	NR	NEG	NA	POS	IT	NEG	NEG
LYM-1177	NY	YES	NO	NEG	NR	POS	NA	IT	IT	NEG	NEG
LYM-1219	NY	YES	NO	NEG	NR	NEG	NA	IT	NEG	NEG	NEG
LYM-1233	NY	YES	NO	NEG	NR	NEG	NA	IT	NEG	NEG	NEG
LYM-1244	NY	YES	NO	NEG	NA	NEG	NEG	NEG	NEG	NEG	NEG
LYM-1248	NY	YES	NO	NEG	NA	NEG	NEG	NEG	NEG	NEG	NEG
LYM-989	CA	YES	NO	NEG	RE	NEG	NA	IT	IT	NEG	NEG
LYM-993	WI	YES	YES (1)	NEG	NA	NEG	NEG	NEG	NEG	NEG	NEG

^a NA: data not available; RE: reactive; NR: not reactive; IT: indeterminate; BL: borderline; POS: positive; NEG: negative; EQUIV: equivocal; NY; New York; WI: Wisconsin; UT: Utah; CA: California.

^b B. burgdorferi qPCR was performed in whole blood samples.

The National Institutes of Health cohort

This cohort consisted of 82 patients diagnosed with Lyme disease and 47 healthy controls from an endemic area without a history of Lyme disease (Table 3). Serum samples were collected under clinical protocols approved by the National Institutes of Health (NIH) Institutional Review Board (ClinicalTrials.gov Identifier: NCT00028080 and NCT00001539), and written informed consent was obtained from all participants. Patients with Lyme disease acquired the infection in the mid-Atlantic region of the United States and fulfilled the 2017 CDC case definition of confirmed or probable Lyme disease (14). Patients were grouped according to their main clinical manifestations and disease stage. Most samples were collected after the start of antibiotic therapy (Table 3). The NIH cohort included 27 patients with single erythema migrans (designated SEM-C), 13 patients with multiple erythema migrans (MEM), 15 patients with acute Lyme neuroborreliosis (ALNB), and 27 patients with Lyme arthritis (LA). There was a male predominance among cases of neuroborreliosis and arthritis.

TABLE 3.

