Severe adult malaria is associated with specificPfEMP1 adhesion types and high parasite biomassMaria Bernabeua, Samuel A. Danzigera, Marion Avrila, Marina Vazb, Prasad H. Babarc,d, Andrew J. Braziera,Thurston Herricksc,d, Jennifer N. Makic,d, Ligia Pereirac,d, Anjali Mascarenhasc,d, Edwin Gomesb, Laura Cheryc,d,John D. Aitchisona, Pradipsinh K. Rathodc,d, and Joseph D. Smitha,1

aCenter for Infectious Disease Research, Seattle, WA 98109; bDepartment of Medicine, Goa Medical College & Hospital, Bambolim, Goa 403202, India;cDepartment of Chemistry, University of Washington, Seattle, WA 98195; and dDepartment of Global Health, University of Washington, Seattle, WA 98195

Edited by Mats Wahlgren, Karolinska Institutet, Stockholm, Sweden, and accepted by the Editorial Board April 7, 2016 (received for review December10, 2015)

The interplay between cellular and molecular determinants thatlead to severe malaria in adults is unexplored. Here, we analyzedparasite virulence factors in an infected adult population in Indiaand investigated whether severe malaria isolates impair endothe-lial protein C receptor (EPCR), a protein involved in coagulationand endothelial barrier permeability. Severe malaria isolates overex-pressed specific members of the Plasmodium falciparum var gene/PfEMP1 (P. falciparum erythrocyte membrane protein 1) family thatbind EPCR, including DC8 var genes that have previously been linkedto severe pediatric malaria. Machine learning analysis revealed thatDC6- and DC8-encoding var transcripts in combination with high par-asite biomass were the strongest indicators of patient hospitalizationand disease severity. We found that DC8 CIDRα1 domains fromsevere malaria isolates had substantial differences in EPCR bindingaffinity and blockade activity for its ligand activated protein C.Additionally, even a low level of inhibition exhibited by domainsfrom two cerebral malaria isolates was sufficient to interfere withactivated protein C-barrier protective activities in human brainendothelial cells. Our findings demonstrate an interplay betweenparasite biomass and specific PfEMP1 adhesion types in the devel-opment of adult severe malaria, and indicate that low impairmentof EPCR function may contribute to parasite virulence.

malaria | Plasmodium falciparum | var | PfEMP1 | EPCR

Severe malaria caused by Plasmodium falciparum is responsi-ble for at least 400,000 deaths every year (1), mainly affectingchildren younger than 5 y old. However, in areas of low and un-stable 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 arethe major severe syndromes in children, multisystem disease is morecommon in adults, including renal impairment, jaundice, respiratorydistress, 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 seques-

tration of infected erythrocytes to microvascular endothelial cells(5). Extensive tissue-specific sequestration results in organ pathol-ogy, such as cerebral malaria and placental malaria, and contributesto metabolic acidosis and endothelial dysfunction (6, 7). Proteinsof the P. falciparum erythrocyte membrane protein 1 (PfEMP1)family, encoded by the var genes, are responsible for infectedred blood cell binding to the microvasculature (8–10). PfEMP1sare classified into three main groups—A, B, and C—based on up-stream sequence (UpsA, UpsB, UpsC) and chromosome location(11). The extracellular domain of PfEMP1s presents a modu-lar 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 arrange-ments known as domain cassettes (DC) (13). Expression ofgroup 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 infectederythrocyte adhesion categories, including “rosetting” and endo-thelial protein C receptor (EPCR) binding. Rosetting involves ad-hesion to uninfected red blood cells (20, 21), possibly leading togreater microvascular obstruction (22). EPCR binding involves in-fected erythrocyte adhesion to vascular endothelial cells (23). Theimportant role of the EPCR-activated protein C (APC) pathway inregulating coagulation, inflammation, and endothelial barrier prop-erties (24) has led to the hypothesis that EPCR-binding parasitesmay drive pathogenic mechanisms by inhibiting the APC–EPCRinteraction (23, 25–28), thus increasing vascular dysfunction andpermeability. Indeed, cerebral swelling is a major risk factor forpediatric death (29) and there is loss of EPCR and fibrin depositionsat sites of cerebral sequestration in pediatric autopsies (30). How-ever, the extent to which severe malaria isolates disrupt EPCRfunction is poorly understood. A better understanding of adhesion-based pathogenic mechanisms may inform novel targeted adjunctivedrug 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

Significance

The clinical presentation of severe malaria differs betweenchildren and adults, but the factors leading to these differencesremain poorly understood. Here, we investigated parasite vir-ulence factors in adult patients in India and show that specificendothelial protein C receptor (EPCR)-binding parasites areassociated with severe adult malaria and act together withparasite biomass in patient hospitalization and disease sever-ity. We found substantial differences in EPCR binding activityfrom severe malaria isolates. However, even parasite domainsthat partially obstructed the interaction between EPCR and itsligand activated protein C were sufficient to interfere withactivated protein C-barrier protective activities in human brainendothelial cells. Thus, restoration of EPCR functions may be akey target for adjunctive malaria drug treatments.

Author contributions: M.B., S.A.D., M.A., and J.D.S. designed research; M.B., M.A., P.H.B.,A.J.B., and T.H. performed research; M.V., J.N.M., L.P., A.M., E.G., L.C., and P.K.R. contrib-uted new reagents/analytic tools; M.B., S.A.D., M.A., J.D.A., and J.D.S. analyzed data; M.B.,S.A.D., and J.D.S. wrote the paper; M.V. and E.G. conducted patient enrollment and datamanagement; J.N.M., A.M., L.C., and P.K.R. designed the clinical study; J.N.M., L.P., A.M.,and L.C. designed the clinical study data management; and P.K.R. was responsible for thestudy site.

The authors declare no conflict of interest.

This article is a PNAS Direct Submission. M.W. is a guest editor invited by the EditorialBoard.

Freely available online through the PNAS open access option.

Data deposition: The sequences reported in this paper have been deposited in the Gen-Bank database (accession nos. KU843600–KU843604).1To whom correspondence should be addressed. Email: joe.smith@cidresearch.org.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental.

E3270–E3279 | PNAS | Published online May 16, 2016 www.pnas.org/cgi/doi/10.1073/pnas.1524294113

Dow

nloa

ded

by g

uest

on

June

4, 2

021

http://crossmark.crossref.org/dialog/?doi=10.1073/pnas.1524294113&domain=pdfhttp://www.ncbi.nlm.nih.gov/nuccore/KU843600http://www.ncbi.nlm.nih.gov/nuccore/KU843604mailto:joe.smith@cidresearch.orghttp://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplementalhttp://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplementalwww.pnas.org/cgi/doi/10.1073/pnas.1524294113

protein 2 (PfHRP2), a surrogate of parasite biomass, can predictdisease severity and fatality rates in both children and adults (31,32), the probability of disease deterioration (33), retinopathy-posi-tive cerebral malaria (34), and whether a fever is caused by malaria(35). Nevertheless, a recent longitudinal study in Tanzanian childrenshowed 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. falciparuminfections treated at the Goa Medical College, and we investigatedwhether severe malaria isolates impair the APC–EPCR pathway.

