1. Immunology and Inflammation
Download icon

Developing a multivariate prediction model of antibody features associated with protection of malaria-infected pregnant women from placental malaria

  1. Elizabeth H Aitken
  2. Timon Damelang
  3. Amaya Ortega-Pajares
  4. Agersew Alemu
  5. Wina Hasang
  6. Saber Dini
  7. Holger W Unger
  8. Maria Ome-Kaius
  9. Morten A Nielsen
  10. Ali Salanti
  11. Joe Smith
  12. Stephen Kent
  13. P Mark Hogarth
  14. Bruce D Wines
  15. Julie A Simpson
  16. Amy W Chung
  17. Stephen J Rogerson  Is a corresponding author
  1. Department of Medicine, University of Melbourne, the Doherty Institute, Australia
  2. Department of Microbiology and Immunology, University of Melbourne, the Doherty Institute, Australia
  3. Centre for Epidemiology and Biostatistics, Melbourne School of Population and Global Health, University of Melbourne, Australia
  4. Department of Obstetrics and Gynaecology, Royal Darwin Hospital, Australia
  5. Menzies School of Health Research, Australia
  6. Walter and Eliza Hall Institute of Medical Research, Australia
  7. Centre for Medical Parasitology, Department of Microbiology and immunology, University of Copenhagen, Denmark
  8. Department of Infectious Disease, Copenhagen University Hospital, Denmark
  9. Seattle Children’s Research Institute, United States
  10. Department of Pediatrics, University of Washington, United States
  11. Immune Therapies Group, Centre for Biomedical Research, Burnet Institute, Australia
  12. Department of Clinical Pathology, University of Melbourne, Australia
  13. Department of Immunology and Pathology, Monash University, Australia
Research Article
  • Cited 0
  • Views 714
  • Annotations
Cite this article as: eLife 2021;10:e65776 doi: 10.7554/eLife.65776

Abstract

Background:

Plasmodium falciparum causes placental malaria, which results in adverse outcomes for mother and child. P. falciparum-infected erythrocytes that express the parasite protein VAR2CSA on their surface can bind to placental chondroitin sulfate A. It has been hypothesized that naturally acquired antibodies towards VAR2CSA protect against placental infection, but it has proven difficult to identify robust antibody correlates of protection from disease. The objective of this study was to develop a prediction model using antibody features that could identify women protected from placental malaria.

Methods:

We used a systems serology approach with elastic net-regularized logistic regression, partial least squares discriminant analysis, and a case-control study design to identify naturally acquired antibody features mid-pregnancy that were associated with protection from placental malaria at delivery in a cohort of 77 pregnant women from Madang, Papua New Guinea.

Results:

The machine learning techniques selected 6 out of 169 measured antibody features towards VAR2CSA that could predict (with 86% accuracy) whether a woman would subsequently have active placental malaria infection at delivery. Selected features included previously described associations with inhibition of placental binding and/or opsonic phagocytosis of infected erythrocytes, and network analysis indicated that there are not one but multiple pathways to protection from placental malaria.

Conclusions:

We have identified candidate antibody features that could accurately identify malaria-infected women as protected from placental infection. It is likely that there are multiple pathways to protection against placental malaria.

Funding:

This study was supported by the National Health and Medical Research Council (Nos. APP1143946, GNT1145303, APP1092789, APP1140509, and APP1104975).

Introduction

The burden of Plasmodium falciparum malaria is greatest in young children and pregnant women (World Health Organisation, 2019). Among women with lifelong malaria exposure, pregnant women have increased susceptibility to malaria, and complications include maternal mortality, severe maternal anemia, miscarriage and stillbirth, and impaired fetal growth or prematurity leading to low birth weight delivery (Steketee et al., 2001; Guyatt and Snow, 2001; Desai et al., 2007; Moore et al., 2017a; Moore et al., 2017b). The prevalence, intensity, and consequences of infection decline with increasing gravidity (Guyatt and Snow, 2001).

The susceptibility of pregnant women to P. falciparum malaria is in part due to the ability of infected erythrocytes (IEs) to sequester in the maternal blood spaces of the placenta (Rogerson et al., 2007), where IEs that express VAR2CSA adhere to chondroitin sulfate A (CSA), a glycosaminoglycan chain on syndecan-1 expressed by the placental syncytiotrophoblast (Salanti et al., 2004; Ayres Pereira et al., 2016). VAR2CSA is a unique 350 kDa protein that is a member of the P. falciparum erythrocyte membrane protein 1 (PfEMP1) family of variant surface antigens (VSA). PfEMP1s are encoded by var genes and are composed of multiple Duffy binding-like (DBL) domains and cysteine-rich interdomain regions (CIDR). VAR2CSA has six DBL domains, with CSA adhesion being strongly associated with the DBL2 domain, which is present in the N-terminal region of the protein (Doritchamou et al., 2013). Two constructs based on this region have entered clinical trials as vaccine candidates to prevent placental malaria (Mordmüller et al., 2019; Sirima et al., 2020).

Antibodies to VAR2CSA, in its native form on IE or expressed as recombinant protein, develop in a parity-dependent manner, block adhesion to CSA, and opsonize IE for phagocytosis by monocytic cells (Fried et al., 1998; Ricke et al., 2000; Salanti et al., 2004; Keen et al., 2007). Some studies report associations between measures of single antibody features and protection from malaria complications such as placental infection, impaired fetal growth, prematurity, or anemia (Duffy and Fried, 2003; Staalsoe et al., 2004; Feng et al., 2009). However, identifying targets and features of protective antibody responses to malaria is challenging because exposure to infection is accompanied by the production of antibody to many parasite antigens, only a fraction of which contributes to protective immunity (Crompton et al., 2010). The large size and multidomain structure of VAR2CSA adds complexity to the identification of targets and features of protective immune response. In a recent systematic review, antibody responses to VAR2CSA were often correlated with exposure to placental malaria, rather than protection from infection (Cutts et al., 2020), although some studies indicate an association between antibody to IE and protection from low birth weight in subsets of pregnant women (Duffy and Fried, 2003; Staalsoe et al., 2004; Feng et al., 2009).

To identify antibody responses that protect women from placental malaria, an alternative approach that examines multiple antibody features and controls for exposure may be needed. Placental malaria mostly occurs when antigenic variation (Roberts et al., 1992) leads to a parasite that has infected a pregnant woman, expressing VAR2CSA. These parasites can flourish by sequestering in the placenta in women who lack VAR2CSA-specific immunity (Hviid, 2004). By contrast, some pregnant women with peripheral blood infection remain free of placental malaria, and we postulate that these women have antibody responses that specifically protect them against VAR2CSA and placental malaria.

Antibody measures towards the IE have traditionally focused on the quantity of bound antibodies and their ability to inhibit IE binding to CSA (Cutts et al., 2020), but there is increasing interest in how antibodies that recognize IE engage with innate immune cells and activate complement (Aitken et al., 2020), which are both determined by biophysical features of the Fab and Fc regions of antibody (reviewed in Arnold and Chung, 2018). In recent years, detailed functional and biophysical characterization of antibody responses has led to the identification of specific antibody determinants that correlate with vaccine-induced protection from HIV (Chung et al., 2015), control of latent tuberculosis infection (Lu et al., 2016), transplacental transfer of antibody (Martinez et al., 2019; Jennewein et al., 2019), and correlates of vaccine-induced protection in human malaria challenge models (Suscovich et al., 2020). This approach, which has been termed systems serology, is well suited to placental malaria because there is a clear relationship between a single antigen (VAR2CSA) and a specific pathology (the sequestration of parasites in the placenta, reviewed in Ataíde et al., 2014).

The objective of this study was develop a prognostic model using antibody features to identify women protected from placental malaria. In this prospective study, we identified features, functions, and targets of naturally acquired antibody to VAR2CSA that contribute to protection against placental malaria. Pregnant Papua New Guinean women who were participating in a trial of intermittent preventive treatment against malaria (Unger et al., 2015) were categorized at delivery as currently uninfected, having placental malaria, or having a non-placental infection. Using plasma samples collected at enrollment in mid-pregnancy, we performed extensive profiling of antibody responses to VAR2CSA recombinant proteins and IEs expressing VAR2CSA. Machine learning techniques revealed that 6 of 169 antibody features measured mid-pregnancy were able to discriminate between women who had placental malaria and non-placental infection at delivery with high accuracy. Five of the six leading antibody features we found were related to inhibition of placental sequestration and/or opsonization for phagocytic clearance.

Materials and methods

Experimental model and subject details

Human subjects

Request a detailed protocol

Study participants were recruited between November 2009 and August 2012 as part of a randomized controlled trial of Intermittent Preventive Treatment in Pregnancy (IPTp) (Unger et al., 2015) (ClinicalTrials.gov NCT01136850), in which pregnant women received either three courses of sulfadoxine pyrimethamine (SP) and azithromycin or one course of SP and chloroquine. Women were recruited at 14–26 gestation weeks and followed up until delivery. All women presenting for their first antenatal visit at one of the nine participating health centers in Madang Papua New Guinea were invited to participate in the original cohort study. Exclusion criteria included a gestation of >26 weeks, hemoglobin <6 g/dl with symptomatic anemia, previous serious adverse reactions to the IPTp study medications, permanent disability or chronic medical conditions, known multiple pregnancy, age <16 years, or known unavailability to follow-up Unger et al., 2015. Demographic data collected at enrollment included maternal gravidity, age, residence, and bed net use. Peripheral blood collected at 14–26 gestation weeks, before initiation of IPTp, was used in the antibody assays. Peripheral blood collected at delivery was used to prepare blood smears for malaria microscopy and to extract DNA for qPCR for malaria parasites (Unger et al., 2015). Placental biopsies were formalin fixed and paraffin embedded. Giemsa stained sections of placental biopsies were examined for malaria infection (Lufele et al., 2017). Women were selected for inclusion in this case-control study based on the presence of P. falciparum IEs in peripheral blood and/or placenta at delivery. Groups included women with no evidence of infection in peripheral blood by PCR or light microscopy nor on examination of placental histology at delivery (n = 50); women with active placental malaria characterized by detection of P. falciparum IEs by placental histology (n = 50); and women with P. falciparum infection by light microscopy and/or PCR in the peripheral blood but with no IEs detected by placental histology examination (n = 27). Sample size was based on published work that used a similar approach to investigate antibody responses to infectious diseases (Lu et al., 2016); in addition, sample size for the non-placental infection group was also limited by sample availability as few women met the criteria for inclusion. All clinical data were collected, and grouping of women by outcome was completed prior to measurements of antibody features included in the model. The three groups were frequency matched for primigravidity, IPTp regime receipt, bed net use, rural residency, and age. See Table 1 for details on clinical characteristics of participants.

Table 1
Clinical characteristics of the three groups of pregnant women at the time of antibody feature measurement at enrollment (14–26 weeks’ gestation) and also at delivery.
Non-infected at deliveryPlacental malaria at deliveryNon-placental infection at delivery
N = 50N = 50N = 27
Enrollment
Mean age (years), SD24.55.324.05.023.14.4
Residence, N (%)
Rural37(74.0)38(76.0)18(66.7)
Non-rural13(26.0)12(24.0)9(33.3)
Ethnicity, N (%)
Sepik6(12.0)11(22.0)3(11.1)
Madang/Morobe39(78.0)30(60.0)22(81.5)
Highlander3(6.0)5(10.0)1(3.7)
Other2(4.0)4(8.0)1(3.7)
Formal schooling, N (%)46(92.0)46(92.0)25(92.6)
Smoking, N (%)9(18.0)11(22.0)6(22.2)
Betel nut user, N (%)†41(82.0)41(82.0)24(88.9)
Alcohol, N (%)2(4.0)2(4.0)2(7.4)
Gravidity, N (%)
Primigravidae26(52.0)29(58.0)14(51.9)
Secundigravidae8(16.0)7(14.0)8(29.6)
Multigravidae16(32.0)14(28.0)5(18.5)
IPTp regime, N (%)
SPCQ27(54.0)30(60.0)18(66.7)
SPAZ23(46.0)20(40.0)9(33.3)
Mean gestational age (days), SD145.931.4147.531.3152.219.8
Mean maternal weight (kg), SD*54.713.153.58.254.17.4
Mean maternal height (cm), SD154.35.9154.66.9154.46.0
Bed net use, N (%)34(68.0)40(80.0)21(77.8)
Hb (g/dL), mean SD9.71.29.32.09.41.2
Light microscopy positive for Pf, N (%)2(4.0)4(8.0)4(14.8)
PCR positive for Pf, N (%)4(8.0)5(10.0)5(18.5)
Delivery
Placenta
light microscopy positive for Pf, N (%)
0(0)50(100)0(0)
Peripheral blood light microscopy positive for Pf, N (%)0(0)10(20)13(48.2)
Peripheral blood PCR positive for Pf, N (%)0(0)13(26)21(77.8)
Placental blood PCR positive for Pf, N (%)§1(2.0)10(20)10(37.04)
Birthweight (g), SD306254628275012840416
Gestation at delivery (days), SD278182791628013
Mean Hb (g/dL), SD10.129.51.910.21.4
  1. * One participant with missing data on betel nut use in placental malaria.

    One participant with missing data on weight in the non-infected group.

  2. Missing Hb data, five in non-infected group, four in placental malaria group, and two in the non-placental infection group.

    § Missing placental PCR data in five non-infected, eight placental malaria, and four non-placental infection women.

  3. One participant with missing data on gestation length at delivery in the non-infected group.

    SD: standard deviation; Hb: hemoglobin; PCR: polymerase chain reaction; IPTp: intermittent preventive treatment in pregnancy; SPAZ: sulfadoxine pyrimethamine-azithromycin; SPCQ: sulfadoxine pyrimethamine-chloroquine.

Parasite cell lines

Request a detailed protocol

P. falciparum IE of the parasite lines CS2 and 3D7 were cultured as previously described (Chandrasiri et al., 2014). Cultures were synchronized as needed by sorbitol lysis (Lambros and Vanderberg, 1979) and IE were regularly selected for expression of knobs by gelatin flotation (Goodyer et al., 1994). Cell cultures were mycoplasma negative (tested for mycoplasma using the MycoAlert kit [Lonza, Mount Waverley, Australia] as per manufacturer’s instructions). For the binding inhibition assays, the parent lines of CS2 and 3D7 (FCR3 and NF54, respectively) were used, cultured, and selected for CSA adhesion as described previously (Nielsen and Salanti, 2015).

