Consensus-based computational mapping of NS1 antigenic determinants in dengue and Zika viruses to improve diagnostic specificity

Abstract

Background

Cross-reactivity between samples from phylogenetically related organisms remains a significant challenge in antigen detection by lateral flow assays. Recent advancements in bioinformatics tools have significantly enhanced our understanding of the complex interactions between antigenic proteins and antibodies, paving the way for the development of more precise and effective solutions to address these challenges.

Methods

In this study, we used two common modeling tools, SWISS-MODEL and AlphaFold (2 and 3), to construct three-dimensional (3D) models of the nonstructural 1 (NS1) protein from dengue (DENV) and Zika viruses (ZIKV) based on their protein sequences. Both linear and 3D structures of the proteins were used to predict antigenicity (Jameson-Wolf and Kolaskar-Tongaonkar methods) and B-cell epitopes (Bepipred-2.0, ElliPro, SEPPA3, and DiscoTope methods), and the results obtained from the two modeling tools were compared. Furthermore, molecular simulations were conducted utilizing FORTE and PROCEEDpKa to estimate binding affinities and predict electrostatic epitopes, respectively.

Results

Our consensus-based pipeline consistently identified two primary antigenic hotspots: 107–127 and 301–319. The antigenicity of the 107–127 region was strongly corroborated by its overlap with the crystallographic binding sites of known antibodies (22NS1, 2B7, and 1G5.3), validating our approach. The 301–319 region emerged as a novel, high-confidence candidate epitope. Furthermore, electrostatic and affinity analyses revealed distinct interaction profiles, providing a molecular basis for discriminating between serotype-specific and cross-reactive antibody responses.

Conclusions

Our computational pipeline identified key NS1 antigenic determinants, validating the 107–127 region and revealing region 301–319 as a novel epitope candidate. These findings provide specific targets for diagnostics designed to circumvent flavivirus cross-reactivity. The validated workflow itself serves as a robust, adaptable framework for rapid antigen prediction, accelerating the development of targeted diagnostics and immunotherapies for future emerging viruses.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-025-12291-6.

Background

Dengue and Zika are tropical arboviruses with significant global impact, causing illnesses with short- and long-term consequences [1]. Dengue has become the most prevalent arboviral infection worldwide, with over 14 million cases reported in 2024, of which 9.8 million occurred in Brazil [2, 3]. Additionally, the incidence of Zika infection peaked in 2016 in the Region of the Americas and decreased substantially thereafter. This region continues to present the highest number of cases worldwide [4]. The distribution of Aedes mosquitoes, the primary vector of these diseases, is now the most extensive ever recorded, reaching all continents. This expansion is largely attributed to anthropogenic climate change, including global warming [5].

The gold standard for detecting dengue virus (DENV) and Zika virus (ZIKV) is real-time reverse transcriptase-polymerase chain reaction (RT-PCR). Despite its accuracy, RT-PCR presents challenges such as long processing times, high costs, and the need for specialized personnel and infrastructure [6]. Alternative techniques have been explored to address these limitations, including spectroscopic methods, electrochemical sensors, and lateral flow assays (LFA) [7–9]. Among these, antigen detection through LFA has shown great potential due to low cost, portability, transport stability, and rapid results.

A critical challenge in antigen detection is cross-reactivity between phylogenetically related viruses such as DENV and ZIKV. The nonstructural protein 1 (NS1), a glycoprotein involved in viral replication, immune evasion, and pathogenesis, is a key diagnostic target due to its high levels in the blood during early infection [10]. However, its high degree of conservation among flaviviruses demands strategies to identify unique protein regions for diagnostic specificity. In addition, the use of LFA for NS1 detection has some limitations. Despite their widespread use, many of these tests suffer from suboptimal sensitivity and specificity, particularly in regions with high co-circulation of DENV and ZIKV. Cross-reactivity often leads to false-positive results, complicating differential diagnosis and public health interventions.

Recent advancements in bioinformatics have enabled rapid structural analysis and epitope prediction, which could significantly accelerate the development of diagnostic tools. This was well exemplified during the COVID-19 pandemic, which showcased the potential of computational approaches [11] and artificial intelligence (AI)-driven bioinformatics tools [12, 13]. In this study, we utilized two different modeling strategies, SWISS-MODEL [14] and AlphaFold [15], to model NS1 proteins of DENV and ZIKV and evaluated their antigenicity and B-cell epitopes. Therefore, this study aims to develop a computational workflow with a dual purpose: first, to identify conserved antigenic regions across flaviviruses that are potential pan-flavivirus targets, and second, to pinpoint unique, serotype-specific epitopes critical for developing diagnostics that can circumvent cross-reactivity.

Materials and methods

To ensure a robust and multifaceted characterization of NS1 epitopes, we employed a complementary suite of computational approaches. These methods span consensus-based structure predictions, empirical overlap analyses, geometric interface mapping, qualitative affinity ranking, and electrostatic epitope mapping. A concise overview of each tool and its principal features is provided in Table 1. Additionally, safety profile assessments for toxicity and allergenicity were performed to evaluate the translational potential of our findings.

