Severe adult malaria is associated with specific PfEMP1 adhesion types and high parasite biomass

Significance

The clinical presentation of severe malaria differs between children and adults, but the factors leading to these differences remain poorly understood. Here, we investigated parasite virulence factors in adult patients in India and show that specific endothelial protein C receptor (EPCR)-binding parasites are associated with severe adult malaria and act together with parasite biomass in patient hospitalization and disease severity. We found substantial differences in EPCR binding activity from severe malaria isolates. However, even parasite domains that partially obstructed the interaction between EPCR and its ligand activated protein C were sufficient to interfere with activated protein C-barrier protective activities in human brain endothelial cells. Thus, restoration of EPCR functions may be a key target for adjunctive malaria drug treatments.

Abstract

The interplay between cellular and molecular determinants that lead to severe malaria in adults is unexplored. Here, we analyzed parasite virulence factors in an infected adult population in India and investigated whether severe malaria isolates impair endothelial protein C receptor (EPCR), a protein involved in coagulation and endothelial barrier permeability. Severe malaria isolates overexpressed specific members of the Plasmodium falciparum var gene/PfEMP1 (P. falciparum erythrocyte membrane protein 1) family that bind EPCR, including DC8 var genes that have previously been linked to severe pediatric malaria. Machine learning analysis revealed that DC6- and DC8-encoding var transcripts in combination with high parasite biomass were the strongest indicators of patient hospitalization and disease severity. We found that DC8 CIDRα1 domains from severe malaria isolates had substantial differences in EPCR binding affinity and blockade activity for its ligand activated protein C. Additionally, even a low level of inhibition exhibited by domains from two cerebral malaria isolates was sufficient to interfere with activated protein C-barrier protective activities in human brain endothelial cells. Our findings demonstrate an interplay between parasite biomass and specific PfEMP1 adhesion types in the development of adult severe malaria, and indicate that low impairment of EPCR function may contribute to parasite virulence.

Severe malaria caused by Plasmodium falciparum is responsible for at least 400,000 deaths every year (1), mainly affecting children younger than 5 y old. However, in areas of low and unstable transmission, severe malaria affects both children and adults (2), although disease symptomatology varies according to patient age. Whereas severe anemia, metabolic acidosis, and cerebral malaria are the major severe syndromes in children, multisystem disease is more common in adults, including renal impairment, jaundice, respiratory distress, metabolic acidosis, and cerebral malaria (3, 4). In addition, disease mortality sharply increases with the age of the patient (4). The factors that drive age-related differences are unknown.

A central process in severe falciparum pathology is the sequestration of infected erythrocytes to microvascular endothelial cells (5). Extensive tissue-specific sequestration results in organ pathology, such as cerebral malaria and placental malaria, and contributes to metabolic acidosis and endothelial dysfunction (6, 7). Proteins of the P. falciparum erythrocyte membrane protein 1 (PfEMP1) family, encoded by the var genes, are responsible for infected red blood cell binding to the microvasculature (8–10). PfEMP1s are classified into three main groups—A, B, and C—based on upstream sequence (UpsA, UpsB, UpsC) and chromosome location (11). The extracellular domain of PfEMP1s presents a modular structure composed of adhesion domains, called Duffy binding-like (DBL) and cysteine-rich interdomain region (CIDR) (12), which sometimes can be found in conserved tandem arrangements known as domain cassettes (DC) (13). Expression of group A PfEMP1 variants (14–18) and PfEMP1 encoding DC8 (15, 19) have been strongly linked with pediatric severe malaria. This subset of PfEMP1s includes mediators of distinct infected erythrocyte adhesion categories, including “rosetting” and endothelial protein C receptor (EPCR) binding. Rosetting involves adhesion to uninfected red blood cells (20, 21), possibly leading to greater microvascular obstruction (22). EPCR binding involves infected erythrocyte adhesion to vascular endothelial cells (23). The important role of the EPCR-activated protein C (APC) pathway in regulating coagulation, inflammation, and endothelial barrier properties (24) has led to the hypothesis that EPCR-binding parasites may drive pathogenic mechanisms by inhibiting the APC–EPCR interaction (23, 25–28), thus increasing vascular dysfunction and permeability. Indeed, cerebral swelling is a major risk factor for pediatric death (29) and there is loss of EPCR and fibrin depositions at sites of cerebral sequestration in pediatric autopsies (30). However, the extent to which severe malaria isolates disrupt EPCR function is poorly understood. A better understanding of adhesion-based pathogenic mechanisms may inform novel targeted adjunctive drug therapies to improve patient survival and outcomes.

Another factor that determines malaria disease severity is total parasite burden. Plasma levels of P. falciparum histidine rich protein 2 (PfHRP2), a surrogate of parasite biomass, can predict disease severity and fatality rates in both children and adults (31, 32), the probability of disease deterioration (33), retinopathy-positive cerebral malaria (34), and whether a fever is caused by malaria (35). Nevertheless, a recent longitudinal study in Tanzanian children showed that high PfHRP2 levels did not necessitate severe disease (36), suggesting severe disease requires additional factors.

In this study, statistical and machine-learning approaches were used to explore the relationship between PfEMP1 expression, parasite biomass, and disease severity in adults with P. falciparum infections treated at the Goa Medical College, and we investigated whether severe malaria isolates impair the APC–EPCR pathway.

Results

Characteristics of the Study Population.

A total of 59 P. falciparum-infected patients from the Goa Medical College were enrolled in the study. Among them, 26 patients had severe malaria (SM) and presented at least one WHO SM criterion (37), 13 patients were hospitalized but did not have any severity criterion for SM (moderately severe malaria, MSM), and 20 were outpatients (OP). As shown in Table 1, the three groups did not present significant differences in age. Previous studies have found a border between pediatric and adult disease symptomatology at 11 y old (4), so four teenagers were included in this study. Among SM patients, 77% presented more than one severity criteria and 57.7% had at least three different severity signs, indicative of multisystem disorders. Because peripheral parasitaemia does not reflect the sequestered parasite population, we measured plasma concentration of PfHRP2 as a surrogate of total parasite biomass (Fig. 1). SM patients had significantly higher plasma PfHRP2 levels than both MSM (P = 0.004) and OP (P < 0.0001) (Fig. 1A). Furthermore, PfHRP2 concentration increased significantly with the number of severe criteria and organ dysfunction (Spearman’s ρ = 0.67, P < 0.0001) (Fig. 1B).

Table 1.

Clinical characteristics of the patients

Patient characteristic	SM (n = 26)	MSM* (n = 13)	OP (n = 20)
Age (mean; IQR), y	33; 22–41	33; 22–47	26; 19–28
Minimum–maximum	18–63	15–62	15–55
Male (%)	88.5	84.6	93.3
P. falciparum and P. vivax coinfection	0	1	0
Parasite density (median), parasites per microliter†	41,410	27,103	51,352
IQR	17,898–93,296	0–47,879	31,811–130,536
Hemoglobin (mean ± SD) (g/dL)	9.9 ± 3.1	11.8 ± 2.4	12.2 ± 2
Glasgow coma score <10 (%)	26.9	0	NA
Glasgow coma score (median)	14	14	NA
Minimum–maximum	3–15	12–15	NA
Respiratory distress (%)	57.7	0	NA
Respiratory rate (mean ± SD), breaths per minute	26.6 ± 9.3	18 ± 0.1	NA
Jaundice (%)	69.2	0	NA
Acute renal failure (%)	46.2	0	NA
Shock (%)	11.5	0	NA
Acidosis (%)	34.6	0	NA
Days in the hospital (median)	7	4	NA
Minimum–maximum	2–31	3–9	NA
Degree of severity (mean)‡	3.2	NA	NA
Percentiles 25 and 75	1.75–5	NA	NA
Mortality (%)	15.4	0	0

^*Patients requiring admission to Goa Medical College who did not fulfill any of the WHO criteria for SM classification.

^†SM (n = 22), MSM (n = 10), OP (n = 12).

^‡Number of WHO SM criteria. IQR, interquartile range; NA, not applicable.

Fig. 1.

Association between PfHRP2 plasma concentrations and disease severity. PfHRP2 levels were compared between patient groups and by number of WHO SM criteria. (A) PfHRP2 plasma concentrations among disease groups. Horizontal lines indicate median for each group. Pairwise comparisons were analyzed using the Mann–Whitney U test. Significant higher concentration is represented by **P < 0.01 and ***P < 0.001. (B) Spearman’s rank correlation coefficient (ρ) and P value for the association between PfHRP2 plasma concentrations and number of severity criteria.

SM Isolates Overexpressed EPCR-Binding var Transcripts.

To identify var genes associated with adult severe disease, we performed quantitative RT-PCR (qRT-PCR) using a set of 5 primers that target groups A, B, or C (VarA, UpsB1, UpsB2, UpsC1, and UpsC2) (16) and 40 degenerate primers that target specific adhesion domains (15). Overall, patients in the study presented a complex population of parasites that transcribed a mixture of A, B, and C var genes (Fig. 2A and Table 2). However, the median VarA transcript level was higher than groups B and C in all patient groups, and SM patients had significantly elevated VarA and UpsB1 transcripts in comparison with OP (Fig. 2A and Table 2).

Fig. 2.