Monocyte cell line THP-1

Request a detailed protocol

THP-1 cells were mycoplasma negative and cultured as previously described (Ataíde et al., 2010).

Primary leukocytes

Request a detailed protocol

Neutrophils were isolated from fresh venous blood collected in lithium heparin vacutainers (BD, Scoresby, Australia) using the EasySep Direct Human Neutrophil Isolation Kit (STEMCELL Technologies, Tullamarine, Australia) as per manufacturer’s instructions. Neutrophil purity was assessed by cell morphology using light microscopy of Giemsa stained smears of the isolated cells, and viability was assessed using trypan blue exclusion. Monocytes were isolated from both fresh venous blood, collected in lithium heparin vacutainers (BD), as well as from buffy coats supplied from the Australian Red Cross Blood Service. Monocytes were isolated by negative selection using the RosetteSep Human Monocyte Enrichment Cocktail (STEMCELL Technologies) as per manufacturer’s instructions. Monocyte purity was assessed by staining (anti-CD14 antibody, BioLegend, San Diego, CA) and measuring CD14+ cells by flow cytometry. Monocytes were either frozen in fetal bovine serum (FBS) in 20% dimethyl sulfoxide (DMSO) in liquid nitrogen for later use or used immediately after isolation. Natural killer (NK) cells were isolated from fresh venous blood collected in sodium heparin vacutainers (BD). NK cells were isolated by negative selection using the RosetteSep Human NK Enrichment Cocktail (STEMCELL Technologies) as per manufacturer’s instructions.

Method details

Samples and controls

Request a detailed protocol

Plasma samples of study participants were obtained from venous blood collected into sodium heparin vacutainers upon enrollment into the trial at 14–26 weeks’ gestation and plasma was separated by centrifugation and stored at −80°C until used. To generate a positive control, plasma samples from pregnant women in Malawi with high levels of IgG that recognizes IEs of the P. falciparum line CS2 (a VAR2CSA-expressing and CSA-binding line; Elliott et al., 2005b) were pooled. Negative controls were sera from individual Melbourne donors obtained from the Australian Red Cross Blood Service or plasma from Melbourne donors collected in lithium heparin vacutainers (BD). All controls were kept at −80°C until use. Antibody features were measured after grouping of women based on outcome. Researchers measuring antibody features were blinded to outcome data of individual samples; this was done by allocating all samples a number (which was not associated with clinical outcomes and which was different from the original cohort study number), samples were decoded after acquisition of the antibody feature was complete. For some experiments, IgG was purified from plasma and serum using Melon Gel purification kits (Thermo Fisher Scientific, Scoresby, Australia) as per manufacturer’s instructions. Purified IgG was quantified before use with a human IgG ELISA development kit (MabTech, Preston, Australia) as per manufacturer’s instructions.

VAR2CSA DBL recombinant antigens

Request a detailed protocol

Twelve recombinant proteins consisting of subunits or DBL domains of VAR2CSA were used in experiments (see Supplementary file 1 for a table detailing proteins used). For some experiments, proteins were biotinylated using EZ-link Sulfo NHS-LC-Biotin kit (Thermo Fisher Scientific) as per manufacturer’s instructions.

Infected erythrocytes

Request a detailed protocol

The CSA-binding phenotypes of CS2 and 3D7 IEs were monitored by measuring binding of trophozoite stage IEs in static binding assays to recombinant human CD36 (R&D systems, Noble Park, Australia) and CSA (Sigma-Aldrich, Macquarie Park, Australia) as previously described (Yosaatmadja et al., 2008). 3D7 was selected for a CSA-binding phenotype by flow cytometry. Purified trophozoite stage 3D7 IEs were incubated with fluorescein isothiocyanate (FITC)-labeled CSA 100 μg/mL (Creative PEGWorks, Durham, NC) for 30 min at 37°C. Purified IE incubated with CSA-FITC and 3D7 without CSA-FITC were used to set the gates. IEs were then gated based on forward and side scatter and sorted based on FITC. For some assays, trophozoite stage IE were purified by Percoll gradient and stained with dihydroethidium (DHE) 25 μg/mL (Sigma-Aldrich). For the binding inhibition assays, FCR3 and NF54 IE were selected for a CSA-binding phenotype by panning on BeWo cells as previously described (Nielsen and Salanti, 2015).

Antibody features to VAR2CSA by multiplex

Request a detailed protocol

Multiplex assays were used to detect plasma levels of antigen-specific antibodies (Abs) towards VAR2CSA DBL domains. Features assessed included total IgG, IgG1, IgG2, IgG3, IgG4, IgA1, IgA2, IgM as well as antibody engagement with complement (C1q) and FcγRs (FcγRI, FcγRIIa, FcγRIIIa, and FcγRIIIb). To conduct the multiplex assays, DBL domains (Supplementary file 1) of the VAR2CSA protein were coupled to Bio-Plex magnetic carboxylated microspheres (Bio-Rad) as per manufacturer’s instructions, blocked with PBS plus 0.1% bovine serum albumin (BSA), then stored in PBS-0.05% sodium azide at −80°C.

On the day of the assay, the DBL-coupled microspheres were resuspended at 1 × 104 beads/mL in 1:100 dilution of plasma in PBS-1% BSA and 50 µL were aliquoted into each well of a 96-well round bottom plate (Greiner Bio-One, Kremsmünster, Austria). Plates were incubated on a shaker overnight at 4°C, then centrifuged and washed with PBS-0.1% Tween20 using a magnetic plate-washer (Bio-Plex Pro wash station). The anti-human Ab detectors for IgG, IgG1, IgG2, IgG3, IgG4, IgA1, IgA2, or IgM conjugated with phycoerythrin (PE) (SouthernBiotech, Birmingham, AL) were added, and the mixture was incubated for 2 hr on a plate shaker. C1q (MP Biomedicals), FcγRI (R&D Systems), and FcγRIIIb (R&D Systems) were biotinylated using EZ-Link Sulfo-NHS-LC-Biotin (Thermo Fisher Scientific) as per manufacturer’s instructions and then labeled streptavidin-PE (SouthernBiotech), before being added to samples. For FcγRIIa and FcγRIIIa, the beads/plasma samples mixture was incubated with biotinylated recombinant dimeric FcγRIIa and FcγRIIIa ectodomains (Wines et al., 2016) for 1 hr to allow binding of FcγRs to antigen-specific IgG, washed, and incubated with streptavidin-PE for 1 hr. After washing with PBS and resuspending in xMAP sheath fluid (Life Technologies, Scoresby, Australia), the plates were read on a Bio-Plex MAGPIX multiplex reader (Bio-Rad, Gladesville, Australia) and analyzed on Bio-Plex Manager software (Bio-Rad). A single measurement of each sample for each antibody feature was done. Data were excluded if the number of beads acquired was too low. Readouts were expressed as the raw median fluorescence index (MFI).

Detection of antibody binding to the IE

Request a detailed protocol

Binding of IgG and IgG subclass antibodies to the surface of CS2 and CSA-selected 3D7 IEs was measured as previously described; all samples were run in duplicate (Aitken et al., 2010; Elliott et al., 2005b).

ADCP of VAR2CSA DBL-coated beads by THP-1

Request a detailed protocol

Antibody-dependent cellular phagocytosis (ADCP) of beads coated with VAR2CSA DBL domains and opsonized with plasma or purified IgG was assessed using THP-1 cells. Individual DBL domains (Supplementary file 1) were biotinylated and incubated with neutravidin 1 μm beads at a ratio of 1 μL of beads (stock 1% v/v) to 3 µL of 1 mg/mL of the respective DBL domain. Beads and proteins were incubated overnight at 4°C on a roller. After incubation, beads were washed twice in PBS-0.1% BSA and then resuspended at 0.01% v/v in PBS-0.1% BSA with 0.02% sodium azide. For the phagocytosis assay, a sterile 96-well U bottom plate (Corning, Mulgrave, Australia) was blocked with high salt PBS buffer (PBS plus 0.5 M NaCl with 1% BSA 0.05% Tween20) for 30 min at room temperature. Then 20 μL of plasma diluted 1:800 in high salt PBS buffer and 10 µL of beads was added and incubated for 1 hr at room temperature. The plate was spun at 4700 × g for 3 min, supernatant was replaced with 30 μL of leukocyte media (RPMI-1640 with 10% FBS and Penicillin-Streptomycin-Glutamine [all from Gibco, Scoresby, Australia]), then 5 × 104 THP-1 cells in 50 µL of leukocyte media were added and incubated for 40 min at 37°C in 5% CO2. After incubation, cells were washed and resuspended in cold PBS-4% paraformaldehyde (PFA). Phagocytosis was measured by gating on THP-1 using forward scatter (FSC) and side scatter (SSC) parameters. Phagocytic score was calculated as (% of THP-1 cells positive for beads × geometric MFI of the bead-positive THP-1 cells). All samples were run in duplicate.

ADCP of IE by THP-1

Request a detailed protocol

ADCP of IE opsonized with plasma by THP-1 cells was assessed as previously described (Ataíde et al., 2010). Phagocytosis of IE opsonized with purified IgG was measured as previously described (Ataíde et al., 2010) with the modification that IE were opsonized with purified IgG at a concentration of equivalent to a 1:20 dilution of plasma. All samples were run in duplicate.

ADCP of VAR2CSA DBL-coated beads by monocytes

Request a detailed protocol

Monocyte phagocytosis of beads coated with VAR2CSA DBL domains and opsonized with plasma was assessed. Beads were coated with DBL domains as described above, except that neutravidin 1 μm beads were incubated with 6 μL of 1 mg/mL of each DBL domain. To measure ADCP, a sterile 96-well U bottom plate (Corning) was blocked with salt PBS buffer (PBS plus 0.25 M NaCl and 1% BSA) for 30 min at room temperature. 40 μL of plasma (diluted 1:300 in salt PBS buffer) was added to each well along with 20 μL of beads (0.01% v/v) and incubated for 1 hr at room temperature. The beads were then washed and resuspended in 60 µL of leukocyte media. In a separate 96-well U bottom plate, 5 × 104 monocytes in leukocyte medium were plated out per well and rested at 37°C for 1 hr prior to phagocytosis. 30 μL of the opsonized beads were then added to each well with monocytes. The monocytes and beads were co-incubated for 50 min at 37°C with 5% CO2. After incubation, the cells were washed and resuspended in cold PBS-2% PFA. Phagocytosis was measured by gating monocytes on the FSC and SSC. All samples were run in duplicate, and assays were repeated with three different primary cell donors. For each donor, a phagocytic score was calculated (% of monocytes cells positive for beads × geometric MFI of the bead-positive monocytes) and then the average of the phagocytic scores for the three donors was taken.

ADCP of IE by monocytes

Request a detailed protocol

ADCP of opsonized IE by primary monocytes was assessed. A sterile 96-well U bottom plate (Corning) was blocked with PBS-1% BSA. In each well, 30 μL of diluted plasma (1:10) along with 3.3 μL of DHE-stained IE in leukocyte media (1.65 × 107 IE/mL) was mixed and left for 1 hr at room temperature, before being washed and resuspended in 50 μL RPMI-1640 with 25 mM HEPES, 0.5% Albumax (w/v), 5% heat-inactivated human serum (HIHS). In a separate 96-well U bottom plate, 5 × 104 monocytes in leukocyte medium were plated out per well and rested at 37°C for 2 hr prior to phagocytosis. Then 25 μL of the opsonized IE were added to each well with monocytes. The monocytes and IE were co-incubated for 40 min at 37°C with 5% CO2. After incubation, the cells were washed and resuspended in cold PBS-2% PFA. Phagocytosis was measured by gating on monocytes using the FSC and SSC parameters. All samples were run in duplicate, and assays were repeated with three different primary cell donors. For each donor, a phagocytic score was calculated (% of monocytes cells positive for DHE × MFI of the DHE-positive monocytes) and then the average of the phagocytic scores for the three donors was used.

ADNP of VAR2CSA DBL-coated beads

Request a detailed protocol

Antibody-dependent neutrophil phagocytosis (ADNP) of beads coated with VAR2CSA DBL domains and opsonized with plasma was assessed. Beads were coated with DBL as described for primary monocytes. Neutravidin 1 μm beads were incubated with 6 μL of 1 mg/mL of the respective DBL domain. To quantitate ADNP, 12 μL of plasma (diluted 1:100 in high salt PBS) was added to each well of a sterile 96-well U bottom plate (Corning) along with 6 μL of beads (0.01% v/v) and incubated for 1 hr at room temperature. The beads were then washed and resuspended in 30 μL of leukocyte medium. In a separate 96-well U bottom plate, 3 × 104 neutrophils in leukocyte medium were plated out per well, 15 μL of the opsonized beads were then added to each well. The plate was incubated for 1 hr at 37°C in 5% CO2, then the cells were washed and resuspended in cold PBS-2% PFA. Phagocytosis was measured by gating neutrophils on the FSC and SSC. All samples were run in duplicate, and assays were repeated with three different primary cell donors. For each donor, a phagocytic score was calculated (% of neutrophil cells positive for beads × geometric MFI of the bead-positive neutrophils) and then the average of the phagocytic scores for the three donors was used.

ADNP of IE

Request a detailed protocol

ADNP of IE opsonized with plasma was assessed. 30 µL of diluted plasma (1:10) and 3.3 μL of DHE-stained IE in neutrophil medium (1.65 × 107 IE/mL) were mixed in a sterile 96-well U bottom plate (Corning), incubated for 1 hr at room temperature, before being washed and resuspended in 50 μL of leukocyte medium. In a separate 96-well U bottom plate, 2.5 × 104 neutrophils in 50 μL of leukocyte medium were plated out per well, 25 μL of the opsonized IE were then added to each well, and the plate was incubated for 1 hr at 37°C with 5% CO2. After incubation, the cells were washed and resuspended in cold PBS-2% PFA. Phagocytosis was measured by gating neutrophils on the FSC and SSC. All samples were run in duplicate, and assays were repeated with three different primary cell donors. For each donor, a phagocytic score was calculated (% of neutrophil cells positive for beads × geometric MFI of the bead-positive neutrophils) and then the average of the phagocytic scores for the three donors was used.