ResultsCharacteristics of the Study Population. A total of 59 P. falciparum-infected patients from the Goa Medical College were enrolled inthe study. Among them, 26 patients had severe malaria (SM) andpresented at least one WHO SM criterion (37), 13 patients werehospitalized 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 significantdifferences in age. Previous studies have found a border betweenpediatric and adult disease symptomatology at 11 y old (4), so fourteenagers were included in this study. Among SM patients, 77%presented more than one severity criteria and 57.7% had at leastthree different severity signs, indicative of multisystem disorders.Because peripheral parasitaemia does not reflect the seques-tered parasite population, we measured plasma concentrationof PfHRP2 as a surrogate of total parasite biomass (Fig. 1). SMpatients had significantly higher plasma PfHRP2 levels thanboth MSM (P = 0.004) and OP (P < 0.0001) (Fig. 1A). Fur-thermore, PfHRP2 concentration increased significantly withthe number of severe criteria and organ dysfunction (Spearman’sρ = 0.67, P < 0.0001) (Fig. 1B).

SM Isolates Overexpressed EPCR-Binding var Transcripts. To identifyvar genes associated with adult severe disease, we performedquantitative RT-PCR (qRT-PCR) using a set of 5 primers that

target groups A, B, or C (VarA, UpsB1, UpsB2, UpsC1, andUpsC2) (16) and 40 degenerate primers that target specific ad-hesion domains (15). Overall, patients in the study presenteda complex population of parasites that transcribed a mixture ofA, B, and C var genes (Fig. 2A and Table 2). However, themedian VarA transcript level was higher than groups B and C inall patient groups, and SM patients had significantly elevatedVarA and UpsB1 transcripts in comparison with OP (Fig. 2A andTable 2).To gain further insight into parasite binding phenotypes as-

sociated with adult SM, domain-specific primers were used toidentify adhesion subtypes expressed in patients, also takingadvantage of the functional specialization of PfEMP1 proteinsto infer parasite binding traits. In particular, the N-terminalPfEMP1 head structure (DBL–CIDR tandem) has diversifiedbetween group A (EPCR binding or rosetting) and groups B andC (CD36 binders). Head structures containing CIDRα1 subtypesbind 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 beamplified by the 40 domain-specific primers in seven annotatedparasite genomes (13) and assigned a predicted binding pheno-type to each gene (EPCR, CD36, or rosetting) depending on itshead structure (Fig. S2). Some primers target var domains thatare associated with more than one type of head structure (15).Using these criteria, we inferred the CD36, EPCR, and rosettingbinding phenotype for each primer (Table 2 and Tables S1 and S2).Compared with outpatients, the SM group presented signifi-

cantly higher transcript levels of var genes with an expectedEPCR-binding phenotype (DC8: DBLα-CIDRα, CIDRα1.1,DBLβ12 and DBLβ3/5, DBLγ4/6; EPCR binders: CIDRα1.4 andCIDRα1) or domains associated with EPCR, rosetting, or CD36binding PfEMP1 (DBLγ of DC6) (Fig. 2 B and C and Table 2).The elevated VarA transcription in SM (Fig. 2A) could be ofeither EPCR or rosetting variants. However, a comparison ofdomain-specific primers suggested EPCR-binding head structuredomains (DBLα1.1/2/4/7, CIDRα1.4, and CIDRα1) were increasedin SM patients (Table 2). By comparison, the primers detected few

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; 19–28Minimum–maximum 18–63 15–62 15–55Male (%) 88.5 84.6 93.3P. falciparum and P. vivax coinfection 0 1 0Parasite density (median), parasites per microliter† 41,410 27,103 51,352IQR 17,898–93,296 0–47,879 31,811–130,536Hemoglobin (mean ± SD) (g/dL) 9.9 ± 3.1 11.8 ± 2.4 12.2 ± 2Glasgow coma score

differences in CD36-binding variants between groups. Transcriptsfor two CD36-binding domains were significantly increased in theSM group (CIDRα3.1–3, DBLα0.1) and for one in the OP group(DBLe2 of DC7), but other CD36-binding transcripts did not differbetween patient groups (Table 2).The MSM group presented a very similar transcription profile

to the SM group and none of the transcripts presented a signifi-cant differential expression after correcting for multiple compar-isons [false-discovery rate (FDR) ≤ 0.2] (Table S1). Furthermore,similar differences in var expression were found when both groupswere 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 (CIDRα1.4and CIDRα1, DBLα2/α1.1/2/4/7) presented a higher expressionin hospitalized patients. In addition, expression of transcriptsamplified by primers DBLγ of DC6 (EPCR, CD36, rosetting) andDBLγ of DC5 (rosetting) was increased among hospitalized pa-tients (Table S2). Taken together, these data show that a higherprevalence of predicted EPCR binding and rosetting parasites wasdetected by multiple primers targeting Group A, DC8, DC6, andDC5 PfEMP1 variants in severe and hospitalized patients (sum-marized in Table S3).

Var Transcript Abundance, Parasite Biomass, and Clinical Disease. PlasmaPfHRP2 concentration presents a close correlation with diseaseseverity in both children and adults (31, 32). To better understandthe interplay between parasite biomass and parasite adhesiontypes in adult SM, we performed correlation analysis. Across allpatients, there was a positive correlation between PfHRP2 plasmaconcentration and transcript abundance of three var domains:DBLγ of DC5 (rosetting) (Spearman’s ρ = 0.3, P = 0.04), DBLγ ofDC6 (EPCR, rosetting, and CD36) (Spearman’s ρ = 0.33, P = 0.02),and CIDRα3.1–3 (CD36) (ρ = 0.32, P = 0.03), suggesting thatparasites expressing these variants might lead to higher parasiteload. In addition, elevated transcription of domains DBLα-CIDRαof 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 DBLe2 of DC7 (CD36) (ρ = −0.33,P = 0.03) and DBLe12 of DC12 (CD36) (ρ = −0.31, P = 0.03) werenegatively associated with PfHRP2. Taking these data together, thisanalysis suggests that specific PfEMP1 domains may contribute tohigher parasite biomass, whereas other PfEMP1 domains may act inconcert with parasite biomass to increase disease severity when theypass 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 decisiontree structure, and random forests (RF) are powerful tools forranking feature importance (39). This methodology generates a

multitude of decision trees and measures the mean decreasein classifier accuracy (MDCA) when a particular feature (e.g.,transcript level detected by a var primer) is removed from themodel. Recent advances in computational techniques have madeit possible to associate P values with features ranked in thismanner (40, 41). First, we used a RF (39) with 1,000,000 trees toselect parasite features that could be used to accurately predictpatient hospitalization and disease severity. Although transcrip-tion of domains DBLα-CIDRα of DC8 and DBLγ of DC6 werethe strongest factors to predict patient hospitalization (SM +MSM vs. OP), the level of PfHRP2 was the most importantfeature for patient severity (SM vs. OP) (Fig. 3A). Moreover, theanalysis suggested that transcription of both DC8 and DC6 do-mains plays a role in patient severity as they ranked second andthird in the prediction of SM [mProbes familywise error rates(FWER) ≤ 0.2]. In addition, all four primers that target the DC8domains ranked in the top 10 of the severity model, highlightingthe importance of DC8-containing PfEMP1 in disease severity.PfEMP1 variants could contribute to disease severity by dis-