Table 1.

Summary of epitope prediction and analysis workflow

Task	Method / Tool	Key Features	Purpose
1. Consensus-Based Prediction	ElliPro [16], DiscoTope [17, 18], SEPPA 3.0 using computer-generated structures by SM [19], AF2 [15], and for DENV1, AF3 [20].	Integration of multiple structure-based algorithms	Consensus criteria to minimize false positives and enhance the reliability of epitope predictions
2. Analysis of Known Experimental Binding Interfaces	NS1 crystal structures	Direct alignment of predicted epitopes with empirically determined antibody-binding interfaces	To confirm predictions by validating overlap with established structural data
3. Geometric Interface Analysis	AlphaFold-Multimer complexes [15]; distance cutoff [21]; SASA [22]	Application of spatial and solvent-accessibility criteria to modeled protein complexes and identified the interaction interface	To delineate credible conformational epitopes at the protein–antibody interface
4. Electrostatic Epitope Prediction	PROCEEDpKa [23]	Mapping of titratable residues exhibiting pKa shifts due to local electrostatic environment	To identify titratable residues critical for the antigen-antibody complexation mechanism
5. Qualitative Affinity Ranking	FORTE [24]	Semi-quantitative estimates of binding affinities using a shared titration engine with the PROCEEDpKa	To provide a relative ranking of antibody affinities in agreement with experimental data and also support the predictions by PROCEEDpKa.

Sequences retrieval

The full amino acid sequences of DENV1, DENV2, DENV3, DENV4, and ZIKV NS1 proteins were retrieved from the National Center for Biotechnology Information (NCBI) protein database [16] (Accession numbers: AMO26075.1, AAD11533.1, AAB52247.1, AHK09949.1, and YP_009430301.1). These sequences were selected as they have been well-characterized in prior studies of NS1 antigenicity [17, 18]. To ensure their continued relevance, the representativeness of these sequences was validated against both NCBI RefSeq [19] entries and recent circulating viral lineages. A detailed analysis confirming their high degree of conservation is presented in the Results section.

Structural modeling of DENV1–4 NS1 proteins and their antibody complexes

The five NS1 proteins were modeled using two primary computational approaches: SWISS-MODEL (SM) [14] and AlphaFold2 (AF2) [15] (v2.2.0). Additionally, AlphaFold3 (AF3) [20] was employed exclusively for the NS1 protein of DENV1, as part of a targeted comparison to validate AF2 predictions. In the case of SM, multiple structural templates were automatically selected, and individual molecular models were generated accordingly. These models will be referred to throughout the manuscript using the subscript (SM). SWISS-MODEL was primarily employed for the initial structural predictions of individual NS1 proteins, particularly in the section titled “Predicting Antigenicity and B-cell Epitopes”. Our computational setup utilized AF2 (v2.2.0) to generate both individual NS1 models and protein–protein complex predictions. For complexes, AF2 was used with the multimer model preset, full database coverage, and the maximum template date set to November 24, 2023. AlphaFold-Multimer incorporates co-evolutionary information between different protein chains and has been shown to predict protein-protein complex structures in community assessments relatively well. Models derived from AF2 will be denoted with the subscript AF2.

Given the timing of this study, the AlphaFold3 (AF3) server [20] became available only in the later stages of manuscript preparation. To strengthen our structural predictions, we conducted a preliminary comparison between the NS1 DENV1 models generated by AF2 and AF3, using the default parameters of the web-based AF3 implementation provided by Google DeepMind and Isomorphic Labs [20]. Models generated with AF3 are referred to using the subscript AF3.

For all modeling approaches, the input data consisted of the complete amino acid sequences of the NS1 proteins, retrieved from the NCBI protein database [21]. UCSF ChimeraX version 1.7 [22] was used for structural analysis and image rendering, using the PDB files generated by each respective modeling method.

The structural stability and reliability of predicted protein models are critical for subsequent analyses, particularly in the absence of experimentally resolved structures. Although molecular dynamics (MD) simulations are a powerful and widely adopted strategy for assessing model stability under such conditions, they were not employed in the present study for specific and well-supported reasons.

The dynamic behavior of NS1 dimeric models (especially those from flaviviruses such as ZIKV and DENV) has already been extensively characterized in the literature through classical MD simulations [23, 24]. Notably, the presence of a second chain in the NS1 dimer imposes natural conformational constraints that significantly attenuate structural fluctuations. Key metrics, including the Radius of Gyration and Solvent Accessible Surface Area, have been shown to remain stable throughout simulation trajectories, confirming the inherent dynamic stability of the dimeric form [24].

Building upon this established foundation, our analysis focused on static structural predictions for epitope mapping and antigenicity assessment as done before [25, 26]. This choice avoided redundant computational effort and enabled the integration of a broader set of complementary computational tools, thereby enhancing the scope and depth of our predictive framework.