Description of the National Institutes of Health cohort^a,b

Sample ID	Group	Gender	Age bracket	Direct microbiological evidence of Bb infection	Interval sample from start of antibiotic therapy (days)	Lyme EIA (C6 Peptide ELISA or VlsE/PepC10)	Western blot IgM	Western blot IgG	Two-tier testing	Cohort differential model	Match observed/predicted
LA_01	Lyme Arthritis	Female	>55	POS	>45	POS	NEG	POS	POS	LA	YES
LA_02	Lyme Arthritis	Male	41–55	POS	pre-therapy	POS	NEG	POS	POS	LA	YES
LA_03	Lyme Arthritis	Female	<20	POS	>45	POS	NEG	POS	POS	LA	YES
LA_04	Lyme Arthritis	Male	>55	POS	>45	POS	POS	POS	POS	LA	YES
LA_05	Lyme Arthritis	Male	20–40	POS	3 to 8	POS	POS	POS	POS	LA	YES
LA_06	Lyme Arthritis	Male	<20	POS	22 to 45	POS	NEG	POS	POS	LA	YES
LA_07	Lyme Arthritis	Male	41–55	NA	>45	POS	NEG	POS	POS	LA	YES
LA_08	Lyme Arthritis	Male	<20	POS	22 to 45	POS	NEG	POS	POS	LA	YES
LA_09	Lyme Arthritis	Male	20–40	NA	>45	POS	NEG	POS	POS	LA	YES
LA_10	Lyme Arthritis	Male	41–55	NA	22 to 45	POS	NEG	POS	POS	LA	YES
LA_11	Lyme Arthritis	Female	41–55	NA	22 to 45	POS	POS	POS	POS	LA	YES
LA_12	Lyme Arthritis	Female	41–55	NA	>45	POS	NEG	POS	POS	LA	YES
LA_13	Lyme Arthritis	Female	>55	NEG	>45	POS	POS	POS	POS	LA	YES
LA_14	Lyme Arthritis	Male	41–55	NEG	>45	POS	NEG	POS	POS	LA	YES
LA_15	Lyme Arthritis	Male	>55	POS	22 to 45	POS	POS	POS	POS	LA	YES
LA_16	Lyme Arthritis	Male	>55	NA	>45	POS	POS	POS	POS	LA	YES
LA_17	Lyme Arthritis	Male	>55	POS	>45	POS	NEG	POS	POS	LA	YES
LA_18	Lyme Arthritis	Female	<20	POS	three to 8	POS	NEG	POS	POS	LA	YES
LA_19	Lyme Arthritis	Male	>55	POS	>45	POS	NEG	POS	POS	LA	YES
LA_20	Lyme Arthritis	Male	>55	NEG	>45	POS	POS	POS	POS	LA	YES
LA_21	Lyme Arthritis	Female	41–55	NA	>45	POS	POS	POS	POS	LA	YES
LA_22	Lyme Arthritis	Male	20–40	POS	22 to 45	POS	NEG	POS	POS	LA	YES
LA_23	Lyme Arthritis	Male	41–55	NA	>45	POS	NEG	POS	POS	LA	YES
LA_24	Lyme Arthritis	Male	>55	POS	>45	POS	POS	POS	POS	LA	YES
LA_25	Lyme Arthritis	Male	41–55	NA	22 to 45	POS	POS	POS	POS	LA	YES
LA_26	Lyme Arthritis	Female	41–55	NEG	>45	POS	NEG	POS	POS	LA	YES
LA_27	Lyme Arthritis	Female	20–40	POS	22 to 45	POS	POS	POS	POS	LA	YES
MEM_01	Multiple EM	Female	41–55	NA	22 to 45	POS	POS	POS	POS	MEM	YES
MEM_02	Multiple EM	Male	>55	NA	>45	POS	POS	NEG	NEG	SEM-C	NO
MEM_03	Multiple EM	Female	>55	POS	22 to 45	POS	POS	POS	POS	SEM-C	NO
MEM_04	Multiple EM	Female	>55	NA	22 to 45	POS	POS	POS	POS	MEM	YES
MEM_05	Multiple EM	Male	41–55	NA	9 to 21	POS	POS	POS	POS	LA	NO
MEM_06	Multiple EM	Female	41–55	NA	22 to 45	POS	POS	POS	POS	MEM	YES
MEM_07	Multiple EM	Female	41–55	NA	>45	POS	POS	POS	POS	MEM	YES
MEM_08	Multiple EM	Male	41–55	NA	22 to 45	POS	POS	NEG	POS	MEM	YES
MEM_09	Multiple EM	Male	20–40	NA	22 to 45	POS	POS	NEG	NEG	MEM	YES
MEM_10	Multiple EM	Female	20–40	NA	22 to 45	POS	POS	NEG	NEG	MEM	YES
MEM_11	Multiple EM	Female	>55	NEG	22 to 45	POS	POS	NEG	NEG	MEM	YES
MEM_12	Multiple EM	Female	41–55	POS	22 to 45	POS	POS	NEG	NEG	SEM-C	NO
MEM_13	Multiple EM	Female	41–55	NA	22 to 45	POS	POS	NEG	POS	ALNB	NO
ALNB_01	Acute LNB	Male	>55	NEG	1 to 2	POS	NEG	POS	POS	LA	NO
ALNB_02	Acute LNB	Male	41–55	NEG	1 to 2	POS	POS	NEG	POS	Acute LNB	YES
ALNB_03	Acute LNB	Male	20–40	NEG	1 to 2	POS	POS	POS	POS	Acute LNB	YES
ALNB_04	Acute LNB	Female	20–40	NEG	3 to 8	POS	POS	NEG	POS	Acute LNB	YES
ALNB_05	Acute LNB	Female	41–55	NEG	9 to 21	POS	POS	POS	POS	Acute LNB	YES
ALNB_06	Acute LNB	Male	20–40	NEG	9 to 21	POS	POS	POS	POS	Acute LNB	YES
ALNB_07	Acute LNB	Male	20–40	NEG	3 to 8	POS	POS	NEG	POS	Acute LNB	YES
ALNB_08	Acute LNB	Male	41–55	NEG	1 to 2	POS	POS	NEG	POS	Acute LNB	YES
ALNB_09	Acute LNB	Male	<20	NEG	Pre-therapy	POS	POS	POS	POS	Acute LNB	YES
ALNB_10	Acute LNB	Male	20–40	NEG	9 to 21	POS	POS	NEG	NEG	Acute LNB	YES
ALNB_11	Acute LNB	Male	>55	NEG	Pre-therapy	POS	POS	POS	POS	MEM	NO
ALNB_12	Acute LNB	Male	41–55	NEG	3 to 8	POS	POS	NEG	POS	Acute LNB	YES
ALNB_13	Acute LNB	Female	41–55	NEG	9 to 21	POS	POS	NEG	POS	Acute LNB	YES
ALNB_14	Acute LNB	Male	20–40	NEG	9 to 21	POS	POS	POS	POS	Acute LNB	YES
ALNB_15	Acute LNB	Male	<20	NEG	Pre-therapy	POS	POS	POS	POS	Acute LNB	YES
SEM-C_01	Single EM	Male	20–40	NA	>45	POS	NEG	NEG	NEG	MEM	NO
SEM-C_02	Single EM	Female	>55	NA	22 to 45	POS	POS	NEG	POS	Single EM	YES
SEM-C_03	Single EM	Female	>55	NA	>45	POS	NEG	NEG	NEG	Single EM	YES
SEM-C_04	Single EM	Male	<20	NA	22 to 45	POS	POS	NEG	NEG	Single EM	YES
SEM-C_05	Single EM	Female	>55	POS	22 to 45	POS	POS	NEG	POS	Single EM	YES
SEM-C_06	Single EM	Female	41–55	NEG	22 to 45	POS	POS	NEG	NEG	Single EM	YES
SEM-C_07	Single EM	Female	20–40	NEG	22 to 45	POS	POS	NEG	NEG	Single EM	YES
SEM-C_08	Single EM	Male	>55	NA	9 to 21	POS	POS	POS	POS	Single EM	YES
SEM-C_09	Single EM	Female	<20	POS	3 to 8	POS	NA	NA	NA	Single EM	YES
SEM-C_10	Single EM	Female	20–40	POS	22 to 45	POS	POS	NEG	NEG	Single EM	YES
SEM-C_11	Single EM	Female	41–55	POS	22 to 45	POS	NEG	POS	POS	Single EM	YES
SEM-C_12	Single EM	Female	>55	NA	22 to 45	POS	NEG	NEG	NEG	Single EM	YES
SEM-C_13	Single EM	Male	20–40	NA	22 to 45	POS	POS	NEG	NEG	Single EM	YES
SEM-C_14	Single EM	Male	>55	NA	>45	POS	POS	NEG	NEG	Single EM	YES
SEM-C_15	Single EM	Female	>55	POS	22 to 45	POS	NEG	POS	POS	Single EM	YES
SEM-C_16	Single EM	Female	41–55	NA	22 to 45	NEG	NA	NA	NA	Single EM	YES
SEM-C_17	Single EM	Female	>55	POS	>45	POS	POS	NEG	NEG	Single EM	YES
SEM-C_18	Single EM	Male	>55	NA	>45	POS	POS	NEG	NEG	MEM	NO
SEM-C_19	Single EM	Female	41–55	POS	22 to 45	POS	POS	NEG	NEG	Single EM	YES
SEM-C_20	Single EM	Male	20–40	POS	22 to 45	NEG	NA	NA	NA	Single EM	YES
SEM-C_21	Single EM	Female	20–40	NEG	22 to 45	POS	POS	POS	POS	LA	NO
SEM-C_22	Single EM	Male	41–55	NEG	22 to 45	NEG	NA	NA	NA	Single EM	YES
SEM-C_23	Single EM	Male	>55	NA	>45	POS	POS	NEG	NEG	Single EM	YES
SEM-C_24	Single EM	Female	20–40	NA	22 to 45	POS	POS	NEG	POS	Single EM	YES
SEM-C_25	Single EM	Female	20–40	NEG	Pre-therapy	NEG	NA	NA	NA	Single EM	YES
SEM-C_26	Single EM	Male	41–55	NA	22 to 45	NEG	NA	NA	NA	SEM-A	NO
SEM-C_27	Single EM	Male	>55	NEG	22 to 45	POS	POS	POS	POS	SEM-A	NO

