Anti-malaria antibody engineering broadens recognition motifs and reveals new homotypic interactions that enhance protective breadth

Abstract

The monoclonal antibody L9 mediates high-level protection against malaria in children for up to 6 months in Africa. L9 preferentially binds with high affinity to a triplicate of NVDP-minor repeats on the P. falciparum circumsporozoite protein (PfCSP). Here, we sought to improve the affinity of L9 to enhance protection against rare strains with two spatially separated minor repeats or a single minor repeat. Site saturation mutagenesis and yeast display-screening identified a panel of affinity-improved variants. In vivo challenge showed one variant, L9_yd19, to be modestly more potent against a transgenic Plasmodium encoding chimeric CSP with two widely spaced minor repeats from a Kenyan parasite strain, with no loss in potency against the benchmark 3D7 strain with its standard complement of minor repeats. L9_yd19 also had high affinity against NANP-major repeats and was protective against transgenic Plasmodium expressing CSP with a NANP₁₂ knock-in lacking NVDP. Cryo-EM studies revealed L9_yd19 to recognize PfCSP with two distinct homotypic interfaces, which combined to yield two trimeric layers of antibodies comprising asymmetric trimers that dimerized in a head-to-head fashion. These data reveal a new antibody mechanism that utilizes interfaces involving dual homotypic symmetry elements, a 2-fold and an asymmetric 3-fold, for improved malaria prevention.

In this study, Chun & Tripathi et al engineered antibody variants against Plasmodium falciparum that showed improved binding and protective breadth. Protection was driven by a new homotypic antibody interaction interface that assembled higher-order antibody multimers against the parasite.

Introduction

Malaria presents a substantial burden on global human public health, causing over 263,000,000 cases and 597,000 deaths in 2023¹. The majority of deaths are in children under 5 years of age and result from Plasmodium falciparum infection, with morbidity and mortality concentrated in Africa. There are a number of tools, such as vaccines, mosquito control measures and preventive drug treatments that are used to control malaria. Monoclonal antibodies are a potential new intervention to prevent malaria infection and transmission^2–4. The circumsporozoite protein (CSP) is presented at high density on the surface of the sporozoite (SPZ) stage of the parasite life cycle and represents an important target for malaria prevention by monoclonal antibodies and vaccines^5–7. Monoclonal antibodies against CSP have been shown to be safe and mediate high-level protection against malaria in a number of recent Phase 2 studies in Africa^8–10. While these studies highlight the potential of monoclonal antibodies to control malaria, a key driver for their broad implementation will be to minimize their cost. Thus, active investigation is underway to generate more potent antibodies.

CSP is mostly comprised of repeated peptide motifs and appears not to have a single unique structure, with different recognizing antibodies inducing distinct conformations. In P. falciparum CSP, the dominant peptide repeats consist of an Asn-Ala-Asn-Pro (NANP) tetrapeptide for the major repeat and an Asn-Val-Asp-Pro (NVDP) for the minor repeating motif. Most P. falciparum strains, including the widely used 3D7 strain, contain three closely spaced minor repeats within the CSP central repeat region¹¹. Each minor repeat is commonly separated by a single major repeat, and CSP often has another distal minor repeat after a dozen or so major repeats. The L9 monoclonal antibody preferentially targets the cluster of three NVDP-minor repeats on P. falciparum CSP, utilizing a homotypic-formed asymmetric trimer^11,12. Thus, L9 mediates protection primarily through high-affinity interactions against the minor repeat. Other potent monoclonal antibodies in clinical development against PfCSP target the major repeat^13–15 or the junction region^5,16–18.

P. falciparum has polymorphisms in the Pfcsp gene, including variation in the numbers of major and minor repeats, as well as the spacing between them. Studies have shown that L9’s protective mechanism is reduced when only a single minor repeat is present, and thus L9 requires at least two minor repeats in the PfCSP molecule for full protection¹⁹. Structural studies also suggest that the L9 homotypic structures optimally recognize minor repeats that are in close proximity¹¹; still, L9 is substantially protective against strains with two spatially separated minor repeats. Very few strains have no NVDP repeats, and 8.4% of strains analyzed in a recent study (32/382 strains registered in GenBank in 2022, which may not reflect actual field distributions) have ≤ 1 NVDP repeat, against which L9 would be less effective¹¹. Thus, a modified version of L9 that could recognize rare strains with only a single minor repeat could enhance breadth across all strains.

Antibody engineering is a useful tool to improve antibody breadth and potency in other systems. Our prior efforts showed that directed evolution via yeast display could be useful to improve antibodies for enhanced performance against infectious diseases, including for antibodies targeting the fusion peptide of HIV-1^20,21 and the junctional peptide motif of PfCSP¹⁶. In this study, we hypothesized that systematic mutagenesis and high-throughput screening could be used to evolve versions of L9 with sufficient affinity to recognize strains with minor repeat variations compared to the benchmark 3D7 strain. Such an evolved variant would have sufficient breadth to protect against malaria variants with potential resistance to the L9 antibody.

Engineering the L9 antibody via yeast display, we found that several mutations on L9 that could improve affinity to PfCSP peptides, which encompassed minor and major motifs, and one of the resulting variants provided improved protection against transgenic Plasmodium strains with one or no minor repeats. Structural and sequence analysis revealed this variant to have increased affinity to major repeats and alterations that introduced a second homotypic interaction leading to head-to-head dimerization. These data reveal a new structural pathway through which to develop effective anti-malaria biologics against the repeat regions of P. falciparum CSP.

Results

Directed evolution to improve L9 affinity against selected PfCSP antigens

L9 binds to PfCSP predominantly through the NVDP minor repeat, with lower affinity contacts to the NANP major repeat that are not sufficient for full protection (Fig. 1A)¹⁹. We sought to further improve L9’s affinity to its targeted epitopes, selecting first for affinity to Peptide22 (Pep22), the sequence motif used to define L9 specificity, which contains two NVDP minor repeats and two NANP major repeats⁷. We also selected libraries for maintained binding to NTDS_5/3, a truncated version of the PfCSP molecule that could present the same peptide antigens in a more native context as probes (Fig. 1B)¹⁶. We employed a stepwise approach to generate L9-variant libraries, using several rounds of yeast display screening to progressively enrich for variants with improved affinity to Pep22, and also for preserved binding to NTDS_5/3 (Fig. 2A). First, we generated site-saturation mutagenesis (SSM) libraries encompassing every possible single amino acid mutation on the heavy (VH) and light (VL) chains of L9²² (Supplementary Data S1, S2). The SSM libraries were cloned into yeast surface display and labeled for FACS with Pep22 (0.5 nM), which contains two major NANP and two minor NVDP motifs. Sorted L9-SSM libraries were analyzed by flow cytometry, with improved affinity validated using a high-affinity gate that measured the ratio of Pep22 binding (PE geometric mean fluorescence intensity, PE gMFI) to Fab display (FITC geometric mean fluorescence intensity, FITC gMFI). This gating strategy normalized antigen fluorescence (antigen PE labels) by the amount of displayed Fabs on each yeast cell (anti-VL FITC), and enabled affinity screening from yeast libraries that contained varying levels of Fab surface display²³. After three rounds of affinity-based FACS, the enriched populations had a higher PE/FITC gMFI compared with the Pre-sort libraries and wild-type (L9_WT), including for binding to peptide antigens and the NTDS_5/3 truncated PfCSP (Fig. 2B and Supplementary Figs. S1–S3).