In this study, we deliberately focused our structural and immunoinformatic analyses on the dimeric form of NS1 to capture the viral processes most intimately tied to genome replication and early virion assembly. Within the endoplasmic reticulum (ER), newly synthesized NS1 monomers rapidly associate into liposoluble dimers that partition into lipid microdomains (rafts) of the ER membrane [27, 28] This membrane-bound NS1 (mNS1) engages directly with the nonstructural protein NS4B, inducing ER membrane invagination and delineating the replication organelle - an essential scaffold for viral RNA synthesis [25, 29].

Concomitant with organelle formation, mNS1 cooperates with the precursor membrane protein (prM) and the envelope glycoprotein (E) to orchestrate the budding and release of infectious virions [30, 31]. These dimer-specific scaffolding functions are central to the flavivirus life cycle yet are entirely absent once NS1 assembles into its secreted hexameric state. By isolating the dimer, we directly interrogate the molecular interfaces that underlie replication‐organelle biogenesis and early virion morphogenesis.

Although NS1 hexamers (sNS1) package lipid cargo within a central channel and are secreted to modulate host immunity [31, 32], they do not contribute to the structuring of replication complexes or the initial stages of particle formation. Moreover, secretion of NS1 can occur independently of full hexamerization, and membrane-bound dimers themselves are detected on infected‐cell surfaces, where they engage B– and T–cell receptors [28, 32]. Hence, the dimeric NS1 embodies both the intracellular replication functions and the cell‐surface immunogenic epitopes that are most relevant to pathogenesis and diagnostic targeting.

Focusing on the NS1 dimer, therefore, allows our computational workflow to resolve the precise surface epitopes and protein–protein interfaces that drive viral RNA replication, replication-organelle formation, and infected‐cell recognition. This targeted approach aligns with prior mutagenesis and molecular dynamics studies that have identified critical β‐roll interactions, salt bridges, and hydrogen‐bond networks stabilizing the dimer interface [33]. By concentrating on this biologically distinct and functionally indispensable state, our work provides novel insights into NS1’s role in flavivirus replication and opens avenues for epitope‐based therapeutics and diagnostics that would be obscured by analyses of the secreted hexamer alone.

Predicting antigenicity and B-cell epitopes

Antigenicity was analyzed using the Jameson-Wolf [34] and Kolaskar-Tongaonkar [35] methods. The DNASTAR Lasergene software (Protean 3D^®, Version 17.5, DNASTAR, Inc., Madison, WI) was used to obtain Jameson-Wolf predictions through the Protean 3D server, with a machine-learning approach to identify antigenicity patterns suggestive of an epitope. This method predicts immunogenic regions by identifying ranges with an increased antigenic index derived from predictions of hydrophilicity, surface accessibility, flexibility, and turn or coil conformations [33]. Kolaskar-Tongaonkar’s predictions were achieved by the Immune Epitope Database (IEDB) server [36], where theoretical immunogenicity scores were generated for each possible antigen, based on a semi-empirical method that uses physicochemical properties of amino acid residues and their frequencies of occurrence in experimentally known segmental epitopes with about 75% accuracy [35]. In both cases, default options were applied.

In order to obtain B-cell epitope prediction, four different methods were employed. The results were generated via online analysis in the IEDB server [36]. The Bepipred-2.0 method predicts linear B-cell epitopes from a protein sequence using a Random Forest Regression algorithm trained on epitopes and non-epitope amino acids determined from crystal structures [37]. The residues with scores above the threshold are predicted to be part of an epitope. To predict continuous and discontinuous epitopes, the ElliPro method utilizes the protrusion index of residues, protein shape approximation, and the final neighboring residue clustering [38]. The minimum score adopted was 0.5, while the maximum distance (in Ångströms) was 6. Alternatively, the Spatial Epitope Prediction of Protein Antigens (SEPPA3) method applies features of glycosylation triangles and glycosylation-related amino acid indexes coupled with a calibration based on neighboring antigenicity [39]. Lastly, the DiscoTope-2.0 method was applied to predict discontinuous epitopes based on amino acid statistics, spatial information, and surface accessibility. The algorithm construction was based on a compiled data set of discontinuous epitopes determined by X-ray crystallography of Ab-antigen protein complexes [40, 41]. The default setting was applied for all analyses.

To evaluate the safety profile of the identified antigenic regions for broader immunogenic applications, all NS1 sequences were assessed for potential toxicity and allergenicity. Toxicity was predicted using the ToxinPred 3.0 server [42]. The analysis was conducted in peptide scanning mode using the Hybrid (ET + MERCI) prediction model, as this approach combines machine learning-based pattern recognition with known toxic motif searches, providing a more robust assessment than either method alone. The peptide fragment length and score threshold were kept at their default settings, which are optimized and validated by the tool’s developers. Allergenicity was evaluated using the AlgPred 2.0 server [43]. For this analysis, the Hybrid (RF + BLAST + MERCI) machine learning model was employed to leverage the combined strengths of machine learning, sequence homology to known allergens, and motif identification, thereby ensuring maximum predictive sensitivity. Proteins were classified as ‘Allergen’ or ‘Non-Allergen’ based on the recommended default score threshold of 0.3, which is established to provide an optimal balance between sensitivity and specificity.