^a Direct microbiological evidence of Borrelia burgdorferi infection by culture and/or PCR in blood, synovial fluid, skin biopsies, and/or cerebrospinal fluid.

^b NA: data not available; POS: positive; NEG: negative. LA: Lyme arthritis; MEM: multiple erythema migrans; ALNB: acute Lyme neuroborreliosis; SEM-C: single erythema migrans convalescent.

Array data analyses

The method for microarray assays is demonstrated in Fig. S1. Sera were tested at a 1:50 dilution. After incubation with sera and fluorescently labeled secondary anti-IgG and anti IgM antibodies, arrays were scanned on a NimbleGen MS 200 Microarray Scanner (Roche) at 2 µm resolution, with an excitation wavelength of 532  nm for Cy3/IgM and 635  nm for Alexa Fluor/IgG. After scanning, a file was generated that included a relative fluorescent unit (RFU) signal for each 12-mer peptide on the array. Next, an aggregate file was generated by combining data files from all subarrays, including 129 samples from the NIH cohort and 107 from the LDB cohort. The final aggregate file included combined data from all Lyme disease cohorts and controls (n = 236), which included 182,676 data points for B. burgdorferi and 91,338 each for IgG and IgM. The analyses were conducted separately for IgG and IgM data sets. The DESeq2 package in R was used to identify peptides with different signal intensity comparing control and case groups (15). Slide-to-slide variation was considered in the differential analysis. The FDR-adjusted P-values ≤ 0.05 were applied to obtain significantly different signal intensity among peptides, and only the peptides with increased signal intensity in the cases were selected. To further narrow down the numbers of potential sero-reactive peptides for the differential analysis, peptides were retained only if its signal intensity was greater than three times of the median signal intensity for all peptides (intensity threshold = 3,000) in at least 30% of the Lyme case samples (signal intensity >3,000 and case prevalence >30%). The peptide-array differential analysis was performed in R version 4.2.2 within RStudio. Data munging was performed by reshape2, dplyR, and tidyR packages in R. The array data have been deposited and are available under the following link: https://datadryad.org/stash/share/Ws_tDf9_WNMl524GfeM6mgYliBSIbCwNByQKZKpsEMA.

Generation of a classification model for Lyme disease

Once both the IgG and IgM sero-reactive peptides were identified, we implemented random forest analyses using the random forest package in R to evaluate their classification performance with subsets of sero-reactive peptides (16). In a random forest model, the measure of importance of a peptide is based on its mean decrease in impurity (MDI) value. For the initial selection of peptides, we calculated the mean MDI and used it as a threshold. We followed an iterative model building approach where peptides with MDI values above the mean MDI threshold were selected to build another model with better accuracy. This process was continued until no further improvement in accuracy was obtained with the subsequent model. Once the minimal number of peptides needed for diagnostic accuracy was selected, we pursued further classification with random forest model using the R package caret (17). For each iteration, our primary data set was randomly split into a training set (80%) and a testing set (20%). In addition, the models were trained with tenfold cross-validation. The receiver operating characteristic (ROC) analysis was conducted to illustrate performances of classification models, using the R package pROC (18). To accurately assess the performance and select the best models with biomarker combinations, the random resampling process was repeated 20 times, and the model with the median AUC score (area under the curve) was obtained to represent the performance of the final model classification.