Fig. 1. Improving affinity to target the CSP minor repeat enhances L9 antibody protective breadth.

A The CSP minor repeat (NVDP) is the principal binding target of the potent anti-malarial antibody L9. Genetic polymorphisms are present among P. falciparum strains in the number and spacing of NVDP motifs contained in the CSP repeat regions, and the loss of NPNV epitopes reduces the in vivo protective potency of the L9 antibody. B To enhance L9 affinity against the CSP antigen, multi-mutation libraries were generated from the L9 template and selectively screened against CSP peptide repeat motifs using yeast display. Libraries were first selected against Peptide22 (Pep22), which contains two minor and two major repeats, and subsequently selected against NTDS_5/3, a truncated CSP molecule that presents peptides in a more native-like conformation. Identified L9 variants were characterized for binding affinity in vitro and for protective potency in vivo against a panel of malaria antigens.

Fig. 2. Library generation and yeast display screening evolves new L9 variants with improved affinity to Peptide22.

A The L9 antibody was used as a template for multiple rounds of directed evolution for affinity improvement against malaria antigens. First, site-saturation mutagenesis (SSM) and sorting was used against 0.5 nM Pep22 (2 NVDPs). Next, enriched variants further diversified using SSM and DNA shuffling, followed by sorting against 1.25 nM Pep22. Finally, the enriched libraries were shuffled again and sorted under increasingly stringent conditions by progressively lowering Pep22 concentrations during FACS, with a final enrichment step against NTDS_5/3 (a truncated CSP with 5 NANP / 3 NVDP repeats). B Representative flow cytometry plots across the screening experiments. Each population was labeled with 62.5 pM Pep22. A high-affinity gate is drawn for comparative purposes, and the percentage of the high-affinity population is shown in the upper-right quadrant. A PE/FITC geometric mean fluorescence intensity ratio (MFI) is calculated as (PE gMFI / FITC gMFI) to quantitatively represent the normalized affinity of Fab molecules. See also Supplementary Data S1, S2 and Supplementary Figs. S1–S3. Source data are provided as a Source Data file.

Next, we performed two rounds of shuffling mutagenesis to the enriched population of single-mutation variants, combining another round of SSM along with DNA shuffling to generate multi-mutant libraries^24,25. Yeast display libraries of the 1^st shuffle mutagenesis were labeled with Pep22 (1.25 nM) for FACS. After the 1^st shuffle mutagenesis, the enriched population showed at least 4-fold higher MFI than Pre-sort populations. Libraries were enriched against Pep22 such that the majority of the Fab-displaying population bound to Pep22 and NTDS_5/3 even when analyzed with antigen at a low concentration of 62.5 pM, where much of the wild-type population was not binding antigen (Fig. 2B and Supplementary Figs. S2, S3). We iterated the shuffle mutagenesis once more to further identify multi-mutation L9 variants. Due to the very large potential library sizes that can be achieved by DNA shuffle mutagenesis²⁶, we first performed VL + MACS to remove non-displaying Fab clones prior to affinity enrichment by FACS. We also lowered the Pep22-labeling concentration again and screened the libraries using NTDS_5/3 in the last round of FACS sorting to ensure that selected variants would bind the major and/or minor repeat peptides when presented in the context of a CSP molecule (Fig. 2A). The libraries were nearly completely enriched with Pep22-binding populations and also showed a higher MFI against NTDS_5/3 than L9_WT, which suggested that the sorted libraries were enhanced in their affinity against both Pep22 and NTDS_5/3 compared to that of the already high-affinity L9_WT molecule (Supplementary Figs. S2, S3).

Bioinformatic mining of affinity-enhanced L9 variants

We assessed the genetic diversity of enriched libraries across each round of screen by next-generation sequencing (NGS). L9 variants were selected for in vitro characterization based on the prevalence of a given mutation or sequence in enriched library populations and also based on the enrichment ratio (ER) of L9 variants in the last round (Round 3) of screening. Every library was sequenced using the Illumina MiSeq platform. For the Round 3 libraries, we also used the Pacific Biosciences Sequel II platform that offers longer read length for complete paired VH:VL sequences, albeit with lower read counts than Illumina MiSeq. We combined the Sequel II data with MiSeq data for higher throughput of separate heavy and light chains²⁷. Due to higher insertions/deletion error rates at homopolymer regions in Sequel II data, we cross-checked all selected sequences by evaluating the separate VH and VL sequences from MiSeq data using the same criteria, and we identified 36 paired VH:VL sequences for expression. There were 32 different VH mutations and 15 different VL mutations present in the selected variants, among which 19 VH mutations and 3 VL mutations were observed only in the last round of DNA shuffling and screening, including 5 unique single amino acid mutations; VH_S25D, VH_T69E, VH_D73R, VL_F80_HW, and VL_G57_LH (template numbering, see Supplementary Data S3). At the single amino acid level, six more mutations (VH_V50A, VH_I58T, VH_M81L, VH_G103R, VL_Q27R, and VL_Q89A) consistently appeared to be most prevalent from the variants screened in DNA shuffling rounds. Interestingly, one VH_F53G mutation was strongly enriched in the first step of DNA shuffling to being contained in 58.7% of sequences, then reduced to 1.8% after the second round of shuffling, which implied that it conferred superior properties to L9_WT but was out-competed in the later screening rounds (Fig. 3A, B). While some of the most prevalent mutations occurred in complementarity-determining regions (CDRs), the mutations identified were not limited to the CDRs (Fig. 3). Because most individual mutations were already showing significant enrichment after the first round of DNA shuffling, the most prevalent mutations going into later screening rounds did not show strong enrichment (Fig. 3A, B). Instead, the later rounds selected for optimal combinations of amino acid mutations and depleted the libraries of mutations that provided an advantage over wild-type L9 but that were outcompeted by other clones, such as VH_F53G. The selected enriched variants after three rounds of directed evolution were then expressed as soluble antibodies for characterization.

Fig. 3. Bioinformatic analysis reveals affinity-enhanced VH:VL sequences after Round 3 (R3) of the 2nd round of shuffle mutagenesis.

The round 3 (R3) enrichment ratio (left) and prevalence (right) for amino acid mutations identified from (A) VH sequences and (B) VL sequences. Wild-type amino acids are indicated with dots (·). On the enrichment ratio panel, mutations highlighted in red are enriched, and those in blue are depleted. C Mutational pairings between single amino acid mutations identified from VH (left half) and VL (right half) sequences. The amino acid residues followed template numbering. See also Supplementary Fig. S4. Source data are provided as a Source Data file.

Binding analysis and in vivo protection by L9 variants

We first tested the 36 selected L9-variant antibodies using biolayer interferometry (BLI) to quantify affinity improvements (Fig. 4A and Supplementary Data S3). In close agreement with the improved gMFI ratios that we observed in flow cytometry (antigen PE / anti-VL FITC), we found that nearly all the selected variants had improved apparent affinity against Pep22, with one variant (yd01) showing enhanced apparent affinity by approximately 100-fold. Nearly all variants maintained binding affinity to the full-length of the benchmark 3D7 Pf strain (fl_CSP), although only a select few variants showed improved affinity to 3D7 fl_CSP.

Fig. 4. In vitro and in vivo characterization reveals functional improvement for apparent affinity-enhanced L9 variants against multiple transgenic strains, including the Kenya KI strain and a major repeat-only transgenic.