Transcription of UpsA, DC6, and DC8 var is elevated in SM patients. The transcript abundances of var gene subtypes were investigated among patients. (A) Transcript levels of A, B, and C var gene groups and (B) domain subtypes of DC8 and DC6 in SM and OP groups. Horizontal lines indicate median for each group. Differences among groups were compared by using the Mann–Whitney U test. Significant higher transcription is represented by *P ≤ 0.05 and FDR ≤ 0.2, ***P ≤ 0.005 and FDR ≤ 0.05. (C) Heat map showing transcription levels of DC8 and DC6 domain subtypes and VarA genes. Maximum transcription levels are represented in red, minimum transcription in blue, and median transcription levels in white. Color equivalents were set by comparing each primer transcript among all patients analyzed. FDR, Benjamini–Hochberg adjustment for false discovery rate.

Table 2.

Transcript levels of var domain subtypes in severe patients and outpatients

Binding phenotype*	Group	Primer	SM (IQR) n = 24	OP (IQR) n = 19	P value	FDR†
Unknown binding	A	DBLα1.1 of DC1	9.2 (0-13.3)	2.8 (0-10.2)	0.23	0.36
	A	DBLε8 of DC3	6.8 (0.4–36.1)	2.1 (0.1–17.6)	0.27	0.38
EPCR	A	DBLα2/α1.1/2/4/7	67.2 (19.5–271.3)	33 (3.7–152.7)	0.06	0.18
	A	DBLα1.4	0.1 (0-1.8)	0.1 (0-2)	0.11	0.28
	A	CIDRα1.6	0 (0-0.1)	0 (0-0.1)	0.34	0.40
	A	CIDRα1.7	2.5 (0-9.6)	0 (0-17.6)	0.28	0.38
	A	DBLβ3	1.5 (0-20.4)	0.2 (0-7.1)	0.14	0.30
	A	DBLα1.7 of DC13	0.5 (0.1–23)	0.2 (0.1–3.8)	0.2	0.35
	A	CIDRα1.4 of DC13	0.9 (0-2.1)	0.5 (0-3.1)	0.32	0.40
	A	CIDRα1.4 and CIDRα1	2.9 (0.2–22.8)	0 (0-3.6)	0.006	0.11
	B/A	DBLα-CIDRα of DC8	4.5 (1.1–9.6)	0.1 (0-3.1)	0.02	0.13
	B/A	CIDRα1.1 of DC8	9.5 (2.8–22)	2 (0.2–11)	0.04	0.16
	B/A,A	DBLβ12 and DBLβ3/5	45.3 (6.2–89.8)	11.8 (0.3–36.3)	0.01	0.11
	B/A	DBLγ4/6 of DC8	7.5 (2.6–30.6)	1.8 (0.6–12.8)	0.03	0.16
EPCR, CD36, and rosetting	B (A,C)	DBLγ of DC6	19.7 (4.1–66.7)	1.4 (0.3–3.1)	0.0003	0.01
	B (A,C)	DBLζ5 of DC6	5 (0.2–15.6)	0.9 (0.1–5.4)	0.1	0.25
	B	DBLζ4 of DC9	20.9 (2.2–50.6)	20.1 (2.3–60.8)	0.49	0.50
	A,B,C	DBLγ9	0 (0-0)	0 (0-0)	0.46	0.49
CD36	B (C)	DBLε2 of DC7	1.2 (0.2–3.5)	4.3 (0.4–23)	0.04	0.16
	B (A,C)	DBLε3 of DC7	1.9 (0.2–9.4)	3.2 (0.1–6.3)	0.5	0.50
	B	DBLγ of DC9	8.7 (0.4–76.5)	24.8 (1.2–71.5)	0.18	0.35
	B	DBLζ6 of DC10	1.5 (0.3–10.6)	1.2 (0.2–4)	0.18	0.35
	B	DBLα0.16 of DC19	4.9 (0.5–32.4)	1 (0.4–22.3)	0.12	0.28
	B,C	CIDRα3.4 of DC19	21.8 (8.2–57.8)	10.9 (4.6–35)	0.06	0.18
	B	DBLα0.9 of DC20	7.3 (1-29)	2.9 (0.5–10.6)	0.13	0.29
	B	DBLα0.1	1 (0.3–4.6)	0.4 (0.3–1.3)	0.05	0.18
	B	DBLα0.6/9	39.3 (10.6–139.1)	42.5 (1.2–93.1)	0.25	0.38
	B	CIDRα2.2	1.2 (0-40.5)	5.3 (0-48.5)	0.32	0.40
	B	CIDRα2.3/5/6/7/9/10	30.7 (9.9–97.7)	38.3 (9.1–57.7)	0.41	0.46
	B,C	CIDRα3.1–3	6.7 (1.6–28)	2.2 (0.6–7.7)	0.02	0.13
	B	CIDRγ2/9	0.4 (0-3.9)	0.2 (0-4.2)	0.45	0.49
	B (A,C)	DBLβ5	0 (0-4.3)	0 (0-0.3)	0.23	0.36
	B,C	CIDRγ	82.9 (34-140.8)	138.1 (24-315.2)	0.21	0.35
CD36 and rosetting	B,A	DBLε12 of DC12	1.9 (0-15.3)	3 (0.2–9.5)	0.33	0.40
	B	CIDRγ1/2	6.5 (1-30.2)	4.4 (0.6–19.8)	0.32	0.40
Rosetting	A	DBLγ of DC5	0 (0-0.2)	0 (0-0)	0.06	0.18
	A	DBLβ7 & 9 of DC5	0 (0-0)	0 (0-0.1)	0.19	0.35
	A	DBLα1.5/6b of DC16	0.6 (0.1–1.7)	0.6 (0.1–4.4)	0.21	0.35
	A	DBLα1.5/6a of DC16	3.1 (0.3–20.1)	8.2 (0.4–16.5)	0.28	0.38
	A	CIDRδ of DC16	1.5 (0.1–13)	2.9 (0.2–14.7)	0.37	0.43
		VarA	118.9 (27.7–298)	39.8 (3.6–112.2)	0.02	0.13
		UpsB1	36 (18-68.4)	21.4 (5.7–47.2)	0.04	0.16
		UpsB2	18.8 (8.1–30)	13.5 (10.5–19)	0.09	0.24
		UpsC1	5.4 (3.3–18.8)	9.8 (3-15.6)	0.49	0.50
		UpsC2	4.7 (2.4–17.1)	2.2 (1.2–7.8)	0.08	0.23

Median Tu level and IQR of var subtypes in SM and OP groups. P values comparing the SM and OP group were calculated using a one-tailed Mann–Whitney U test.

^*Predicted binding phenotype of PfEMP1 head structure.

^†Benjamini–Hochberg adjustment for FDR. Differential expression of transcripts with P ≤ 0.05 and FDR ≤ 0.2 is represented in boldface and considered significant.

To gain further insight into parasite binding phenotypes associated with adult SM, domain-specific primers were used to identify adhesion subtypes expressed in patients, also taking advantage of the functional specialization of PfEMP1 proteins to infer parasite binding traits. In particular, the N-terminal PfEMP1 head structure (DBL–CIDR tandem) has diversified between group A (EPCR binding or rosetting) and groups B and C (CD36 binders). Head structures containing CIDRα1 subtypes bind EPCR, CIDRβ/γ/δ subtypes are associated with rosetting, and CIDRα2–6 subtypes bind CD36 (Fig. S1) (reviewed in ref. 38). Using in silico analysis, we predicted var genes that would be amplified by the 40 domain-specific primers in seven annotated parasite genomes (13) and assigned a predicted binding phenotype to each gene (EPCR, CD36, or rosetting) depending on its head structure (Fig. S2). Some primers target var domains that are associated with more than one type of head structure (15). Using these criteria, we inferred the CD36, EPCR, and rosetting binding phenotype for each primer (Table 2 and Tables S1 and S2).

Fig. S1.

The relationship between the C-terminal DC6 cassette and PfEMP1 head structures. (A) Schematic showing the chromosomal location of group A, B, and C var genes. The classification and predicted binding properties of PfEMP1 in 3D7 genome reference isolate are shown based on the PfEMP1 head structures. PfEMP1-containing DC8 are classified as B/A. (B) DC6 C-terminal subtypes are found in combination with rosetting (group A), EPCR-binding (group B and B\A), and CD36-binding (group B and C) head structures among the seven Plasmodium falciparum annotated genomes (13).

Fig. S2.

In silico predicted performance of the var domain primers on seven annotated parasite genotypes. Genes that presented annealing of both primers (with no more than one mismatch, which was not present at the 3′ end of the primer) are shown. Rosetting (yellow), EPCR (blue and violet), and CD36-binding (orange) phenotypes were predicted by examining the CIDR domain in the PfEMP1 semiconserved head structure (rosetting = CIDRβ/γ/δ, EPCR-binding = CIDRα1.1/4–8, CD36 binding = CIDRα2–6). DBLβ3 and DBLβ5 domains are associated with ICAM-1 binding (62–64) and some DBLε/ζ subtypes with IgM/α2-macroglobulin binding (54, 55).

Table S1.