RESULTS

Peptide selection—Lyme disease diagnostic model

We pursued a machine learning approach to identify reactive 12-mer linear peptides of B. burgdorferi that could be used in a stepwise fashion to (i) identify serologic signatures unique to Lyme disease and (ii) distinguish cohorts with different stages and/or manifestations of Lyme disease. We first used a case/control data set to identify the minimum set of peptides that could differentiate sera of patients with early Lyme disease from healthy controls. The Lyme disease cases consisted of 38 sera samples from confirmed early Lyme disease patients presenting with erythema migrans (LDB cohort), collected at the time of the diagnosis (acute sera) (Table 1). For controls, we used a merged data set of 85 sera samples from LDB (N = 38) and NIH (N = 47) cohorts. The combined case and control data set consisted of 123 samples. The initial differential analysis identified 1,169 (12.8%) IgG or IgM-reactive peptides with a significantly higher expression in cases vs controls. We used the random forest method to downselect this peptide panel into the minimum number of peptides with the lowest degree of predictive error. The final panel consisted of 62 reactive peptides (31 IgG and 31 IgM) and generated an error rate of 7.3% (Table 4). By using this panel, a total of 31 out of 38 early acute Lyme disease samples were predicted as cases. Of the 85 healthy controls, two were also classified as cases.

TABLE 4.

List of peptides and their importance for the Lyme disease diagnostic model

Mean decrease Gini	Peptide sequence	Antigen	Antibody class
4.34	QIAAAIALRGRA	VlsE C6	IgG
3.55	NQIAAAIALRGM	VlsE C6	IgG
3.14	QIAAAIALRGMA	VlsE C6	IgG
2.62	HIAAAIALRGMA	VlsE C6	IgG
1.98	DNQIAAAIALRG	VlsE C6	IgG
1.79	PIAAAIALRGMA	VlsE C6	IgG
1.72	NPIAAAIALRGM	VlsE C6	IgG
1.63	DQIAAAIALRGM	VlsE C6	IgG
1.57	DQIAAAIALRGR	VlsE C6	IgG
1.22	PAQEGAQQEGVQ	FlaB	IgG
1.09	AAMNGNDKIAAA	VlsE C6	IgM
1.04	VQQEGAQQPALA	FlaB	IgM
0.9	QSAPVQEGVQQE	FlaB	IgG
0.81	DDHIAAAIALRG	VlsE C6	IgG
0.81	DAGKLFAKKNDA	VlsE C3	IgG
0.77	DAGKLFAKKNDE	VlsE C3	IgG
0.76	AGDGGEKAGVKA	VlsE	IgM
0.74	LFGKAGAGGDSE	VlsE	IgM
0.73	DAGKLFAKKNDD	VlsE C3	IgG
0.72	QEGAQQPALATA	FlaB	IgM
0.71	KDGKFAVKSNDE	VlsE C6	IgM
0.71	GKLFAKKNDDGD	VlsE C3	IgM
0.7	DDQIAAAIALRG	VlsE C6	IgG
0.66	GVQQEGAQQPAL	FlaB	IgM
0.65	AGMNGNDKIAAA	VlsE C6	IgM
0.64	IGEGNGDAEFNQ	VlsE	IgM
0.64	QAAPVQEGAQQE	FlaB	IgG
0.63	GKLFAKKNDDGD	VlsE C3	IgG
0.6	VQQEGAQQPAPV	FlaB	IgM
0.6	GCNLDDNSKMER	S2	IgG
0.57	CNLDDNSKMERE	S2	IgG
0.55	QEGVQQEGAQQQ	FlaB	IgM
0.55	VKLTISDDLNKT	OspA	IgM
0.54	GCNLDDNSKIER	S2	IgG
0.53	GGMNGNDKIAAA	VlsE C6	IgM
0.53	LKNSEELNKKIE	OspC	IgM
0.52	QEGAQQEGVQAA	FlaB	IgM
0.51	EGAQQEGAQQPT	FlaB	IgM
0.51	KDGKFAVKKDEE	VlsE C6	IgM
0.5	IKAIVDAAGNGG	VlsE	IgM
0.49	KDKDGKYSLDAT	OspA	IgM
0.49	CNLDDNSKMERK	S2	IgG
0.47	NEDAGKLFAAKN	VlsE C3	IgG
0.45	KGLNAKIDSLDV	BdrK	IgM
0.44	IVDAAGGGEQDG	VlsE	IgM
0.43	QQEGAQQPALAT	FlaB	IgM
0.42	CNLDDNSKIERK	S2	IgG
0.41	QEGVQQEGAQQS	FlaB	IgM
0.41	EKQFGIKFDNLI	BdrN	IgM
0.41	QSAPVQEGVQQE	FlaB	IgM
0.4	VQDGVQQEGAQQ	FlaB	IgM
0.38	KDGKFAVKSDGD	VlsE C6	IgG
0.36	QEGVQQEGAQQP	FlaB	IgM
0.35	DAGKLFAAKNAN	VlsE C3	IgG
0.35	TNPIAAAIALRG	VlsE C6	IgG
0.35	EGVQQEGAQQPA	FlaB	IgM
0.32	QAAPVQEGVQQE	FlaB	IgM
0.32	QEGVQQEGARQP	FlaB	IgM
0.3	PVQEGVQQEGAR	FlaB	IgG
0.29	KDGKFAVKDERE	VlsE C6	IgG
0.23	QVAPVQEGVQQE	FlaB	IgG
0.22	VQEGVQQEGAQQ	FlaB	IgG