tinct mechanisms, including by promoting the rapid multiplica-tion of parasites (biomass) or by encoding dangerous adhesiontraits (impairing critical endothelial functions). Univariate cor-relation analysis showed an association between PfHRP2 plasmalevels and transcription of domain DBLγ of DC6. To investigatethe possibility that the correlation of the DC6 domain in diseaseseverity was simply because of an association with parasite bio-mass, we applied conditional mutual information (CMI) algo-rithms. CMI is an information theory technique to test if onevariable’s predictive power is weakened by the presence of oth-ers. After PfHRP2 filtration (i.e., filtering out those primers thatwere uninformative after PfHRP2 levels were accounted for),transcription of domains DBLγ of DC6 and CIDRα1.1 of DC8remained important to disease severity and patient hospitalization(CMI PfHRP2 filtration P ≤ 0.05), showing that the virulence ofthese DCs is not simply explained by a major contribution in par-asite biomass (Fig. 3A).To provide a visualization of the logic used by decision trees,

we generated evolutionary decision trees to understand howparasite factors might interact to determine disease severity inadults (Fig. 3B). In the tree that tests patient hospitalization(MSM + SM vs. OP), a combination of high PfHRP2 levels andelevated transcription of domains CIDRα1.1 of DC8 and DBLγof DC6 were sufficient to correctly classify 91% of patients ashospitalized or not. Furthermore, a combination of these threefeatures was sufficient to classify 100% of SM patients (SM vs.OP). Taken together, machine-learning approaches determinedthat both parasite biomass and transcription of DC8- and DC6-containing PfEMP1 variants are critical for adult malaria severity.

Fig. 1. Association between PfHRP2 plasma concentrations and disease severity. PfHRP2 levels were compared between patient groups and by number ofWHO SM criteria. (A) PfHRP2 plasma concentrations among disease groups. Horizontal lines indicate median for each group. Pairwise comparisons wereanalyzed using the Mann–Whitney U test. Significant higher concentration is represented by **P < 0.01 and ***P < 0.001. (B) Spearman’s rank correlationcoefficient (ρ) and P value for the association between PfHRP2 plasma concentrations and number of severity criteria.

E3272 | www.pnas.org/cgi/doi/10.1073/pnas.1524294113 Bernabeu et al.

Dow

nloa

ded

by g

uest

on

June

4, 2

021

http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=ST1http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=ST2http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=ST2http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=ST3www.pnas.org/cgi/doi/10.1073/pnas.1524294113

The machine-learning analysis implicated both an N-terminalPfEMP1 domain (DC8 cassette: DBLα2-CIDRα1.1/1.8-DBLβ12-DBLγ4/6) and a C-terminal PfEMP1 domain (DC6 cassette:DBLγ14-DBLζ5-DBLe4) in severe disease. DC6 can be presentin 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 struc-tural linkage between DC8 and DC6 in SM patients, we per-formed correlation analysis. As expected, the four DC8 domainswere highly correlated (Spearman’s ρ = 0.51–0.71 for all pairwisecomparisons). By comparison, the DBLγ of DC6 domain wasequally correlated to other DC6 domains (ρ = 0.34) and DC8domains and group A variants (ρ > 0.3) (Fig. S3). These resultssuggest that DC8 and DC6 may be linked in the same PfEMP1,but DC6 may also be present in non-DC8 PfEMP1. Thus, it ispossible that different subsets of DC6-containing PfEMP1 vari-ants may contribute to SM.Many of the var primers match different parts of the same DC

(e.g., DC8) or share similar functional annotations (e.g., EPCRbinding). Consequently, var primers with the same annotationmay artificially lower MDCA values in the RF analysis by act-ing as proxies for each other. To account for this effect, set-enrichments were performed to combine primers targeting thesame var group or adhesion types (Fig. S2 and Table S4). For thisanalysis, a Mann–Whitney U test was used to compare theMDCA values of set-enriched primer groups versus primers nottargeting that annotation (SI Materials and Methods). In both thepatient hospitalization and disease-severity models, differentcombinations of DC8 primers (DC8 pure and DC8 all) presentedenrichment P values lower than 0.005 that remained significantafter FDR correction (FDR ≤ 0.2) (Fig. 3C). The associationwith patient hospitalization and disease severity was still signif-icant but reduced in the group that contained all EPCR binders(DC8 EPCR + group A EPCR). Despite playing an importantrole in the univariate analysis, VarA transcriptional levels werenot necessary to predict hospitalization or disease severity. Anenrichment group that excluded DC8 and incorporated bothtypes of UpsA adhesion traits (group A = EPCR and rosettingvariants combined) failed to reach significance, reinforcing theimportance of DC8. Rosetting, IgM, or intercellular adhesionmolecule-1 (ICAM-1) binding domains groups, which have beenpreviously associated with SM (reviewed in ref. 42) did notpresent any significance in disease predictability of adult SM. Asexpected, the CD36 binding group showed no significance forpredicting either hospitalization or severity. Taken together,these data show that machine-learning approaches highlight theimportance of DC8 in adult SM, and determine that in combi-nation with the DBLγ of DC6 domain and elevated parasitebiomass, is strongly associated with disease severity.

DC8 CIDRα1 Domains Expressed by SM Isolates Inhibit the APC–EPCRInteraction.DC8 CIDRα1 differ in sequence, binding affinity, andability to inhibit the APC–EPCR interaction (25, 27, 28), butthere have been no in-depth functional characterizations of DC8CIDRα1 from SM isolates. To perform a deeper phenotypiccharacterization of DC8 variants in severe isolates, we amplifiedand sequenced the full-length DC8 CIDRα1 domain (Fig. S4)from five patients who met WHO SM criteria: patient 24 hadanemia, patient 25 had jaundice, and patients 62, 87, and 95presented multiorgan complications, including cerebral malariain patients 62 and 87. In total, seven CIDRα transcripts wereamplified from the five patients. All of the sequences clusteredtogether with domains from subgroups CIDRα1.1 and CIDRα1.8,which are diagnostic for DC8-like (Fig. 4A). All five patients pre-sented a single DC8 sequence except for patient 62, who presentedthree distinct DC8 CIDRα1 domains. Notably, an identical DC8CIDRα1 transcript was amplified from three patients with multi-organ complications, including both cerebral malaria patients

Fig. 2. Transcription of UpsA, DC6, and DC8 var is elevated in SM pa-tients. The transcript abundances of var gene subtypes were investigatedamong 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. Horizontallines indicate median for each group. Differences among groups werecompared by using the Mann–Whitney U test. Significant higher tran-scription is represented by *P ≤ 0.05 and FDR ≤ 0.2, ***P ≤ 0.005 andFDR ≤ 0.05. (C) Heat map showing transcription levels of DC8 and DC6 domainsubtypes and VarA genes. Maximum transcription levels are represented inred, minimum transcription in blue, and median transcription levelsin white. Color equivalents were set by comparing each primer tran-script among all patients analyzed. FDR, Benjamini–Hochberg adjust-ment for false discovery rate.