Transcript levels of var domain subtypes in severe and moderately severe patients

Binding phenotype*	Group	Primer	SM (IQR) n = 24	MSM (IQR) n = 9	P value	FDR†
Unknown binding	A	DBLα1.1 of DC1	9.2 (0-13.3)	3 (0.5–9.4)	0.24	0.44
	A	DBLε8 of DC3	6.8 (0.4–36.1)	1.7 (0.2–23.7)	0.27	0.44
EPCR	A	DBLα2/α1.1/2/4/7	67.2 (19.5–271.3)	66.5 (45.6–231.9)	0.29	0.44
	A	DBLα1.4	0.1 (0-1.8)	0.5 (0-2.2)	0.38	0.44
	A	CIDRα1.6	0 (0-0.1)	0 (0-0.3)	0.43	0.44
	A	CIDRα1.7	2.5 (0-9.6)	0.9 (0-23.1)	0.29	0.44
	A	DBLβ3	1.5 (0-20.4)	9.1 (0.6–58.8)	0.21	0.44
	A	DBLα1.7 of DC13	0.5 (0.1–23)	4.3 (0.4–19.9)	0.11	0.44
	A	CIDRα1.4 of DC13	0.9 (0-2.1)	0.1 (0.1–1.9)	0.46	0.46
	A	CIDRα1.4 and CIDRα1	2.9 (0.2–22.8)	4.6 (1.3–25.5)	0.41	0.44
	B/A	DBLα-CIDRα of DC8	4.5 (1.1–9.6)	5 (0.9–19.5)	0.22	0.44
	B/A	CIDRα1.1 of DC8	9.5 (2.8–22)	15.3 (4.8–38.3)	0.15	0.44
	B/A,A	DBLβ12 and DBLβ3/5	45.3 (6.2–89.8)	21.3 (28.9–54.2)	0.07	0.39
	B/A	DBLγ4/6 of DC8	7.5 (2.6–30.6)	27.4 (4.9–83.6)	0.28	0.44
EPCR, CD36, and rosetting	B (A, C)	DBLγ of DC6	19.7 (4.1–66.7)	3.3 (1.8–9.5)	0.07	0.39
	B (A, C)	DBLζ5 of DC6	5 (0.2–15.6)	2.1 (0.7–3.7)	0.32	0.44
	B	DBLζ4 of DC9	20.9 (2.2–50.6)	8.7 (3.2–16.9)	0.18	0.44
	A,B,C	DBLγ9	0 (0-0)	0 (0-0.8)	0.05	0.39
CD36	B (C)	DBLε2 of DC7	1.2 (0.2–3.5)	2.3 (0.06–9)	0.06	0.39
	B (A, C)	DBLε3 of DC7	1.9 (0.2–9.4)	3 (0.1–14.7)	0.06	0.39
	B	DBLγ of DC9	8.7 (0.4–76.5)	3.6 (1.8–21.3)	0.38	0.44
	B	DBLζ6 of DC10	1.5 (0.3–10.6)	20 (3-23.9)	0.01	0.23
	B	DBLα0.16 of DC19	4.9 (0.5–32.4)	7.2 (2.4–35.1)	0.4	0.44
	B,C	CIDRα3.4 of DC19	21.8 (8.2–57.8)	24.8 (11.9–130.2)	0.33	0.44
	B	DBLα0.9 of DC20	7.3 (1-29)	1.1 (0.4–58)	0.19	0.44
	B	DBLα0.1	1 (0.3–4.6)	1.4 (0.4–6.3)	0.26	0.44
	B	DBLα0.6/9	39.3 (10.6–139.1)	18.8 (1.5–43.5)	0.21	0.44
	B	CIDRα2.2	1.2 (0-40.5)	13 (1.3–120.2)	0.007	0.23
	B	CIDRα2.3/5/6/7/9/10	30.7 (9.9–97.7)	33.7 (24-137.7)	0.25	0.44
	B,C	CIDRα3.1–3	6.7 (1.6–28)	8.3 (4-26)	0.14	0.44
	B	CIDRγ2/9	0.4 (0-3.9)	1.2 (0-3.7)	0.11	0.44
	B (A, C)	DBLβ5	0 (0-4.3)	2.9 (0-14.7)	0.23	0.44
	B, C	CIDRγ	82.9 (34-140.8)	98.7 (24.9–239.4)	0.19	0.44
CD36 and rosetting	B, A	DBLε12 of DC12	1.9 (0-15.3)	0 (0-2.3)	0.43	0.44
	B	CIDRγ1/2	6.5 (1-30.2)	19.9 (4.5–40.2)	0.04	0.39
Rosetting	A	DBLγ of DC5	0 (0-0.2)	0.1 (0-0.5)	0.26	0.44
	A	DBLβ7 & 9 of DC5	0 (0-0)	0.1 (0-0.1)	0.43	0.44
	A	DBLα1.5/6b of DC16	0.6 (0.1–1.7)	0.6 (0.2–7.2)	0.31	0.44
	A	DBLα1.5/6a of DC16	3.1 (0.3–20.1)	2.8 (0.3–23.8)	0.3	0.44
	A	CIDRδ of DC16	1.5 (0.1–13)	1.6 (0-34.1)	0.39	0.44
		VarA	118.9 (27.7–298)	46.1 (24.2–258.7)	0.14	0.44
		UpsB1	36 (18-68.4)	28.4 (12.9–32.2)	0.14	0.44
		UpsB2	18.8 (8.1–30)	13.6 (8.5–38.2)	0.38	0.44
		UpsC1	5.4 (3.3–18.8)	8.5 (1.9–14)	0.34	0.44
		UpsC2	4.7 (2.4–17.1)	4.6 (3.2–9)	0.35	0.44

Median Tu level and IQR of var subtypes in SM and MSM groups. P values comparing the SM and MSM group were calculated using a one-tailed Mann–Whitney U test.

^*Predicted binding phenotype of PfEMP1 head structure.

^†Benjamini–Hochberg adjustment for FDR. Differential expression of transcripts with P ≤ 0.05 and FDR ≤ 0.2 was considered significant.

Table S2.

Transcript levels of var domain subtypes in hospitalized patients and outpatients

Binding phenotype*	Group	Primer	SM + MSM (IQR) n = 33	OP (IQR) n = 19	P value	FDR†
Unknown binding	A	DBLα1.1 of DC1	4.4 (0-12.6)	2.8 (0-10.2)	0.30	0.39
	A	DBLε8 of DC3	3.8 (0.3–27.9)	2.1 (0.1–17.6)	0.29	0.39
EPCR	A	DBLα2/α1.1/2/4/7	66.5 (24.8–236)	33 (3.7–152.7)	0.02	0.10
	A	DBLα1.4	0.1 (0-1.8)	0.1 (0-2)	0.21	0.34
	A	CIDRα1.6	0 (0-0.13)	0 (0-0.1)	0.45	0.45
	A	CIDRα1.7	2.1 (0-8.8)	0 (0-17.6)	0.22	0.34
	A	DBLβ3	3.5 (0-33.1)	0.2 (0-7.1)	0.06	0.17
	A	DBLα1.7 of DC13	1 (0.1–22.3)	0.2 (0.1–3.8)	0.14	0.27
	A	CIDRα1.4 of DC13	0.7 (0-2)	0.5 (0-3.1)	0.31	0.39
	A	CIDRα1.4 and CIDRα1	3 (0.6–22.1)	0 (0-3.6)	0.002	0.05
	B/A	DBLα-CIDRα of DC8	4.9 (1-9.6)	0.1 (0-3.1)	0.006	0.06
	B/A	CIDRα1.1 of DC8	7.9 (3.2–22.2)	2 (0.2–11)	0.01	0.06
	B/A,A	DBLβ12 and DBLβ3/5	40.5 (7.2–77.3)	11.8 (0.3–36.3)	0.03	0.12
	B/A	DBLγ4/6 of DC8	8.4 (3.7–32.7)	1.8 (0.6–12.8)	0.01	0.06
EPCR, CD36, and rosetting	B (A,C)	DBLγ of DC6	13.4 (2.2–44.1)	1.4 (0.3–3.1)	0.001	0.05
	B (A,C)	DBLζ5 DC6	2.5 (0.2–13.4)	0.9 (0.1–5.4)	0.13	0.27
	B	DBLζ4 of DC9	10.4 (2.8–36.2)	20.1 (2.3–60.8)	0.40	0.44
	A,B,C	DBLγ9	0 (0-0)	0 (0-0)	0.20	0.34
CD36	B (C)	DBLε2 DC7	1.3 (0.1–3.8)	4.3 (0.4–23)	0.02	0.10
	B (A,C)	DBLε3 DC7	2.1 (0.2–10.4)	3.2 (0.1–6.3)	0.29	0.39
	B	DBLγ of DC9	7.4 (0.9–54.2)	24.8 (1.2–71.5)	0.15	0.28
	B	DBLζ6 of DC10	2.3 (0.6–15.5)	1.2 (0.2–4)	0.08	0.20
	B	DBLα0.16 of DC19	5 (1.2–30.9)	1 (0.4–22.3)	0.08	0.20
	B,C	CIDRα3.4 of DC19	22.7 (8.5–71.1)	10.9 (4.6–35)	0.06	0.17
	B	DBLα0.9 of DC20	5.7 (0.9–24.3)	2.9 (0.5–10.6)	0.14	0.27
	B	DBLα0.1	1.2 (0.3–4.5)	0.4 (0.3–1.3)	0.04	0.15
	B	DBLα0.6/9	26.3 (7.1–105.6)	42.5 (1.2–93.1)	0.37	0.44
	B	CIDRα2.2	4.6 (0-60.8)	5.3 (0-48.5)	0.38	0.44
	B	CIDRα2.3/5/6/7/9/10	33.7 (14.1–93)	38.3 (9.1–57.7)	0.43	0.45
	B,C	CIDRα3.1–3	8.1 (2.9–26)	2.2 (0.6–7.7)	0.01	0.06
	B	CIDRγ2/9	0.5 (0-3.9)	0.2 (0-4.2)	0.31	0.39
	B (A,C)	DBLβ5	0 (0-7.6)	0 (0-0.3)	0.16	0.29
	B,C	CIDRγ	86.5 (35.2–150.4)	138.1 (24-315.2)	0.25	0.36
CD36 and rosetting	B,A	DBLε12 of DC12	0.4 (0-10.6)	3 (0.2–9.5)	0.21	0.34
	B	CIDRγ1/2	9.4 (1.2–30.5)	4.4 (0.6–19.8)	0.12	0.27
Rosetting	A	DBLγ of DC5	0 (0-0.3)	0 (0-0)	0.03	0.12
	A	DBLβ7 & 9 of DC5	0 (0-0.1)	0 (0-0.1)	0.25	0.36
	A	DBLα1.5/6b of DC16	0.6 (0.1–3.6)	0.6 (0.1–4.4)	0.37	0.44
	A	DBLα1.5/6a of DC16	2.9 (0.3–20.4)	8.2 (0.4–16.5)	0.41	0.44
	A	CIDRδ of DC16	1.7 (0.1–14.4)	2.9 (0.2–14.7)	0.44	0.45
		VarA	113.8 (28.4–273)	39.8 (3.6–112.2)	0.01	0.06
		UpsB1	30.1 (15.6–61.9)	21.4 (5.7–47.2)	0.06	0.17
		UpsB2	16 (8.5–33)	13.5 (10.5–19)	0.14	0.27
		UpsC1	5.4 (3.2–18)	9.8 (3-15.6)	0.41	0.44
		UpsC2	4.6 (2.5–12.6)	2.2 (1.2–7.8)	0.05	0.17