Characterization of peptides in the diagnostic model

The selected 62 peptides mapped to 14 different regions within the B. burgdorferi proteome and often included multiple versions of the same 12-mer fragment, with each version containing variations in the amino acid (aa) sequence associated with differences in strain origin (Fig. 1). The majority of the 62 peptides originated from VlsE and FlaB and included the key peptides driving the diagnostic model (Table 4). Most of the VlsE peptides mapped to two invariable (IR) domains. Six IgG-reactive peptides and one IgM-reactive peptide were mapped to a IR3 and partial variable (VR3) fragment corresponding to aa 197 to 212 of the B31 strain (Fig. 1A). Twelve IgG-reactive peptides clustered within a 14-aa portion that corresponds to aa 4–17 (shown in bold) of the B31 IR6 (C6) sequence MKKDDQIAAAIALRGMAKDGKFAVKD. All these IgG-reactive peptides contained a conserved internal IAAAIALRG motif (Fig. 1B). In addition, three IgM-reactive peptides mapped 5 aa upstream of the IgG peptides, and all included a MNGNDKIAAA motif. Two IgG and two IgM peptides mapped to the C-terminal part of the C6, and they all contained a KDGKFAVK motif. All IgG (N = 15) and IgM (N = 6) FlaB peptides mapped to a highly reactive 23-aa fragment located within residues 207 and 229 (Fig. 1C). Combined, 46 out of the 62 peptides in our model included peptides within these three fragments in VlsE and FlaB. The remaining 16 peptides included five peptides that clustered within a 13-aa portion of the N terminus of the S2 antigen, as well as peptides from within Borrelia direct repeat proteins K and N, OspA, OspC, and other regions within VlsE (Table 4).

Fig 1.

Mapping of the key VlsE (panels A and B) and FlaB (panel C) peptides identified by the diagnostic model, which differentiates between patients with Lyme disease and controls. The peptides were mapped to the B31 sequence. The numbers above the sequence correspond to the amino acid position in the protein. IgG peptides are indicated in blue and IgM in orange.

A closer examination of individual peptides within the C6 revealed predominantly IgG reactivity, which was mostly confined within the N-terminal half of the 26-aa sequence of the C6 (Fig. 2). Using the B31 C6 sequence as a reference, we noted that the fourth, fifth, and sixth peptides (DDQIAAAIALRG, DQIAAAIALRGM, and QIAAAIALRGMA) were the predominant reactive peptides in the samples from the LDB cohort. All three peptides were identified as key predictive drivers of our differential peptide panel.

Fig 2.

IgG and IgM reactivity of individual peptides that make up the C6 fragment. Shown is the reactivity of sera from SEM-A samples. Each number on the X axis corresponds to a unique 12-mer peptide sequence displayed on the right. Panel A shows the average from 30 C6 ELISA positive samples. Panel B demonstrates the average from eight C6 ELISA negative samples. The three dominant reactive peptides are indicated with *.

Two healthy control samples (LYM-518 and LYM-1211) were classified as Lyme disease by our model. We examined the array data to determine if these two samples yielded antibody signatures consistent with Lyme disease. Sample LYM-518, part of the NIH control cohort, generated elevated IgG reactivity against multiple peptides within the 207–229 FlaB fragment (Fig. S3). The misclassification of sample LYM-1211, from the LDB cohort, was less clear. Nonetheless, we did note slight (>3 fold) elevated reactivity to multiple VlsE and FlaB peptides in our model compared to controls that could account for the positive classification.

Prediction algorithm results

Using the same 123-sample data set, we trained our model on a set of randomly selected 99 samples (80% of the data set) and then used it on a tester set that consisted of the remaining 24 samples (20% of the data set). The tester set included seven Lyme disease cases and 17 healthy controls. The clinical status of five out of seven Lyme disease cases and all controls were correctly predicted (AUC = 0.96). Of the five predicted Lyme disease cases, four were positive by the standard two-tiered (STT) algorithm. The two Lyme disease samples classified as controls by our model (LYM-1099 and LYM-1114) were B. burgdorferi PCR-positive and STT-negative, respectively, although LYM-1114 was positive by a commercial C6 peptide ELISA. However, neither sample displayed any significant reactivity with any of the C6 peptides on the array.

We next employed our model on 31 sera of patients with probable Lyme disease from the LDB cohort. All samples were classified as STT-negative, although four samples had a positive C6-peptide ELISA, and several others had a single positive whole-cell ELISA or Western Blot IgM test (Table 2). Our model predicted nine samples (29%) as representing subjects with Lyme disease. These included two samples (LYM-991 and LYM-1081) that were positive on the C6-peptide ELISA.

We also determined if clinical features could be used as a predictor of positive serology in both confirmed and probable cases of early Lyme disease. There was no significant correlation between the size of the erythema migrans rash or the presence of multiple symptoms with positive results as determined by our model (Wilcoxon test, P = 0.65 and 0.52, respectively).

Next, we applied our model to predict the Lyme disease status of subjects in the NIH cohort comprising 82 cases and 47 healthy controls. Of the 82 cases, 77 were positive by STT or C6 alone with the five negative samples, all from the SEM-C group. Our model correctly identified all controls and 81 out of 82 Lyme disease samples. The lone misclassified sample, LYM-465, was an SEM-C sample, which, upon review, was non-reactive for all B. burgdorferi antigens on the array. This sample was also negative by a commercial C6 ELISA and STT.

FlaB and C6 include the major immunodominant linear epitopes of B. burgdorferi

Because VlsE and FlaB peptides were the key peptides in our diagnostic model, we examined the array intensity to determine antibody abundance to these targets relative to other antigens. For each of the five Lyme disease groups, we identified peptides reactive in the cases vs all 85 controls and sorted them by intensity. As anticipated, we observed a wide range of redundancy among the reactive peptides. Nonetheless, intensity data revealed that the key peptides from the VlsE C6, VlsE IR3, and the 207–229 FlaB fragments that were driving diagnosis in our model were also among the most immunoreactive peptides on the array throughout all five groups (Fig. 3). The lone exception was in the ALNB group, where the FlaB 207–229 peptides displayed lower IgG reactivity, but instead were the highest reactive IgM targets (Fig. 4; Fig. S4).