A Biolayer interferometry (BLI) analysis against Peptide22 (containing 2 NVDPs) and full-length CSP (fl_CSP) of P. falciparum 3D7. L9_WT affinity is marked with a vertical line on the legend bar. L9 variants marked with an asterisk (*) were selected for detailed in vivo characterization. B In the transgenic sporozoite intravenous (IV) challenge model, mice were administered IgG from L9 variants or control antibodies (− 2 h), followed by IV challenge with P. berghei sporozoites (PbSPZ) encoding different PfCSP knock-in sequences (0 h), and subsequent luciferin-based evaluation (42 h). Each group comprised n = 5 mice, with animals receiving the indicated antibody dose. C Transgenic PbSPZ challenge (n = 10). Liver infection burden was quantified from two independently replicated assays (n = 5), and combined data are shown as mean ± standard deviation after log-transformation; mean liver burden total flux for L9_WT at 12.5 µg/mouse is indicated by a dashed line. Statistical significance between L9_WT and each L9 variant group at 12.5 µg were determined using the ordinary one-way ANOVA test with a two-tailed P value calculation and a Dunnett’s multiple test correction (L9_WT vs L9_yd19: p = 0.050 for Kenya KI, p = 0.025 for 1 NVDP KI, p = 0.00004 for NANP₁₂ KI; L9_WT vs L9_yd07: p = 0.0003 for NANP₁₂ KI). Statistical significance was also analyzed between L9_yd19 and MAM01 using an unpaired t test with a two-tailed P- value calculation; the two groups were significantly different when tested against 1 NVDP KI PbSPZ (p = 0.041). See also Data S3 and Fig. S5. Source data are provided as a Source Data file.

Based on these affinity data, we selected a panel of seven of the improved variants for in vivo challenge analysis. In these models, a luciferase-expressing transgenic P. berghei strain was engineered to express 3D7 fl_CSP or a chimeric CSP containing knock-in (KI) sequences from PfCSP. Mice were challenged intravenously with the engineered transgenic P. berghei sporozoites (PbSPZs) labeled with luciferase, and parasite burden in the liver was quantified^16,19,28. For initial analysis we used two transgenic PbSPZ strains: (i) a Kenya KI strain harboring PfCSP-derived repeat sequences within the endogenous PbCSP, resulting in two NVDP minor repeats separated by six NANP major repeats (NVDP-NANP₆-NVDP), modeled on Kenyan isolates, and (ii) a strain in which the endogenous PbCSP was replaced with full-length PfCSP from the benchmark 3D7 isolate, which contains three closely spaced NVDP minor repeats and an additional widely spaced NVDP repeat (Fig. 4B). With a 25 microgram/mouse antibody dose, L9 variants were comparable to L9_WT in mediating reduction of liver burden against the Kenya or 3D7 knock-in strains (Supplementary Fig. S5).

We selected three variants (yd01, yd07, and yd19) for broader characterization against a panel of transgenic strains based on their ~1-log reduction in liver burden in 25-microgram dose assays, and for their diversity of affinity phenotypes. yd01 had the highest Pep22 apparent affinity and lower fl_CSP apparent affinity than L9_WT; yd07 had moderately improved apparent affinity to both Pep22 and fl_CSP; and yd19 had some improved apparent affinity to Pep22 and the very highest improvement against fl_CSP (Supplementary Data S3). These three variants all showed in vivo protection against the Kenya KI PbSPZ at 12.5 microgram/mouse, with a modest yet statistically significant improvement for L9_yd19 compared to L9_WT (Fig. 4C). As an additional comparison, MAM01, an antibody with high affinity binding to major repeats, showed a comparable reduction of liver burden at the same dose.

We next sought to understand how the reduction in liver burden by yd19 and the other selected variants was affected by changes in the number and arrangement of the NVDP minor repeats. The Kenya KI strain has two separated NVDP motifs. Thus, we also tested the antibody variants against transgenic chimeric parasites bearing two adjacent NVDP motifs (2 NVDP KI) or with just a single NVDP motif (1 NVDP KI) (Fig. 4B). While there were no significant differences in liver burden against the 2 NVDP KI strain across the L9 variants, L9_yd19 showed the lowest liver burden amongst all variants tested against the 1 NVDP KI strain (p = 0.025 vs. L9_WT), including improved performance as compared to MAM01 (p = 0.041), indicating that mutations in yd19 alleviated the loss in potency known to occur for L9 against parasites with only a single NVDP motif (Fig. 4C). Finally, we analyzed the performance of all three variants against the NANP₁₂ KI strain, which contains only NANP motifs and no NVDP motifs (Fig. 4B). L9_yd19 showed significantly reduced liver burden against the NANP₁₂ KI strain compared to L9_WT (p < 0.001), with no statistically significant difference compared to that of the malaria-protective antibody MAM01, which binds preferentially to the major repeat (Fig. 4C)¹³. These in vivo challenge data demonstrated that the mutations selected by yeast display were effective in enhancing L9 affinity, and also in broadening its targeted epitope to include both the NVDP minor repeat, as well as the NANP major repeat regions.

Functional correlates of antigen affinity with protective potency across malaria knock-in strains

We sought to understand the functional and affinity-based correlates of protection for the knock-in strains used. In these affinity assays, we evaluated L9_WT and the three variants yd01, yd07, and yd19 (selected across a broad affinity range for detailed testing in challenge studies). We used a broad panel of CSP antigens that included PfCSP-Kenya, NTDS_5/3, PfCSP-3D7 (Fig. 5 and Supplementary Data S4), and also the antigens Pep22 (major and minor repeat), Pep21 (junction region, major repeat, minor repeat), and Pep29 (major repeat only)⁷ (Supplementary Fig. S6 and Supplementary Data S4). We found that protection against the NANP₁₂ KI strain was correlated with affinity to PfCSP-Kenya (p = 0.036) and with K_D2,app of PfCSP-3D7 (p = 0.015), while protection against 2 NVDP KI PbSPZ was correlated with PfCSP-3D7 K_D2,app (p = 0.012). Interestingly, protection against these transgenic strains was not as closely correlated with affinity to NTDS_5/3, or with K_D1,app of PfCSP-3D7 (Fig. 5 and Supplementary Data S4). Peptide affinity correlates did not show significance, although several correlations had strong R² values (Supplementary Fig. S6 and Supplementary Data S4). Neither peptide affinities nor PfCSP affinities were effectively correlated with protection against the Kenya KI strain (Supplementary Fig. S7 and Supplementary Data S4).

Fig. 5. Apparent affinity correlations reveal that improved protection breadth correlates with affinity to multiple malaria antigens.

Correlation between apparent affinity to PfCSP Kenya strain, NTDS_5/3 or PfCSP 3D7 strain (x-axis, linear scale) and normalized liver burden against various transgenic PbSPZs (y-axis, linear scale). Pearson’s correlation coefficient (R) was calculated using the mean normalized liver burden values, with corresponding two-tailed P-values and 95% confidence intervals. A linear regression was fit based on a Pearson’s correlation analysis. See also Supplementary Figs. S6, S7 and Supplementary Data S4. Source data are provided as a Source Data file.

Cryo-EM structures for L9_yd19 in complexes with PfCSP antigens

To gain mechanistic insights into the enhanced protective efficacy of L9_yd19 against the Kenya KI strain, we sought to obtain its cryo-EM structure in complex with both PfCSP-3D7 and PfCSP-Kenya. L9_yd19 antigen binding fragment (Fab) was complexed with PfCSP (3D7 or Kenya) at a 2:1 molar ratio, and single particle cryo-EM data collected (Supplementary Figs. S8, S9 and Supplementary Data S5).