Bernabeu et al. PNAS | Published online May 16, 2016 | E3273

MICRO

BIOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

June

4, 2

021

http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=SF1http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=SF3http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=SF2http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=ST4http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=STXThttp://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=SF4

(62 and 87). By comparison, the five distinct CIDRα1 sequenceshad 32% sequence identity.To investigate whether SM isolates interfere with the APC–

EPCR interaction, the five unique Indian CIDRα1 domains wereexpressed as recombinant proteins (Fig. S4 and Table S5). Fourof five bound EPCR with low to moderate nanomolar bindingconstants (Kds = 3.81 nM to 63 nM) (Fig. 4B and Fig. S5), similarin range to other CIDRα1.1/8 domains (27, 29). Although r62-2-22CIDRα1.1 did not bind recombinant EPCR, this domain exhibiteddose-dependent binding to Chinese hamster ovary (CHO) cellsexpressing EPCR in a flow-cytometry binding assay (Fig. S5),suggesting it also possesses low EPCR binding activity. Asexpected, in binding competition studies, there was limited or no

inhibition of APC binding to CHO745-EPCR cells coincubatedwith either the negative-control PFE1640w CIDRα1.3 domain orthe weak binder domain r62-2-22 (Fig. 4B). By comparison, twoCIDR domains from Indian isolates (r25-2-4 and r62-2-1) pre-sented low inhibition of APC binding and the two remainingCIDR domains from Indian isolates (r62-2-23 and r24-2-4)presented moderate APC inhibition (Fig. 4B). However, all ofthe DC8 domains expressed by SM isolates competed less effec-tively for the APC–EPCR interaction than a positive control, groupA variant (IT4var07 CIDRα1.4 domain), which strongly blocks theAPC–EPCR interaction (25, 26, 28).To study if SM isolates interfere with the endothelial bar-

rier protective activity of APC, we investigated the ability of

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 0.36A DBLe8 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 0.18A DBLα1.4 0.1 (0-1.8) 0.1 (0-2) 0.11 0.28A CIDRα1.6 0 (0-0.1) 0 (0-0.1) 0.34 0.40A CIDRα1.7 2.5 (0-9.6) 0 (0-17.6) 0.28 0.38A DBLβ3 1.5 (0-20.4) 0.2 (0-7.1) 0.14 0.30A DBLα1.7 of DC13 0.5 (0.1–23) 0.2 (0.1–3.8) 0.2 0.35A CIDRα1.4 of DC13 0.9 (0-2.1) 0.5 (0-3.1) 0.32 0.40A CIDRα1.4 and CIDRα1 2.9 (0.2–22.8) 0 (0-3.6) 0.006 0.11B/A DBLα-CIDRα of DC8 4.5 (1.1–9.6) 0.1 (0-3.1) 0.02 0.13B/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 0.11B/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 0.01B (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 0.50A,B,C DBLγ9 0 (0-0) 0 (0-0) 0.46 0.49

CD36 B (C) DBLe2 of DC7 1.2 (0.2–3.5) 4.3 (0.4–23) 0.04 0.16B (A,C) DBLe3 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 0.35B DBLζ6 of DC10 1.5 (0.3–10.6) 1.2 (0.2–4) 0.18 0.35B DBLα0.16 of DC19 4.9 (0.5–32.4) 1 (0.4–22.3) 0.12 0.28B,C CIDRα3.4 of DC19 21.8 (8.2–57.8) 10.9 (4.6–35) 0.06 0.18B DBLα0.9 of DC20 7.3 (1-29) 2.9 (0.5–10.6) 0.13 0.29B DBLα0.1 1 (0.3–4.6) 0.4 (0.3–1.3) 0.05 0.18B DBLα0.6/9 39.3 (10.6–139.1) 42.5 (1.2–93.1) 0.25 0.38B CIDRα2.2 1.2 (0-40.5) 5.3 (0-48.5) 0.32 0.40B CIDRα2.3/5/6/7/9/10 30.7 (9.9–97.7) 38.3 (9.1–57.7) 0.41 0.46B,C CIDRα3.1–3 6.7 (1.6–28) 2.2 (0.6–7.7) 0.02 0.13B 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 0.36B,C CIDRγ 82.9 (34-140.8) 138.1 (24-315.2) 0.21 0.35

CD36 and rosetting B,A DBLe12 of DC12 1.9 (0-15.3) 3 (0.2–9.5) 0.33 0.40B 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 0.18A DBLβ7 & 9 of DC5 0 (0-0) 0 (0-0.1) 0.19 0.35A DBLα1.5/6b of DC16 0.6 (0.1–1.7) 0.6 (0.1–4.4) 0.21 0.35A DBLα1.5/6a of DC16 3.1 (0.3–20.1) 8.2 (0.4–16.5) 0.28 0.38A 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 0.13UpsB1 36 (18-68.4) 21.4 (5.7–47.2) 0.04 0.16UpsB2 18.8 (8.1–30) 13.5 (10.5–19) 0.09 0.24UpsC1 5.4 (3.3–18.8) 9.8 (3-15.6) 0.49 0.50UpsC2 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 aone-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 boldfaceand considered significant.

E3274 | www.pnas.org/cgi/doi/10.1073/pnas.1524294113 Bernabeu et al.

Dow

nloa

ded

by g

uest

on

June

4, 2

021

http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=SF4http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=ST5http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=SF5http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=SF5www.pnas.org/cgi/doi/10.1073/pnas.1524294113

CIDR domains to inhibit APC protective activity in thrombin-induced endothelial barrier permeability assays. APC diminishedthrombin-induced barrier disruption by 24% in EA.hy926 humanumbilical vein endothelial cells and 35% in cultured primary hu-man brain endothelial cells (Fig. 4C). In agreement with thebinding competition studies, r62-2-22 and r25-2-4 did not inhibitAPC barrier protective function, r62-2-1 and r62-2-23 partially

inhibited (∼30–45% reduction relative to APC), and IT4var07 andr24-2-4 strongly inhibited the APC protective pathway in bothEA.hy926 cells (40–48% reduction) and primary brain endothelialcells (60–75% reduction). In general, CIDR domains tended tohave slightly higher inhibition on brain endothelial cells thanEA.hy926 cells (Fig. 4C), although the difference did not reachstatistical significance. Taking these data together, this analysis

Fig. 3. Machine-learning approach to understand disease severity. Parasite factors associated with a higher risk of patient hospitalization and disease severity wererevealed by machine-learning approaches. (A) Summary of RF feature selection strategy to identify parasite virulence factors that discriminate between hospitalizedpatients (SM + MSM) (Left), SM patients (Right), and OP. The top 10 parasite factors with the highest MDCA are shown. Positive correlation with disease severity isshown 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 usingmProbes 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 tofind 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 boxesrepresent the probability of each pathway to classify patients into hospitalized (H), severe malaria (S), or outpatients (OP). The number of patients in each pathway isindicated below. The percentages beneath the lines show the proportion of the total severe patients classified by each pathway. (C) Var primers were groupedaccording to binding phenotype or var group (Fig. S2 and Table S4) and ranked by MDCA (15, 16). The association with patient hospitalization (Left) and diseaseseverity (Right) was determined using a Mann–Whitney U. P ≤ 0.05 and FDR ≤ 0.2 are considered significant. FDR, Benjamini–Hochberg adjustment for FDR.