Fig 3.

List of top five reactive peptides from each Lyme disease group. A single representative peptide was selected for each reactive epitope. The C6 representative peptide are shown in blue, the FlaB peptides are shown in red; the rest of the peptides are shown in gray. The SEM-A cohort included only confirmed Lyme disease cases.

Fig 4.

IgG reactivity to FlaB peptides in the five sample types. The Y axis indicates the relative amino acid coordinates of the peptides within the full-length FlaB protein. Reactivity is shown in yellow. For clarity, only peptides with the reactivity above 10,000 RFU are shown. Samples are indicated on the X axis. To illustrate baseline reactivity, ten random normal control samples were selected and shown on the right. The red asterisks indicate the position of the key regions identified by our models. * indicates the peptides within residues 147–158; ** - indicates the highly immunoreactive region encompassing residues 207–229; ^ - includes only confirmed Lyme disease cases.

Cohort differential model

Next, we used the random forest approach to identify a panel of peptides that could differentiate between different clinical manifestations of Lyme disease. Our combined data set consisted of 107 samples and included 25 SEM-A samples that had an erythema migrans >5 cm from the LDB cohort (Table 1) and the four Lyme disease types (SEM-C, MEM, ALNB, and LA) from the NIH cohort. By downselecting the number of differential peptides with a random forest model, we selected 20 peptides as the optimal combination with an OOB error of 12.15% (Table 5). The model provided 100% accuracy in predictions for SEM-A (25/25) and LA (27/27) samples. The predictive accuracy for SEM-C samples was 81.5% (22/27), with two samples classified as SEM-A, two samples as MEM, and one sample as LA. For patients with ALNB, the prediction was 87% (13/15), with one sample classified as LA and another as MEM. The lowest accuracy was observed for MEM samples, with six out of 13 samples misclassified. One sample was misclassified as ALNB, two as LA, and three as SEM-C. We generated a three-dimensional principal component analysis (PCA) plot using the IgG raw intensity values of the 20 peptides selected by our model to visualize the separation of the five groups (Fig. 5). We observed a clear separation for the LA group and an association of the selected peptides with late disease.

TABLE 5.

List of peptides and their importance in the differential model

Mean decrease gini	Sequence	Antibody class	Antigen	Sequence origin - Annotation
5.330794088	MIINHNTSAINA	IgG	FlaB	NP_212281.1 flagellin [Borreliella burgdorferi B31]
5.177076454	GKDDPFSAYIKV	IgG	p66	NP_212737.1 integral outer membrane protein p66 [Borreliella burgdorferi B31]
4.857495488	NNQTEQSSTSTK	IgG	p66	NP_212737.1 integral outer membrane protein p66 [Borreliella burgdorferi B31]
4.669973934	DKDDPTNKFYQS	IgG	VlsE N	YP_004940414.1 outer surface protein VlsE1 (plasmid) [Borreliella burgdorferi B31]
4.512979257	TAEELGMQPAKT	IgG	FlaB	ABW79842.1 flagellin, partial [Borreliella burgdorferi]
4.501736876	SGKDDPTNKFYQ	IgG	VlsE N	ADQ30189.1 vlsE protein (truncated), partial (plasmid) [Borreliella burgdorferi JD1]
4.446413545	LGKDDPFSAYIK	IgG	p66	NP_212737.1 integral outer membrane protein p66 [Borreliella burgdorferi B31]
4.325870964	ENSGKDDPTNKF	IgG	VlsE N	ADQ30189.1 vlsE protein (truncated), partial (plasmid) [Borreliella burgdorferi JD1]
4.268780356	ESIKNEFLNKGF	IgM	BdrK	ADQ44869.1 BdrK (plasmid) [Borreliella burgdorferi 297]
4.124715706	KDDPTNKFYQSV	IgG	VlsE N	YP_004940414.1 outer surface protein VlsE1 (plasmid) [Borreliella burgdorferi B31]
4.025256035	MAKDGKFAVKKG	IgG	VlsE C6	ACD00653.1 VlsE, partial (plasmid) [Borreliella burgdorferi]
3.971709872	KDGKFAVKSGGG	IgG	VlsE C6	ACD00984.1 VlsE, partial (plasmid) [Borreliella burgdorferi]
3.968590071	KDDDAKAFGKGK	IgG	VlsE	ACN55594.1 outer surface protein VlsE (plasmid) [Borreliella burgdorferi WI91-23]
3.842963268	GKKPADAKNPIA	IgM	VlsE V5 C5	ACD00940.1 VlsE, partial (plasmid) [Borreliella burgdorferi]
3.744512829	ILKAIVEAAGVS	IgG	VlsE	ACN55594.1 outer surface protein VlsE (plasmid) [Borreliella burgdorferi WI91-23]
3.738322943	NAAAFGGNMKKK	IgG	VlsE V6-C6	WP_002662199.1 variable large family protein [Borreliella burgdorferi]
3.736929574	ANGDAGHLFAAA	IgM	VlsE	ACO38545.1 borrelia lipoprotein (plasmid) [Borreliella burgdorferi 29805]
3.339991462	TAEELGMQPAKI	IgG	FlaB	NP_212281.1 flagellin [Borreliella burgdorferi B31]
3.313878237	DGAEFNKEGMKK	IgM	VlsE	ACC99986.1 VlsE, partial (plasmid) [Borreliella burgdorferi]
3.168644553	KKPGDAKNPIAA	IgM	VlsE V5-C5	ACD01023.1 VlsE, partial (plasmid) [Borreliella burgdorferi]

Fig 5.