ADRB using VAR2CSA DBL domains

Request a detailed protocol

Antibody-dependent respiratory burst (ADRB) was measured using an assay of reactive oxygen species (ROS) production. Firstly, 25 μL of 4 μg/mL of individual DBL domains (Supplementary file 1) in PBS were coated on 96-well, white, flat bottom plates (NUNC MaxiSorp flat bottom; Thermo Fisher Scientific) and left overnight at 4°C. The liquid was removed, and plates were then washed with PBS and blocked with PBS-0.1% BSA for 1 hr at room temperature. Then 25 μL of each diluted sample (1:10 in PBS) was added to a single well and incubated at room temperature for 2 hr, the plates were then washed twice with PBS. Neutrophils were resuspended in Hanks’ buffered saline solution at 2 × 106/mL and 20 μL of neutrophils were added to each well followed by 20 μL luminol horseradish peroxidase (HRP) solution (PBS with 33 ng/mL of HRP and 4 mM luminol; all from Sigma-Aldrich). The plate was spun briefly to settle the contents and read immediately on a FLUOstar plate reader. Luminescence was measured in each well for 1 s every 2 min for 1 hr and was calculated as the average luminescence 5 min either side of the peak of the curve. Luminescence score was standardized to the no serum controls that were run on every plate. All assays were repeated with three different primary cell donors, and the average luminescence score from the three donors was used.

ADRB to IE

Request a detailed protocol

To measure ADRB in response to IEs, 96-well U bottom plates (Corning) were coated with PBS-1% FBS for 1 hr. 1 μL of plasma from each donor was aliquoted into a single well and along with 20 μL of purified IEs at 2 × 107 cells/mL in PBS-1% FBS were mixed and incubated for 1 hr at room temperature. The IE were then washed and resuspended in 20 μL of PBS. 10 μL of neutrophils in Hanks’ buffered saline solution at 2 × 106/mL, 10 μL of luminol HRP solution, and 5 μL of opsonized IEs were added to each well of a 384-well white, flat bottom plate (NUNC MaxiSorp flat bottom; Thermo Fisher Scientific). The plate was spun briefly to settle the contents and then read immediately on a FLUOstar plate reader (BMG LABTECH, Mornington, Australia). Luminescence was measured as described above. Luminescence score was calculated as the % of the positive plasma control, which was run on every plate. All assays were repeated with three different primary cell donors, and the average luminescence score from the three donors was used.

ADCC using VAR2CSA DBL domains

Request a detailed protocol

Antibody-dependent cellular cytotoxicity (ADCC) assays using human NK cells were modified for use with DBL antigens (Supplementary file 1; Jegaskanda et al., 2013; Lu et al., 2016). NUNC MaxiSorp flat bottom plates (Thermo Fisher Scientific) were coated with DBL proteins (200 ng/well) at 4°C for 12 hr. After washing with PBS, the plate was blocked with PBS-1% BSA for 1 hr. Purified IgG (0.5 mg/mL) was added to each well and incubated at 37°C for 2 hr. NK cells (0.25 × 106 cells/mL), anti-CD107a-allophycocyanin (APC)-H7 (BD) brefeldin A (10 mg/mL; Sigma-Aldrich), and GolgiStop (BD) were added to each well, and the plates were incubated for 5 hr at 37°C. NK cells were then stained for surface markers using anti-CD16-Brilliant violet (BV)-605 (BD), anti-CD56-Brilliant ultraviolet (BD), and anti-CD3-peridinin-chlorophyll-protein (PerCP; BD), and then stained intracellularly with anti-IFNγ-PE (BD) and anti-TNFα-BV-785 (BD) after fixation (10% PFA) and Perm B solutions (Thermo Fisher Scientific). NK cells were analyzed via flow cytometry and defined as CD3- and CD56highCD16± and CD56lowCD16high. Boolean gates (FlowJo) were used to include all NK cells that expressed degranulation marker CD107a or produced cytokines, IFNγ and TNFα. Levels of CD107a, IFNγ, and TNFα expression or production were calculated as % of NK cells positive for APC-H7, PE, and BV-785, respectively. All assays were repeated with three different primary cell donors, and the average score from the three donors was used.

CSA binding inhibition antibody to IE

Request a detailed protocol

CSA binding inhibition was measured as previously described (Nielsen and Salanti, 2015). The assays were all run twice and averages of the two runs were used.

IgG levels to non-pregnancy-specific antigens

Request a detailed protocol

IgG levels towards schizont extract and MSP-1 were measured by ELISA as previously described (Barua et al., 2019) using goat anti-human IgG biotinylated (Mabtech 3820-4-250).

Quantification and statistical analysis

Processing of data

Request a detailed protocol

Prior to analysis, the data were processed. The right skewness of the distribution of the features was reduced by log-transformation (log(x + 1)). Four antibody features that had negative values were right-shifted to have their minimum at zero prior to log-transformation. Next, the distributions of the features were centered and scaled to have zero mean and unit standard deviation.

Demographic and clinical characteristics of the pregnant women were described using the mean (SD) for continuous variables and frequency (%) for categorical variables.

Univariate analysis

Request a detailed protocol

Antibody levels for individual antibody features were compared between groups using the Welch’s t-test. The fold-change in the volcano plot was defined as the exponential of difference between the means of log-transformed and standardized data across the groups. Antibody features were correlated using Pearson pairwise correlation and correlation networks were plotted using the qgraph package (Epskamp et al., 2012). Associations between select antibody features and gravidity were investigated using linear regression.

Identification of key antibody features

Request a detailed protocol

Non-infected women were excluded from the multivariate analyses. Multivariate imputations by chained equations (Buuren and Groothuis-Oudshoorn, 2011) with predictive mean matching was used to impute any missing values (0.82% of observations). The imputation process was repeated five times, and the median of the imputed values across the five generated imputed datasets was finally used for each missing value. An elastic net-regularized logistic regression (ENLR) model was initially used to identify the subset of antibody features that best discriminate between the pregnant women with non-placental infection and women with placental infection. Elastic net (Zou and Hastie, 2005) allows to attain sparsity (keeping only a small subset of important variables), but unlike the least absolute shrinkage and selection operator (LASSO) penalization, groups of highly correlated important variables can as well be selected (a property of ridge regression). The parameter α, which alters the nature of penalization between ridge regression and LASSO, was set to 0.5 to achieve a balance between sparsity and group selection; the final selected antibody features did not alter when α was varied over [0, 1], as detailed below. We used a resampling approach in which the ENLR model was repeatedly fitted to subsets of data, as used by Gunn et al., 2018. This provided a means to take account of the uncertainty in the feature selection by ENLR and identify the features that are consistently selected when the model is fitted to different resampled data. The selection stability of an antibody feature was defined as the proportion of times that it was picked in the selected set of important features by elastic net when the model was repeatedly fitted to 50,000 resampled subsets of data (5000 repeats of 10-fold cross-validation). In each of the 5000 repetitions, the dataset was divided into 10-folds, of which 9-folds were used to perform an inner 10-fold cross-validation to find the best value of λ (shrinkage parameter) that maximizes the area under the receiver operating characteristic (AUROC) and in turn select the antibody features; using other performance measures such as binomial deviance gave the same ranking of the features according to selection frequencies. In each resampling iteration, the majority group (50 women with placental malaria) was downsampled to have an equal number of observations in both groups, thus avoiding overfitting the model to the majority group. The resampling process was repeated where α was also tuned over the set {0, 0.25, 0.5, 0.75, 1} (instead of only α=0.5) in addition to λ, and the results showed that the final top frequently selected variables do not change (see Figure 2—figure supplement 1).

Partial least squares discriminant analysis (PLSDA) (Barker and Rayens, 2003) was subsequently applied on the results of the ENLR to reassert the set of selected antibody features and find a minimal set of antibody features that accurately classifies the pregnant women; two components were selected for the PLSDA model (exception: one component was used where the model included only one feature). The antibody features were added to the PLSDA model one by one, from the highest to lowest selection frequency as estimated by the ENLR, and the minimal set of features beyond which the accuracy did not improve significantly were selected as the final set of key antibody features (Gunn et al., 2018); 500 repeats of 10-fold cross-validation were performed to compute the accuracies. The statistical significance of the results was assessed by comparing the performance of the model with that of two random permutation tests (null cases): (1) the PLSDA model was fitted to six randomly selected antibody features and the performance was computed for 500 repeats of 10-fold cross-validation resampling; (2) 100 datasets were generated by randomly permuting the group labels (PM and NPI) and the same analysis performed for the original dataset (i.e., building PLSDA models using the top six frequently selected antibody features found by resampling of elastic net) was repeated for each dataset.

All of the computations were performed in the R software (R Development Core Team, 2019). The resampling was performed using the caret package (Kuhn, 2008). The glmnet package (Zou and Hastie, 2005) was utilized for implementing the ENLR. The PLSDA method was implemented via the PLS package (Mevik and Wehrens, 2007), and some visualizations of the results of the PLSDA method were carried out using the mixOmics package (Rohart et al., 2017).

Results

Cohort description

We measured antibody features in plasma samples collected at enrollment (mean 21 weeks’ gestation) in 127 women and assessed the influence of antibody features on the presence of placental malaria at delivery. All women resided in malaria-exposed areas, and 17 were positive for P. falciparum at enrollment (by PCR and/or light microscopy). Women were selected based on P. falciparum infection status at delivery, using histology to define placental malaria (Bulmer et al., 1993). The cohort comprised 50 women who subsequently had no P. falciparum parasites detected on placental histology and were PCR negative for peripheral parasitemia at delivery (non-infected), 50 women who had P. falciparum parasites in the placenta by histology (placental malaria), and 27 women who had no P. falciparum parasites in the placenta by histology but who had peripheral blood parasitemia by PCR and/or light microscopy (non-placental infection) at delivery. Of the latter group, 10 were also PCR positive in placental blood, consistent with free circulation of early-stage parasites throughout the body. SP and azithromycin was more efficacious against placental malaria than SP-chloroquine (Unger et al., 2015), but antimalarial regimes were similarly distributed between groups. The three groups were frequency matched for primigravidity, IPTp regime receipt, bed net use, rural residency, and age, and all three groups had similar clinical characteristics at enrollment (Table 1).

Measurement of antibody features

We initially studied antibody to an array of recombinant VAR2CSA DBL proteins (see Figure 1—figure supplement 1 for a schematic of VAR2CSA). Thirteen features of antibody towards recombinant forms of each of the six DBL domains were measured by multiplex in the 127 women. In order to consider antigenic diversity (Benavente et al., 2018), four of the DBLs were also represented by an additional recombinant protein from a heterologous parasite strain (see Supplementary file 1 for details of recombinant proteins). Antibody features measured by multiplex included antibody isotypes (IgG, IgA, IgM) and subclasses (IgG1, IgG2, IgG3, IgG4, IgA1, IgA2), and the ability of antigen-specific antibody to bind Fcy receptors (FcyR) and complement protein C1q (which assesses the ability of antibodies to activate the classical complement cascade). Thus, using samples from the 127 women and measuring 13 antibody features directed at 10 DBL domains, 15,340 out of 16,510 possible data points were acquired (see  Figure 1—figure supplement 2 for a flow chart of data acquisition).

Levels of antibody features measured by multiplex were compared between non-infected and placental malaria, and between non-placental infection and placental malaria groups. Volcano plots summarizing the two comparisons of the multiplex data are given in Figure 1 (see Supplementary file 2 for individual univariate analyses). Notably, levels of antibody to DBL domains at enrollment were similar between non-infected women at delivery and those with placental malaria (Figure 1A), whereas multiple antibody features were higher in women with non-placental infection than in placental malaria, indicating potential roles in protection (Figure 1B). Antibody features that were higher in women with non-placental infection compared to those with placental malaria included DBL2-specific IgG3, IgA2, IgG2, and IgG4 (IgG3.DBL2(ID1-ID2a).FCR3, IgA2.DBL2.1010, IgG2.DBL2(ID1-ID2a).FCR3, and IgG4.DBL2.1010, respectively) as well as DBL3-specific antibodies that bind FcγRIIIb (FcγRIIIB.DBL3.FCR3) and DBL5-specific IgG2 and IgG4 (IgG2.DBL5.0466 and IgG4.DBL5.0466). Thus, we selected DBL2, DBL3, and DBL5 for the assessment of functional antibody features using cell-based assays in order to provide insights into the mechanistic contribution of these enhanced antibody populations. To investigate whether the higher antibody levels seen in the non-placental infection group were just correlates of exposure, we also measured IgG levels to non-pregnancy-specific P. falciparum antigens merozoite surface protein-1 (MSP-1) and schizont extract (which are associated with exposure Barua et al., 2019) by ELISA and compared the levels between groups. IgG levels to neither MSP-1 nor schizont extract were higher in women without placental malaria (Supplementary file 2), suggesting that differences in antibody to pregnancy specific antigens were not simply due to exposure. As antibody responses did not significantly differ between non-infected and placental malaria groups, further data analysis was restricted to those 77 women with non-placental infection or with placental malaria.

Figure 1 with 2 supplements see all
Individual antibody features to recombinant VAR2CSA Duffy binding-like (DBL) domain proteins measured by multiplex comparing women with placental malaria at delivery to (A) non-infected women and (B) women with non-placental infection.

Fold-change (log2 transformed), characterizing the magnitude of difference between the antibody levels of two groups (x-axis), is plotted against the -log10 p-value, characterizing the statistical significance of the difference (y-axis). The vertical dotted lines (log2(2) and log2(0.5)) mark a threshold for a twofold change, and the horizontal dotted lines (log10(0.05)) mark the statistical significance threshold (p≤0.05, Welch’s t-test). Antibody features beyond the thresholds are shown in blue and labeled.

We next measured responses to native protein on the surface of IEs of P. falciparum lines CS2 (which is isogenic with FCR3) and 3D7 that had been selected for high levels of CSA adhesion and studied whether antibody-opsonized IE and beads coated with recombinant DBL domains could activate leukocytes or promote phagocytosis to identify which antibody features associated with innate immune cells play important roles in parasite clearance or killing. We adapted existing cellular assays to measure ADCP of DBL protein-coated fluorescent beads and IE using THP-1 cells (Ackerman et al., 2011; Teo et al., 2015) and to measure NK cell activation following IgG binding to DBL proteins in plate-based assays (Jegaskanda et al., 2014; Lu et al., 2016). We also developed novel assays of ADCP by primary monocytes, and ADNP using primary human neutrophils, with IE and DBL-coated beads as their targets, as well as assays to measure ADRB with IEs or DBL proteins coated onto plates as their targets. Thus up to nine different cellular responses towards DBL2, DBL3, and DBL5 were measured. Additionally, six features of antibody to VSA of two strains of CSA-binding IE were measured, namely IgM, IgG, and IgG subclass binding to VSA, and the ability of antibodies to block CSA binding was assessed (see Supplementary file 3 for a table listing all antibody features measured). With 39 variables for 77 women, 2997 measurements (99.8% of possible 3003) were acquired. When added to the multiplex data, 169 variables with a total of 12,906 measurements (99.2% of 13,013) were used in subsequent analyses (see Figure 1—figure supplement 1 for the flow chart). None of the women had any data missing for the outcome variable, and 11 out of 77 women had data missing for one or more of the 169 predictor variables.