The 2D classification of L9_yd19:PfCSP-3D7 showed a major class consisting of three L9_yd19 Fabs bound to PfCSP-3D7 (Fig. 6A, upper left panel, Supplementary Fig. S8), similar to L9_WT in complex with PfCSP-3D7^11,12. A small number of classes were also observed that showed more than three L9_yd19 Fabs bound to PfCSP-3D7, however, a high-resolution structure could not be obtained because of the low overall number of particles. The structure of the three Fabs bound complex was obtained at a resolution of 3.6 Å (Fig. 6A, upper right panel). The density for 27 residues of PfCSP-3D7 was observed, and the rest of the CSP was disordered. L9_yd19 had five mutations relative to L9_WT; two in heavy chain (V50A, I57T) and three in light chain (Q27R, Q89A, K103T) (Supplementary Fig. S10). The V50 in CDRH2 and Q89 in CDRL3 surround the CDRL3 loop, which is critical for the interactions of L9’s preferred epitope (NPNV) with the light chain. The mutation of both residues to smaller residues (V50A, Q89A) may provide additional flexibility to the CDRL3 loop for a more favorable interaction with the NPNV epitope (Fig. 6A, lower panel).

Fig. 6. Structural analyses of L9_yd19 and PfCSP (3D7 or Kenya) reveals a new interaction mode.

A Cryo-EM structure of L9_yd19 in complex with PfCSP (3D7) at 3.6 Å resolution. Heavy chain mutations are shown in red, and light chain mutations are shown in blue. A 27-residue stretch of PfCSP is ordered, and the rest of the PfCSP is disordered. B Cryo-EM structure of L9_yd19 in complex with PfCSP (Kenya) at 3.3 Å resolution. 2D classes show both three L9_yd19 Fabs (orange arrow) and six L9_yd19 Fabs (red arrow) bound to PfCSP (Kenya). 3D reconstruction for six L9_yd19 Fabs bound to PfCSP (Kenya) is shown. See also Supplementary Figs. S8–S10, and Supplementary Data S5.

The 2D classification of L9_yd19 in complex with PfCSP-Kenya showed an alternate binding mode with the majority of 2D classes containing more than the previously observed triplicate of three Fabs, although a few classes of Fab triplicates were also observed (Fig. 6B, upper panel, Supplementary Fig. S9). 3D reconstruction yielded a 3.3 Å resolution map with six L9_yd19 Fabs bound to a 53-residue stretch of PfCSP-Kenya. Since the preferred high-affinity epitope (NPNV) is separated by 6 NANPs in PfCSP-Kenya, this higher-order complex utilized the lower-affinity epitope (NPNA) to form two layers of Fab trimers stacked on top of each other (Fig. 6B, bottom panel), one layer recognizing minor-major-major repeats and the other layer recognizing another sequence of minor-major-major repeats. This unique binding modality of L9_yd19 thus utilizes additional Fab-Fab homotypic interactions that enabled enhanced affinity and protective efficacy against the Kenya transgenic strain. Overall, while L9_yd19 displayed a canonical three Fab binding mode with PfCSP-3D7, the cryo-EM analysis also revealed a new binding mode in which six L9-yd19 Fabs bound to a combination of minor and major repeats.

L9_yd19 engineered mutations enable new Fab-Fab homotypic interactions

To identify the critical residues in L9_yd19 that enabled additional homotypic interactions, we analyzed structural differences between the canonical three Fab complex (3D7) versus the newly observed six Fab-CSP complex (Kenya) (Fig. 7A). The analyses revealed Fab-1 (numbered based on distance from the N-term of CSP) to bind the preferred high-affinity epitope (NPNV), anchoring the complex. The subsequent Fabs (2 and 3) bound to a secondary low-affinity epitope (NPNA) separated by a four-residue linker (D/NPNA) and potentially stabilized through side-to-side Fab-Fab homotypic interactions in the atypical trimer context. These interactions utilize homotypic contacts between the heavy chain of one Fab to the light chain of an adjacent Fab (Fig. 7B, upper left panel).

Fig. 7. L9_yd19 engineered mutations stabilize a new head-to-head homotypic interface allowing up to six Fabs to bind to PfCSP (Kenya).

A Structural comparison of L9_yd19 bound to PfCSP-3D7 and PfCSP-Kenya. PfCSP is shown in red sticks with the N- and C- termini labeled. B L9_yd19 in complex with PfCSP-Kenya. Fabs (3,4), (2,5), and (1,6) are related by 2-fold (C2) symmetry. Critical residues generating the head-to-head homotypic interface are shown in insets. See also Supplementary Fig. S11.

The engineered mutations in L9_yd19, specifically VL_Q27R and VH_I57T, facilitated a second layer of Fab trimer comprising of three Fabs (4–6) to stack on top of the first layer of Fab trimer (1–3). The Fab-3 and Fab-4 were separated by an eight-residue linker (NPNPNPNA), and this arrangement allowed Fab-4 to bind the primary high-affinity epitope (NPNV) and to act as an anchor for the upper layer of Fab trimer. In the upper layer, Fabs (4–6) interacted in the same fashion as Fabs in the bottom layer trimer. A unique head-to-head Fab-Fab interaction was observed between the Fabs from the upper and the lower layer (Fig. 7B, top right panel). This novel arrangement allowed additional inter-Fab contacts and stabilized the higher-order complex.

Symmetry analysis revealed that the upper layer Fabs exhibited a two-fold (C2) symmetry with the lower layer Fabs (Fab-4 and -3, Fab-5 and −2, Fab-6 and -1) (Fig. 7B, lower left panel). This head-to-head homotypic interface was stabilized by polar interactions between both heavy and light chains of the two Fabs. VL_D1 and VH_D61 from Fab-3 formed a reciprocal salt bridge with VH_K64 on Fab-4, which created a symmetric pair of salt bridges between Fabs 3 and 4 and further stabilized the interface. The introduced mutations in L9_yd19 (Q27R_L and I57T_H) also stabilized this interface via hydrogen bond interactions between VL_Q27R and the backbone carbonyl of VH_N56 and a polar interaction between VL-Q27R and VH_I57T (Fig. 7B, lower right panel). Similar set of interactions are observed between Fabs (2 and 5), and Fabs (1 and 6).

While structural analysis defined the location of Fabs, the location of the connecting constant regions needed to be inferred using geometric constraints. Within each recognition layer, the three Fabs exhibited distances between the C-terminal ends of the Fab heavy chains (C-terminus of CH1) ranging from 93–107 Å. These separations exceed the span typically accommodated by a single IgG1 constant region^11,12, although they may be accessible given sufficient hinge extension or the longer hinge of IgG3²⁹. Similarly, CH1-CH1 distances between Fabs 3-4, 2-5, and 1-6 ranged from 106-108 Å and would be difficult to accommodate within a single IgG1 molecule without substantial hinge extension.

To assess whether the observed Fab arrangements are compatible with intact antibodies, we modeled full-length IgG1 molecules (PDB: 1HZH) onto each Fab position by rigid-body alignment. This analysis revealed that six intact IgG1 molecules can be accommodated simultaneously without steric clashes (Supplementary Fig. S11). Incorporating hinge flexibility and variable Fc orientations, this geometric modeling data revealed that the observed multimeric Fab-antigen complex structures could be generated by full-length IgG1.