Following this broad screening, a targeted, peptide-specific toxicity analysis was performed on the lead epitope candidates (regions 103–130 and 301–319). Each peptide sequence was individually submitted to the ToxinPred 3.0 server [42] to obtain a quantitative toxicity score. This focused analysis was performed using the same Hybrid (ET + MERCI) model, applying the server’s default score threshold of 0.38 for classification.

The sequence alignment of the full-length proteins and predicted regions was conducted using the Clustal Omega online server [44]. The input data for epitope predictions, which consider the 3D structure of proteins and the molecular simulations with both FORTE [21] and PROCEEDpKa [25], were in PDB format, generated using SM [14] and AF2 [15] as discussed above. Prior to molecular simulations, these atomistic PDB models were converted to the coarse-grained representation required by the FORTE [21] and PROCEEDpKa [25] methods.

Epitope prediction for known anti-NS1 monoclonal antibodies (mAbs)

Epitope characterization can be performed either by relying solely on sequences, without considering structural features, or by incorporating the structural properties of known Abs. In the previous section, we focused on sequence-based predictions that exclude structural information about known antigen-antibody interactions, even when crystallographic data for deposited Abs is available. Consequently, sequences identified as epitopes in these predictions may or may not be part of an epitope that is experimentally known to bind to an existing Ab. To address this issue, we performed an additional analysis on Ab-NS1 complex structures available in the protein database [45, 46].

Experimental antibody-antigen complexes were extracted from the RCSB Protein Data Bank (PDB) under PDB IDs 4OII, 6WEQ, 6WER, 7BSD, and 7BSC. We have selected three Abs against ortho-flavivirus NS1 proteins: 22NS1, 2B7, and 1G5.3. From the crystallographic data, possible complexations between these Abs and the NS1 proteins (for DENV1-4 and ZIKV) were generated using AF2. For each system, the 10 best-ranked complexes were selected for further analysis, adopting two different approaches: solvent-accessible surface area (SASA) [47] and a 15 Å separation distance criterion [48].

In order to measure the protein-protein – or rather protein-Ab – interface, the software “Scoring Protein Interaction Decoys using Exposed Residues” (SPIDER) [49] was used. This algorithm uses a score function that captures geometric and compositional residue interaction patterns formed at the interfaces of X-ray-characterized protein complexes and then evaluates how “geometrically” similar the interacting residues are with “native” patterns derived from native poses [49]. Following the approach suggested by Poveda-Cuevas, Etchebest, and Barroso da Silva (2020), changes in solvent-accessible surface area (SASA) were calculated for the entire isolated proteins and their complexes to define the biological interface between them. Additionally, amino acids from one chain with a separation distance of less than 15 Å from those in another chain were considered to be part of the Ab-Ag interface and defined as epitopes [48].

Lastly, we mapped the EE using the “Prediction of Electrostatic Epitopes Based on pK_a Shifts” (PROCEEDpK_a) method [25]. In this case, we selected 2B7 to be analyzed against all viral proteins in this study. This analysis is grounded in the physicochemical principle that, within a molecular structure, the pK_a values of titratable groups are modulated by the local electrostatic environment, particularly through interactions with proximal charged residues or functional groups. Identifying the groups exhibiting pronounced pK_a shifts suggests their greater contribution to stabilizing electrostatic interactions during molecular complexation pK_a [25].

To complement this analysis, we employed the constant-pH “Fast cOarse-grained pRotein-proTein modEl” (FORTE). The FORTE model was employed not to reproduce absolute experimental binding affinities, but to provide a semi-quantitative, relative ranking of antibody interactions. Simulations were performed under physiological conditions (pH 7.0 and 150 mM salt concentration), with all other parameters set as described in [21]. This approach enabled the establishment of a qualitative ranking that is in agreement with available experimental data for known antibody affinities against different flaviviruses [50, 51]. Importantly, since both FORTE [21] and PROCEEDpKa [25] rely on the same titration engine [52] and equivalent electrostatic models [53], their combined use offers an additional layer of internal validation for our computational strategy.

Relative binding affinities

Although the methods discussed above suggest possible structural binding interfaces, they do not provide estimations of binding affinities. These estimations were computed in molecular simulations using FORTE [21] for all the NS1 proteins from DENV1-4 and ZIKV with 2B7 Fab and 1G5.3 Fab. The latter is known to cross-react with both DENV2 and ZIKV [54]. No experimental data are available for 1G5.3 Fab with DENV1, DENV3, and DENV4.

Results

Sequence conservation analysis of the selected proteins

Initially, we validated the representativeness of the selected sequences through alignment against the NCBI RefSeq database [19]. As demonstrated in Supplementary Material (Table S1), the pairwise alignment revealed an exceptionally high degree of conservation, with sequence identity exceeding 97.7% across all four DENV serotypes. Crucially, our analysis extended beyond simple percent identity to assess the physicochemical nature of the divergent residues. We found that the vast majority of amino acid substitutions were evolutionarily conservative. This profound conservation, particularly the scarcity of alterations that would drastically remodel the local protein chemistry, confirms that the major B-cell epitope regions are fully conserved across the analyzed sequences (Supplementary Figure S1). Consequently, the antigenic landscape of the NS1 protein appears stable among the represented variants, providing a solid foundation for the subsequent epitope predictions.