Median Tu level and IQR of var subtypes in hospitalized and outpatients. P values comparing the hospitalized (SM + MSM) and OP group were calculated using a one-tailed Mann–Whitney U test.

^*Predicted binding phenotype of PfEMP1 head structure.

^†Benjamini–Hochberg adjustment for FDR. Differential expression of transcripts with P ≤ 0.05 and FDR ≤ 0.2 is represented in boldface and considered significant.

Compared with outpatients, the SM group presented significantly higher transcript levels of var genes with an expected EPCR-binding phenotype (DC8: DBLα-CIDRα, CIDRα1.1, DBLβ12 and DBLβ3/5, DBLγ4/6; EPCR binders: CIDRα1.4 and CIDRα1) or domains associated with EPCR, rosetting, or CD36 binding PfEMP1 (DBLγ of DC6) (Fig. 2 B and C and Table 2). The elevated VarA transcription in SM (Fig. 2A) could be of either EPCR or rosetting variants. However, a comparison of domain-specific primers suggested EPCR-binding head structure domains (DBLα1.1/2/4/7, CIDRα1.4, and CIDRα1) were increased in SM patients (Table 2). By comparison, the primers detected few differences in CD36-binding variants between groups. Transcripts for two CD36-binding domains were significantly increased in the SM group (CIDRα3.1–3, DBLα0.1) and for one in the OP group (DBLε2 of DC7), but other CD36-binding transcripts did not differ between patient groups (Table 2).

The MSM group presented a very similar transcription profile to the SM group and none of the transcripts presented a significant differential expression after correcting for multiple comparisons [false-discovery rate (FDR) ≤ 0.2] (Table S1). Furthermore, similar differences in var expression were found when both groups were combined in a hospitalized group and compared with OP (Table S2). Transcripts from DC8 (DBLα-CIDRα, CIDRα1.1, DBLβ12 and DBLβ3/5, DBLγ4/6) and EPCR binders (CIDRα1.4 and CIDRα1, DBLα2/α1.1/2/4/7) presented a higher expression in hospitalized patients. In addition, expression of transcripts amplified by primers DBLγ of DC6 (EPCR, CD36, rosetting) and DBLγ of DC5 (rosetting) was increased among hospitalized patients (Table S2). Taken together, these data show that a higher prevalence of predicted EPCR binding and rosetting parasites was detected by multiple primers targeting Group A, DC8, DC6, and DC5 PfEMP1 variants in severe and hospitalized patients (summarized in Table S3).

Table S3.

Summary of differences in transcript levels among patient groups

Differential expression of transcripts with P ≤ 0.05 and FDR ≤ 0.2 is represented in gray and considered significant.

^*Predicted binding phenotype of PfEMP1 head structure.

Var Transcript Abundance, Parasite Biomass, and Clinical Disease.

Plasma PfHRP2 concentration presents a close correlation with disease severity in both children and adults (31, 32). To better understand the interplay between parasite biomass and parasite adhesion types in adult SM, we performed correlation analysis. Across all patients, there was a positive correlation between PfHRP2 plasma concentration and transcript abundance of three var domains: DBLγ of DC5 (rosetting) (Spearman’s ρ = 0.3, P = 0.04), DBLγ of DC6 (EPCR, rosetting, and CD36) (Spearman’s ρ = 0.33, P = 0.02), and CIDRα3.1–3 (CD36) (ρ = 0.32, P = 0.03), suggesting that parasites expressing these variants might lead to higher parasite load. In addition, elevated transcription of domains DBLα-CIDRα of DC8 (ρ = 0.3, P = 0.04) and DBLγ of DC6 (ρ = 0.46, P = 0.001) was correlated with the number of severity criteria. Conversely, transcripts targeted by primer DBLε2 of DC7 (CD36) (ρ = −0.33, P = 0.03) and DBLε12 of DC12 (CD36) (ρ = −0.31, P = 0.03) were negatively associated with PfHRP2. Taking these data together, this analysis suggests that specific PfEMP1 domains may contribute to higher parasite biomass, whereas other PfEMP1 domains may act in concert with parasite biomass to increase disease severity when they pass a certain threshold.

To test this hypothesis, we built machine-learning models to investigate disease causation and performed statistical analysis. Multivariate, threshold-based logic maps nicely on to a decision tree structure, and random forests (RF) are powerful tools for ranking feature importance (39). This methodology generates a multitude of decision trees and measures the mean decrease in classifier accuracy (MDCA) when a particular feature (e.g., transcript level detected by a var primer) is removed from the model. Recent advances in computational techniques have made it possible to associate P values with features ranked in this manner (40, 41). First, we used a RF (39) with 1,000,000 trees to select parasite features that could be used to accurately predict patient hospitalization and disease severity. Although transcription of domains DBLα-CIDRα of DC8 and DBLγ of DC6 were the strongest factors to predict patient hospitalization (SM + MSM vs. OP), the level of PfHRP2 was the most important feature for patient severity (SM vs. OP) (Fig. 3A). Moreover, the analysis suggested that transcription of both DC8 and DC6 domains plays a role in patient severity as they ranked second and third in the prediction of SM [mProbes familywise error rates (FWER) ≤ 0.2]. In addition, all four primers that target the DC8 domains ranked in the top 10 of the severity model, highlighting the importance of DC8-containing PfEMP1 in disease severity.

Fig. 3.

Machine-learning approach to understand disease severity. Parasite factors associated with a higher risk of patient hospitalization and disease severity were revealed by machine-learning approaches. (A) Summary of RF feature selection strategy to identify parasite virulence factors that discriminate between hospitalized patients (SM + MSM) (Left), SM patients (Right), and OP. The top 10 parasite factors with the highest MDCA are shown. Positive correlation with disease severity is shown with a 1, negative with a −1, and no association with a 0. To adjust for false discovery, familywise error rates (RF mProbes FWER) were estimated using mProbes algorithm and values ≤ 0.2 were considered significant. The predicted binding phenotype was determined as described in Fig. S2. The CMI P value is used to find primers that are significantly informative even after PfHRP2 is accounted for. Var features with a P ≤ 0.05 presented virulence not explained by parasite biomass. (B) evTrees illustrate disease pathways to patient hospitalization (Left) or severe disease (Right) after PfHRP2 filtration (P ≤ 0.20). The percentages in the boxes represent the probability of each pathway to classify patients into hospitalized (H), severe malaria (S), or outpatients (OP). The number of patients in each pathway is indicated below. The percentages beneath the lines show the proportion of the total severe patients classified by each pathway. (C) Var primers were grouped according to binding phenotype or var group (Fig. S2 and Table S4) and ranked by MDCA (15, 16). The association with patient hospitalization (Left) and disease severity (Right) was determined using a Mann–Whitney U. P ≤ 0.05 and FDR ≤ 0.2 are considered significant. FDR, Benjamini–Hochberg adjustment for FDR.