Altogether, these findings reveal how L9_yd19 circumvents the spatial requirements of L9_WT, which requires two high-affinity epitopes (NPNV) to be separated by a four-residue linker for optimal potency, via the formation of higher-order complexes through head-to-head homotypic interactions.

Discussion

Malaria continues to be a global threat to public health, and efficient, low-cost interventions are urgently needed to complement existing control measures. Here, using yeast display and site-saturation mutagenesis, we engineered the potent L9 antibody variant, L9_yd19 for improved apparent affinity against PfCSP. We found that L9_yd19 had enhanced apparent affinity to the NANP major repeat compared with L9_WT, while also showing enhanced apparent affinity to the NVDP minor repeat. L9_yd19 mediated reduction in liver burden with transgenic P. berghei parasites with chimeric CSP containing only NANP repeats, which may expand breadth against rare strains with a single NVDP minor repeat or improve potency against strains with two minor repeats that are spaced.

Cryo-EM analyses provided a mechanistic framework to study the structural modifications associated with the broader observed protection of L9_yd19. Enhanced NANP-only epitope recognition appears to result from a unique binding mode that relies on a new form of head-to-head homotypic Fab-Fab interactions. These previously unreported Fab-Fab interactions complex two of the previously reported trimeric Fab structures with CSP in between them¹¹. Cryo-EM data show that higher-order complexes can be formed with the canonical 3D7 strain, or even with entirely major repeat epitopes, as in the protective efficacy of L9_yd19 against the NANP₁₂ KI strain. The engineered mutations in L9_yd19 promoted formation of higher-order complexes by stabilizing the head-to-head homotypic interface via polar interactions. This unique recognition mode, with two distinct homotypic interfaces, enabled L9_yd19 to achieve enhanced protective efficacy against transgenic P. berghei strains that lacked the optimal high-affinity minor repeat epitope or preferred spacing. In cryo-EM reconstructions from 2D class averages containing the full complement of six bound Fabs, interpretable density was observed for a 53-residue segment of CSP, whereas the remaining CSP regions were not resolved because of conformational disorder. However, reconstructions for incomplete complexes with fewer than six bound Fabs were limited by the well-described disordered regions of the CSP central repeats, and thus some uncertainty remains related to the conformations of the unbound portions of the CSP repeats in structures with fewer than six bound Fabs.

An important consideration is how short screening antigens such as Pep22 and NTDS_5/3, which do not present sufficient epitopes for full hexameric assemblies like those observed in complex with Kenya CSP, could enrich for variants capable of forming extended homotypic complexes. In the yeast display system used in this study, Fab fragments are linked to the cell surface through Aga1p–Aga2p interactions, where the Fab is fused to Aga2p via a flexible linker³⁰. The displayed Fabs are therefore not fixed to a single orientation and remain relatively flexible on the cell surface. Moreover, each yeast cell displays a high density of Fab molecules, typically on the order of 10⁴–10⁵ copies per cell³¹. Given these display characteristics, although Pep22 can accommodate a maximum of two bound Fabs and NTDS_5/3 up to three, the improved affinity of the selected variants to the major repeat enables homotypic interactions against these antigens in both side-to-side and head-to-head orientations. Furthermore, even for two homotypically interacting Fabs bound to separate short peptide antigens, homotypic contacts could stabilize the bound structural conformations, providing an enthalpic benefit to higher-order complex assembly on the yeast surface¹¹. Thus, the high Fab display density and orientational flexibility of the yeast surface display system may have enabled selection for variants with homotypic interaction propensity even when screening against short model peptides.

Although high apparent affinity to the PfCSP core epitope has been identified as a prerequisite for in vivo potency³², protective efficacy by monoclonal antibodies against various epitopes of CSP may not be directly correlated with apparent affinity^14,32, suggesting that distinct structural binding modes may be important to inhibit sporozoite activity. The present study provides further information for how avidity and conformation dynamics influence antibody protection against sporozoite challenge. The importance of conformation is illustrated in our data related to antibody variants L9_yd07 and L9_yd19. Compared to L9_yd19, the antibody L9_yd07 had: tighter Peptide 22 affinity (nearly 10-fold), tighter Peptide 21 affinity ( ~ 2-fold), and a slightly tighter but comparable affinity to Peptide 29. Despite its tighter or equivalent affinity to all linear peptide epitopes investigated, L9_yd07 showed generally lower protection than L9_yd19. Remarkably, despite its tighter affinity to the shorter peptides, L9_yd07 had 20-fold lower apparent affinity to the Kenya CSP compared to L9_yd19, and lower protection accordingly. L9_yd07 protection against the Kenya KI strain was similar to wild-type L9, despite L9_yd07 also having markedly improved affinity compared to wild-type against constituent Peptides 22/21/29 by 30-, 6-, and 14-fold, respectively, and also 3-fold tighter affinity against Kenya CSP compared to WT. For L9_yd07, tighter affinity to all three peptides and also full-length CSP did not translate to higher protection.

In a similar case for the NANP₁₂ antigen, L9_yd01 showed a 14-fold improved apparent affinity compared to L9_WT against Peptide 29 (approximately equivalent Peptide 29 affinity as L9_yd19), and yet was not as protective as L9_yd19 against the NANP₁₂ KI strain, nor did L9_yd01 show a statistically significant difference in protection compared to L9_WT. The protection data for L9_yd01 against the NANP₁₂ KI strain indicate that the improved protective breadth observed for L9_yd19 cannot be recapitulated by enhanced affinity to Peptide 29 alone. The unique ability of L9_yd19 to recognize full-length CSP of diverse strains with very high avidity, potentially as a result of its unique multi-axis homotypic binding interactions, enabled the broad performance of the L9_yd19 molecule that was not observed in other molecules that showed equivalent or higher avidity against all three linear CSP epitopes.

Our conclusions regarding differences in protection are limited to those with detectable statistical significance at the group sizes used in the study (n = 10 mice per passive transfer mAb group). All in vivo protection data in this study were derived from mouse models using transgenic P. berghei expressing P. falciparum CSP. This mouse model has become a standard preclinical approach to evaluate the anti-sporozoite infection process, but may not fully recapitulate human malaria infection dynamics, immune responses, or pharmacokinetics. Follow-up studies using a mosquito bite challenge to ensure sterile protection to prevent blood stage infection and assessment in humanized models using P. falciparum infection would provide important additional data to support clinical development of any new variant monoclonal antibody. Clinical development for any improved variants will also require evaluation for potential immunogenicity, stability, safety/off-target effects, and pharmacokinetics, similar to the developmental progress already reported for the template antibody L9⁸.

We note that strains with fewer and/or more sparse NVDP epitopes are rare in sequenced malaria genotypes¹¹. It may be significant that the L9_yd19 antibody can reduce liver burden in mouse models with NANP repeat-only KI strains, or modestly reduce liver burden with chimeric transgenic strains containing a single NVDP minor repeat, potentially allowing for expanded breadth across all strains. L9 variants utilizing the dual homotypic mechanism of recognition revealed by L9-yd19 could potentially be further improved by increasing major repeat recognition or by utilizing IgG3-linkers instead of IgG1.