Bernabeu et al. PNAS | Published online May 16, 2016 | E3275

MICRO

BIOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

June

4, 2

021

http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=SF2http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=SF2http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=ST4

Fig. 4. Inhibition of the APC–EPCR interaction by DC8 CIDRα from severe isolates. DC8 CIDRα domains expressed from SM isolates were analyzed for EPCRbinding 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 (Kd) for rCIDRα1.1-EPCRmeasured by biolayer interferometry (see Fig. S5 for detailed kinetics). Histograms showAPC binding to CHO-EPCR cells in the presence or absence of 250 μg/mLIndian 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 andSD, 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 notinhibit 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 werecalculated 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.

E3276 | www.pnas.org/cgi/doi/10.1073/pnas.1524294113 Bernabeu et al.

Dow

nloa

ded

by g

uest

on

June

4, 2

021

http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=SF5www.pnas.org/cgi/doi/10.1073/pnas.1524294113

indicates that DC8 CIDRα1 domains expressed from SM isolatesbind EPCR and may inhibit the APC–EPCR pathway.

DiscussionStudies to understand the role of PfEMP1 in SM pathogenesishave been mainly focused on children in Africa (15, 18, 19) andinformation on adult severe patients still remains scarce (43).However, in areas of unstable transmission, SM occurs across agegroups. The different symptomatology in adult SM, the presence ofmultiorgan complications, and the higher fatality rate urge researchon this population. Here, we investigated the relationship betweenparasite 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 SMpatients. Prior studies suggest group A and DC8-containing vargenes are preferentially expressed in young African children withlimited immunity (15, 17, 44) and nonimmune European travelers(45), suggesting these variants confer a parasite growth advantagein malaria naïve hosts, and in some circumstances increase the riskfor SM (14, 15, 17, 23). Machine-learning prediction models sug-gest that high transcription of DC8 and DC6 domains in combi-nation with high parasite biomass is associated with adult patienthospitalization and severity. Conversely, the importance of groupA dropped in the RF analysis, possibly because group A transcriptswere highly expressed in all patient groups, as might be expectedfor a lower transmission setting. It also remains possible that groupA may be more closely linked to specific adult disease syndromes.Previously, logistic regression analysis was used to assign differentbinding variants to respiratory distress and impaired consciousnessin pediatric malaria (46), but to our knowledge, this is the first timethat more advanced machine-learning approaches have been usedto understand parasite factors associated with disease progressionand malaria severity. We expect that powerful machine-learningalgorithms, such as those presented here, would be useful foranalyzing var expression data from African children (15, 18, 19)and may shed light into pathogenesis mechanisms that drive pe-diatric cerebral malaria, respiratory distress, or anemia.Endothelial dysfunction is thought to play an important role in

SM pathology. Indeed, microthrombi are a common finding inpediatric cerebral malaria autopsies (47, 48), and cerebral swellingis a major risk factor for death (29). Although fibrin deposits andcerebral swelling are more variable in adult cerebral malaria au-topsies (49, 50), alterations in blood–brain barrier integrity havebeen associated with infected erythrocyte sequestration in adultSoutheast Asian and pediatric African autopsies (51, 52). It hasbeen hypothesized that EPCR binding parasites may drive diseasepathogenesis by blocking the anticoagulation, anti-inflammatory,and barrier protective functions of the APC–EPCR pathway (24).However, recent studies of CIDRα1 domains from long-termcultured adapted parasites have revealed large differences in theirability to inhibit the APC–EPCR interaction (25–28). The con-sequences of these differences for disease severity remain un-known. Here, we showed that DC8 CIDRα1 domains expressed bySM isolates possess differential activity to disrupt the EPCR–APCprotective pathway. Some DC8 domains had low activity andothers had nearly equivalent activity to a group A CIDRα1 do-main that strongly blocks the APC–EPCR interaction (25, 26, 28).Notably, an identical DC8 CIDRα1 domain (r62-2-1) isolatedfrom two patients with cerebral malaria and a third patient withmultiorgan complications produced a small and significant in-hibition of APC barrier protective activity on brain endothelialcells. This finding raises the possibility that even parasite domainswith weaker APC inhibition activity may interfere with this impor-tant host regulatory pathway, especially in microvascular beds wherethere is loss of EPCR expression associated with parasite seques-tration (30). Unexpectedly, a DC8 CIDRα1 domain (r24-2-4) iso-lated 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 symp-tomatology. Therefore, the relation between CIDRα1 phenotypesand disease symptoms reinforces the notion that a certain thresholdof parasite biomass in combination with virulent PfEMP1 variants isassociated with overlapping severe symptomatology in adults.The clear association of DC8 with both children (15) and adult

SM spotlights the CIDRα–EPCR interaction. However, DC8variants encode multiple endothelial binding domains (53), andit is possible that other undefined coadhesion traits may also in-crease the risk for SM. Furthermore, our study, to our knowledgefor the first time, implicates specific C-terminal PfEMP1 domainsin SM. DC6 is characterized by DBLγ14-DBLζ5-DBLe4 domainsand can be present in combination with rosetting, EPCR, orCD36-binding head structures (13). Thus, it will be important toexplore if both rosetting and nonrosetting PfEMP1 variants arecontributing to adult SM. Although DC6 binding properties areuncharacterized, recent studies have shown that DBLζ and DBLemediate binding to IgM and α2-macroglobulin, and it has beenhypothesized that binding to these serum factors can cross-linkmultiple PfEMP1 to increase the binding affinity of N-terminaldomains (54, 55). Further sequencing of DC6 variants will benecessary to assess whether DC6 and DC8 are part of the same ordifferent proteins in severe isolates, and to determine the bindingproperties of DBLζ and DBLe Indian domains to serum factors.A limitation of our study is the relatively small population

studied after 3 y of patient recruitment. Future machine-learningapproaches with larger sample groups and examining differentgeographic settings will be important to understand whetherdifferences in var expressions are responsible for distinct adultsevere symptoms. A second limitation is that the primers used forvar profiling and sequencing (15) might fail to recognize some vardomains important for disease severity or might be biased towardcertain variants. For example, the DC6 cassette is found in bothrosetting and nonrosetting PfEMP1 variants and rosetting variantsmay have been underestimated by the PCR typing approach. It willbe valuable to use independent methodologies to investigate thecontribution of rosetting parasite variants in our adult India pop-ulation. Other parasite adhesins expressed at the surface of infectederythrocytes, such as RIFINs and STEVOR, play a role in rosetting(56, 57) but were not studied here. In the future, it would be in-teresting to study the interplay between var and other parasiteadhesins in disease severity. Nevertheless, the degenerate primersagainst var adhesion domains remain a sensitive and cost-effectivetool to cover the var geographical diversity and, combined withmachine-learning approaches, provide a powerful methodology toinvestigate pathogenic mechanisms.In summary, our data show that elevated DC8 and DC6 var

transcripts, along with high parasite biomass, promote diseaseprogression in adult SM. In addition, our findings raise the pos-sibility that DC8 CIDRα1 domains with low or moderate APCblockade activity interfere with APC–EPCR protective pathways,highlighting attention on this pathway for disease interventions andthe future development of SM adjunctive therapies.