Predicting antigenicity and B-cell epitopes

By applying our consensus threshold (support by ≥ 4 of 6 prediction tools), we identified five discrete antigenic regions in NS1 (Fig. 1). These consensus epitopes span 10–20 residues each and cluster into two major surface patches: three lie within the wing domain (residues 30–48, 76–94 and 110–128) and two map to exposed loops of the β-ladder (residues 154–171 and 249–265). Structural mapping confirms that all five regions are solvent-accessible on the NS1 dimer, highlighting them as prime candidates for B-cell recognition in both diagnostic and vaccine applications.

Fig. 1.

Consensus analysis for NS1 DENV1-4 and NS1 ZIKV based on SWISS-MODEL and AlphaFold2 structural predictions. Dots represent predicted amino acid residues categorized by consensus across computational tools: (a) Red: Residues predicted by > 3 independent software tools in both 3D modeling methods (SM and AF2); (b) Yellow: Residues predicted by > 3 tools in one 3D modeling method (SM or AF2); (c) Gray: Residues predicted by ≤ 3 tools in both 3D modeling methods. Detailed analysis for each case is provided in the Supplementary Material

All predicted antigenic regions of NS1 proteins are detailed in the Supplementary Material (Figures S2-S11). Certain regions emerged as antigenicity and epitope “hotspots”, aligning with findings from previous studies [17, 55, 56]. For instance, the 103–130 region highlighted in all five proteins has been extensively described as having high antigenicity scores and low homology with other flaviviruses, as Japanese encephalitis virus (JEV), West Nile virus (WNV), and yellow fever virus (YFV) [18, 57–59], making it a highly suitable candidate for antibody (Ab) generation. Similarly, the 301–319 region also stands out as a promising target for peptide synthesis and application.

To evaluate the translational potential of our epitope candidates beyond diagnostics, we assessed their safety profiles. As detailed in the Supplementary Material (Figures S11 and S12), both the 103–130 and 301–319 regions contain peptide segments predicted as potentially toxic by the ToxinPred 3.0 server [42]. While this computational flag has minimal bearing on the use of these regions for in vitro diagnostic tools, it highlights a critical safety parameter that warrants careful consideration and experimental de-risking for any future vaccine design efforts centered on these epitopes.

To facilitate the prioritization of these candidates and provide a consolidated overview of our findings, the key properties of the two lead epitope regions are summarized in Table 2. This synthesis integrates the results from the consensus prediction pipeline with the peptide-specific toxicity analysis, offering a clear comparative framework to guide forthcoming experimental validation for either diagnostic or therapeutic applications.

Table 2.

Summary of lead epitope candidates and their predicted properties. The two primary antigenic regions identified for each NS1 protein are listed, along with their quantitative toxicity scores and the consensus of prediction tools that identified each region. Toxicity scores were generated for each specific peptide sequence using the toxinpred 3.0 server [42]., as described in the Methods. Default parameters were applied, including a threshold of 0.38 and the ET plus MERCI hybrid prediction approach

NS1 source	Sequence	Length (amino acids)	Toxicity score	Predicted by
DENV1	103MIRPQPMEYKYSWKSWGKAKIIGADVQN130	28	0 (non-toxin)	Jameson-Wolf, Bepipred 2.0, ElliPro, SEPPA 3.0
	301TVTGKIIHEWCCRSCTLPP319	19	0.815 (toxin)	Jameson-Wolf, Kolaskar-Tongaonkar, Bepipred 2.0, ElliPro
DENV2	103SLRPQPTELKYSWKAWGKAKMLSTESHN130	28	0.245 (non-toxin)	Jameson-Wolf, Bepipred 2.0, ElliPro, DiscoTope, SEPPA 3.0
	301TASGKLITEWCCRSCTLPP319	19	0.25 (non-toxin)	Jameson-Wolf, Kolaskar-Tongaonkar, Bepipred 2.0, ElliPro
DENV3	103TLTPQPMELKYSWKTWGKAKIVTAETQN130	28	0.315 (non-toxin)	Jameson-Wolf, Bepipred 2.0, ElliPro, SEPPA 3.0
	301TVSGKLIHEWCCRSCTLPP319	19	0.81 (toxin)	Jameson-Wolf, Kolaskar-Tongaonkar, Bepipred 2.0, ElliPro
DENV4	103ALTPPVNDLKYSWKTWGKAKIFTPEARN130	28	0 (non-toxin)	Jameson-Wolf, Bepipred 2.0, ElliPro, SEPPA 3.0
	301TASGKLVTQWCCRSCTMPP319	19	0.315 (non-toxin)	Jameson-Wolf, Kolaskar-Tongaonkar, Bepipred 2.0, ElliPro
ZIKV	103RLPVPVNELPHGWKAWGKSYFVRAAKTN130	28	0 (non-toxin)	Jameson-Wolf, Bepipred 2.0, ElliPro, SEPPA 3.0
	301TTASGRVIEEWCCRECTMP319	19	0.33 (non-toxin)	Jameson-Wolf, Bepipred 2.0, ElliPro