In conclusion, we show here that directed evolution via yeast display provided improved variants for a very high-affinity antibody, and unveiled a new structural pathway for anti-malarial antibodies with improved protection against Plasmodium expressing globally circulating PfCSP variants. Because prior studies showed no detectable protective combinations of mAbs against the central repeat region³³, this study focused on improved L9 variants as single monoclonal treatments. We note that scalability and cost of producing engineered antibodies are critical considerations for global therapeutic use, and full clinical benefits of anti-CSP mAbs can only be achieved by sufficiently low-cost production to support availability in malaria-endemic regions. The conformational affinity- and structure-based correlates of protection identified here can be used to further accelerate progress in antibody engineering against global P. falciparum strains, especially for potency improvements that could reduce the quantity of active product or extend the duration of protection conferred by a single dose. The data shown here provide a new window of opportunity for antibody development against PfCSP and introduce new broadly protective antibodies that could efficiently prevent new malaria cases around the globe.

Methods

Site-saturation mutagenesis library generation and cloning

DNA libraries containing all possible single mutations to the L9 template antibody were constructed on the VH and VL (kappa) variable region genes using a previously reported one-pot site saturation mutagenesis (SSM) protocol^16,20–22. Pools of mutagenic primers containing degenerate single codons were designed using QuikChange® Primer Design Program (Agilent) to cover all possible amino acids at each residue (Data S1-S2) and synthesized through Integrated DNA Technologies. As previously reported, a single-strand nick was introduced, one strand of the plasmid was digested, and the remaining single-stranded circular DNA was used as a template for codon-targeted site mutagenesis. After primer annealing, the second strand was generated to yield dsDNA plasmids ready for transformation of electrocompetent bacterial cells. E. coli libraries were transformed, and colony counts were used to estimate the transformed library size. Plasmid DNA libraries were isolated using a Maxi-Prep kit (Zymo Research), and expression cassettes and the yeast display pCT plasmid backbone were transformed into the yeast strain AWY101 using a high-efficiency homologous recombination, as previously reported¹⁶. The resulting yeast libraries were passaged twice in SDCAA media (20 g/L dextrose, 6.7 g/L yeast nitrogen base, 5 g/L casamino acids, 8.6 g/L NaH₂PO₄.H₂O, and 5.4 g/L Na₂HPO₄; SDCAA from Teknova) at 30 ^oC and 225 rpm for 12 to 24 h to ensure that individual yeast cells maintain a single copy of the plasmid. These initial libraries expressed in yeast cells are referred to as the Pre-sort libraries.

Single-mutation yeast display library screening via FACS

Expression was induced in yeast antibody libraries, and the yeast were stained with antigen for FACS as in prior studies^16,20. Affinity evaluation and fluorescence-activated cell sorting (FACS) were performed using a SONY MA900 flow cytometer, as in prior reports^23,27. Yeast display libraries were incubated in SGDCAA induction medium (2 g/L dextrose supplemented to 20 g/L D-(+)-galactose, 6.7 g/L yeast nitrogen base, 5 g/L casamino acids, 8.6 g/L NaH₂PO₄.H₂O, and 5.4 g/L Na₂HPO₄; SGCAA from Teknova) at 20 ^°C and 225 rpm for 36 hours to induce expression and surface display of Fabs. We refer to Fab-displaying yeast cells in a library as VL + populations, because the FLAG tag is located on the same polypeptide as the VL region. Induced yeast libraries were washed twice with ice-cold staining buffer (1 x PBS with 0.5% BSA and 2 mM EDTA) and labeled with an anti-FLAG FITC Clone M2 monoclonal antibody (F4049, Sigma-Aldrich) at 1 μg/mL, and with 0.5 nM biotinylated Peptide22 (Pep22, NANPNVDPNANPNVD)⁷ conjugated to streptavidin-PE (S21388, Invitrogen) at a 2:1 molar ratio, to enable VL + population quantification and affinity evaluation, respectively. In the first round of FACS, 3E + 07 yeast cells from Pre-sort libraries were gated into high-affinity groups based on the ratio of Fab display to antigen binding, as described previously^16,20. High-affinity populations were collected for subsequent rounds of affinity enrichment. Sorted yeast cells were expanded in low-pH SDCAA (20 g/L dextrose, 6.7 g/L yeast nitrogen base, 5 g/L casamino acids, 10.4 g/L trisodium citrate, 7.4 g/L citric acid monohydrate, pH 4.5) at 30 ^°C and 225 rpm for 24–48 h, followed by two additional rounds of FACS with reduced input of 1E + 07 yeast cells.

Shuffled multi-mutation library generation and screening

Plasmid DNAs from the screened single-mutation libraries were isolated using a Maxi-prep kit (Zymo Research) to generate multi-mutation libraries via SSM as described above, and using the DNA shuffling methodology as described previously²⁵. Input DNA from the VH and VL genes were amplified from the library plasmids using Kapa Hifi HotStart ReadyMix (Roche)²⁷, and amplicons were randomly fragmented by introducing double- stranded breaks with DNase I in the presence of MnCl₂ at 15 ^oC for 35–45 s. The resulting DNA fragments were reassembled using Platinum^TM Taq DNA polymerase (family A polymerase, Invitrogen) and AccuPrime^TM Pfx DNA polymerase (family B polymerase, Invitrogen), followed by re-amplification with Kapa Hifi HotStart ReadyMix (Roche). The separate shuffled-VH and shuffled-VL genes were combined into a single shuffled-VH:VL library via step-wise cloning¹⁶. Subsequently, yeast libraries were generated by homologous recombination with VL-BDP-VH expression cassette and pCT backbone as described previously^16,20,21; library sizes exceeding 1E + 07 were maintained for all multi-mutation libraries. 3E + 07 yeast cells were screened in the first round of screening for all multi-mutation libraries.

Pre-sort libraries from the first round of DNA shuffle mutagenesis (1^st shuffle mutagenesis) were labeled with 1.25 nM biotinylated Pep22 conjugated to streptavidin-PE (S21388, Invitrogen) at a 2:1 molar ratio, and sorted for three rounds by FACS, as described above. Plasmids of the enriched populations from the 1^st shuffle mutagenesis Round 3 were isolated to serve as the template for the second round of DNA shuffle mutagenesis (2^nd shuffle mutagenesis). Due to the large potential diversity, Pre-sort libraries from the 2^nd shuffle mutagenesis were labeled with an anti-FLAG FITC Clone M2 monoclonal antibody (F4049, Sigma-Aldrich) at 1 μg/mL for down-selection of Fab-displaying (VL + ) populations using magnetic-activated cell sorting (MACS). Subsequently, VL+ libraries from the 2^nd shuffle mutagenesis were labeled with monovalent antigen molecules, instead of antigen molecules pre-conjugated with streptavidin-PE (S21388, Invitrogen). Libraries were enriched against 0.5 nM of biotinylated Pep22 in the first round of FACS, 0.25 nM of biotinylated Pep22 in the second round of FACS, and 0.06 nM of biotinylated truncated-PfCSP (N-terminal domain stabilized truncated-PfCSP, NTDS_5/3)⁷ in the third round of FACS. NTDS_5/3 is a N-terminal domain stabilized truncated CSP, with three adjacent NPNV motifs. It was created by introducing four N-terminal domain stabilizing (NTDS) mutations, C25S and ₆₆KKNSR₇₀ to ₆₆SSNSA₇₀, to the 3D7 clone PfCSP, and reducing the extensive central repeat region by displaying five major repeat peptides (NANP) instead of 38, and three minor repeat peptides (NVDP) instead of four, while keeping the C-terminus unaltered^7,16.