Identifying antibody features at enrollment that differentiate women with non-placental infection and placental malaria at delivery

We used a combination of ENLR and PLSDA to identify a set of antibody features that could best differentiate women with non-placental infection from those who had placental malaria at delivery (Gunn et al., 2018). Initially, the ENLR model was repeatedly fitted to randomly resampled subsets of the data and the selection frequency and the effect size (odds ratio) of the antibody features were calculated (Figure 2); only the top 20 antibody features with highest selection frequency are shown. The results show the antibody features that were consistently selected (Figure 2A) and have a high effect on discriminating the group of pregnant women (Figure 2B). The frequently selected antibody features were positively associated with non-placental infection (Figure 2B), and the effect of the antibody features on the odds of belonging to the non-placental infection group increased as the selection frequency increased.

Figure 2 with 1 supplement see all
Antibody features that are influential in distinguishing between malaria-infected women with and without placental infection identified by the elastic net-regularized logistic regression model.

Resampling (5000 repeats of 10-fold cross-validation) was used to obtain the selection frequencies and the odds ratios. (A) The top 20 antibody features are ranked in ascending order of selection frequency. (B) Boxplots of the estimated odds ratios, an odds ratio >1 indicates the antibody feature is positively associated with non-placental infection at delivery. Boxplots are median, IQR, whiskers (the lowest data point that falls between Q1 and 1.5 * Q1 IQR, the highest data point that falls between Q3 and 1.5 * Q3 IQR). IQR: interquartile range.

A minimal subset of antibody features that could accurately separate the two groups was then identified by adding the antibody features into a PLSDA model one by one (based on the selection frequency; Figure 2A). The mean AUROC and accuracy of classifying women with non-placental infection do not increase significantly after the inclusion of the six top features (Figure 3A). The performance of the model with these six selected features was then compared with the results of two random permutation tests (Figure 3B). The results show that the model with the six selected features performed significantly better than the PLSDA models with randomly selected features. The AUROC is also significantly higher than that of the permutation test where the group labels were randomly permuted (Figure 3B). The AUROC for the six features corresponds to a median accuracy of 0.86 (Figure 3—figure supplement 1) and means that the model with these six antibody features can accurately classify, on average, 86% of the women.

Figure 3 with 1 supplement see all
Selecting a minimal set of antibody features by partial least squares discriminant analysis (PLSDA).

(A) Performance of PLSDA at classifying women as having placental malaria (PM) or non-placental infection (NPI) when the features (ranked by selection frequency using the elastic net; Figure 2) are added one by one to the model from highest to lowest rank (500 repeats of 10-fold cross-validation were performed to estimate accuracy for each model). The three lines represent the accuracy of classification of all women in the cohort (black), those with NPI (blue), and those with PM (red). The vertical dashed line denotes the cutoff beyond which the accuracy does not change significantly by adding more antibody features (used for selecting the minimal set of features). (B) Comparing the performance of the elastic net-regularized logistic regression (ENLR) + PLSDA applied on the original data of the top six variables (rightmost boxplot) with two permutation tests (null cases): (1) the PLSDA model was fitted to six randomly selected antibody features and the performance was computed for 500 repeats of 10-fold cross-validation resampling; (2) 100 datasets were generated by randomly permuting the group labels (PM and NPI) and the same analysis performed for the original dataset (i.e., building PLSDA models using the top six frequently selected antibody features found by resampling of elastic net) was repeated for each dataset. (C) Segregation of women with NPI (blue) and PM (red) using the scores of the two components of the PLSDA model with data from the selected six antibody features. The background colors show the predicted classification of the women for all the possible score values in the depicted range. (D) Feature loadings on the components of the PLSDA of the six selected antibody features (see Figure 3—source data 1 for more details about the factor loadings and group prediction using the PLSDA). Boxplots are median, IQR, whiskers (the lowest data point that falls between Q1 and 1.5 * Q1 IQR, the highest data point that falls between Q3 and 1.5 * Q3 IQR). IQR: interquartile range; AUROC: area under the receiver operating characteristic curve.

Figure 3—source data 1

Partial least squares discriminant analysis (PLSDA) prediction model.

A PLSDA model with two components and using the six selected antibody features can be formed for prediction of the group of the pregnant women. The estimated loading factors of the model for the two components (shown in Figure 3) are listed here. The softmax technique was used to normalize the scores for each class (placental malaria and non-placental infection) that works as the probability of an observation belonging to a certain class (Kuhn, 2008). The predicted class is the one with the largest model prediction or, equivalently, the largest class probability.

https://cdn.elifesciences.org/articles/65776/elife-65776-fig3-data1-v3.docx

The segregation of the pregnant women and the loadings on the components of the PLSDA model for the six selected features are illustrated in Figure 3C, D. The loading components illustrate that the segregation of the observations by the PLSDA was largely driven by the top two variables identified by ENLR, DBL2-specific IgG3 (IgG3.DBL2(ID1-ID2a).FCR3), and ADCP of the CSA-binding IE CS2 by THP1 (THP1.Phago.CS2).

As immunity against placental malaria increases with gravidity (Fried et al., 1998), we also conducted a univariate linear regression for these six selected variables with gravidity (Supplementary file 4). Ability of antibodies to inhibit binding of IE to CSA (CSA.Binding.Inhibition.FCR3) was the only antibody feature with evidence of a positive association with gravidity in this cohort (p=0.046). In addition, as a small subset of women were infected with P. falciparum at enrollment (by light microscopy and/or PCR, n = 17, Table 1), to investigate the influence of these early infections on the identified features we conducted a subanalysis, excluding these infected women, comparing the six identified antibody features between non-placental and placental malaria cohorts (Supplementary file 4). Importantly, four of the six antibody features were significantly higher in the non-placental infection cohort (p≤0.013), and the remaining two features trended to be higher in the non-placental cohort (p≤0.064). This suggested that the associations seen with these six variables were not greatly affected by the presence of the individuals with infections at enrollment.

Closer inspection of the features using a correlation network of the 77 women provided information about the relationships between different antibody features (Figure 4). Distinct clustering of some antibody features was apparent; for example, IgA1, IgA2, IgG4, and IgM to recombinant DBL domains formed four distinct clusters. Similar functions sometimes also clustered together (e.g., primary monocyte and THP1 cell phagocytosis of DBL domain-coated beads). Multiplex measurement of IgG, IgG1, IgG3, and ability of antibody to bind FcyRs and C1q for all the different DBLs formed one large cluster, highlighting the collinearity of many of the antibody features measured towards recombinant proteins. Importantly, the six features identified by ENLR + PLSDA were spread throughout the correlation network and were not well correlated with each other (Figure 4, Figure 4—figure supplement 1).

Figure 4 with 1 supplement see all
Correlation network of antibody features.

Correlation network of all antibody features in both women with non-placental infection and those with placental malaria. The six selected antibody features do not cluster together. Antibody features with similar functions (denoted by the same letter) tend to correlate with each other. Blue: positive correlation; red: negative correlation; line width and closeness of variables increase with increasing correlation coefficients; only significant correlations (after Bonferroni correction for multiple comparisons) are shown. Selected antibody features identified by elastic net are highlighted and labeled. See supplementary file 3 for a full list of feature names.

Antibody features selected highlight the importance of functional antibodies towards the infected erythrocyte and DBL2

The six selected features included IgG3 to DBL2 (IgG3.DBL2(ID1-ID2a).FCR3), ADCP, and ADNP of CSA-binding IE (THP1.Phago.CS2 and Neutrophil.Phago.CS2, respectively), binding inhibition of IE to CSA (CSA.Binding.Inhibition.FCR3), IgA2 to DBL2 (IgA2.DBL2.1010), and IgG3 to CSA binding IE (IgG3.3D7). For all six features, antibody levels were significantly higher in the non-placental infection group compared to the placental malaria group (Figure 5A–F), but the patterns of antibody responses distinguishing the two groups were not uniform. For example, some women in the non-placental infection group had no measurable levels of THP1.phago.CS2, IgG3.3D7 or IgG2.DBL5.0466 antibody. This indicated that there are multiple mechanisms for antibodies to confer protection. To investigate this further, we constructed a heat map of the Z-score of each feature against individual women and compared the groups with non-placental infection and placental malaria (Figure 5G). Within these two cohorts, there was significant variation in individual profiles and no evidence for a unique and consistent profile of antibody functions in all individuals with non-placental infection. Overall, the women with non-placental infection had significantly higher levels of most of the identified antibody features; 81% of women with non-placental infection had a Z-score >0 for at least four antibody features (compared to 26% of those with placental malaria). Looking at just the three functional measures, 52% of women with non-placental infection had elevated antibodies (a Z-score >0) that both promoted IE phagocytosis by THP-1 cells and/or neutrophils and inhibited binding of IE to CSA, compared to 18% of women with placental malaria (Figure 5G).

Distribution of antibody features in women with non-placental infection (NPI) and placental malaria (PM).

(A–F) Levels of each of the selected antibody features in individual women in the two groups (G) No single antibody feature was present in all individuals with NPI (or was absent in all those with PM). Errors bars are mean (SD), p-values derived from Welch’s t-test. IE: infected erythrocyte; Z-score: distribution of the features was centered and scaled to have zero mean and a standard deviation of 1.

Discussion

We used a systems serology approach to broadly characterize antibody responses to the VAR2CSA protein that mediates protection from placental malaria. This involved measuring a wide range of antibody features to this protein followed by employing the machine learning techniques elastic net-regularized regression and PLSDA to identify those antibody features that protect pregnant women with P. falciparum infection from placental malaria. The combination of a prospective study design and a systems serology approach provided an unbiased way to identify the key features of the antibody response to VAR2CSA that mediates protection. Using this approach, we identified six antibody features of 169 tested, which were able to correctly differentiate, on average, 86% of the pregnant women with placental malaria and non-placental infection. The antibody features identified fell into broad functional groups: antibody that blocks placental binding of IEs and opsonizing antibody that result in phagocytic clearance of IEs (see Figure 6 for a schematic of mechanisms of protection of the six selected antibody features).

The six selected antibody features may protect women from placental malaria by (A) inhibiting infected red blood cells from binding chondroitin sulfate A (CSA) and sequestering in the placenta and/or (B) promoting phagocytosis of infected red blood cells by monocytes and/or neutrophils.

Selected features that may inhibit placental sequestration include IgG3 to the whole infected red blood cell and IgG3 and IgA2 to VAR2CSA's CSA binding domain DBL2. Selected features that may promote parasite clearance by antibody dependent phagocytosis include IgG3 to the whole infected red blood cell and to DBL2, IgA2 to DBL2, and neutrophil and monocyte phagocytosis of whole infected red blood cells. ADNP: antibody-dependent neutrophil phagocytosis; ADCP: antibody-dependent cellular phagocytosis; VAR2CSA: a parasite protein expressed on the surface of the infected red blood cell, made up of Duffy binding-like domains (DBL).

We hypothesized that pregnant women who had P. falciparum peripheral blood parasitemia at delivery but who did not have placental sequestration were protected from placental malaria by antibody immunity targeting VAR2CSA, and we sought to determine the characteristics of this protective antibody response. We frequency matched groups on participant characteristics such as gravidity and urban or rural residence to minimize confounding by these variables that are known to affect exposure and immunity. Antibody features at enrollment did not differ between women with placental malaria and those who were uninfected at delivery, but they did differ between infected women with and without placental malaria at delivery.

Unlike previously published studies (Aitken et al., 2010; Fried et al., 2018), the inclusion of an infected group without placental malaria allowed us to accurately control for exposure to the parasite at the time point of outcome. The lack of differences in antibody features between the non-infected and placental malaria groups was quite striking and mirrors previous observations (reviewed by Cutts et al., 2020). It may in part be attributable to the heterogeneous nature of the non-infected group, which probably includes both women who are exposed but protected by antibody and women who have not been exposed to malaria in pregnancy and therefore lack pregnancy-specific antibodies. It is possible that the non-infected group is protected by antibodies towards other antigens. That differences were seen only when comparing the two infected groups, clearly highlighting the importance for controlling for exposure when studying antibody responses and protection from malaria. In Cameroon, multigravid women studied in early pregnancy who did not go on to have placental malaria had a greater breadth and intensity of antibody responses to DBL domains than those with placental malaria at delivery (Tutterrow et al., 2012), and higher avidity antibodies to full-length VAR2CSA, suggesting that a broad, high-affinity IgG response to DBL proteins may also correlate with protection.

In this study, fitting the ENLR model to resampled subsets of data allowed us to rank antibody features by their selection frequency. We generated a ranking of the 20 most frequently selected features, and then used PLSDA to determine a minimal set of antibody features that classify the pregnant women with a high accuracy. These six features were selected by ENLR more than 70% of the time, and the inclusion of these six features in the PLSDA showed an ability to accurately distinguish 86% of the women in the two groups. Notably this ability to distinguish the two groups did not increase if more than six features were included in the model. The six antibody features that best distinguished the groups were all more common in women with non-placental infection, and they included previously identified and novel responses.

Among the features we identified as associated with protection against placental malaria were adhesion inhibition and cytophilic antibodies that mediate ADCP of IE. The ability to block IE binding to CSA has been associated with protection against placental malaria, maternal anemia, low birth weight, or prematurity in subsets of pregnant women (Duffy and Fried, 2003; Ndam et al., 2015). IgG1 and IgG3 antibody responses to VAR2CSA are dominant and mediate opsonic clearance of IE (Tebo et al., 2002; Elliott et al., 2005a; Megnekou et al., 2005; Keen et al., 2007; Damelang et al., 2019). ADCP using THP-1 cells has been correlated with protection from malaria or its complications in pregnant women and in adults (reviewed in Teo et al., 2016). ADNP of IE, which has not previously been systematically studied in malaria immunity, also correlated with placental infection status. Antibodies that block adhesion and mediate ADCP by THP-1 cells or ADNP are important correlates of protection that should be included in future studies of naturally acquired and vaccine-induced antibody to malaria in pregnancy.