PCA plot for individual samples from the five groups. Each point on the plot represents an individual sample. The different colors represent samples from subjects with different types of disease. SEM-A: , SEM-C: The full 3D PCA plot can be accessed at https://magical-muffin-f665c4.netlify.app/.

Characterization of peptides in a differential model

The differential model was driven by IgG-reactive peptides from FlaB, p66, and VlsE. Similar to the diagnostic model, there was redundancy within the selected peptide panel, with 12 distinct regions represented within the 20 peptides. The three FlaB peptides mapped to two regions; the key peptide driving the model was the peptide MIINHNTSAINA encompassing the first 12 aa residues of FlaB (Fig. 5). This peptide was not reactive in SEM-A and SEM-C samples. Two redundant peptides mapped to coordinates 147–158 and were reactive primarily in LA samples (Fig. 5). The three p66 peptides originated from two dominant reactive regions (Fig. 6). Two peptides were mapped to aa 78–90 and consisted of the aa sequence LGKDDPFSAYIKV that was highly reactive in most of LA sera. Another p66 peptide mapped to aa 497–508 and consisted of the sequence NNQTEQSSTSTK that was highly reactive in the majority of LA samples and, overall, was among the highest reactive peptides in this cohort (Fig. 3). These p66 peptides were mostly nonreactive in SEM-A, SEM-C, and ALNB sera and reactive in only four of 13 MEM sera.

Fig 6.

IgG reactivity to p66 peptides in the five sample types. The Y axis indicates the relative amino acid coordinates of the peptides within the full-length p66 protein. The reactivity is shown in yellow. For clarity, only peptides with reactivity above 10,000 RFU are shown. Samples are indicated on the X axis. To illustrate baseline reactivity, ten random control samples were selected and shown on the right. The red asterisks indicate the position of the key peptides identified by our differential model, at position 78–90 and 497–508. ^ - includes only confirmed Lyme disease cases.

Thirteen peptides from seven different fragments mapped to VlsE. Four of them were from the N terminal region of the protein (VlsE 18–38), and all included a conserved KDDPTNKF motif (Fig. 7). The immunoreactivity to this region was strongly associated with late disease (Fig. 8). Along with C6, peptides within this region were the most reactive of all Borrelia peptides in the LA samples (Fig. 3). Other VlsE peptides consisted of peptides within the IR5 region, VR5-IR6, and IR6 (Fig. 7).

Fig 7.

Mapping of the main VlsE peptides identified by the differential model. Panel A shows the peptides that mapped to a fragment within the N terminus, and panel B indicates peptides within the C5/VR5 and C6/VR6 region. The numbers above the sequence correspond to the amino acid positions in the protein. IgG peptides are indicated in blue. IgM peptides are indicated in orange.

Fig 8.

IgG reactivity to VlsE peptides in the five sample types. The Y axis indicates the relative amino acid coordinates of the peptides within the full-length VlsE protein. Reactivity is shown in yellow. For clarity, only peptides with reactivity above 10,000 RFU are shown. Samples are indicated on the X axis. To illustrate baseline reactivity, ten random control samples were selected and shown on the right. The asterisks indicate the regions encompassing the peptides within aa 18–38 (*) and the C6 (**). ^ - includes only confirmed Lyme disease cases.

DISCUSSION

Our aim in this study was to identify diagnostic immune signatures for progressive stages of Lyme disease. We used a combination of high-density peptide arrays and machine learning in a two-step approach. In the first step, we used a diagnostic model for selection of diagnostic Lyme disease antibody-reactive peptides. In the second step, we used a differential model to select reactive peptides associated with disease stage.

FlaB and VlsE were the major antibody targets throughout all stages of disease and contained peptides with the foremost predictive value in both of our models. The diagnostic model was driven primarily by peptides originating from a FlaB fragment located within residues 191–231 and two invariable regions (IR) of VlsE. The sequence encompassing the FlaB 191–231 fragment is a major immunodominant region of FlaB in B. burgdorferi and other Borreliae (19). IgM immunoreactivity was detected to peptides from throughout the length of this fragment, whereas IgG reactivity was confined to approximately 33 aa within residues 207–229. The FlaB 191–231 fragment included some of the most immunoreactive peptides for both IgG and IgM in all Lyme disease stages, although IgM reactivity waned in LA patients. We also detected intermittent, mostly IgM reactivity of peptides within this fragment to control sera. FlaB is a key component of both IgM and IgG Western blots used for Lyme disease serodiagnosis, and cross-reactivity to FlaB on both blots is not uncommon in patients without a documented history of Lyme disease (20, 21). The reactivity we observed in our control samples may explain the source of the false-positive Western blot results. Therefore, despite clear diagnostic utility of this large immunoreactive fragment, only a focus on select smaller peptides like the ones identified by our model is likely to provide the required specificity.