PfEMP1 variants could contribute to disease severity by distinct mechanisms, including by promoting the rapid multiplication of parasites (biomass) or by encoding dangerous adhesion traits (impairing critical endothelial functions). Univariate correlation analysis showed an association between PfHRP2 plasma levels and transcription of domain DBLγ of DC6. To investigate the possibility that the correlation of the DC6 domain in disease severity was simply because of an association with parasite biomass, we applied conditional mutual information (CMI) algorithms. CMI is an information theory technique to test if one variable's predictive power is weakened by the presence of others. After PfHRP2 filtration (i.e., filtering out those primers that were uninformative after PfHRP2 levels were accounted for), transcription of domains DBLγ of DC6 and CIDRα1.1 of DC8 remained important to disease severity and patient hospitalization (CMI PfHRP2 filtration P ≤ 0.05), showing that the virulence of these DCs is not simply explained by a major contribution in parasite biomass (Fig. 3A).

To provide a visualization of the logic used by decision trees, we generated evolutionary decision trees to understand how parasite factors might interact to determine disease severity in adults (Fig. 3B). In the tree that tests patient hospitalization (MSM + SM vs. OP), a combination of high PfHRP2 levels and elevated transcription of domains CIDRα1.1 of DC8 and DBLγ of DC6 were sufficient to correctly classify 91% of patients as hospitalized or not. Furthermore, a combination of these three features was sufficient to classify 100% of SM patients (SM vs. OP). Taken together, machine-learning approaches determined that both parasite biomass and transcription of DC8- and DC6-containing PfEMP1 variants are critical for adult malaria severity.

The machine-learning analysis implicated both an N-terminal PfEMP1 domain (DC8 cassette: DBLα2-CIDRα1.1/1.8-DBLβ12-DBLγ4/6) and a C-terminal PfEMP1 domain (DC6 cassette: DBLγ14-DBLζ5-DBLε4) in severe disease. DC6 can be present in combination with all four types of PfEMP1 head structures: group A rosetting, group A EPCR-binding, DC8 EPCR-binding, or CD36 binding (Fig. S1) (13). To investigate a possible structural linkage between DC8 and DC6 in SM patients, we performed correlation analysis. As expected, the four DC8 domains were highly correlated (Spearman’s ρ = 0.51–0.71 for all pairwise comparisons). By comparison, the DBLγ of DC6 domain was equally correlated to other DC6 domains (ρ = 0.34) and DC8 domains and group A variants (ρ > 0.3) (Fig. S3). These results suggest that DC8 and DC6 may be linked in the same PfEMP1, but DC6 may also be present in non-DC8 PfEMP1. Thus, it is possible that different subsets of DC6-containing PfEMP1 variants may contribute to SM.

Fig. S3.

Pairwise comparison correlation of DC8 and DC6 subtypes expression among all patients. Spearman ρ values are depicted and ranked from highest (black) to lowest (light gray). All of the ρ values presented a P ≤ 0.05 except for the correlation DBLγ of DC6-DBLγ4/6 of DC8.

Many of the var primers match different parts of the same DC (e.g., DC8) or share similar functional annotations (e.g., EPCR binding). Consequently, var primers with the same annotation may artificially lower MDCA values in the RF analysis by acting as proxies for each other. To account for this effect, set-enrichments were performed to combine primers targeting the same var group or adhesion types (Fig. S2 and Table S4). For this analysis, a Mann–Whitney U test was used to compare the MDCA values of set-enriched primer groups versus primers not targeting that annotation (SI Materials and Methods). In both the patient hospitalization and disease-severity models, different combinations of DC8 primers (DC8 pure and DC8 all) presented enrichment P values lower than 0.005 that remained significant after FDR correction (FDR ≤ 0.2) (Fig. 3C). The association with patient hospitalization and disease severity was still significant but reduced in the group that contained all EPCR binders (DC8 EPCR + group A EPCR). Despite playing an important role in the univariate analysis, VarA transcriptional levels were not necessary to predict hospitalization or disease severity. An enrichment group that excluded DC8 and incorporated both types of UpsA adhesion traits (group A = EPCR and rosetting variants combined) failed to reach significance, reinforcing the importance of DC8. Rosetting, IgM, or intercellular adhesion molecule-1 (ICAM-1) binding domains groups, which have been previously associated with SM (reviewed in ref. 42) did not present any significance in disease predictability of adult SM. As expected, the CD36 binding group showed no significance for predicting either hospitalization or severity. Taken together, these data show that machine-learning approaches highlight the importance of DC8 in adult SM, and determine that in combination with the DBLγ of DC6 domain and elevated parasite biomass, is strongly associated with disease severity.

Table S4.

Primer grouping for RF enrichment analysis

No.	EPCR binders	CD36 binders	Rosetting	IgM binders	ICAM-1 binders	Group A	Groups B and C	Group B	Group C	DC8 pure	DC8 all
1	DBLα2/α1.1/2/4/7	DBLβ5	DBLγ of DC5	DBLγ of DC6	DBLβ3	DBLα1.1 of DC1	DBLγ of DC6	UpsB1	UpsC1	DBLα-CIDRα of DC8	DBLα-CIDRα of DC8
2	DBLα1.4	DBLζ4 of DC9	DBLβ7 & 9 of DC5	DBLζ5 of DC6	DBLβ5	DBLε8 of DC3	DBLζ5 of DC6	UpsB2	UpsC2	CIDRα1.1 of DC8	CIDRα1.1 of DC8
3	CIDRα1.6	CIDRγ	DBLα1.5/6b of DC16	DBLε2 of DC7		DBLβ7 & 9 of DC5	DBLε2 of DC7			DBLγ4/6 of DC8	DBLγ4/6 of DC8
4	CIDRα1.7	DBLε2 of DC7	DBLα1.5/6a of DC16	DBLε3 of DC7		DBLγ of DC5	DBLε3 of DC7				DBLβ12 and DBLβ3/5
5	DBLβ3	DBLε3 of DC7	CIDRδ of DC16	DBLε12 of DC12		DBLα1.7 of DC13	DBLγ of DC9
6	DBLα1.7 of DC13	DBLγ of DC9				CIDRα1.4 of DC13	DBLζ4 of DC9
7	CIDRα1.4 of DC13	DBLζ6 of DC10				CIDRα1.4 and CIDRα1	DBLζ6 of DC10
8	CIDRα1.4 and CIDRα1	DBLα0.16 of DC19				DBLα1.5/6b of DC16	DBLε12 of DC12
9	DBLα-CIDRα of DC8	CIDRα3.4 of DC19				DBLα1.5/6a of DC16	DBLα0.16 of DC19
10	CIDRα1.1 of DC8	DBLα0.9 of DC20				CIDRδ of DC16	CIDRα3.4 of DC19
11	DBLγ4/6 of DC8	DBLα0.1				DBLα2/α1.1/2/4/7	DBLα0.9 of DC20
12	DBLβ12 and DBLβ3/5	DBLα0.6/9				DBLα1.4	DBLα0.1
13		CIDRα2.2				CIDRα1.6	DBLα0.6/9
14		CIDRα2.3/5/6/7/9/10				CIDRα1.7	DBLβ5
15		CIDRα3.1–3				DBLβ3	DBLγ9
16		CIDRγ2/9				VarA	CIDRα2.2
17							CIDRα2.3/5/6/7/9/10
18							CIDRα3.1–3
19							CIDRγ2/9
20							CIDRγ1/2
21							CIDRγ
22							DBLβ12 and DBLβ3/5

DC8 CIDRα1 Domains Expressed by SM Isolates Inhibit the APC–EPCR Interaction.

DC8 CIDRα1 differ in sequence, binding affinity, and ability to inhibit the APC–EPCR interaction (25, 27, 28), but there have been no in-depth functional characterizations of DC8 CIDRα1 from SM isolates. To perform a deeper phenotypic characterization of DC8 variants in severe isolates, we amplified and sequenced the full-length DC8 CIDRα1 domain (Fig. S4) from five patients who met WHO SM criteria: patient 24 had anemia, patient 25 had jaundice, and patients 62, 87, and 95 presented multiorgan complications, including cerebral malaria in patients 62 and 87. In total, seven CIDRα transcripts were amplified from the five patients. All of the sequences clustered together with domains from subgroups CIDRα1.1 and CIDRα1.8, which are diagnostic for DC8-like (Fig. 4A). All five patients presented a single DC8 sequence except for patient 62, who presented three distinct DC8 CIDRα1 domains. Notably, an identical DC8 CIDRα1 transcript was amplified from three patients with multiorgan complications, including both cerebral malaria patients (62 and 87). By comparison, the five distinct CIDRα1 sequences had 32% sequence identity.

Fig. S4.

Identification and expression of DC8 CIDRα1 expressed by SM isolates. (A) PCR strategy to amplify the full-length DC8 CIDRα1 domain from SM patients using var domain primers (15). (B) CIDRα1 recombinant proteins were analyzed under nonreducing conditions in a SDS/PAGE gel and stained by GelCode Blue Protein Stain.

Fig. 4.