Although only 11% of features measured antibody to IE, four of the six top antibody features involved assays using IE. Similarly, both features towards recombinant proteins involved the DBL2 domain, which is the principal DBL domain involved in placental adhesion. The overrepresentation of both IE and DBL2 in the selected features supports the notion that antibody features directed towards functionally important targets are key to protection from placental malaria.

Novel features identified using the system serology approach include the antibody feature IgA2. Of note, IgA2 levels in serum are generally low (most serum IgA is IgA1) (Macpherson et al., 2008). Interestingly a recent study looking for correlates of malaria protection from vaccine-induced immunity also identified IgA2 as a variable that helped predict those protected from P. falciparum infection (Suscovich et al., 2020). IgA2 can elicit myeloid cell effector functions, including opsonic phagocytosis (mediated by FcαRI) (Breedveld and van Egmond, 2019), and as it recognized the CSA binding DBL domain DBL2, it may also inhibit IE binding to CSA. However, since IgA2 levels in serum are low, we hypothesize that rather than being directly involved in protection IgA2 might instead be a marker of a protective antibody response. Further investigation of VAR2CSA-specific B cells and whether a protective response is associated with increased IgA2 is warranted.

Our analysis suggests that features determined by measuring antibody to VAR2CSA-expressing IE are important indicators of protection from placental malaria. Four features measured antibody to IE using representatives of the dominant 3D7-like and FCR3-like clades (Patel et al., 2017; Benavente et al., 2018) whose N-terminal regions are used in two VAR2CSA-based pregnancy malaria vaccines (Sirima et al., 2020; Mordmüller et al., 2019). While three of these four assays targeted FCR3 or the isogenic CS2 parasite line, studies of global diversity in VAR2CSA to date have included few sequences from Papua New Guinea (Benavente et al., 2018), and the relative importance of immunity to specific clades may vary by region. Vaccination offered limited heterologous protection (Sirima et al., 2020), and it is unclear whether effective naturally acquired protective immunity needs to target multiple clades or different clades in different regions.

Antibody responses to VAR2CSA domains serve as excellent markers of exposure to malaria in pregnant women (Cutts et al., 2020), but few were selected as correlates of protection in our analysis. This may reflect the fact that the responses to different VARCSA domains were generally highly correlated, as revealed by their clustering together in the center of our network analysis, or it might be because of the longevity of antibody that promotes ADCP of IE compared to that of antibody responses to recombinant proteins (Teo et al., 2014; Teo et al., 2015; Chandrasiri et al., 2014). Antibody to VAR2CSA recombinant proteins, and particularly to specific VAR2CSA peptides, appears to be a good correlate of exposure and transmission intensity (Fonseca et al., 2019), but these responses appear to be less useful as correlates of protection.

The correlation network showed close correlations between different antibody responses to recombinant DBL domains (which are predominantly located in the central cluster) but the protective antibody features were dispersed, suggesting that distinct pathways contribute to the development of protection against placental malaria. The analysis of functions identified in the selected features further supports this, with most protected women (and few unprotected women) having significant antibody that promoted both binding inhibition to CSA and phagocytosis; however, no consistent pattern of antibody functions in protected women was seen. Similar associations have been seen in the relationships between antibody features and protection from viral pathogens, and there is a growing awareness that polyfunctional antibody responses, rather than a single functional correlate, contribute to protection from disease (Gunn et al., 2018; Chung et al., 2014). None of the antibody features targeted more than one variant of the protein or parasite strain. One possible explanation is that epitopes recognized by antibodies mediating each function may vary between strains (as suggested by Doritchamou et al., 2019). This could in turn mean that protective functional features could vary with different populations (as there is variation in global distribution of the different VAR2CSA domain clades; Benavente et al., 2018). Future studies should ideally include multiple VAR2CSA clades and study diverse populations.

We used systems serology to identify protective features of the naturally acquired antibody response to VAR2CSA, a PfEMP1 variant that is specifically associated with placental malaria, complementing a recent study that used systems serology to identify functional antibody correlates of vaccine-induced immunity, in subjects vaccinated with RTS,S (Suscovich et al., 2020). A systems serology approach could also be used to identify the characteristics of protective immunity to other malaria syndromes, such as cerebral malaria, which is often associated with a subset of PfEMP1s that are potential vaccine candidates (reviewed in Jensen et al., 2020). It will be important to determine whether protective features of the cerebral malaria-associated PfEMP1 antibody response analogous to the ones that we identified in this study can be identified.

We were interested in measuring antibodies to placental binding parasites and defined placental malaria by the detection of IE in the placenta by histology because previous studies have shown that these are the infections associated with adverse pregnancy outcomes (Menendez et al., 2000; Rogerson et al., 2003). Ring stage parasites in the peripheral blood will circulate through the placenta, and, in some cases, these could be detected by PCR of placental blood; however, we did not consider these submicroscopic infections as placental malaria as they were not shown to be sequestering there.

Study strengths include the application of 169 novel and established assays; minimal missing predictor and no missing outcome data, the ability to control for exposure with the identification of pregnant women with P. falciparum infection with, and without, placental malaria; the prospective study design; and the use of native antigens on IEs to complement the data we acquired using recombinant protein. Together, this allowed us to discover the determinants of naturally acquired protective immunity. Weaknesses include the lack of inclusion of full-length VAR2CSA in our Luminex and bead-based assays, and an imbalance in numbers between women with parasitemia but no placental infection relative to the other groups. Samples from all women in the former group were selected for this study (and statistical analysis controlled for this imbalance between groups), but this group is relatively rare, with placental infection in the absence of peripheral infection being substantially more common (Desai et al., 2007). High levels of collinearity between some variables make it difficult to dissect out individual protective responses. Demonstration that our selected variables correlate with lack of placental malaria in other cohorts would further strengthen the findings. Antibody glycosylation (Lu et al., 2016) is an additional feature that could be included in future studies.

The present study identified leading candidate antibody features that mediate naturally acquired protection from placental malaria and demonstrated we could use these features to identify women in this cohort protected from placental malaria with high accuracy (though this should be validated in a second cohort). Our results suggest that multiple pathways exist that lead to the development of protective immunity against placental malaria. Next steps will include validating this set of features in other sample sets from different geographical settings and also determining whether these antibody features predict protection from adverse pregnancy outcomes such as low birth weight. The immune responses identified are relevant to further development of pregnancy malaria-specific therapies such as vaccines or monoclonal antibodies and possibly to the identification of protective antibody against malaria in non-pregnant populations. Multiple antibody functions and characteristics may be required to develop robust protective immunity to placental malaria.

Dataset used for Figures 16 has been uploaded to Dryad: doi:10.5061/dryad.wpzgmsbkx.

Appendix 1

Appendix 1—key resources table
Reagent type (species) or resourceDesignationSource or referenceIdentifiersAdditional information
Strain, strain background (Plasmodium falciparum, CS2)CS2Chandrasiri et al., 2014
Strain, strain background (Plasmodium falciparum, 3D7)3D7Chandrasiri et al., 2014Selected for chondroitin sulfate A (CSA) adhesion
Strain, strain background (Plasmodium falciparum, FCR3)FCR3Nielsen and Salanti, 2015Selected for CSA adhesion (parent of CS2)
Strain, strain background (Plasmodium falciparum, NF54)NF54Nielsen and Salanti, 2015Selected for CSA adhesion
(parent of 3D7)
Biological sample (Homo sapiens)PlasmaUnger et al., 2015
Biological sample (Homo sapiens)Primary monocytesThis paperFreshly isolated cells (see Materials and methods – Primary leukocytes)
Biological sample (Homo sapiens)Primary neutrophilsThis paperFreshly isolated cells (see Materials and methods – Primary leukocytes)
Biological sample (Homo sapiens)Primary NK cellsThis paperFreshly isolated cells (see Materials and methods – Primary leukocytes)
Cell line (Homo sapiens)THP1Ataíde et al., 2010RRID:CVCL_0006Monocytic cell line
Peptide, recombinant proteinDBL1-7G8Avril et al., 2011MV-1398Parasite line 7G8
Peptide, recombinant proteinDBL2(ID1-ID2a)-FCR3Doritchamou et al., 2016MV1942Parasite line FCR3
Peptide, recombinant proteinDBL2(ID1-ID2a)-FCR3Mordmüller et al., 2019Parasite line FCR3
Peptide, recombinant proteinDBL2-isolateDoritchamou et al., 2016MV 1940Parasite isolate 1010
Peptide, recombinant proteinDBL3- FCR3Nielsen et al., 2009MP1028Parasite line FCR3
Peptide, recombinant proteinDBL3- 7G8Avril et al., 2011MV-1914Parasite line 7G8
Peptide, recombinant proteinDB4-FCR3Fried et al., 2018MP2369Parasite line FCR3
Peptide, recombinant proteinDBL4-isolateDoritchamou et al., 2016MV1700Parasite isolate I 0711
Peptide, recombinant proteinDBL5-3D7Avril et al., 20111218Parasite line 3D7
Peptide, recombinant proteinDBL5-7G8Avril et al., 20111269Parasite line 7G8
Peptide, recombinant proteinDBL5-isolateDoritchamou et al., 2016MV 1749Parasite isolate I 0466
Peptide, recombinant proteinDBL6-IT4Avril et al., 2011MV-1137Parasite line IT4
Peptide, recombinant proteinMSP-1Barua et al., 2019
Biological sample (Plasmodium falciparum)Schizont extractBarua et al., 2019
AntibodyGoat anti-human IgG (polyclonal)Mabtech3820-4-250(1:2000)
AntibodyMouse anti-human IgG-PE (polyclonal)SouthernBiotech9040-09
RRID:AB_2796601
(1:77)
AntibodyMouse anti-human IgG1-PE (monoclonal)SouthernBiotech9052-09
RRID:AB_2796621
(1:77)
AntibodyMouse anti-human IgG2-PE (monoclonal)SouthernBiotech9070-09
RRID:AB_2796639
(1:77)
AntibodyMouse anti-human IgG3-PE (monoclonal)SouthernBiotech9210-09
RRID:AB_2796701
(1:77)
AntibodyMouse anti-human IgG4-PE (monoclonal)SouthernBiotech9200-09
RRID:AB_2796693
(1:77)
AntibodyMouse anti-human IgA1-PE (monoclonal)SouthernBiotech9130-09
RRID:AB_2796656
(1:77)
AntibodyMouse anti-human IgA2-PE (monoclonal)SouthernBiotech9140-09
RRID:AB_2796664
(1:77)
AntibodyMouse anti-human IgM-PE (monoclonal)SouthernBiotech9020-09
RRID:AB_2796577
(1:77)
AntibodyRabbit anti-human IgG (polyclonal)DAKOA0425(1:100)
AntibodyMouse anti-human IgG1 HP6069 (monoclonal)Merck Millipore411451(1:50)
AntibodyMouse anti-human IgG2 HP6002 (monoclonal)Merck MilliporeMAB1308(1:50)
AntibodyMouse anti-human IgG3
HP6050
(monoclonal)
SigmaI7260.2ml(1:50)
AntibodyMouse anti-human IgG4
HP6023
(monoclonal)
Merck MilliporeMAB1312-K(1:50)
AntibodyGoat anti-mouse IgG AlexaFluor 647 (polyclonal)Life TechnologiesA-21235
RRID:AB_2535804
(1:500)
AntibodyDonkey anti-rabbit 647 (polyclonal)Life TechnologiesA-31573
RRID:AB_2536183
(1:500)
Peptide, recombinant proteinC1qMP Biomedicals80295-33-61.3 µg/ml
Peptide, recombinant proteinFcγRIR&D Systems1257-FC
Peptide, recombinant proteinFcγRIIaWines et al., 2016
Peptide, recombinant proteinFcγRIIIaWines et al., 2016
Peptide, recombinant proteinFcγRIIIbR&D Systems1875 CD
AntibodyMouse anti-human CD16-Brilliantviolet 605
(monoclonal)
BD563172(1:50)
AntibodyMouse anti-human CD56-Brilliantultraviolet 737
(monoclonal)
BD347344(1:800)
AntibodyMouse anti-human CD3-peridinin-chlorophyll-protein
(monoclonal)
BD552127(1:200)
AntibodyMouse anti-human IFNγ-PE
(monoclonal)
BD554701(1:200)
AntibodyMouse anti-human TNFα-BV-785
(monoclonal)
BioLegend502947
RRID:AB_2565857
(1:200)
Software, algorithmR software, caret packageR software
Kuhn, 2008
Software, algorithmR software, glmnet packageR software
Zou and Hastie, 2005
Software, algorithmR software, PLS packageR software
Mevik and Wehrens, 2007
Software, algorithmR software, mixOmics packageR software
Rohart et al., 2017
Software, algorithmR software, qgraph packageR software
Epskamp et al., 2012
Commercial assay or kitEasySep Direct Human Neutrophil Isolation KitSTEMCELL Technologies19666
Commercial assay or kitRosetteSep Human Monocyte Enrichment CocktailSTEMCELL Technologies15068
Commercial assay or kitRosetteSep Human NK Enrichment CocktailSTEMCELL Technologies15065
Commercial assay or kitMelon Gel purification kitsThermo Fisher Scientific45212
Commercial assay or kitEZ-link Sulfo NHS-LC-Biotin kitThermo Fisher Scientific24520
OtherBio-Plex magnetic carboxylated microspheresBio-Rad#MC100xx-01
OtherStreptavidin-PESouthernBiotech7105-09

Data availability

All antibody feature data has been deposited in Dryad.

The following data sets were generated

References

    1. Buuren S
    2. Groothuis-Oudshoorn K
    (2011)
    Mice : multivariate imputation by chained equations in R
    Journal of Statistical Software 45:1–67.
    1. Kuhn M
    (2008)
    Building predictive models in R Using the caret Package
    Journal of Statistical Software 28:1–26.
    1. Mevik B-H
    2. Wehrens R
    (2007)
    The pls package: principal component and partial least squares regression in R
    Journal of Statistical Software 18:1–23.
    1. Zou H
    2. Hastie T
    (2005)
    Regularization and variable selection via the elastic net
    Journal of the Royal Statistical Society: Series B 301:20.