Because L9 is already a very high-affinity molecule, depletion of antigen from solution can be observed at very low staining concentrations, for example, during the later rounds of screening (Supplementary Figs. S2, S3). Antigen depletion effects limit the overall MFI observed in enriched populations for high-affinity antibodies like L9 at low antigen concentrations; still, affinity-gate sorting was used successfully to improve antibody affinity, as we showed in this study.

NGS and bioinformatic analysis of sorted yeast display libraries

We sequenced Pre-sort libraries from 1^st shuffle mutagenesis and VL + libraries from 2^nd shuffle mutagenesis with the Illumina MiSeq platform, which provided high sequence coverage. Due to the shorter length of the MiSeq reads, heavy and light antibody chains were sequenced separately. For sorted library populations with reduced diversity, the long-read Pacific Biosciences Sequel II platform was also used to sequence paired heavy and light antibody chains together. These two sets of sequencing data were first processed separately, and their information was mined together to identify enriched variants for experimental validation.

In the MiSeq data pipeline, paired-end reads were first merged to assemble the entire heavy or light antibody chain via FLASH (v1.2.11)³⁴. Then, the linked reads were quality filtered with FASTX-Toolkit (v0.0.14) to retain sequences where at least 90% of the bases had a Phred Quality Score of at least 30 (corresponding to 99.9% base call accuracy). IgBLAST (v1.16.0)³⁵ was then used to annotate the antibody V-(D)-J genes and identify CDR3 sequences, with only productive sequence alignments retained in the output files.

For the Sequel II data, the raw sequence data was demultiplexed using Lima (v2.7.1). Once individual sample sequence files were recovered, the long reads were split into two reads, one with the heavy chain and the other with the light chain; pairing information was preserved via sequence headers. For separate heavy and light chain sequences, IgBLAST (v1.16.0)³⁵ was used to map V-(D)-J genes and identify CDR3 regions. Subsequently, paired heavy and light chain alignments were organized based on unique header information. Each heavy and light chain sequence was compared to the wild-type antibody sequence to identify substitution mutations.

Bioinformatic analysis was performed using MiSeq and Sequel II sequence data to identify antibody variants with improved antigen binding. Prevalence for each variant in a sorted library was calculated as the read count for that variant divided by the total read number in that library. The enrichment ratio (ER) quantifies the extent to which a variant is enriched after antigen-binding selection, and ER was defined as variant prevalence in a sorted library divided by prevalence in the Pre-sort or VL + library^16,20,21,36. Together, prevalence and ER metrics provided a framework for defining selection criteria for high-affinity variants in enriched libraries.

For the 2^nd shuffle mutagenesis libraries, the VL + libraries were sequenced as separate heavy and light chain reads using MiSeq. Round 3 libraries were sequenced as both separate heavy/light reads using MiSeq and paired VH:VL reads using Sequel II. To determine ER for variants in the paired VH:VL sequences, the prevalence of each VH:VL pair in the VL + library was estimated as the product of the corresponding VH prevalence and VL prevalence measured from the VL + MiSeq dataset.

Initially, ER values for separate heavy and light chain reads were calculated independently, and heavy and light chain variants of interest were identified based on either high prevalence or high ER. Sequel II sorted-library data from Round 3 were then filtered for sequences with exact matches to a selected heavy or light amino acid sequence. Nucleotide consensus sequences were generated for the complementary light or heavy chain pair from quality-filtered library data via USEARCH (v 5.2.236)³⁷, enforcing a 90% sequence identity for clusters. Finally, full candidate antibody sequences were assembled by pairing the exact match search sequence (VH or VL) and the corresponding consensus sequence (VL or VH) together. In Sequel II data, paired VH:VL sequences from FACS R3 with prevalence > 0.1% and ER > 1 were classified as high-affinity variants.

Antibody expression and purification

Antibody variable heavy chain and light chain sequences were codon optimized, synthesized and cloned into a pcDNA3.4 (CMV expression vector)-based IgG1 vector which incorporates the LS mutation in the Fc region. The resulting plasmids of heavy and light chain pairs were transiently co-transfected in CHO-S cells (GenScript, Piscataway, NJ). Expressed antibodies were harvested and purified by using Protein G MagBeads MX (L00673, GenScript), followed by buffer exchange to the formulation buffer (PBS, pH 7.4). Purified antibodies were assayed by CE-SDS and SEC-HPLC, and were validated to be free of endotoxin contamination. Antigen-binding fragments (Fabs) were generated by overnight digestion of intact IgG with Lys-C at 37 ^oC using an IgG:Lys-C ratio of 4000:1 (w/w), followed by sequential purification using Protein A affinity chromatography and size-exclusion chromatography.

Affinity measurement of L9-variants by Biolayer Interferometry (BLI)

Antibody-binding affinities to various ligands were measured using biolayer interferometry on an Octet Red384 instrument (fortéBio, Menlo Park, CA) with streptavidin capture biosensors (fortéBio) in solid black tilt-well 96-well plates (Greiner Bio-One, Kremsmünster, Austria). Assays were performed with agitation at 30°C. Immobilization of biotinylated fl_PfCSP_3D7, NTDS_5/3_CSP, fl_PfCSP_Kenya (GenBank accession code AAN87596)³⁸, Pep21, Pep22, and Pep29 was performed for 5 s, followed by a 60-second baseline in buffer (PBS + 1% BSA). Association with IgG (serially diluted from 1000 to 1.3 nM) was done for 240 s, followed by a dissociation step in buffer for 1200 seconds. In all Octet measurements, parallel correction to subtract systematic baseline drift was carried out by subtracting the measurements recorded for a loaded sensor incubated in PBS. Data analysis was carried out using Octet software, version 9.0. Experimental data were fitted globally with a 1:1 Langmuir model of binding for all the antigens with the exception of fl_PfCSP_3D7, which was fitted with a 2:1 Langmuir model of binding.

Evaluation of protective potency of L9-variants against sporozoite challenge in mouse model

Female 6- to 8-week-old C57BL/6 J mice were obtained from The Jackson Laboratory (Bar Harbor, ME). All animals were maintained at the animal facility of the Johns Hopkins Bloomberg School of Public Health and cared for in accordance with the American Association for Accreditation of Laboratory Animal Care Standards. Mouse housing was maintained at 40–60% relative humidity and a temperature of 68–79 ºF, with at least ten room air changes per hour and a 14/10 h light/dark cycle. All mouse procedures were performed according to protocols approved by the Institutional Animal Care and Use Ethics Committees at Johns Hopkins University, Protocol number MO24H329.

To generate sporozoites, transgenic P. berghei (strain ANKA 676m1C11, MRA-868) expressing P. falciparum 3D7 full-length CSP or chimeric CSP and a green fluorescent protein/luciferase fusion protein (Pb-PfCSP-GFP/Luc-SPZ or Pb-PfCSP-SPZ) were obtained from salivary glands of infected mosquitoes, as previously described²⁸. Briefly, Anopheles stephensi mosquitoes were obtained and reared from a colony maintained at the insectary at Johns Hopkins University, Malaria Research Institute. Female mosquitoes were allowed to feed on 6- to 8-week-old female Swiss Webster mice infected with blood-stage Pb-PfCSP- GFP/LUC parasites. After infection, mosquitoes were maintained in an incubator at 19−20 °C and supplied with a sterile cotton pad soaked in 10% sucrose, changed every 48 h. 20−23 days following mosquito infections, salivary glands were dissected and ground in 2% FBS-HBSS, SPZs were counted in a Neubauer chamber.