Epitope prediction for known anti-NS1 mAbs

Structural analysis

Mapping of the three available NS1–antibody crystal complexes (PDB IDs 4OII, 6WEQ and 7BSD) revealed that each antibody engages a discrete but partially overlapping surface patch on the NS1 dimer (Fig. 2). The 22NS1 Fab primarily contacts residues in the wing domain, 2B7 binds an adjacent loop at the wing–β-ladder interface, and 1G5.3 targets an epitope on the β-ladder itself. Together, these footprints define two contiguous antigenic hotspots on NS1 that coincide with our sequence- and structure-based consensus predictions, providing experimental validation of the regions most likely to elicit B-cell responses.

Fig. 2.

Molecular representation of NS1_DENV1-antibody interfaces. (A–C) Predicted 3D structures of the dengue virus NS1 protein (serotype 1) in complex with monoclonal antibodies: (a) A (blue): 22NS1, (b) B (purple): 2B7, (c) C (orange): 1G5.3. Ab-Ag complexes were generated with AlphaFold-Multimer (see the text for details). Images were produced in UCSF ChimeraX (v1.7). Panel (D) shows the consensus antigenic regions (red) derived from integrating all B-cell epitope predictions, highlighting residues with cross-method agreement across SM and AlphaFold-Multimer structural models

The observed epitopes in these analyses with the crystallographic structures indicate that all three known Abs bind to the same region of viral NS1 protein, which corresponds to part of the 103–130 sequence predicted by general bioinformatic tools. This pattern is observed even for different flaviviruses, such as WNV (PDB id 4OII), DENV (PDB id 6WEQ), and ZIKV (PDB id 7BSD). This result validates the approach adopted in this study, as these theoretical predictions align with practical observations. Furthermore, the 301–319 region, also predicted by bioinformatical general tools, should not be prematurely dismissed, as the potential for Ab binding in this region could still be possible, despite a lack of documented experimental evidence to date. In fact, recent experimental structural data, which became available during the preparation of this manuscript, confirmed the existence of another class of anti-ZIKV Abs that bind to distinct regions [51].

An ensemble of predicted NS1–antibody complexes revealed a consistent binding hotspot within residues 107–127 of the larger 103–130 region (Fig. 3). Across ten AlphaFold-Multimer models generated for each of the three monoclonal antibodies and NS1 from DENV1–4 and ZIKV, this segment repeatedly showed the highest interfacial contact density and solvent-accessible surface engagement by both distance‐ and SASA‐based criteria. Its recurrent involvement in all antigen–Ab pairs underscores 107–127 as the prime candidate for experimental epitope validation (Supplementary Figs. S14, S15).

Fig. 3.

Consensus analysis of the separation distance and SASA criteria for NS1 DENV1-4 and NS1 ZIKV. Dots represent predicted amino acid residues categorized by cross-criteria agreement: (a) Red: Residues identified as epitope components in > 5 AlphaFold-Multimer models for both criteria (distance and SASA); (b) Yellow: Residues identified in > 5 models for one criterion (distance or SASA); (c) Gray: Residues identified in ≤ 5 models for both criteria. Detailed analysis for each case is provided in the Supplementary Material

Electrostatic epitopes

We also employed these 3D structures together with the NS1 proteins in complexation studies to map critical titratable amino acids [known as “electrostatic epitopes” (EE)] that are important for complexation [25]. Among all the evaluated Abs here, 2B7 was the only one that had its crystallographic structure deposited in complexes with more than one NS1_DENV protein. Specifically, it was complexed with NS1_DENV1 (PDB id 6WEQ) and NS1_DENV2 (PDB id 6WER). The titratable residues most affected by the presence of the 2B7 Ab were considered part of EE (Fig. 4).

Fig. 4.

Electrostatic epitope (EE) prediction for NS1 DENV1-4 and NS1 ZIKV. Red dots represent predicted amino acids identified as part of an EE, while gray dots indicate residues classified as non-EE. Calculations were performed using the PROCEEDpKa method with a 0.01 threshold for the pKa shifts [25]

Contrary to the common biological view based on the simple “lock-and-key” mechanism and in agreement with the long-range nature of electrostatic interactions [60], the EE are not predominantly concentrated in the regions experimentally known to serve as binding sites; instead, they are dispersed throughout the proteins, as shown in Fig. 5. Therefore, this analysis reveals a much more complex electrostatic coupling of the ionizable amino acids and a larger number of these residues contributing to the complexation process. It is important to note that this analysis was carried out with the lowest threshold used in PROCEEDpKa [25], meaning the EE represented in this figure corresponds to the slightest disturbance suffered when exposed to the Ab. The initial assumption was that increasing the threshold would reduce the number of titratable residues classified as EE scattered throughout the protein and restrict them solely to the complexation known interfaces or their immediate vicinity. However, this was not the case. Even at higher thresholds, the EE remained dispersed, with only a small portion located in the Ab-binding region. A larger number of ionizable amino acids important for the complexation mechanism keeps open the possibility of the existence of alternative binding regions, as suggested above and seen in the recent work by Pan and co-authors [51].