Decision letter

  1. Urszula Krzych
    Reviewing Editor; Walter Reed Army Institute of Research, United States
  2. Dominique Soldati-Favre
    Senior Editor; University of Geneva, Switzerland
  3. Sidhartha Chaudhury
    Reviewer; Walter Reed Army Institute of Research, United States

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

This is the very first study to apply systems serology approaches to study antibody responses directed to specific Plasmodium (malaria) antigens and the results demonstrate that these responses correlate with protection from placental malaria. The analyses of samples from approximately 200 women revealed that these antibody features do discriminate between women with placental malaria at delivery and women with malaria detectable in the peripheral blood, but not in placenta. These findings support the effort to develop vaccines against the certain protein antigens of placental malaria parasites.

Decision letter after peer review:

Thank you for submitting your article "Identifying women protected from placental malaria using systems serology" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Dominique Soldati-Favre as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Sidhartha Chaudhury (Reviewer #3).

The Editors and reviewers have discussed your revision plan with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

1) Your responses to the revisions as proposed by the Reviewers have, in general, been considered as sufficient for addressing their concerns.

2) However, apart from the specific critical points as indicated in the initial decision letter and which are included here once again, the revised manuscript should clearly state your definition of the primary group used to analyze correlates of protection – the "non-placental infection at delivery" group. At present, it appears to be defined as "women with P. falciparum infection by light microscopy and/or PCR in the peripheral blood but with no IEs visible in the placenta by histology". This response ("Line 85") leaves open whether this definition means "no IEs" or "no mature IEs" by histology. Thus, it seems that a full clear definition is missing in the manuscript, and it is not clear from your response what the planned revision includes. In addition, please indicate in Table 1 how many of the 27 women in this group were positive by microscopy versus PCR of their peripheral blood.

Reviewer #1:

The authors have conducted a systems serology study of the relationship between antibody responses to VAR2CSA and protection from placental malaria (PM) in pregnant women from Papua New Guinea.

Based on initial analysis of 130 variables and samples from 127 women (15,340 data points) and subsequent analysis of 169 variables from 77 women (12,906 data points), they identified a list of six selected serological features that discriminated with 86% accuracy between women with placental parasitaemia at delivery and women with peripheral but not placental parasitaemia at the same time point. It is an important conclusion from the work performed that it underscores the likelihood that multiple pathways to clinical protection from PM exists – a conclusion that can safely be extended to other clinical categories of malaria – and probably even to infectious diseases in general. While this may seem obvious to some, it is nevertheless largely ignored in almost all existing reports. In particular, current efforts to develop vaccines against PM are completely focussed on inhibition of adhesion of infected erythrocytes (IEs), completely ignoring the antibody effector functions that rely on opsonization of IEs for cell-mediated destruction, and which the current analysis suggest are important.

The manuscript is very well written, based on careful and in-depth analysis, and constitutes a valuable addition to the existing literature.

While I am not able to ascertain the finer points of the advanced statistical analyses conducted, I have no reason to suspect that anything is amiss.

I have only a couple of specific comments:

P. 14 ("ADNP of IE… contributed to protection"): I think it would be more correct to write "ADNP of IE… correlated with placental infection status" as no causal relationship was established.

P. 15: The authors list weaknesses of their study, which is commendable. However, they fail to mention what I consider a more prominent weakness than those listed, namely the absence of a test of their six-features-list on an independent data/sample set. While such a test would be optimal, at least it should be mentioned as a weakness if it is not done.

Reviewer #2:

The authors are the first to use systems serology to examine antibody responses associated with protection from placental malaria. They determine that functional antibodies against CSA-binding IE are associated with protection from placental malaria, in general confirming findings from earlier studies. While the consistency with earlier studies engenders confidence in their findings, it does lessen the novelty. Important aspects of the clinical study design and study cohort need better description and justification. The study would be strengthened considerably if the top features identified here were tested for their predictive capacity in a second confirmatory cohort, albeit the fact that these associations have previously been made in other studies give general confidence in the findings. Protective immunity against placental malaria increases with gravidity, and therefore the authors should report how the protective features that they identify are related to gravidity.

Abstract concludes "inhibition of placental binding and/or opsonic phagocytosis of infected erythrocytes" identify women who are protected from placental malaria. This is consistent with earlier studies using non-systems approaches that previously identified these associations.

Line 85 "some pregnant women with peripheral blood infection remain free of placental malaria" – while such cases have been sporadically described, the "placental malaria-negative infection" phenotype is an uncommon phenomenon in studies that carefully document placental and peripheral Pf infections, and requires better definition here. Who has used such a study design before? "Placental malaria-negative" can mean that mature parasites do not sequester in the placenta, but does not necessarily mean that all parasites are absent from the placenta. Practically speaking, how could peripheral ring stage parasites be excluded from circulating through the intervillous spaces at the same density they are found in the peripheral blood? Related to this, the authors should provide detailed information on the pregnancy outcomes for the women in the different groups: especially Pf quantification in peripheral blood and placental blood by microscopy and by PCR, etc, but also BW, gestational age. It may be that the "PM-negative infection" phenotype is related to highly sensitive detection of sparse Pf in peripheral blood by PCR versus less sensitive detection of sparse Pf in placental blood using Giemsa-stained histology (see similar comment below).

Line 108 "Five of the six leading antibody features we found were related to inhibition of placental sequestration and/or opsonisation for phagocytic clearance" Can the authors comment whether the same antibodies might be mediating both effects, or do they surmise they are unlikely to overlap?

Line 112 Please state briefly but clearly here how the samples/women were selected for inclusion in the cohort (in addition to the more detailed description in the Methods).

Line 115 "All women resided in heavily malaria exposed areas" is not consistent with the evidence provided in Table 1 that only 14/127 women had Pf detected by PCR at their baseline visit. The authors should explain this. Had these women previously received IPTp, or was this their baseline sample collected before first IPTp?

Line 119 "no P. falciparum parasites in the placenta but who had peripheral blood parasitemia by PCR and/or light microscopy" This wording and the Methods section indicate peripheral blood was tested for Pf by Giemsa and PCR, but placental blood was tested by Giemsa but not PCR. This is not a fair comparison. What is the sensitivity of their PCR versus their Giemsa stained blood or tissue?

Line 142 "using samples from the 127 women and measuring 13 antibody features directed at ten DBL domains, 15,340 out of 16,510 possible data points were acquired". This is an impressive dataset to examine correlates of protection.

All pregnant women in this study have received IPTp and this can affect the PM outcome at delivery. Looks like this has not been really considered in the interpretation of the data, particularly in view of the fact that two types of IPTp combinations were used, and readers are not informed of their potentially different efficacies.

Among the identified antibody features in this study it appeared that none of the activities targeted more than one variant of protein or parasite. Can the authors discuss this lack of strain-transcending pattern in these activities? A similar pattern of parasite variant dependent activity of PM antibody has previously been described by Doritchamou et al. IAI 2019. The authors should discuss their findings in that context and also comment whether the six antibody features identified in this study might vary in populations from different geographical areas.

Lines 214 – 215 "… the six features identified by elastic net were spread throughout the correlation network and were not well correlated with each other (Figure 4 and Supplementary Figure 3)". Based on this statement it is confusing when the authors conclude that the six features fall into two major functional groups. In particular the inverse correlation between IgG3.DBL2(ID1-ID2a).FCR3 with CSA.Binding.Inhibition.FCR3 is inconsistent with Figure 6 suggesting that these represent related responses.

The authors indicated that 17 samples were Pf positive at enrolment. What is the impact of early infection on these identified antibody features in relation to protection against PM and to antibody responses?

Reviewer #3:

The authors hypothesize that protection from placental malaria requires specific humoral immunity to malaria protein VAR2CSA. The authors obtained plasma samples from 127 pregnant women and grouped them into three categories based on malaria diagnosis at delivery: non-infected at delivery, placental malaria at delivery, and non-placental malaria at delivery. They use systems serology to measure 169 variables related to antibody binding, isotype, and functional activity and used machine learning to identify 6 variables that could accurately distinguish between placental and non-placental malaria cases. They delve into these 6 variables to provide some insight into the mechanisms of protection against placental malaria.

Strengths: The cohorts in this study are large and robust, with 50 placental and 27 non-placental cases with similar demographic profiles. They assess a wide range of serology measures for binding, isotype, isotype subclass, and functional responses against both infected erythrocyte and recombinant antigens. They find significant group-level differences in serology measures between placental and non-placental malaria groups suggesting that there is great potential in this data set to identify correlates of immunity and potential mechanisms of protection.

Weaknesses: They used a elastic network-regularized linear regression method to identify from 169 variables, the top variables that distinguish placental and non-placental malaria cases, and then use a Partial Least Square Discriminant Analysis method to assess how accurately a subset of these variables could be used to predict placental malaria susceptibility. While this overall approach is reasonable, it contains a substantial risk of overfitting that is not properly assessed in the study. Overfit models produce seemingly highly accurate predictions that result from incidental/random variation in the data rather than biologically significant differences, and jeopardize the scientific conclusions of the study.

Furthermore, the authors assume that differences between the placental malaria and non-placental malaria groups reflect correlates of protection, thus justifying an exploration into mechanisms of protection. However, these two groups could reflect differences in malaria exposure, despite demographic similarities in age/residence. If correlates of protection cannot be distinguished from correlates of exposure, this limits what can be inferred about mechanisms of protection.

Appraisal: The authors conclusions rest substantially on the six variables identified through machine learning. Ultimately the study conclusions cannot be supported by the data without demonstrating that these six variables reflect biologically significant differences and do not simply arise from analysis choices and model overfitting.

1) It is important to demonstrate that malaria exposure is similar in the placental and non-placental malaria groups because a key assumption made here is that a comparison between these groups identifies correlates of protection, and not simply correlates of exposure. Indeed, the authors find that the non-placental malaria group had almost double the PCR positive rate of the placental malaria group at enrolment, suggesting it is possible they had higher exposure, despite similar demographic profiles.

One way to address this would be to take serological markers in the data that others in literature have shown to be linked to exposure, and show that there is no significant differences between the three groups with respect to these serological markers of exposure.

2) Much of the conclusions of this study rests on the outcome of the machine learning analysis, so it is critical that each step of this analysis is properly justified.

In the elastic network-regularized logistic regression (ENLR), more detail is needed to explain how correlated variables are handled. For example, why was α set to 0.50, as opposed to any other value between 0 and 1? To what degree are the findings that a small number of non-correlated variables predictive of placental malaria susceptibility simply a reflection of the analysis approach rather than the underlying biology?

3) Only three variables out of the top 20 identified by logistic regression have odds-ratios whose lower bound IQR exceeds 1.0 (Figure 2). Does this suggest the observed effects are not significant? How was it decided that 20 variables would be selected for further study?

4) In the Partial Least Squares Discriminant Analysis (PLSDA), was any cross-validation done? Or was the model trained on the entire data set? PLSDA can be highly susceptible to overfitting, and cross-validation should be done to assess prediction accuracy.

Also, given the imbalance in the data set (2:1 ratio of placental to non-placental malaria cases), other measures of accuracy should be considered, such as sensitivity, specificity, Cohen's kappa value, and/or ROC AUC.

5) An assessment of the this approach (ENLR + PLSDA) in terms of its risk for overfitting must be done. Guided feature reduction (ENLR) combined with PLSDA carries a high risk of overfitting. A comparison with 6 random variables in PLSDA is not sufficient to assess overfitting because it doesn't account for the degree to which guided feature reduction (ENLR) step contributes to overfitting.

Authors should consider doing a permutation test, where the data set is randomized with respect to outcome (placental vs. non-placental malaria), and then the entire ENLR + PLSDA method is carried out to determine prediction accuracy. This can be carried out for N different permutations (for example N=100 times) to determine the average prediction accuracy for this randomized outcome. Because the outcome is randomized, it is decoupled from the immune data, and thus acts as a 'negative control' for machine learning and it should fail to predict the outcome with any accuracy greater than chance. The degree to which it does 'accurately' predict this randomized outcome is a reflection of the overfitting of the model.

https://doi.org/10.7554/eLife.65776.sa1

Author response

Essential revisions:

The revised manuscript should clearly state your definition of the primary group used to analyze correlates of protection – the "non-placental infection at delivery" group. At present, it appears to be defined as "women with P. falciparum infection by light microscopy and/or PCR in the peripheral blood but with no IEs visible in the placenta by histology". This response ("Line 85") leaves open whether this definition means "no IEs" or "no mature IEs" by histology. Thus, it seems that a full clear definition is missing in the manuscript, and it is not clear from your response what the planned revision includes. In addition, please indicate in Table 1 how many of the 27 women in this group were positive by microscopy versus PCR of their peripheral blood.

We have added further clarification to the text under Methods, lines 148 onwards and Results, Cohort description, starting at line 467, to indicate that placental malaria was defined as no IEs visible by histology (JN Bulmer, Histopathology 1993; PR Walter et al., Am J Path 1982). As those authors observed, histology is most sensitive at detecting mature trophozoites and schizonts, in part because they actively accumulate in the intervillous space, and in part because they are both larger and contain visible hemozoin. We did record presence of any IEs on histology examination irrespective of stage as placental malaria.

We have expanded Table 1 to add the results of peripheral blood PCR and microscopy. See also response to reviewer 2.

Reviewer #1:

The authors have conducted a systems serology study of the relationship between antibody responses to VAR2CSA and protection from placental malaria (PM) in pregnant women from Papua New Guinea.

Based on initial analysis of 130 variables and samples from 127 women (15,340 data points) and subsequent analysis of 169 variables from 77 women (12,906 data points), they identified a list of six selected serological features that discriminated with 86% accuracy between women with placental parasitaemia at delivery and women with peripheral but not placental parasitaemia at the same time point. It is an important conclusion from the work performed that it underscores the likelihood that multiple pathways to clinical protection from PM exists – a conclusion that can safely be extended to other clinical categories of malaria – and probably even to infectious diseases in general. While this may seem obvious to some, it is nevertheless largely ignored in almost all existing reports. In particular, current efforts to develop vaccines against PM are completely focussed on inhibition of adhesion of infected erythrocytes (IEs), completely ignoring the antibody effector functions that rely on opsonization of IEs for cell-mediated destruction, and which the current analysis suggest are important.

The manuscript is very well written, based on careful and in-depth analysis, and constitutes a valuable addition to the existing literature.

While I am not able to ascertain the finer points of the advanced statistical analyses conducted, I have no reason to suspect that anything is amiss.