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

Patient Recruitment and Samples Collection. Subjects were recruited betweenApril 2012 to October 2014 from the hospital admission or outpatient wardsfrom patients presenting at Goa Medical College and Hospital (Goa, India).Subjects were enrolled by project staff who explained the project. Followinginformed consent, blood samples from P. falciparum-positive patients werecollected in acid citrate dextrose vacutainers and separated into plasma andred blood cells in RNALater. Samples were stored at –80 °C. Infections were

Bernabeu et al. PNAS | Published online May 16, 2016 | E3277

MICRO

BIOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

June

4, 2

021

confirmed by study staff using Giemsa-stained thin and thick smears forparasitemia determination and Plasmodium species identification. Rapiddiagnostic test (Zephyr Biomedicals) was additionally used for the diagnosisof parasite species. Parasites per milliliter was calculated from thin filmsmears (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 < 80mmHg with cold extremities), (vi) metabolic acidosis (peripheral venous bi-carbonate < 15 mmol/L), (vii) respiratory distress (respiratory rate > 20breaths per minute or PaO2 < 75 mmHg), and (viii) hypoglycemia (bloodglucose < 40 mg/dL). Patients admitted to the hospital without any of thesecriteria were considered as MSM and nonadmitted patients were consideredas OP. Patients with mono P. falciparum infections were treated with oralartesunate and mefloquine, and intravenous artesunate was used for SMpatients. One study patient was coinfected with P. falciparum and Plasmo-dium 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-sitesandwich 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 reactionsubstrate, and measured spectrophotometrically at 450 nm. Positive andnegative controls were included in each plate. A standard curve wasestablished using purified recombinant PfHRP2 protein (kindly donated byDavid Sullivan, Johns Hopkins Bloomberg School of Public Health, Baltimore)diluted 0.25–55 ng/mL in PBS. Five patient samples that presented a lowerconcentration than the detection limit were excluded from the analysis.

Determination of var Transcription by qRT-PCR. Thawed red blood cells inRNAlater were dissolved in 12 volumes of TRIzol and RNAwas extracted usingan RNeasy micro kit (Qiagen). cDNA was synthesized using random hexamersand a MultiScribe reverse transcriptase (Thermo Fisher). qRT-PCR was per-formed using QuantiTect SYBR in a Mastercycler Realplex2 following pub-lished amplification conditions (15, 16). Data were acquired after theelongation step of each cycle. Absence of DNA in RNA samples was con-firmed by running a reverse-transcriptase negative sample with the house-keeping gene primer adenylosuccinate lyase (ASL) (PFB0295w). Levels of vargene expression were determined by relative quantification of the averageexpression of ASL and seryl-tRNA synthetase housekeeping genes (ΔCtvar_primer = Ct var_primer − Ct average_housekeeping primers). The levelof var expression was represented as Transcript units (Tu) and calculated asTu = 2(5−ΔCt) (15). Samples were included in the univariate statistical analysisonly if the Ct average of the housekeeping genes was below 30.

CIDRα1 Sequencing and Recombinant Protein Expression. DC8 transcripts wereamplified from patients using the strategy depicted in Fig. S4. Bands of theexpected size were excised, purified, and cloned in a Zero Blunt TOPO vector. Atleast 15 different colonies were sequenced per patient and analyzed usingGeneious (7.1.7). Sequences were deposited in GenBank with accession numbersKU843600–KU843604. Recombinant CIDRα1 were synthesized as GBlocks genefragments (IDT) and produced as His6-MBP-TEV-PfEMP1 insert-StrepII–taggedproteins, as previously described (58). Recombinant proteins were purified in atwo-step process using an amino-terminal His tag and a carboxyl-terminal StrepIItag and analyzed by SDS/PAGE according to standard procedures (28).

Biolayer Interferometry Analysis. Binding of the CIDR domains to biotinylatedEPCR was determined on the Octet QKe instrument (ForteBio), as describedpreviously (28). Mean Kon and Koff and apparent Kd values were determinedfrom double-reference subtracted data from three concentrations that werefitted globally to a 1:1 Langmuir binding model using the data analysis software.

rCIDRα Binding Titration to CHO745-EPCR and APC Competition Assay. For therCIDRα binding titration to CHO745-EPCR and APC competition assay, 105

CHO745-EPCR cells were lifted and washed in complete HBSS (HBSS with3 mM CaCl2, 0.6 mM MgCl2, 1% BSA) (28). A five-point titration curve was de-termined with 1–250 μg/mL of recombinant CIDR for 30 min. CIDR binding wasdetected with a rabbit polyclonal anti-StrepII tag antibody followed by a goatanti-rabbit Alexa488-coupled antibody. Control samples were labeled with pri-mary and secondary antibodies alone to set gates. APC competition assays weredone as described previously (28). Briefly, 105 CHO745-EPCR cells were coincu-bated with 50 μg/mL APC (Sigma) and 50 or 250 μg/mL CIDR recombinant pro-teins 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 coupledantibody. The percentage (%) APC binding was calculated relative to the value ofcells incubated with APC alone. Labeled cells were analyzed by flow cytometryusing LSRII (Becton Dickinson) and data were analyzed by FlowJO 10 software(Tree Star).

Measurement of the Monolayer Permeability. Barrier function was moni-tored using a real-time cell analyzer (xCELLigence System, ACEA Biosciences).This system measures electrical impedance across the cell monolayer, cell im-pedance (CI), via gold microelectrodes integrated on the bottom of a 96-wellplate. Next, 10,000 EA.hy926 HUVEC (ATCC) or primary human brain endo-thelial cells (Cells Systems) were seeded in each well. Cell proliferation wasassessed for 72 h, atwhich time the cells reached a sustainedmaximumCI value.For the experiment, cells were incubated with rCIDRα1 (0.05 mg/mL) or culturemedium. After 30 min, 100 nM of human APC (Haematologic Technologies)was added and incubated for 2 h. Barrier disruption was induced with 5 nMthrombin (Sigma) and compared with untreated cells (baseline). Cells weretreated in triplicate and CI was measured every minute up to 120 min, thenevery 5 min up to 245 min after thrombin challenge. The baseline-normalizedcell index was calculated by comparing the CI values of treated cells to the CIvalues for baseline control wells of untreated cells at the time point immedi-ately before the thrombin challenge. The level of barrier protection achievedby APC + thrombin treatment relative to thrombin treatment alone was set to100% to calculate the binding inhibitory activity of CIDR domains.