The majority of the approximate 350-aa sequence of VlsE is divided into alternating fragments of genetically heterogenous (or variable, VR) and invariable regions (IR) (22, 23). The 26-aa-long IR6, or C6, region is a well-known target of specific B. burgdorferi antibodies and has been exploited in Lyme disease serodiagnostic assays (24, 25). Peptides within the C6 were typically among the first and most reactive B. burgdorferi peptides in patients with early disease, and peptides located at both the N and C termini of this region were selected in our diagnostic model. However, in agreement with other studies, we observed that the N terminal portion constitutes the primary immunoreactive linear antigenic portion within the C6 (26). We also found that the 9-aa fragment IAAAIALRG serves as the key antigenic motif in C6, and 12-mer peptides that included this sequence were the primary peptides driving the diagnostic model. Additional VlsE peptides, particularly within the IR3 and containing a GKLF motif, were also included in the diagnostic model. However, the diagnostic utility of the IR3 peptides may be partially compromised by their higher degree of sequence divergence relative to C6 in different strains of B. burgdorferi. In addition, we surprisingly found substantial IgM reactivity to VlsE peptides, including within the C6 region. Other peptides in the diagnostic model included peptides within the S2 (BB_RS05130, old designation as BBA04), Bdr, and OspA antigens. Both S2 and Bdr are plasmid-encoded antigens that are expressed at higher levels in B. burgdorferi during vertebrate infection. The selection of the two OspA IgM-reactive peptides in the model was surprising, as OspA is a tick-associated lipoprotein that is not expressed by B. burgdorferi during vertebrate infection (27). Nonetheless, the presence of anti-OspA antibodies and their potential utility for diagnosis have previously been demonstrated (28, 29). We propose the reactivity to these peptides could stem from the immune interaction with a limited number of spirochetes that did not clear OspA from their surface during early infection.

Our differential model was utilized to determine if temporal antigen expression and the subsequent development of the antibody response could be associated with a particular disease presentation. Despite examining a wide range of antigens, the optimal predictive model was built with peptides only from VlsE, FlaB, and p66. The predictive accuracy of the differential model was most pronounced for SEM-A (early disease) and LA (late disease) samples, mostly because the primary drivers of the model were peptides from epitopes reactive predominately in LA sera and nonreactive in SEM-A.

The p66 is one of the 10 antigens recognized on the Lyme disease serodiagnostic IgG Western blot (20). Previous epitope mapping efforts by Arnaboldi et al. revealed a lack of specific regions within p66 that were useful for serodiagnosis of early Lyme disease (30). We also did not identify consistently reactive p66 epitopes in samples from early disease. Although we did identify reactive peptides within several p66 fragments, including regions located at aa 223–271 and 331–361, they were reactive in <50% of sera from each group. The p66 peptides selected in our differential model originated from two distinct reactive portions of p66, located at aa 78–90 and 497–508, and were reactive almost exclusively in the LA samples. Thus, our results indicate that antibodies to p66 78–90 and 497–508 arise late in disease and represent IgG fragments of p66 that can be useful for serologic differentiation between early and late disease.

Similarly, the 18–38 N terminal region of VlsE is also a major target of antibodies late in disease. Along with C6, peptides from the 18–38 fragment included the most immunoreactive peptides in sera from LA patients. However, unlike the C6, this region was largely nonreactive in non-LA sera. A strong antibody response to this region was uncovered in patients with posttreatment Lyme disease syndrome (31). Our data suggest that the 18–38 region and the C6 represent two major sequence-conserved VlsE targets of antibodies during disease, with the C6 antibodies arising first, and the antibodies to the N terminal region only arising during the latter progressive stages of infection.

In agreement with prior studies, our findings indicate that epitopes within VlsE and FlaB are key targets for Lyme disease antibody detection assays. Accordingly, both of these antigens have been utilized in the majority of Lyme disease serologic tests. Similar to our work here, other studies that employed epitope mapping have identified peptides within the IR6 of VlsE and the FlaB 191–231 fragment as targets with high utility for serologic diagnosis (26, 32) Consequently, these shorter peptide fragments, either separately or combined, have been included in several peptide-based serologic assays. The utility of a concatemer using both the partial IR6 and a FlaB-13 mer for all Lyme disease stages has been demonstrated (33) (34). Our finding that these peptides represent the optimum serologic targets throughout the course of disease adds further validity to these earlier studies.

Of the 31 samples listed as “probable Lyme disease,” only nine were predicted as serologically positive by our model. In the absence of conclusive laboratory molecular or serologic findings, the primary rationale for diagnosis of this cohort as probable Lyme disease was the presence of an EM >5 cm. However, EM rashes can be heterogenous in appearance, and skin lesions originating from other, often noninfectious causes can be erroneously characterized as erythema migrans (35). One potential cause of misdiagnosis is the skin lesion associated with the bites of the Lone Star ticks, called Southern Tick-Associated Rash Illness (STARI), a condition of unknown etiology (36). Since Lone Star ticks are increasingly found in Lyme disease endemic areas, there is a growing likelihood of STARI rashes being misdiagnosed as EM (37, 38). It is possible that some of these probable Lyme disease cases may not be caused by B. burgdorferi infection.

A limitation of our study was that we used a partial B. burgdorferi proteome. Although we included the major antigens known to elicit an antibody response, we cannot exclude that other proteins could also improve predictive diagnosis. In addition, our approach only applies to non-conformational epitopes. Nonetheless, our study provides insights into antibody responses at different stages of disease and identified peptides with diagnostic utility.

SUPPLEMENTAL MATERIAL

The following material is available online at https://doi.org/10.1128/mbio.02360-24.

Supplemental Figures. mbio.02360-24-s0001.pdf.

Figures S1 to S4.

ASM does not own the copyrights to Supplemental Material that may be linked to, or accessed through, an article. The authors have granted ASM a non-exclusive, world-wide license to publish the Supplemental Material files. Please contact the corresponding author directly for reuse.