Inhibition of the APC–EPCR interaction by DC8 CIDRα from severe isolates. DC8 CIDRα domains expressed from SM isolates were analyzed for EPCR binding affinity and blocking the interaction with its ligand APC. (A) Neighbor-joining tree (bootstrap n = 100) of 66 previously classified CIDRα1 sequences (14) and 7 Indian CIDRα1 transcripts amplified from adult SM patients in this study (black dots). (B) The first column shows the dissociation constant (K_d) for rCIDRα1.1-EPCR measured by biolayer interferometry (see Fig. S5 for detailed kinetics). Histograms show APC binding to CHO-EPCR cells in the presence or absence of 250 μg/mL Indian CIDRα1 domains. The vertical line shows the primary and secondary antibody background used to set the gate for APC⁺ cells. Red: strong inhibition; blue: medium; green: low; light gray: no inhibition. The bar graphs show the percentage of APC binding in the presence of CIDR domains relative to APC alone (mean and SD, n = 4 independent experiments). (C) Inhibition by rCIDRα1.1 of APC-dependent protection of endothelial barrier properties. (C, Left) Kinetics showing APC (50 nM)-mediated protection of thrombin (2 nM) induced barrier disruption in human brain endothelial cell monolayers, and examples of rCIDR1.1 that do or do not inhibit APC barrier protection activity. (Right) Bar graph showing the barrier protection (%) activity of APC on human brain endothelial cells and HUVEC cells (EA.hy926 cells) pretreated with rCIDRα1.1 (mean and SD, n = 6 independent experiments for all CIDR domains, except n = 3 for rPFE1640wCIDRα1.3). P values were calculated using a one-way ANOVA and Dunnet’s multiple comparison test. Significant values are represented by *P ≤ 0.05, **P ≤ 0.01, and ***P ≤ 0.001.

To investigate whether SM isolates interfere with the APC–EPCR interaction, the five unique Indian CIDRα1 domains were expressed as recombinant proteins (Fig. S4 and Table S5). Four of five bound EPCR with low to moderate nanomolar binding constants (K_ds = 3.81 nM to 63 nM) (Fig. 4B and Fig. S5), similar in range to other CIDRα1.1/8 domains (27, 29). Although r62-2-22 CIDRα1.1 did not bind recombinant EPCR, this domain exhibited dose-dependent binding to Chinese hamster ovary (CHO) cells expressing EPCR in a flow-cytometry binding assay (Fig. S5), suggesting it also possesses low EPCR binding activity. As expected, in binding competition studies, there was limited or no inhibition of APC binding to CHO745-EPCR cells coincubated with either the negative-control PFE1640w CIDRα1.3 domain or the weak binder domain r62-2-22 (Fig. 4B). By comparison, two CIDR domains from Indian isolates (r25-2-4 and r62-2-1) presented low inhibition of APC binding and the two remaining CIDR domains from Indian isolates (r62-2-23 and r24-2-4) presented moderate APC inhibition (Fig. 4B). However, all of the DC8 domains expressed by SM isolates competed less effectively for the APC–EPCR interaction than a positive control, group A variant (IT4var07 CIDRα1.4 domain), which strongly blocks the APC–EPCR interaction (25, 26, 28).

Table S5.

Sequence of recombinant CIDRα1 domains used in this study

CIDR domain	Sequence
IT4var07 CIDRα1.4	PDCGVICENGKCVVKENGSNCRHYNIYEPAPDVKTTEINVIVSGDEQGIITKKLQDFCMNPNNENGTNNQIWKCYYKDEKENKCKVETKSGNSTYKEKITSFDEFFDFWVRKLLIDTIKWETELTYCINNTTNADCNNECNKNCVCFDKWVKQKEKEWKNIMDLFTNKHDIPKKYYLNINDLFNSFFFQVIYKFNEGEAKWNKLKENLKKKTESSKKNKGTKDSEAAIKVLFDHLKETATICKDNNTNEAC
24-2-4 CIDRα1.8	PICGVKCENKSCTEKENDDDCKNKKKYDPPKGVTPIDIPILYSGDKQGDITKKLEDFCYNPTKENEKTYQNWKCYYKDSEFNKCKMESKSGKSTTQEKIISFDEFFYLWVNNLLIDSIMWENDIKHCINNTNVTNCNNGCNENCICFEKWVGQKEKEWENVKKVLKNPSKNLNNYYNKLNGIFSGFFFEVVYKFNNKEEKWNQFTEDLKKKIEASQKNKGTENSQDAIELLLDHLKDNAITCKDNNSLKEDKNC
25-2-4 CIDRα1.1	PNCVVKCDGGTCEQDENDENCRSKIIQKILEKEKDTDIDVLNSDDKQGDITKKLKDFCSSTTKVDVKNVQKWKCYNKNNDYNNCEMNISSYKDATDPNVMLSIKCFDSWAKNLLIDTIKWEHQLKNCINNTNVTYCESKCNKNCECYEKWINRKKDEWEKLKEVLKKKDENSDNYYNKLSSVFDSFLFQVKGALEEDEKVKWDQFTEDLEKKFEPSEKKTRTTDSQDAIEFLLDHLKDNATTCKDNNSNESCDVS
62-2-1 CIDRα1.8	PDCVVDCNGAKCQQKMKRDGTCEKPQIYTRPKDVTPKKIKVLFSGENQEDITEKLSSFCSNPKSKIDRNYQTLQCYYKKPDHNNCEMKGSSYKDKHDPNIIISDECFHLWVKNLLIDTIKWETKLKKCINNTNVTNCNSGCNENCECFENWVEQKKKEWENVNDVYKDQKEILGIYYKKLENLFDSYFFEVMGALENEEKHGKWNQLTAKLKQIIKPSKENRDSGNSQDTIKLLLDHLKETATTCIDNNSLESDENC
62-2-22 CIDRα1.1	PDCLVVCDNRGCKENENGDNCRSKIIEEILKSEQHTEIDVLYSDDKQGVITEKLDDFCKYPNNDKGKNYKKWKCYNKNSDYDKCEMISWLYEDPNESNLMLSIKCFDSWAQNLLIDIIRWEHQLKDCINNTNVTDCESKCIKNCQCYEAWIKQKQKEWQQVKEVLKKKDENSHNYYNKLSSVFDSFLYQVMNALKKEDKDGKWDQFTKDLKKKFEPSKDKASTANSQDAIELLLDHLKDNATTCKDNNSNEACDVS
62-2-23 CIDRα1.8	PDCIVKCNGGKCTENTIEENCKSKRTYSLPPGVNSTEIEVLFSGDNQKDITEKLNSFCKNTNNENGENVEKWECYYQNEYNNKCQMTSPKLEDKKRPTVMIFDEFFYLWVNNLLIDSIMWENDIKHCINNTNVTNCNNGCNENCICFEKWVGQKEKEWENVKKVLKNPSKNLNNYYNKLNGIFSGFFFEVVYKFNNKEEKWNQLTAKLKQIIEPSEKNTRTTDSQDAIKLLLDHLKDNAITCKDNNSLEPC
PFE1640w CIDRα1.3	PHCEVDCENGNCEVKNKPDGNCGKNVKYKPPYGVKPTEITVLYSGNEKGDISKKLSEFCSNKNNINVKNNETWKCYYKNSDNNKCKMESNSENNKGAEKITSFHEFFELWVKNLLKDTMKWENEIKDCINNTNITDCNDECNKNCVCFDKWVKQKEEEWKNVKKVFENKKYIQDKYYLDINKLFESFLFKVISELDQGEAKWNQLKEELKKKIESSKANEGIKDSESAIELLLDHLKESATTCKDNNANEAC

Fig. S5.

Binding kinetics between recombinant CIDRα1 domains and EPCR determined by BLI and flow cytometry. (A) Representative sensograms with corresponding kinetic fits to the data (red). (B) Summary of binding kinetics. *Binding not detected. (C) Dose-dependent binding titration of CIDRα1 domains to CHO745-EPCR cells analyzed by flow cytometry, median levels depicted (n = 4). Binding levels to the negative control, untransfected CHO745 cells is shown only for the highest CIDRα1 concentration analyzed.

To study if SM isolates interfere with the endothelial barrier protective activity of APC, we investigated the ability of CIDR domains to inhibit APC protective activity in thrombin-induced endothelial barrier permeability assays. APC diminished thrombin-induced barrier disruption by 24% in EA.hy926 human umbilical vein endothelial cells and 35% in cultured primary human brain endothelial cells (Fig. 4C). In agreement with the binding competition studies, r62-2-22 and r25-2-4 did not inhibit APC barrier protective function, r62-2-1 and r62-2-23 partially inhibited (∼30–45% reduction relative to APC), and IT4var07 and r24-2-4 strongly inhibited the APC protective pathway in both EA.hy926 cells (40–48% reduction) and primary brain endothelial cells (60–75% reduction). In general, CIDR domains tended to have slightly higher inhibition on brain endothelial cells than EA.hy926 cells (Fig. 4C), although the difference did not reach statistical significance. Taking these data together, this analysis indicates that DC8 CIDRα1 domains expressed from SM isolates bind EPCR and may inhibit the APC–EPCR pathway.

Discussion

Studies to understand the role of PfEMP1 in SM pathogenesis have been mainly focused on children in Africa (15, 18, 19) and information on adult severe patients still remains scarce (43). However, in areas of unstable transmission, SM occurs across age groups. The different symptomatology in adult SM, the presence of multiorgan complications, and the higher fatality rate urge research on this population. Here, we investigated the relationship between parasite biomass and PfEMP1 in adult SM.