Machine-Learning Models. Decision trees for understanding parasite factorsassociated with patient hospitalization and disease severity were analyzed byRF (39) and evolutionary Trees (59) or by CMI analysis (60). Detailed methodscan be found in SI Materials and Methods.

Statistical Analysis. Univariate analyses were performed using GraphPadPrism v5.02 for Windows. Correlations between variables were tested using theSpearman’s rank correlation coefficient. Differences between groups wereevaluated using the Mann–Whitney U test with a Benjamini–Hochberg adjust-ment for FDR or the one-way ANOVAwith a Dunnett’s multiple comparison test.

ACKNOWLEDGMENTS.We thank the patients who participated in this study;Dr. David Sullivan for PfHRP-2 recombinant protein; and Profs. Panda andPatankar for the use of their ELISA reader at IIT, Bombay. This work wassupported by funds from NIH Grants U19 AI 089688 (to P.K.R. and J.D.S.) andP41 GM109824 (to J.D.A.).

1. World Health Organization (2015) WHO World Malaria Report 2015. Avaliableat www.who.int/malaria/publications/world-malaria-report-2015/report/en/. AccessedFebruary 2016.

2. Anonymous (2014) Severe malaria. Trop Med Int Health 19(Suppl 1):7–131.3. White NJ, et al. (2014) Malaria. Lancet 383(9918):723–735.4. Dondorp AM, et al. (2008) The relationship between age and the manifesta-

tions of and mortality associated with severe malaria. Clin Infect Dis 47(2):151–157.

5. Miller LH, Baruch DI, Marsh K, Doumbo OK (2002) The pathogenic basis of malaria.Nature 415(6872):673–679.

6. Dondorp AM, et al. (2008) Direct in vivo assessment of microcirculatory dysfunction insevere falciparum malaria. J Infect Dis 197(1):79–84.

7. Yeo TW, et al. (2008) Angiopoietin-2 is associated with decreased endothelial nitricoxide and poor clinical outcome in severe falciparum malaria. Proc Natl Acad Sci USA105(44):17097–17102.

8. Baruch DI, et al. (1995) Cloning the P. falciparum gene encoding PfEMP1, a malarialvariant antigen and adherence receptor on the surface of parasitized human eryth-rocytes. Cell 82(1):77–87.

9. Smith JD, et al. (1995) Switches in expression of Plasmodium falciparum var genescorrelate with changes in antigenic and cytoadherent phenotypes of infectederythrocytes. Cell 82(1):101–110.

10. Su XZ, et al. (1995) The large diverse gene family var encodes proteins involved incytoadherence and antigenic variation of Plasmodium falciparum-infected erythro-cytes. Cell 82(1):89–100.

11. Lavstsen T, Salanti A, Jensen AT, Arnot DE, Theander TG (2003) Sub-grouping ofPlasmodium falciparum 3D7 var genes based on sequence analysis of coding and non-coding regions. Malar J 2(1):27.

12. Smith JD, Subramanian G, Gamain B, Baruch DI, Miller LH (2000) Classification ofadhesive domains in the Plasmodium falciparum erythrocyte membrane protein1 family. Mol Biochem Parasitol 110(2):293–310.

E3278 | www.pnas.org/cgi/doi/10.1073/pnas.1524294113 Bernabeu et al.

Dow

nloa

ded

by g

uest

on

June

4, 2

021

http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=SF4http://www.pnas.org/lookup/suppl/doi:10.1073/pnas.1524294113/-/DCSupplemental/pnas.201524294SI.pdf?targetid=nameddest=STXThttp://www.who.int/malaria/publications/world-malaria-report-2015/report/en/www.pnas.org/cgi/doi/10.1073/pnas.1524294113

13. Rask TS, Hansen DA, Theander TG, Gorm Pedersen A, Lavstsen T (2010) Plasmodiumfalciparum erythrocyte membrane protein 1 diversity in seven genomes—Divide andconquer. PLOS Comput Biol 6(9):e1000933.

14. Kyriacou HM, et al. (2006) Differential var gene transcription in Plasmodium falci-parum isolates from patients with cerebral malaria compared to hyperparasitaemia.Mol Biochem Parasitol 150(2):211–218.

15. Lavstsen T, et al. (2012) Plasmodium falciparum erythrocyte membrane protein1 domain cassettes 8 and 13 are associated with severe malaria in children. Proc NatlAcad Sci USA 109(26):E1791–E1800.

16. Rottmann M, et al. (2006) Differential expression of var gene groups is associatedwith morbidity caused by Plasmodium falciparum infection in Tanzanian children.Infect Immun 74(7):3904–3911.

17. Warimwe GM, et al. (2009) Plasmodium falciparum var gene expression is modified byhost immunity. Proc Natl Acad Sci USA 106(51):21801–21806.

18. Warimwe GM, et al. (2012) Prognostic indicators of life-threatening malaria are as-sociated with distinct parasite variant antigen profiles. Sci Transl Med 4(129):129ra45.

19. Bertin GI, et al. (2013) Expression of the domain cassette 8 Plasmodium falciparumerythrocyte membrane protein 1 is associated with cerebral malaria in Benin. PLoSOne 8(7):e68368.

20. Carlson J, et al. (1990) Human cerebral malaria: Association with erythrocyte rosettingand lack of anti-rosetting antibodies. Lancet 336(8729):1457–1460.

21. Ghumra A, et al. (2012) Induction of strain-transcending antibodies against GroupA PfEMP1 surface antigens from virulent malaria parasites. PLoS Pathog 8(4):e1002665.

22. Kaul DK, Roth EF, Jr, Nagel RL, Howard RJ, Handunnetti SM (1991) Rosetting ofPlasmodium falciparum-infected red blood cells with uninfected red blood cells en-hances microvascular obstruction under flow conditions. Blood 78(3):812–819.

23. Turner L, et al. (2013) Severe malaria is associated with parasite binding to endo-thelial protein C receptor. Nature 498(7455):502–505.

24. Mosnier LO, Zlokovic BV, Griffin JH (2007) The cytoprotective protein C pathway.Blood 109(8):3161–3172.

25. Gillrie MR, et al. (2015) Diverse functional outcomes of Plasmodium falciparum liga-tion of EPCR: Potential implications for malarial pathogenesis. Cell Microbiol 17(12):1883–1899.

26. Lau CK, et al. (2015) Structural conservation despite huge sequence diversity allowsEPCR binding by the PfEMP1 family implicated in severe childhood malaria. Cell HostMicrobe 17(1):118–129.

27. Petersen JE, et al. (2015) Protein C system defects inflicted by the malaria parasiteprotein PfEMP1 can be overcome by a soluble EPCR variant. Thromb Haemost 114(5):1038–1048.

28. Sampath S, et al. (2015) Plasmodium falciparum adhesion domains linked to severemalaria differ in blockade of endothelial protein C receptor. Cell Microbiol 17(12):1868–1882.