Antibody selection for in vivo testing integrated (i) sequence-level enrichment metrics (prevalence and ER), (ii) representation of distinct mutation combinations, and (iii) BLI apparent affinity profiles across antigens. We selected variants for functional evaluation that had both high apparent affinity and spanned distinct apparent affinity phenotypes from the BLI panel. To assess the protective efficacy of anti-PfCSP mAbs in vivo, mAbs were diluted in sterile filtered 1 x PBS and administered into the tail veins of 6- to 8-week-old female C57BL/6 J mice (The Jackson Laboratory, Bar Harbor, ME). After 2 h, the mice were intravenously challenged in the tail vein with 2000 freshly harvested Pb-PfCSPGFP / Luc-SPZ in 2% FBS-HBSS (Thermo Fisher Scientific, Waltham, MA, USA).

To assess Plasmodium infection, mice were injected intraperitoneally (i.p.) with 100 μL d-luciferin (30 mg/mL, PerkinElmer, Waltham, MA), anesthetized with isoflurane and imaged with an IVIS® Spectrum in vivo imaging system (PerkinElmer) 5 min after luciferin injection. Parasite liver load was assessed 42 h after challenge, whereas parasitemia was measured 6 days following challenge. Parasite load was quantified by analyzing a region of interest (ROI) in the upper abdominal region for liver stage, or whole animal for parasitemia; bioluminescence or the total flux (photons/second; p/s) was measured using the manufacturer’s software (Living Image 4.5, PerkinElmer).

Cryo-EM Sample preparation, grid preparation, and data collection

The sample for cryo-EM was prepared by mixing antigen- binding fragment (Fab) to PfCSP (3D7) or PfCSP (Kenya) at 2:1 molar ratio in 1 x PBS pH 7.4 buffer. The complex was incubated overnight at 4 °C and flash frozen until ready for use. To prepare grids of the complex, 3 µL of protein sample was applied to freshly glow-discharged (easiGLow) C-flat grids (Protochips, CF1.2/1.3-3Au). Blotting was done using a Vitrobot Mark IV (Thermo-Fisher), with 5 s blotting time and 8 pN blotting force at 6 °C in 100% humidity. Grids were vitrified by plunging into liquid ethane and stored in liquid nitrogen before examination by cryo-EM. Images were recorded on a Glacios TEM (Thermo Fisher) at 200 kV and recorded at 165,000X magnification with a defocus range of − 0.6 to − 2.0 µm on a K3 direct electron detector (Gatan) in super-resolution mode.

Cryo-EM Data processing and refinement

Motion correction, contrast transfer function (CTF) estimation, particle picking, extraction, 2D classification, ab initio model generation, 3D refinements and local resolution estimation were carried out in cryoSPARC (v3.3.1)³⁹. The 3D reconstructions were performed using C1 symmetry for both classes. The coordinates of the L9-PfCSP structure, PDB entry (8EK1) was employed as the initial model for fitting the sharpened cryo-EM map of the L9-PfCSP structures. Manual and automated model building were iteratively performed using Coot (v0.9.6)⁴⁰ and real space refinement in Phenix (v1.20.4459)⁴¹ to accurately fit the coordinates to the electron density map. Molprobity⁴² and EMRinger⁴³ was used to validate geometry and check structure quality. UCSF ChimeraX (v1.7.1)⁴⁴ was used for map-fitting cross correlation calculation (Fit-in-Map tool) and for figure preparation.

Quantification and statistical analysis

L9 variants were compared to an equivalent dose of L9_WT as a benchmark control, using a two-tailed ordinary one-way ANOVA test with Dunnett’s multiple test correction. Statistical significance between L9_yd19 and MAM01 was analyzed using a two-tailed unpaired t test. Statistical differences were estimated on log₁₀(liver burden) values. To enable comparison of the correlation between parasite infection and antigen/epitope apparent affinity across different studies, the liver burden of each group was normalized based on the mean of the liver burden values from the untreated (max burden) mice in the same experiment. Pearson’s correlation coefficient (R) was calculated using the mean normalized liver burden values with corresponding two-tailed P-values. A linear regression was fit based on a Pearson’s correlation analysis. Data were plotted and graphed using GraphPad Prism, unless otherwise stated. P-values ≤ 0.05 were considered significant as indicated in the figures.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Supplementary information

Supplementary Information ^{(6.2MB, pdf)}

41467_2026_75840_MOESM2_ESM.pdf ^{(13KB, pdf)}

Description of Additional Supplementary Files

Data S1 ^{(18.7KB, xlsx)}

Data S2 ^{(18KB, xlsx)}

Data S3 ^{(16.2KB, xlsx)}

Data S4 ^{(17.4KB, xlsx)}

Data S5 ^{(12.9KB, xlsx)}

Reporting Summary ^{(112.2KB, pdf)}

Peer Review file ^{(561.7KB, pdf)}

Source data

Source Data ^{(5.9MB, xlsx)}

Author contributions

Designed experiments: J.C., P.T., Y.F-G., B.M., M.P., K.M., T.Z., A.H.I., F.Z., R.A.S., P.D.K., and B.J.D. Performed experiments: J.C., P.T., Y.F-G., B.M., H.L., I-T.T., N.K.H., and B.J.F. Analyzed the data: J.C., P.T., Y.F-G., G.A.L., A.S.F., K.M., and B.J.D. Writing: J.C., P.T., P.D.K., and B.J.D. Reviewing and editing: all authors.

Peer review

Peer review information

Nature Communications thanks Dimuthu Angage, Jürgen Bosch and Stephen Scally for their contribution to the peer review of this work. A peer review file is available.

Funding

Funding was provided by NIH grants DP5OD023118, R01AI181684, and R01AI192975, the Mark and Lisa Schwartz AI/ML Initiative, the MIT Research Support Committee, the MIT Department of Chemical Engineering, and the Ragon Institute of Massachusetts General Brigham, MIT, and Harvard. This study used the Office of Cyber Infrastructure and Computational Biology High Performance Computing cluster at the National Institute of Allergy and Infectious Diseases, Bethesda, MD. This research was supported in part by the Intramural Research Program of the National Institutes of Health (NIH). The contributions of the NIH authors are considered Works of the United States Government. The findings and conclusions presented in this paper are those of the authors and do not necessarily reflect the views of the NIH or the U.S. Department of Health and Human Services.

Data availability

Raw FASTQ files from Illumina MiSeq and Pacific Biosciences Sequel II sequencing have been deposited in the NCBI Sequence Read Archive with the BioProject accession number PRJNA1438926. The VH and VL sequences of the L9 variants have been deposited in the NCBI GenBank with accession numbers PZ196424 – PZ196495. The atomic coordinates and cryo-EM density maps generated in this study have been deposited in the Protein Data Bank (PDB) under accession codes 9ZNT and 9ZNU, and in the Electron Microscopy Data Bank (EMDB) under accession codes EMD-74456 and EMD-74457. The NGS analysis data, apparent affinity data, mouse liver burden data are provided in the Supplementary Information and Source Data files. This paper does not report original code. Any additional information required to analyze the raw data reported in this paper or any requests for resources should be directed to the lead contact, Brandon DeKosky (dekosky@mit.edu). Source data are provided in this paper.

Competing interests

The authors have filed a U.S. provisional patent application (63/890,352, pending) describing the sequences of protective antibodies introduced here.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-75840-w.