Although adult malaria patients presented a complex mixture of parasites in peripheral blood, parasites expressing group A, DC8-, and DC6-containing var transcripts were elevated in SM patients. Prior studies suggest group A and DC8-containing var genes are preferentially expressed in young African children with limited immunity (15, 17, 44) and nonimmune European travelers (45), suggesting these variants confer a parasite growth advantage in malaria naïve hosts, and in some circumstances increase the risk for SM (14, 15, 17, 23). Machine-learning prediction models suggest that high transcription of DC8 and DC6 domains in combination with high parasite biomass is associated with adult patient hospitalization and severity. Conversely, the importance of group A dropped in the RF analysis, possibly because group A transcripts were highly expressed in all patient groups, as might be expected for a lower transmission setting. It also remains possible that group A may be more closely linked to specific adult disease syndromes. Previously, logistic regression analysis was used to assign different binding variants to respiratory distress and impaired consciousness in pediatric malaria (46), but to our knowledge, this is the first time that more advanced machine-learning approaches have been used to understand parasite factors associated with disease progression and malaria severity. We expect that powerful machine-learning algorithms, such as those presented here, would be useful for analyzing var expression data from African children (15, 18, 19) and may shed light into pathogenesis mechanisms that drive pediatric cerebral malaria, respiratory distress, or anemia.

Endothelial dysfunction is thought to play an important role in SM pathology. Indeed, microthrombi are a common finding in pediatric cerebral malaria autopsies (47, 48), and cerebral swelling is a major risk factor for death (29). Although fibrin deposits and cerebral swelling are more variable in adult cerebral malaria autopsies (49, 50), alterations in blood–brain barrier integrity have been associated with infected erythrocyte sequestration in adult Southeast Asian and pediatric African autopsies (51, 52). It has been hypothesized that EPCR binding parasites may drive disease pathogenesis by blocking the anticoagulation, anti-inflammatory, and barrier protective functions of the APC–EPCR pathway (24). However, recent studies of CIDRα1 domains from long-term cultured adapted parasites have revealed large differences in their ability to inhibit the APC–EPCR interaction (25–28). The consequences of these differences for disease severity remain unknown. Here, we showed that DC8 CIDRα1 domains expressed by SM isolates possess differential activity to disrupt the EPCR–APC protective pathway. Some DC8 domains had low activity and others had nearly equivalent activity to a group A CIDRα1 domain that strongly blocks the APC–EPCR interaction (25, 26, 28). Notably, an identical DC8 CIDRα1 domain (r62-2-1) isolated from two patients with cerebral malaria and a third patient with multiorgan complications produced a small and significant inhibition of APC barrier protective activity on brain endothelial cells. This finding raises the possibility that even parasite domains with weaker APC inhibition activity may interfere with this important host regulatory pathway, especially in microvascular beds where there is loss of EPCR expression associated with parasite sequestration (30). Unexpectedly, a DC8 CIDRα1 domain (r24-2-4) isolated from an anemic patient had the highest APC blockade activity. However, the low parasite biomass in this patient (plasma PfHRP2 = 15.86 ng/mL) might explain the lack of a life-threatening symptomatology. Therefore, the relation between CIDRα1 phenotypes and disease symptoms reinforces the notion that a certain threshold of parasite biomass in combination with virulent PfEMP1 variants is associated with overlapping severe symptomatology in adults.

The clear association of DC8 with both children (15) and adult SM spotlights the CIDRα–EPCR interaction. However, DC8 variants encode multiple endothelial binding domains (53), and it is possible that other undefined coadhesion traits may also increase the risk for SM. Furthermore, our study, to our knowledge for the first time, implicates specific C-terminal PfEMP1 domains in SM. DC6 is characterized by DBLγ14-DBLζ5-DBLε4 domains and can be present in combination with rosetting, EPCR, or CD36-binding head structures (13). Thus, it will be important to explore if both rosetting and nonrosetting PfEMP1 variants are contributing to adult SM. Although DC6 binding properties are uncharacterized, recent studies have shown that DBLζ and DBLε mediate binding to IgM and α2-macroglobulin, and it has been hypothesized that binding to these serum factors can cross-link multiple PfEMP1 to increase the binding affinity of N-terminal domains (54, 55). Further sequencing of DC6 variants will be necessary to assess whether DC6 and DC8 are part of the same or different proteins in severe isolates, and to determine the binding properties of DBLζ and DBLε Indian domains to serum factors.

A limitation of our study is the relatively small population studied after 3 y of patient recruitment. Future machine-learning approaches with larger sample groups and examining different geographic settings will be important to understand whether differences in var expressions are responsible for distinct adult severe symptoms. A second limitation is that the primers used for var profiling and sequencing (15) might fail to recognize some var domains important for disease severity or might be biased toward certain variants. For example, the DC6 cassette is found in both rosetting and nonrosetting PfEMP1 variants and rosetting variants may have been underestimated by the PCR typing approach. It will be valuable to use independent methodologies to investigate the contribution of rosetting parasite variants in our adult India population. Other parasite adhesins expressed at the surface of infected erythrocytes, such as RIFINs and STEVOR, play a role in rosetting (56, 57) but were not studied here. In the future, it would be interesting to study the interplay between var and other parasite adhesins in disease severity. Nevertheless, the degenerate primers against var adhesion domains remain a sensitive and cost-effective tool to cover the var geographical diversity and, combined with machine-learning approaches, provide a powerful methodology to investigate pathogenic mechanisms.

In summary, our data show that elevated DC8 and DC6 var transcripts, along with high parasite biomass, promote disease progression in adult SM. In addition, our findings raise the possibility that DC8 CIDRα1 domains with low or moderate APC blockade activity interfere with APC–EPCR protective pathways, highlighting attention on this pathway for disease interventions and the future development of SM adjunctive therapies.

SI Materials and Methods

Machine-Learning Models.

Correlations between var transcripts and disease severity were analyzed using Spearman’s rank correlation coefficient with a permutation test for P value. Features with P < 0.2 were assigned to be 1 if correlated or −1 if anticorrelated, and all others were assigned a value of 0. Thus, a var primer scored 1 if its presence is positively correlated with severe disease. However, we expect that univariate methods (especially those backed by linear models) are insufficiently powerful for determining which var transcript levels are associated with disease phenotypes because we expect the relationships to be neither univariate nor linear. Thus, we expect factors to act in concert and to become relevant when they passed a certain threshold. To test this hypothesis, we built machine learning models to investigate disease causation by RF (consisting of many decisions trees) for ranking feature importance (39). Therefore, to investigate the role of var transcripts and parasite biomass in disease, features (i.e., log-transformed Tu var transcript levels detected by the 45 domain primers) were ranked using the MDCA as calculated by the [R] RF package with 1,000,000 decision trees (39). To control for multiple comparisons in the RF analysis, FWER were estimated for features using the [R] rFerns package (61) with 1,000,000 ferns and a modified version of the mProbes algorithm (40), which allows features to be repeatedly shuffled to ensure a minimum number of trials in excess of the number of features (i.e., 1,000,000 trials instead of the 45 domain primers).

Set Enrichment.

Because some var primers share the same information as others (e.g., four DC8 primers), this may lower MDCA values in the RF analysis because other primers can act as noisy proxies. Therefore, var primers with shared annotations were grouped into set-enrichments. To perform the set-enrichment grouping, we used an in silico primer analysis to determine the set of var genes targeted by each primer and to validate the lack of cross-reactivity on nonspecific var adhesion domains. The in silico PCR prediction required annealing of both primers with no more than one mismatch that was never present at the 3′ end of the primer. Primer-grouping set-enrichments were made for: (i) adhesion properties, (ii) var groups, and (iii) DC structures. For the three categories of PfEMP1 head structures (EPCR, CD36, rosetting), primers were assigned based on the predicted head structure binding phenotype of the gene targeted by the primer in the in silico PCR (CIDRα1 = EPCR, CIDRα2–6 = CD36, CIDRβ/γ/δ = rosetting). For example, DBLε2 of DC7 is always linked to CD36-binding head structures. In the case of promiscuous C-terminal domains, we determined a cut-off of 70% for being included in a certain binding group (Fig. S2). This means that at least 70% of the var genes detected by that primer have the predicted head structure binding property. DBLβ3 and DBLβ5 were included in the ICAM-1 binders group (62–64) and primers that targeted domains with DBLε/ζ subtypes in the IgM-binders group (54, 55). For var group set-enrichments, assignments were based on var domain primers (15) that specifically targeted the group A, a mixed group B and C, or primers that targeted the 5′ upstream region of the group B or the group C (16). The analysis also included two set-enrichments for the DC8 group. “DC8 pure” included three of the four DC8 primer sets with higher specificity for DC8 (Fig. S2). “DC8 all” included a fourth DC8 primer that cross-reacts on non-DC8 var genes (DBLβ12 and DBLβ3/5) (Fig. S2). A given var primer could be present in more than one set-enrichment. For example, DBLβ3 binds ICAM-1 and is present in group A var genes, so it was included in both the ICAM-1 binders and the group A set-enrichments. The complete list of primers for each set-enrichment is listed in Table S2.

Conditional Mutual Information.