Fig. 5.

Electrostatic epitope analysis of NS1 complexes with the 2B7 antibody across DENV1-4 and ZIKV. Data for NS1_DENV1 (A), NS1_DENV2 (B), NS1_DENV3 (C), NS1_DENV4 (D), and NS1_ZIKV (E) in complexation with 2B7 antibody. NS1 is represented in gray, and 2B7 is in purple. Residues that are part of the EE as predicted by PROCEEDpKa [25] are marked in red. For visualization purposes, the EEs were mapped in the 3D structures of the complexes generated by AlphaFold-Multimer. Panel F shows the crystallographic structure of the complex between NS1_DENV1 and 2B7, as deposited in the RCSB PDB (PDB ID: 6WEQ). All images were created with UCSF ChimeraX (v1.7)

As homologous proteins, all NS1 viral protein variants exhibited similar EE, including residues in the Ab 2B7 binding region. When compared to the crystallographic structure, strong similarities can be observed in the binding region, indicating that the AlphaFold2 prediction is reasonable and aligns well, though not exactly.

Relative binding affinities

Simulated binding affinities of the 2B7 and 1G5.3 Fabs against NS1 from DENV1–4 and ZIKV reveal that both antibodies bind DENV1 and DENV2 more tightly than ZIKV (Fig. 6). This trend qualitatively matches experimental KDs for 1G5.3—0.10 µM for DENV2 and 0.27 µM for ZIKV (Natal strain)—and mirrors the similar crystallographic complexes observed for 2B7 with NS1_DENV1 and NS1_DENV2. The agreement between simulation and experiment validates our approach and indicates that 2B7 is likely to exhibit stronger cross-serotype binding than 1G5.3 under these conditions. Altogether, these affinity profiles complement our structural epitope mapping and highlight antibody–antigen pairs with the highest potential for cross-reactivity.

Fig. 6.

Comparative estimated binding affinities of monoclonal antibodies 2B7 and 1G5.3 against NS1 DENV1-4 and NS1 ZIKV. The relative binding affinities were calculated from the minimum free energy values of interactions given in kT units (k is the Boltzmann constant and T is the temperature) measured for NS1 viral protein-Ab complexes under physiological conditions (pH 7.0, 150 mM NaCl) using constant-pH Monte Carlo simulations (FORTE). Error bars represent the standard deviation, derived from three independent replicate runs per simulated system. Other details as in [50]

Discussion

Notably, most antigenic regions predicted for the four DENV serotypes overlap with those predicted for NS1_ZIKV. Identifying regions with minimal sequence similarity between NS1_{DENV1 − 4} and NS1_ZIKV is crucial for diagnostic applications aimed at minimizing cross-reactivity. These findings indicate that selecting an antigenic region for NS1_{DENV1 − 4} precludes its use for NS1_ZIKV, and vice versa. Consequently, the number of unique peptides available for these diseases is more limited than the total number of predicted antigenic regions. This limitation extends to NS1 proteins from other flaviviruses, providing a strategic approach to reduce cross-reactivity among homologous proteins [61].

When producing Abs against predicted antigens, it is important to take precautions and identify common regions predicted by different strategies. Since each method employs hypotheses that may differ from one another, a consensus approach is recommended to better incorporate a wider range of predictions. Many studies rely on only one tool to predict antigenicity and/or epitopes [17, 18, 57, 62]. This leads to low reliability in the in silico results, since each theoretical approach has an intrinsic hypothesis and an error rate. Using more than one method might increase the chances of ensuring that regions overlooked by a certain tool are identified by another.

Remarkably, most of the tools used here identified the 107–127 region as highly antigenic. This was also observed in the known Ab-binding complexes, both from X-ray structures and those produced by AF2. Electrostatic epitopes were dispersed throughout the protein, failing to reinforce any specific region previously identified. Conversely, a large number of ionizable amino acids have been suggested as critical for the interaction between these two macromolecules.

Whether sequence-based or structure-based, all the aforementioned traditional tools predict which residues in an antigen could be recognized by an Ab as antigenic regions. As more epitopes are discovered, it is becoming apparent that essentially any surface-accessible region of an antigen can be the target of an Ab – see ref. 51. With that in mind, for a given Ab, it is possible to determine which amino acid residues are part of binding regions [63]. Such predictions are very valuable for monoclonal Abs that are intended to be used as diagnostics, given that knowing the epitope is crucial for understanding possible cross-reactivity [64].