I have only a couple of specific comments:

P. 14 ("ADNP of IE… contributed to protection"): I think it would be more correct to write "ADNP of IE… correlated with placental infection status" as no causal relationship was established.

We agree and have revised the sentence as suggested.

Line 678-680 “ADNP of IE, which has not previously been systematically studied in malaria immunity also correlated with placental infection status”P. 15: The authors list weaknesses of their study, which is commendable. However, they fail to mention what I consider a more prominent weakness than those listed, namely the absence of a test of their six-features-list on an independent data/sample set. While such a test would be optimal, at least it should be mentioned as a weakness if it is not done.

We have added the following to the section on study weaknesses:

Line 760-2 “Demonstration that our selected variables correlate with lack of placental malaria in other cohorts would further strengthen the findings.”

Reviewer #2:The authors are the first to use systems serology to examine antibody responses associated with protection from placental malaria. They determine that functional antibodies against CSA-binding IE are associated with protection from placental malaria, in general confirming findings from earlier studies. While the consistency with earlier studies engenders confidence in their findings, it does lessen the novelty. Important aspects of the clinical study design and study cohort need better description and justification. The study would be strengthened considerably if the top features identified here were tested for their predictive capacity in a second confirmatory cohort, albeit the fact that these associations have previously been made in other studies give general confidence in the findings. Protective immunity against placental malaria increases with gravidity, and therefore the authors should report how the protective features that they identify are related to gravidity.

We thank the reviewer for their helpful comments. We have added a comment to the discussion under study weaknesses regarding the need to confirm our findings in further cohorts (Lines 760-2, 765-767), see also response to Reviewer 1. Our plans to examine responses in a second confirmatory cohort have been severely disrupted by COVID-19, especially in regard to sample shipments which have not been possible for the past 12 months.

Regarding associations of protective features and gravidity, we have conducted a univariate analysis using linear regression looking at how the antibody levels vary with gravidity in our cohort and have added this to the paper in the form of Supplementary File 4-table 1.

In brief: Antibody feature CSA.Binding.Inhibition.FCR3 was positively associated with gravidity, and there was weak evidence that THP1.Phago.CS2 and Neutrophil.Phago.CS2 were also positively associated with gravidity. There was no evidence that IgG3.DBL2(ID1-ID2a).FCR3, IgA2.DBL2.1010 and IgG3.3D7 were associated with gravidity in this cohort. We now discuss these observations in the manuscript (lines 573-585).

Abstract concludes "inhibition of placental binding and/or opsonic phagocytosis of infected erythrocytes" identify women who are protected from placental malaria. This is consistent with earlier studies using non-systems approaches that previously identified these associations.

We made small edits to the abstract to reflect that selected protective responses include features previously described:

Line 50-51 “Selected features included previously described associations with inhibition of placental binding and/or opsonic phagocytosis of infected erythrocytes.”

Line 85 "some pregnant women with peripheral blood infection remain free of placental malaria" – while such cases have been sporadically described, the "placental malaria-negative infection" phenotype is an uncommon phenomenon in studies that carefully document placental and peripheral Pf infections, and requires better definition here. Who has used such a study design before? "Placental malaria-negative" can mean that mature parasites do not sequester in the placenta, but does not necessarily mean that all parasites are absent from the placenta. Practically speaking, how could peripheral ring stage parasites be excluded from circulating through the intervillous spaces at the same density they are found in the peripheral blood? Related to this, the authors should provide detailed information on the pregnancy outcomes for the women in the different groups: especially Pf quantification in peripheral blood and placental blood by microscopy and by PCR, etc, but also BW, gestational age. It may be that the "PM-negative infection" phenotype is related to highly sensitive detection of sparse Pf in peripheral blood by PCR versus less sensitive detection of sparse Pf in placental blood using Giemsa-stained histology (see similar comment below).

We agree with the reviewer that women with peripheral blood infection without evidence of placental sequestration form a relatively rare group, and we only identified 27 in our cohort. They are analogous to infections that are found in non-pregnant women in malarious areas. When women conceive, some of these infections progress to higher density, especially in younger, less immune women (Ndam et al. Clin Infect Dis 2018; Hounkonnou et al. Clin Infect Dis 2020), likely due to emergence of variants than can sequester in the placenta and avoid existing antibody immunity.

Previous studies (reviewed by Cutts et al., BMC Med 2020) highlight the difficulty of distinguishing effects of exposure from those of immunity using the more common comparison of women with and without placental malaria, which was also not informative in our hands (Figure 1 a), compared to our novel inclusion of this additional group (Figure 1 b).

The reason we (and many other studies) use detection of IEs in the placenta by histology to define placental malaria, is because many studies show that these infections are associated with worse pregnancy outcomes such as prematurity, low birth weight or maternal anemia (Menendez C et al., J Infect Dis 2000; Rogerson SJ et al., Am J Trop Med Hyg 2003). Identifying women whose infection has not progressed to placental malaria, and the antibody responses that might limit that progression by inhibiting placental sequestration, is our specific aim.

The circulating ring stage parasites will pass through the placenta although they do not sequester there, and placental blood PCR did detect them in some women (added to Table 1). These circulating rings are understood not to have any deleterious effects on the placenta, and submicroscopic infections, including in placental blood, appear to be only weakly associated with adverse birth outcomes, unlike active placental infection on histology.

We have included further information as requested on the pregnancy outcomes, including BW, hemoglobin and gestational age at delivery and peripheral and placental PCR in Table 1. Original groupings of the cohorts were not done using placental PCR, however subsequent analysis revealed that one individual classified as non-infected (which was based on PCR and light microscopy of peripheral blood as well as placental histology) was actually PCR positive in the placental blood. As our analysis identifying protective features did not include the non-infected cohorts, this individual’s results do not affect the conclusions in this paper.

Line 108 "Five of the six leading antibody features we found were related to inhibition of placental sequestration and/or opsonisation for phagocytic clearance" Can the authors comment whether the same antibodies might be mediating both effects, or do they surmise they are unlikely to overlap?

The diversity of responses in “protected” women in Figure 5G suggests that these often do not overlap, or the overlap is heterogenous in different women. However, determining whether individual antibodies have overlapping functions would ideally require monoclonals and therefore we cannot confidently comment on whether the same antibody might be able to mediate both functions.

Line 112 Please state briefly but clearly here how the samples/women were selected for inclusion in the cohort (in addition to the more detailed description in the Methods).

As suggested, in addition to the detailed description in the methods, we have briefly described the selection of the cohort in the Results section (which now follows methods, as requested by the Editors), see below:

LINE 467 onwards: “Women were selected based on P. falciparum infection status at delivery […] The three groups were frequency matched for primigravidity, IPTp regime receipt, bed net use, rural residency and age”

Line 115 "All women resided in heavily malaria exposed areas" is not consistent with the evidence provided in Table 1 that only 14/127 women had Pf detected by PCR at their baseline visit. The authors should explain this. Had these women previously received IPTp, or was this their baseline sample collected before first IPTp?

Line 466: We have removed the word “heavily”, given the relatively low parasite prevalence. In an earlier study at one of our nine sites we found 64% of women to have current or past placental malaria on histology and 34% and 49% to have P. falciparum at first antenatal visit by microscopy and PCR respectively (Stanisic et al. Trans Roy Soc Trop Med Hyg 2015), and transmission in the area had only recently declined (Ome-Kaius et al. BMC Med 2019).

All women were recruited at first antenatal visit before receiving any IPTp. We have clarified this under “Human Subjects” in the methods in line 143-4.

“Peripheral blood collected at 14-26 gestation weeks, before initiation of IPTp, was used in the antibody assays.”Line 119 "no P. falciparum parasites in the placenta but who had peripheral blood parasitemia by PCR and/or light microscopy" This wording and the Methods section indicate peripheral blood was tested for Pf by Giemsa and PCR, but placental blood was tested by Giemsa but not PCR. This is not a fair comparison. What is the sensitivity of their PCR versus their Giemsa stained blood or tissue?

See our earlier response to line 85. We specifically grouped women with malaria infection into those with and without placental malaria on histology, as classically defined (Walter Am J Path 1982; Bulmer, Histopathology 1993). Our question was, are there antibody responses that might explain why some infected women have not progressed to placental malaria, when others have. Antibody responses that prevent infection crossing a threshold (to symptomatic infection, to placental infection or to severe disease) are more relevant to blood stage immunity than antibodies that simply prevent infection, as may be sought for pre-erythrocytic infection.

Line 142 "using samples from the 127 women and measuring 13 antibody features directed at ten DBL domains, 15,340 out of 16,510 possible data points were acquired". This is an impressive dataset to examine correlates of protection.

All pregnant women in this study have received IPTp and this can affect the PM outcome at delivery. Looks like this has not been really considered in the interpretation of the data, particularly in view of the fact that two types of IPTp combinations were used, and readers are not informed of their potentially different efficacies.

We added a statement to the results highlighting the different efficacy of SPAZ and SPCQ against placental malaria (Line 476-7). As shown in Table 1, a slightly higher proportion of women in the non-placental malaria group received the less efficacious SPCQ.

“Sulfadoxine pyrimethamine (SP) and azithromycin was more efficacious against placental malaria than SP-chloroquine(Unger et al., 2015).”

Among the identified antibody features in this study it appeared that none of the activities targeted more than one variant of protein or parasite. Can the authors discuss this lack of strain-transcending pattern in these activities? A similar pattern of parasite variant dependent activity of PM antibody has previously been described by Doritchamou et al. IAI 2019. The authors should discuss their findings in that context and also comment whether the six antibody features identified in this study might vary in populations from different geographical areas.

As suggested we have discussed our findings in relation to those described by Doritchamou et al. IAI 2019

Line 726 onwards: None of the antibody features targeted more than one variant of the protein or parasite strain. One possible explanation is that epitopes recognised by antibodies mediating each function may vary between strains (as suggested by Doritchamou et al., 2019). This could in turn mean that protective functional features could vary with different populations (as there is variation in global distribution of the different VAR2CSA domain clades (Benavente et al., 2018)). Future studies should ideally include multiple VAR2CSA clades, and study diverse populations.

Lines 214 – 215 "… the six features identified by elastic net were spread throughout the correlation network and were not well correlated with each other (Figure 4 and Supplementary Figure 3)". Based on this statement it is confusing when the authors conclude that the six features fall into two major functional groups. In particular the inverse correlation between IgG3.DBL2(ID1-ID2a).FCR3 with CSA.Binding.Inhibition.FCR3 is inconsistent with Figure 6 suggesting that these represent related responses.

The broad functional mechanisms of antibody can fall into similar groups, even when the development of responses to different members of each group is only weakly correlated. As the reviewer notes, there was no correlation between IgG3.DBL2(ID1-ID2a).FCR3 with CSA.Binding.Inhibition in Figure 4 and a weak, non-significant negative correlation between them (Figure 4—figure supplement 1). Although binding inhibition is likely to be primarily mediated by the more-abundant IgG1, we included IgG3 to DBL2 in both Figure 6A and Figure 6B because its principal functional role could fall in either group, as total IgG antibody to this domain is known to block adhesion.

The authors indicated that 17 samples were Pf positive at enrolment. What is the impact of early infection on these identified antibody features in relation to protection against PM and to antibody responses?

We thank the reviewer for asking this important question, as it has improved the quality of the manuscript. We have now conducted a sub-analysis comparing the identified antibody features at enrolment in the non-placental infection cohort with those in the placental malaria cohort, excluding women who were infected (by light microscopy and/or PCR) at enrolment. In this sub-analysis we saw strong evidence that 4 of the 6 features were higher in the non-placental infection cohort (P≤0.013) and weak evidence that the other two features were higher in the non-placental infection cohort (≤0.064). The maintenance of the protective associations in the smaller, non-infected cohort suggests to us that the protective association seen in these variables was not greatly affected by the presence of the positive samples.

These data are included in the manuscript in Supplementary File 4 – table 2. These results are discussed in lines 577 onwards.

Reviewer #3:1) It is important to demonstrate that malaria exposure is similar in the placental and non-placental malaria groups because a key assumption made here is that a comparison between these groups identifies correlates of protection, and not simply correlates of exposure. Indeed, the authors find that the non-placental malaria group had almost double the PCR positive rate of the placental malaria group at enrolment, suggesting it is possible they had higher exposure, despite similar demographic profiles.

One way to address this would be to take serological markers in the data that others in literature have shown to be linked to exposure, and show that there is no significant differences between the three groups with respect to these serological markers of exposure.

Thank you for this suggestion. We measured IgG levels by ELISA to non-pregnancy specific malaria antigens merozoite surface protein-1 and schizont extract which we have previously shown to be strongly associated with exposure (Barua 2019 Malaria Journal) and have done univariate analyses comparing non-infected v placental malaria and non-placental infection v placental malaria. There was no association with absence of placental infection for IgG towards either antigen, suggesting that the features we have identified are not simply correlates of exposure.

We have included this in the Results section (line 510 onwards) and in Supplementary File 2.

2) Much of the conclusions of this study rests on the outcome of the machine learning analysis, so it is critical that each step of this analysis is properly justified.

In the elastic network-regularized logistic regression (ENLR), more detail is needed to explain how correlated variables are handled. For example, why was α set to 0.50, as opposed to any other value between 0 and 1? To what degree are the findings that a small number of non-correlated variables predictive of placental malaria susceptibility simply a reflection of the analysis approach rather than the underlying biology?

To address the dependence of the results on α, we have now computed the selection frequencies across all α values {0,0.25,0.5,0.75,1} instead of only α= 0.5, using the resampling of ENLR explained in Section Identification of Key Antibody Features of the previous manuscript. The results showed that the top six variables do not change (please see Figure 2—figure supplement 1). We have now modified the revised manuscript to address this point.

Line 411 onwards now states:

Identification of Key Antibody Features

“The parameter α, which alters the nature of penalization between ridge regression and LASSO, was set to 0.5 to achieve a balance between sparsity and group selection; the final selected antibody features did not alter when α was varied over [0, 1], as detailed below. […] The resampling process was repeated where α was also tuned over the set {0,0.25,0.5,0.75,1} (instead of only α=0.5) in addition to λ and the results showed that the final top frequently selected variables do not change. (Figure 2—figure supplement 1)”

3) Only three variables out of the top 20 identified by logistic regression have odds-ratios whose lower bound IQR exceeds 1.0 (Figure 2). Does this suggest the observed effects are not significant? How was it decided that 20 variables would be selected for further study?