Enrichment P values were calculated using a one-tailed Mann–Whitney U test and a Benjamini–Hochberg adjustment for FDR. This test evaluates the null hypothesis that an annotation was not enriched in primers with higher MDCAs than primers outside the enrichment. Thus, an annotation (e.g., DC8) would have a lower P value if primers with DC8 annotations tended to have higher MDCAs than other var primers. As expected, parasite biomass (as measured by PfHRP2 levels) was a powerful predictor of disease severity; however, to test if any of the var primers presented inherent virulence beyond parasite biomass, we used CMI analysis (60). CMI is an information theory framework for measuring how much information variable X adds about class Y given that variable Z is known. In this case: How much information does each of the var primers add about disease severity given that HRP2 levels are known? CMI was calculated using the [R] infotheo package (with empirical probability distribution) and the unit “nats” as a measure of information. Primers were associated with P values by testing the null hypothesis that the information added by each primer was the same as information added by randomly generated expression values. These P values were calculated using a permutation test where the score (in nats) for each primer was compared with a distribution of scores generated from 10,000 random primers containing sampled expression values from all primers appearing in the dataset. To provide a visualization of the logic used by decision trees, evolutionary Trees (59) were calculated on data including PfHRP2 levels and primers that added significant (P ≤ 0.2) information to PfHRP2 levels using 10,000 iterations of 100 trees to produce a final tree.

Materials and Methods

Ethical Approval.

Informed consent was obtained from all study participants. The study was approved by the ethics boards at Goa Medical College and Hospital, the University of Washington, the Western Institutional Review Board used on behalf of the Center for Infectious Disease Research, as well as by the Government of India Health Ministry Screening Committee.

Patient Recruitment and Samples Collection.

Subjects were recruited between April 2012 to October 2014 from the hospital admission or outpatient wards from patients presenting at Goa Medical College and Hospital (Goa, India). Subjects were enrolled by project staff who explained the project. Following informed consent, blood samples from P. falciparum-positive patients were collected in acid citrate dextrose vacutainers and separated into plasma and red blood cells in RNALater. Samples were stored at –80 °C. Infections were confirmed by study staff using Giemsa-stained thin and thick smears for parasitemia determination and Plasmodium species identification. Rapid diagnostic test (Zephyr Biomedicals) was additionally used for the diagnosis of parasite species. Parasites per milliliter was calculated from thin film smears (count/1,000 RBC/125.6/Hematocrit). SM was defined as: (i) coma (Glasgow Coma Score < 10), (ii) severe anemia (Hb < 7 g/dL), (iii) jaundice (bilirubin > 3 mg/dL), (iv) renal compromise [serum creatinine > 3 mg/dL or (blood urea nitrogen > 17 mmol/L), (v) shock (systolic blood pressure < 80 mmHg with cold extremities), (vi) metabolic acidosis (peripheral venous bicarbonate < 15 mmol/L), (vii) respiratory distress (respiratory rate > 20 breaths per minute or PaO₂ < 75 mmHg), and (viii) hypoglycemia (blood glucose < 40 mg/dL). Patients admitted to the hospital without any of these criteria were considered as MSM and nonadmitted patients were considered as OP. Patients with mono P. falciparum infections were treated with oral artesunate and mefloquine, and intravenous artesunate was used for SM patients. One study patient was coinfected with P. falciparum and Plasmodium vivax and treated with oral artesunate, mefloquine, and primaquine. All patient care was managed according to hospital standard procedures.

PfHRP2 Plasma Quantification.

PfHRP2 was quantified using double-site sandwich ELISA according to published methodologies (31, 36). In brief, plates were coated overnight with mouse anti-PfHRP2 IgM antibody (MPFM-55A, ICL) at 1 μg/mL in PBS and blocked for 4 h with 2% (wt/vol) BSA-PBS. Patient plasma samples were diluted to the desired detectable dilutions (1:10–1:200) and tested in triplicate for 1 h. For detection, mouse anti-PfHRP2 IgG antibody (MPFG-55P, ICL) was added at 0.2 μg/mL in 2% (wt/vol) BSA-1% Tween 20-PBS for 1 h, incubated for 5 min with TMB reaction substrate, and measured spectrophotometrically at 450 nm. Positive and negative controls were included in each plate. A standard curve was established using purified recombinant PfHRP2 protein (kindly donated by David Sullivan, Johns Hopkins Bloomberg School of Public Health, Baltimore) diluted 0.25–55 ng/mL in PBS. Five patient samples that presented a lower concentration than the detection limit were excluded from the analysis.

Determination of var Transcription by qRT-PCR.

Thawed red blood cells in RNAlater were dissolved in 12 volumes of TRIzol and RNA was extracted using an RNeasy micro kit (Qiagen). cDNA was synthesized using random hexamers and a MultiScribe reverse transcriptase (Thermo Fisher). qRT-PCR was performed using QuantiTect SYBR in a Mastercycler Realplex² following published amplification conditions (15, 16). Data were acquired after the elongation step of each cycle. Absence of DNA in RNA samples was confirmed by running a reverse-transcriptase negative sample with the housekeeping gene primer adenylosuccinate lyase (ASL) (PFB0295w). Levels of var gene expression were determined by relative quantification of the average expression of ASL and seryl-tRNA synthetase housekeeping genes (ΔCt var_primer = Ct var_primer − Ct average_housekeeping primers). The level of var expression was represented as Transcript units (Tu) and calculated as Tu = 2^(5−ΔCt) (15). Samples were included in the univariate statistical analysis only if the Ct average of the housekeeping genes was below 30.

CIDRα1 Sequencing and Recombinant Protein Expression.

DC8 transcripts were amplified from patients using the strategy depicted in Fig. S4. Bands of the expected size were excised, purified, and cloned in a Zero Blunt TOPO vector. At least 15 different colonies were sequenced per patient and analyzed using Geneious (7.1.7). Sequences were deposited in GenBank with accession numbers KU843600–KU843604. Recombinant CIDRα1 were synthesized as GBlocks gene fragments (IDT) and produced as His6-MBP-TEV-PfEMP1 insert-StrepII–tagged proteins, as previously described (58). Recombinant proteins were purified in a two-step process using an amino-terminal His tag and a carboxyl-terminal StrepII tag and analyzed by SDS/PAGE according to standard procedures (28).

Biolayer Interferometry Analysis.

Binding of the CIDR domains to biotinylated EPCR was determined on the Octet QKe instrument (ForteBio), as described previously (28). Mean K_on and K_off and apparent K_d values were determined from double-reference subtracted data from three concentrations that were fitted globally to a 1:1 Langmuir binding model using the data analysis software.

rCIDRα Binding Titration to CHO745-EPCR and APC Competition Assay.

For the rCIDRα binding titration to CHO745-EPCR and APC competition assay, 10⁵ CHO745-EPCR cells were lifted and washed in complete HBSS (HBSS with 3 mM CaCl₂, 0.6 mM MgCl₂, 1% BSA) (28). A five-point titration curve was determined with 1–250 μg/mL of recombinant CIDR for 30 min. CIDR binding was detected with a rabbit polyclonal anti-StrepII tag antibody followed by a goat anti-rabbit Alexa488-coupled antibody. Control samples were labeled with primary and secondary antibodies alone to set gates. APC competition assays were done as described previously (28). Briefly, 10⁵ CHO745-EPCR cells were coincubated with 50 μg/mL APC (Sigma) and 50 or 250 μg/mL CIDR recombinant proteins for 30 min on ice. APC binding was detected with a goat anti-APC mAb (1:100; Affinity Biologicals) followed by a chicken anti-goat Alexa488 coupled antibody. The percentage (%) APC binding was calculated relative to the value of cells incubated with APC alone. Labeled cells were analyzed by flow cytometry using LSRII (Becton Dickinson) and data were analyzed by FlowJO 10 software (Tree Star).

Measurement of the Monolayer Permeability.

Barrier function was monitored using a real-time cell analyzer (xCELLigence System, ACEA Biosciences). This system measures electrical impedance across the cell monolayer, cell impedance (CI), via gold microelectrodes integrated on the bottom of a 96-well plate. Next, 10,000 EA.hy926 HUVEC (ATCC) or primary human brain endothelial cells (Cells Systems) were seeded in each well. Cell proliferation was assessed for 72 h, at which time the cells reached a sustained maximum CI value. For the experiment, cells were incubated with rCIDRα1 (0.05 mg/mL) or culture medium. After 30 min, 100 nM of human APC (Haematologic Technologies) was added and incubated for 2 h. Barrier disruption was induced with 5 nM thrombin (Sigma) and compared with untreated cells (baseline). Cells were treated in triplicate and CI was measured every minute up to 120 min, then every 5 min up to 245 min after thrombin challenge. The baseline-normalized cell index was calculated by comparing the CI values of treated cells to the CI values for baseline control wells of untreated cells at the time point immediately before the thrombin challenge. The level of barrier protection achieved by APC + thrombin treatment relative to thrombin treatment alone was set to 100% to calculate the binding inhibitory activity of CIDR domains.

Machine-Learning Models.

Decision trees for understanding parasite factors associated with patient hospitalization and disease severity were analyzed by RF (39) and evolutionary Trees (59) or by CMI analysis (60). Detailed methods can be found in SI Materials and Methods.

Statistical Analysis.

Univariate analyses were performed using GraphPad Prism v5.02 for Windows. Correlations between variables were tested using the Spearman’s rank correlation coefficient. Differences between groups were evaluated using the Mann–Whitney U test with a Benjamini–Hochberg adjustment for FDR or the one-way ANOVA with a Dunnett's multiple comparison test.