29. Seydel KB, et al. (2015) Brain swelling and death in children with cerebral malaria.N Engl J Med 372(12):1126–1137.

30. Moxon CA, et al. (2013) Loss of endothelial protein C receptors links coagulation andinflammation to parasite sequestration in cerebral malaria in African children. Blood122(5):842–851.

31. Dondorp AM, et al. (2005) Estimation of the total parasite biomass in acute falcipa-rum malaria from plasma PfHRP2. PLoS Med 2(8):e204.

32. Hendriksen IC, et al. (2012) Diagnosing severe falciparum malaria in parasitaemicAfrican children: A prospective evaluation of plasma PfHRP2 measurement. PLoS Med9(8):e1001297.

33. Fox LL, et al. (2013) Histidine-rich protein 2 plasma levels predict progression to ce-rebral malaria in Malawian children with Plasmodium falciparum infection. J InfectDis 208(3):500–503.

34. Seydel KB, et al. (2012) Plasma concentrations of parasite histidine-rich protein2 distinguish between retinopathy-positive and retinopathy-negative cerebral ma-laria in Malawian children. J Infect Dis 206(3):309–318.

35. Hendriksen IC, et al. (2013) Defining falciparum-malaria-attributable severe febrileillness in moderate-to-high transmission settings on the basis of plasma PfHRP2concentration. J Infect Dis 207(2):351–361.

36. Gonçalves BP, et al. (2014) Parasite burden and severity of malaria in Tanzanianchildren. N Engl J Med 370(19):1799–1808.

37. World Health Organization (2000) Severe falciparum malaria. World Health Organi-zation, communicable diseases cluster. Trans R Soc Trop Med Hyg 94(Suppl 1):S1–S90.

38. Smith JD, Rowe JA, Higgins MK, Lavstsen T (2013) Malaria’s deadly grip: Cytoadhesionof Plasmodium falciparum-infected erythrocytes. Cell Microbiol 15(12):1976–1983.

39. Liaw A, Wiener M (2002) Classification and regression by randomForest. R News 2(3):18–22.

40. Huynh-Thu VA, Saeys Y, Wehenkel L, Geurts P (2012) Statistical interpretationof machine learning-based feature importance scores for biomarker discovery.Bioinformatics 28(13):1766–1774.

41. Finney OC, et al. (2014) Predicting antidisease immunity using proteome arrays andsera from children naturally exposed to malaria. Mol Cell Proteomics 13(10):2646–2660.

42. Rowe JA, Claessens A, Corrigan RA, Arman M (2009) Adhesion of Plasmodium falci-parum-infected erythrocytes to human cells: Molecular mechanisms and therapeuticimplications. Expert Rev Mol Med 11:e16.

43. Subudhi AK, et al. (2015) Plasmodium falciparum complicated malaria: Modulationand connectivity between exportome and variant surface antigen gene families. MolBiochem Parasitol 201(1):31–46.

44. Cham GK, et al. (2009) Sequential, ordered acquisition of antibodies to Plasmodiumfalciparum erythrocyte membrane protein 1 domains. J Immunol 183(5):3356–3363.

45. Bachmann A, et al. (2011) Highly co-ordinated var gene expression and switching inclinical Plasmodium falciparum isolates from non-immune malaria patients. CellMicrobiol 13(9):1397–1409.

46. Abdi AI, et al. (2014) Plasmodium falciparum antigenic variation: Relationships be-tween widespread endothelial activation, parasite PfEMP1 expression and severemalaria. BMC Infect Dis 14(1):170.

47. Dorovini-Zis K, et al. (2011) The neuropathology of fatal cerebral malaria in Malawianchildren. Am J Pathol 178(5):2146–2158.

48. Milner DA, Jr, et al. (2014) The systemic pathology of cerebral malaria in Africanchildren. Front Cell Infect Microbiol 4:104.

49. MacPherson GG, Warrell MJ, White NJ, Looareesuwan S, Warrell DA (1985) Humancerebral malaria. A quantitative ultrastructural analysis of parasitized erythrocytesequestration. Am J Pathol 119(3):385–401.

50. Spitz S (1946) The pathology of acute falciparum malaria. Mil Surg 99(5):555–572.51. Brown H, et al. (1999) Evidence of blood-brain barrier dysfunction in human cerebral

malaria. Neuropathol Appl Neurobiol 25(4):331–340.52. Brown H, et al. (2001) Blood-brain barrier function in cerebral malaria in Malawian

children. Am J Trop Med Hyg 64(3-4):207–213.53. Avril M, Brazier AJ, Melcher M, Sampath S, Smith JD (2013) DC8 and DC13 var genes

associated with severe malaria bind avidly to diverse endothelial cells. PLoS Pathog9(6):e1003430.

54. Stevenson L, et al. (2015) α 2-Macroglobulin can crosslink multiple Plasmodium fal-ciparum erythrocyte membrane protein 1 (PfEMP1) molecules and may facilitateadhesion of parasitized erythrocytes. PLoS Pathog 11(7):e1005022.

55. Jeppesen A, et al. (2015) Multiple Plasmodium falciparum erythrocyte membraneprotein 1 variants per genome can bind IgM via its Fc fragment Fcμ. Infect Immun83(10):3972–3981.

56. Niang M, et al. (2014) STEVOR is a Plasmodium falciparum erythrocyte binding pro-tein that mediates merozoite invasion and rosetting. Cell Host Microbe 16(1):81–93.

57. Goel S, et al. (2015) RIFINs are adhesins implicated in severe Plasmodium falciparummalaria. Nat Med 21(4):314–317.

58. Avril M, et al. (2012) A restricted subset of var genes mediates adherence of Plas-modium falciparum-infected erythrocytes to brain endothelial cells. Proc Natl Acad SciUSA 109(26):E1782–E1790.

59. Grubinger T, Zeileis A, Pfeiffer K-P (2014) evtree: Evolutionary learning of globallyoptimal classification and regression trees in R. Journal of Statistical Software, 61(1):1–29.

60. Wyner AD (1978) A definition of conditional mutual information for arbitrary en-sembles. Inf Control 38(1):51–59.

61. Fisher R (1921) On the «probable error» of a coefficient of correlation deduced from asmall sample. Metron 1(Pt 4):1–32.

62. Bengtsson A, et al. (2013) A novel domain cassette identifies Plasmodium falciparumPfEMP1 proteins binding ICAM-1 and is a target of cross-reactive, adhesion-inhibitoryantibodies. J Immunol 190(1):240–249.

63. Howell DPG, et al. (2008) Mapping a common interaction site used by Plasmodiumfalciparum Duffy binding-like domains to bind diverse host receptors. Mol Microbiol67(1):78–87.

64. Oleinikov AV, et al. (2009) High throughput functional assays of the variant antigenPfEMP1 reveal a single domain in the 3D7 Plasmodium falciparum genome that bindsICAM1 with high affinity and is targeted by naturally acquired neutralizingantibodies. PLoS Pathog 5(4):e1000386.

Bernabeu et al. PNAS | Published online May 16, 2016 | E3279

MICRO

BIOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

June

4, 2

021