Due to the penalisation in the elastic net, the unselected variables are assigned a coefficient of zero (i.e. odds ratio = 1). Here we are repeating ENLR, resampling from the dataset 50,000 times, and therefore, the estimated coefficient for every variable is zero in a certain proportion of resampled datasets (even though a variable may frequently be selected in a high proportion of the resampled datasets). This pushes the lower bounds of the IQRs towards zero, hence lower bound of odds ratio will be very close to 1 for many of the variables. Therefore, taking the penalisation into consideration, this does not necessarily suggest that the effects are insignificant for those variables. Figure 2 presented the top 20 variables for ease of illustration.

4) In the Partial Least Squares Discriminant Analysis (PLSDA), was any cross-validation done? Or was the model trained on the entire data set? PLSDA can be highly susceptible to overfitting, and cross-validation should be done to assess prediction accuracy.

Yes, 500 repeats of 10-fold cross-validation was performed and the reported performance measures were computed from the cross-validation. This is now added to the “Identification of Key Antibody Features” section and the caption of Figure 3 is edited for clarity.

Identification of Key Antibody Features (line 440)

“…; 500 repeats of 10-fold cross validation were performed to compute the accuracies.”

Figure Legend 3

“… (500 repeats of 10-fold cross-validation were performed to estimate accuracy for each model)”

Also, given the imbalance in the data set (2:1 ratio of placental to non-placental malaria cases), other measures of accuracy should be considered, such as sensitivity, specificity, Cohen's kappa value, and/or ROC AUC.

We have now changed figures 3A and 3B and the text (line 555 onwards) by presenting AUROC as the performance measure, as the reviewer suggests.

5) An assessment of the this approach (ENLR + PLSDA) in terms of its risk for overfitting must be done. Guided feature reduction (ENLR) combined with PLSDA carries a high risk of overfitting. A comparison with 6 random variables in PLSDA is not sufficient to assess overfitting because it doesn't account for the degree to which guided feature reduction (ENLR) step contributes to overfitting.

Authors should consider doing a permutation test, where the data set is randomized with respect to outcome (placental vs. non-placental malaria), and then the entire ENLR + PLSDA method is carried out to determine prediction accuracy. This can be carried out for N different permutations (for example N=100 times) to determine the average prediction accuracy for this randomized outcome. Because the outcome is randomized, it is decoupled from the immune data, and thus acts as a 'negative control' for machine learning and it should fail to predict the outcome with any accuracy greater than chance. The degree to which it does 'accurately' predict this randomized outcome is a reflection of the overfitting of the model.

We thank the reviewer for raising this important point and agree that the estimated performances are probably slightly optimistic and the accuracy may slightly decrease if this method is applied on external validation data. We performed the analysis suggested by the reviewer, i.e. repeating the whole ENLR + PLSDA process for 100 datasets whose group labels (PM and NPM) were randomly permuted. For each of the 100 datasets, 500 repeats of 10-fold cross validation were performed on the EN and the top six frequently selected variables were fed into the PLSDA, on which 5 repeats of 10-fold cross-validation were performed to compute the AUROC. The AUROCs of the 100 datasets were aggregated to produce the boxplot in Figure 3B.

The results (Figure 3B) show that the AUROC for our six selected variables is still significantly higher than that for the permutation test.

Line 441 onwards now state:

Identification of Key Antibody Features

“The statistical significance of the results was assessed by comparing the performance of the model with that of two random permutation tests (null cases): 1) the PLSDA model was fitted to six randomly selected antibody features and the performance was computed for 500 repeats of 10-fold cross-validation resampling; 2) 100 datasets were generated by randomly permuting the group labels (PM and NPM) and the same analysis performed for the original data set (i.e. building PLSDA models using the top six frequently selected antibody features found by resampling of elastic net) was repeated for each dataset.”

https://doi.org/10.7554/eLife.65776.sa2

Article and author information

Author details

  1. Elizabeth H Aitken

    Department of Medicine, University of Melbourne, the Doherty Institute, Melbourne, Australia
    Contribution
    Conceptualization, Data curation, Formal analysis, Supervision, Funding acquisition, Investigation, Visualization, Methodology, Writing - original draft, Project administration, Writing - review and editing
    Contributed equally with
    Timon Damelang and Amaya Ortega-Pajares
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-2677-6208
  2. Timon Damelang

    Department of Microbiology and Immunology, University of Melbourne, the Doherty Institute, Melbourne, Australia
    Contribution
    Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - review and editing
    Contributed equally with
    Elizabeth H Aitken and Amaya Ortega-Pajares
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-6150-4435
  3. Amaya Ortega-Pajares

    Department of Medicine, University of Melbourne, the Doherty Institute, Melbourne, Australia
    Contribution
    Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - review and editing
    Contributed equally with
    Elizabeth H Aitken and Timon Damelang
    Competing interests
    No competing interests declared
  4. Agersew Alemu

    Department of Medicine, University of Melbourne, the Doherty Institute, Melbourne, Australia
    Contribution
    Validation, Investigation, Methodology
    Competing interests
    No competing interests declared
  5. Wina Hasang

    Department of Medicine, University of Melbourne, the Doherty Institute, Melbourne, Australia
    Contribution
    Supervision, Validation, Investigation, Methodology, Project administration
    Competing interests
    No competing interests declared
  6. Saber Dini

    Centre for Epidemiology and Biostatistics, Melbourne School of Population and Global Health, University of Melbourne, Melbourne, Australia
    Contribution
    Data curation, Formal analysis, Investigation, Visualization, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared
  7. Holger W Unger

    1. Department of Medicine, University of Melbourne, the Doherty Institute, Melbourne, Australia
    2. Department of Obstetrics and Gynaecology, Royal Darwin Hospital, Darwin, Australia
    3. Menzies School of Health Research, Darwin, Australia
    Contribution
    Data curation, Writing - review and editing
    Competing interests
    No competing interests declared
  8. Maria Ome-Kaius

    Walter and Eliza Hall Institute of Medical Research, Parkville, Australia
    Contribution
    Data curation
    Competing interests
    No competing interests declared
  9. Morten A Nielsen

    Centre for Medical Parasitology, Department of Microbiology and immunology, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Resources, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-2668-4992
  10. Ali Salanti

    1. Centre for Medical Parasitology, Department of Microbiology and immunology, University of Copenhagen, Copenhagen, Denmark
    2. Department of Infectious Disease, Copenhagen University Hospital, Copenhagen, Denmark
    Contribution
    Resources, Writing - review and editing
    Competing interests
    No competing interests declared
  11. Joe Smith

    1. Seattle Children’s Research Institute, Seattle, United States
    2. Department of Pediatrics, University of Washington, Seattle, United States
    Contribution
    Resources, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7915-6360
  12. Stephen Kent

    Department of Microbiology and Immunology, University of Melbourne, the Doherty Institute, Melbourne, Australia
    Contribution
    Resources, Supervision, Writing - review and editing
    Competing interests
    No competing interests declared
  13. P Mark Hogarth

    1. Seattle Children’s Research Institute, Seattle, United States
    2. Immune Therapies Group, Centre for Biomedical Research, Burnet Institute, Melbourne, Australia
    3. Department of Clinical Pathology, University of Melbourne, Melbourne, Australia
    4. Department of Immunology and Pathology, Monash University, Melbourne, Australia
    Contribution
    Resources, Writing - review and editing
    Competing interests
    No competing interests declared
  14. Bruce D Wines

    1. Immune Therapies Group, Centre for Biomedical Research, Burnet Institute, Melbourne, Australia
    2. Department of Clinical Pathology, University of Melbourne, Melbourne, Australia
    3. Department of Immunology and Pathology, Monash University, Melbourne, Australia
    Contribution
    Resources, Writing - review and editing
    Competing interests
    No competing interests declared
  15. Julie A Simpson

    Centre for Epidemiology and Biostatistics, Melbourne School of Population and Global Health, University of Melbourne, Melbourne, Australia
    Contribution
    Formal analysis, Supervision, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-2660-2013
  16. Amy W Chung

    Department of Microbiology and Immunology, University of Melbourne, the Doherty Institute, Melbourne, Australia
    Contribution
    Conceptualization, Resources, Supervision, Funding acquisition, Methodology, Project administration, Writing - review and editing
    Contributed equally with
    Stephen J Rogerson
    Competing interests
    No competing interests declared
  17. Stephen J Rogerson

    Department of Medicine, University of Melbourne, the Doherty Institute, Melbourne, Australia
    Contribution
    Conceptualization, Resources, Supervision, Funding acquisition, Methodology, Writing - original draft, Project administration, Writing - review and editing
    Contributed equally with
    Amy W Chung
    For correspondence
    sroger@unimelb.edu.au
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-4287-1982

Funding

National Health and Medical Research Council (APP1143946)

  • Elizabeth H Aitken
  • Amy W Chung
  • Stephen J Rogerson

National Health and Medical Research Council (GNT1145303)

  • P Mark Hogarth
  • Bruce D Wines

National Health and Medical Research Council (APP1092789)

  • Stephen J Rogerson

National Health and Medical Research Council (APP1140509)

  • Amy W Chung

University of Melbourne

  • Amaya Ortega-Pajares

Australian Society for Parasitology

  • Elizabeth H Aitken

National Health and Medical Research Council (APP1104975)

  • Julie A Simpson

Bill and Melinda Gates Foundation (46099)

  • Stephen J Rogerson

Miller Foundation Australia

  • Amaya Ortega-Pajares

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

This work was supported by Project Grants from the National Health and Medical Research Council of Australia (NHMRC) to SJR, EHA, and AWC (APP1143946) and to PMH and BDW (GNT1145303) and a Program Grant (APP1092789) also from the NHMRC. AWC is supported by an NHMRC Career Development Fellowship (APP1140509). AO-P is supported by scholarships from the University of Melbourne and the Miller Foundation. Work by EHA while visiting the University of Copenhagen was supported by a grant from the Australian Society for Parasitology. JAS is supported by a NHMRC Senior Research Fellowship (APP1104975). Sample collection was primarily supported by the Malaria in Pregnancy Consortium, through a grant from the Bill and Melinda Gates Foundation (46099). We thank Patrick Duffy and David Narum, Laboratory of Malaria Immunology and Vaccinology (LMIV) at NIAID, NIH, for provision of recombinant VAR2CSA proteins, Anne Corfitz, Centre for Medical Parasitology, Copenhagen, for technical assistance, and Isobel Walker, Graham Brown, Siddhartha Mahanty, and Louise Randall for constructive feedback on the work. We acknowledge Vanta Jameson and Alexis Gonzalez from the Melbourne Cytometry Platform (Melbourne Brain Centre node) for assistance with flow cytometry. The funders had no role in the study design, data collection, analysis, and interpretation of the data, or in the decision to submit the work for publication.

Ethics

Clinical trial registration NCT01136850.

Human subjects: Collection and use of plasma samples from women in PNG was approved by the PNG Institute of Medical Research Institutional Review Board, the PNG Medical Research Advisory Council and the Melbourne Health Human Research Ethics Committee. All participants provided informed written consent. The use of blood products from donors in Melbourne for isolation of primary cells, culture of parasites and leukocytes and for use as negative controls was approved by the Melbourne Health Human Research Ethics committee and the University of Melbourne Human Research Ethics committee.

Senior Editor

  1. Dominique Soldati-Favre, University of Geneva, Switzerland

Reviewing Editor

  1. Urszula Krzych, Walter Reed Army Institute of Research, United States

Reviewer

  1. Sidhartha Chaudhury, Walter Reed Army Institute of Research, United States

Publication history

  1. Received: December 15, 2020
  2. Accepted: June 16, 2021
  3. Accepted Manuscript published: June 29, 2021 (version 1)
  4. Version of Record published: June 29, 2021 (version 2)
  5. Version of Record updated: August 16, 2021 (version 3)

Copyright

© 2021, Aitken et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 714
    Page views
  • 82
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Cell Biology
    2. Immunology and Inflammation
    Shannon Rausser et al.
    Research Article

    Using a high-throughput mitochondrial phenotyping platform to quantify multiple mitochondrial features among molecularly-defined immune cell subtypes, we quantify the natural variation in citrate synthase, mitochondrial DNA copy number (mtDNAcn), and respiratory chain enzymatic activities in human neutrophils, monocytes, B cells, and naïve and memory T lymphocyte subtypes. In mixed peripheral blood mononuclear cells (PBMCs) from the same individuals, we show to what extent mitochondrial measures are confounded by both cell type distributions and contaminating platelets. Cell subtype-specific measures among women and men spanning 4 decades of life indicate potential age- and sex-related differences, including an age-related elevation in mtDNAcn, which are masked or blunted in mixed PBMCs. Finally, a proof-of-concept, repeated-measures study in a single individual validates cell type differences and also reveals week-to-week changes in mitochondrial activities. Larger studies are required to validate and mechanistically extend these findings. These mitochondrial phenotyping data build upon established immunometabolic differences among leukocyte sub-populations, and provide foundational quantitative knowledge to develop interpretable blood-based assays of mitochondrial health.

    1. Immunology and Inflammation
    2. Microbiology and Infectious Disease
    Jason Neidleman et al.
    Research Article Updated

    While mRNA vaccines are proving highly efficacious against SARS-CoV-2, it is important to determine how booster doses and prior infection influence the immune defense they elicit, and whether they protect against variants. Focusing on the T cell response, we conducted a longitudinal study of infection-naïve and COVID-19 convalescent donors before vaccination and after their first and second vaccine doses, using a high-parameter CyTOF analysis to phenotype their SARS-CoV-2-specific T cells. Vaccine-elicited spike-specific T cells responded similarly to stimulation by spike epitopes from the ancestral, B.1.1.7 and B.1.351 variant strains, both in terms of cell numbers and phenotypes. In infection-naïve individuals, the second dose boosted the quantity and altered the phenotypic properties of SARS-CoV-2-specific T cells, while in convalescents the second dose changed neither. Spike-specific T cells from convalescent vaccinees differed strikingly from those of infection-naïve vaccinees, with phenotypic features suggesting superior long-term persistence and ability to home to the respiratory tract including the nasopharynx. These results provide reassurance that vaccine-elicited T cells respond robustly to emerging viral variants, confirm that convalescents may not need a second vaccine dose, and suggest that vaccinated convalescents may have more persistent nasopharynx-homing SARS-CoV-2-specific T cells compared to their infection-naïve counterparts.