Antibodies generated against DENV NS1 may cross-react with homologous ZIKV proteins, potentially contributing to antibody-dependent enhancement (ADE) and yielding false-positive serological results. ADE arises when non-neutralizing or sub-neutralizing antibody–virus complexes engage Fcγ receptors on myeloid cells, facilitating viral entry and amplifying replication. This phenomenon is well documented for DENV and has also been reported for ZIKV in regions of co-circulation. Therefore, computational epitope-prediction strategies that pinpoint NS1-specific regions are essential not only for differential diagnostics but also for guiding the design of vaccines and therapeutics that mitigate ADE-related risks. Furthermore, because NS1 is absent in the virion, NS1-based vaccine candidates inherently pose a low risk of ADE, highlighting their promise as safe immunogens [65, 66].

Conclusions

In this study, we aimed to gather as much information as possible on the prediction of antigenic sequences in the NS1_{DENV1 − 4} and NS1_ZIKV proteins. By constructing a ‘consensus’ approach, we expected to identify regions that appeared most frequently in our theoretical predictions. The goal was not to predict static sequences that would work for all strains regardless of mutations, but rather to create a prediction workflow based on these flaviviruses. This approach ensures that when new mutations arise in existing sequences, it will be easier to make predictions using an established workflow.

From this perspective, the 107–127 sequence stood out in most predictions and was located within the previously reported Ab-binding region. The 301–319 sequence also emerged as relevant in the predictions; however, unlike 107–127, no experimentally known Ab binds to this region. Nevertheless, 301–319 remains noteworthy and could be selected as a target for synthesis, as its presence was consistent across general predictors of antigenicity, B-cell epitopes, and electrostatic epitopes.

The approach followed in the present study proved valuable by combining general tools that predict antigenicity and B-cell epitopes for full proteins (both sequence- and structure-based) with predictions that consider known Abs. In addition to identifying consensus regions, our pipeline incorporated multiple layers of validation to enhance the reliability of the predicted epitopes. Structure-based tools such as ElliPro, DiscoTope, and SEPPA 3.0 were used in a consensus strategy to minimize false positives. The predicted epitopes were further validated by their overlap with known antibody binding sites from NS1 crystal structures, and by applying geometric criteria to AlphaFold-Multimer-generated complexes. Moreover, binding affinity estimations using the constant-pH FORTE model provided a qualitative ranking of antibody interactions, consistent with experimental data. These complementary approaches collectively strengthen the credibility of our predictions.

Finally, it is important to clarify that the primary objective of our pipeline is not only to identify conserved antigenic regions, but also to detect serotype-specific and non-cross-reactive epitopes. This dual-purpose strategy is essential to circumvent diagnostic cross-reactivity, a known challenge in flavivirus serology. By analyzing sequence and structural differences among NS1 proteins, our framework successfully pinpointed discriminative regions with potential for specific antibody binding. This approach aligns with evidence from the literature and reinforces the translational relevance of our findings for future diagnostic and therapeutic applications.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

Ab

Antibody

AI

Artificial intelligence

COVID-19

Coronavirus disease 2019

DENV

Dengue virus

EE

Electrostatic epitope

FORTE

Fast coarse-grained protein-protein model

IEDB

Immune epitope database

JEV

Japanese encephalitis virus

LFA

Lateral flow assay

mAbs

Monoclonal antibodies

NS1

Nonstructural protein 1

NCBI

National Center for Biotechnology Information

PROCEEDpKa

Prediction of electrostatic epitopes based on pK_a shifts

PDB

Protein Data Bank

RT-PCR

Reverse-transcriptase polymerase chain reaction

SASA

Solvent-accessible surface area

SEPPA3

Spatial epitope prediction of protein antigens 3

SPIDER

Scoring protein interaction decoys using exposed residues

WNV

West Nile virus

YFV

Yellow fever virus

ZIKV

Zika virus

Author contributions

M.C.C.G. and F.L.B.D.S. conceptualized and supervised the study, acquired funding, and took overall responsibility. R.A.M.A. and F.L.B.D.S. conceptualized and designed the computational experiments. F.L.B.D.S. carried out AlphaFold2 and 3 structure predictions, SPIDER-based analyses, and molecular simulations. R.S.V. and D.S.C. performed the prediction assays of antigenicity and B-cell epitopes. R.S.V. generated the 3D images of NS1 proteins and NS1-antibody complexes. P.C.L.J. performed the circle concept and analysis. R.S.V. produced and wrote the manuscript. F.V.C. and F.L.B.D.S. wrote and revised the manuscript. All of the authors contributed to the writing of the manuscript and approved the final draft.

Funding

This research was supported by Espírito Santo Research and Innovation Support Foundation (FAPES) grants (592/2018 and 1030/2022) (M.C.C.G). R.V.S. is also supported by the Espírito Santo Research and Innovation Support Foundation (FAPES). F.L.B.d.S. is supported by National Council for Scientific and Technological Development (CNPq) grants 305393/2020-0 and 307461/2025-4, and M.C.C.G. by grant 306433/2021-4. This work was also funded in part by the São Paulo Research Foundation (FAPESP) grant 2020/07158-2 (F.L.B.d.S.).

Data availability

Data is provided within the manuscript.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.