Differentiation between mpox infection and MVA immunization by a novel machine learning-supported serological multiplex assay

doi:10.21203/rs.3.rs-4822502/v1

Download PDF

Article

Differentiation between mpox infection and MVA immunization by a novel machine learning-supported serological multiplex assay

https://doi.org/10.21203/rs.3.rs-4822502/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

With case numbers exceeding 97,000 worldwide, the 2022 global mpox outbreak underscored the potential for zoonotic diseases with limited human-to-human transmission to trigger a widespread health crisis. Primarily men who have sex with men (MSM) were affected. Monitoring mpox-specific seroprevalences through epidemiological studies is essential, but challenging due to the cross-reactive antibody immune response which is induced by several orthopoxviruses including modified vaccinia virus Ankara (MVA)-based vaccines, which were used to help bring the outbreak under control. Here we show how machine learning (ML)-guided analysis of a serological multiplex assay that targets 15 immunogenic poxvirus proteins derived from monkeypox virus, vaccinia, and cowpox virus, can confidently discern between sera from patients post-mpox infection, post-MVA immunization, and pre-immunization or infection. Mean F1 scores representing the geometric means between precision and recall were calculated as metrics for the performance of six different ML models. The models were trained and tested on panels containing both sera taken in the early phase of seroconversion as well as sera taken six months after the peak of the mpox outbreak from individuals in an at-risk MSM population in Berlin. Scores ranged between 0.60 ± 0.05 and 0.81 ± 0.02 with Gradient Boosting Classifier (GBC) being the best performing algorithm. In order to ensure high confidence in our results, which is imperative in epidemiological studies, we excluded ambiguous results by using the robustly performing linear discriminant analysis’ (mean F1 scores 0.80 ± 0.02) classification confidence as a threshold. Hereby, sera with uncertain serostatus were segregated, leading to confident predictions with F1 scores above 0.90, at the cost of more inconclusive results for samples below the threshold. Beyond providing a valuable tool for monitoring mpox-specific antibodies, our work demonstrates how the combination of machine learning and multiplexing enables precise differentiation — and a deepened understanding — of complex antibody responses to closely related viruses.

Biological sciences/Microbiology/Virology/Pox virus

Biological sciences/Immunology/Vaccines/Live attenuated vaccines

Biological sciences/Microbiology/Vaccines/Live attenuated vaccines

Biological sciences/Computational biology and bioinformatics/Machine learning

Biological sciences/Biochemistry/Proteins/Viral proteins

Mpox disease is caused by infections with the monkeypox virus (MPXV), a member of the Orthopoxvirus genus within the Poxviridae family. While variola virus, the most well-known Orthopoxvirus and the causative agent of smallpox, was declared eradicated globally by the WHO in 1980, zoonotic transmissions of other orthopoxviruses, including MPXV, cowpox virus (CPXV), and vaccinia virus (VACV) were still being reported worldwide in recent years ¹. Until 2022 mpox infections were leading to sporadic outbreaks from both zoonotic and limited human-to-human transmissions. Initially, these outbreaks were confined to Western and Central Africa, with the notable exception of a significant outbreak in the USA in 2003 ^2,3 and single to few exported cases over the last 10 years ⁴. Then, an unprecedented outbreak occurred which has affected over 97,745 individuals in 116 countries (as of July 4th, 2024) ⁵. During this outbreak, human-to-human transmission through sexual activity was observed predominantly among men who have sex with men (MSM)⁶. The peak of the multinational outbreak occurred in mid-2022, also in Europe⁷ and Germany⁸. Since then, several factors including effective risk communication and fast set-up of vaccination offers for groups at risk, a high level of community engagement leading to behavioural changes, and increasing immunity in vulnerable populations by either natural infection or immunization, led to a steady decline in daily cases to a low — but still discernible — level of circulation ^9–11.

Although the WHO officially declared the 2022 mpox outbreak to no longer be a Public Health Emergency of International Concern (PHEIC) in May 2023, long-term efforts are still needed to improve surveillance, research, and control measures, including targeted surveillance and laboratory testing ¹². Additionally, mpox remains highly relevant as exemplified by the ongoing outbreak in the Democratic Republic of Congo, which is driven by a novel strain of clade I, which is characterized by higher case fatalities as compared to the clade IIb strain, which caused the 2022 worldwide outbreak ¹³.

Serological assays specific for antibodies induced by infection or immunization could play a crucial role in addressing pressing research questions, such as the level and duration of protection after natural infection and vaccination, monitoring of infection control strategies, and understanding the natural history of infection in at-risk populations ². However, distinguishing between the humoral immune responses induced by infection and immunization poses a significant challenge. The use of attenuated viruses for immunization leads to the production of polyclonal cross-reactive antibodies that are similar to the response after infection ^14–17. While this is desirable for vaccine-induced immunity, as exemplified by the smallpox vaccination campaign using vaccinia virus, it complicates mpox-specific sero-epidemiological studies and diagnostics two-fold^18–20: firstly many people vaccinated against smallpox in their early childhood retain detectable levels of antibodies into adulthood^21,22, and secondly during the 2022 outbreak, large-scale vaccination efforts were made using the Modified Vaccinia Ankara vaccine (MVA; marketed as Imvanex, Jynneos or Imvamune in Europe, the US and Canada, respectively), a safer attenuated virus originally developed for smallpox vaccination and now also approved for mpox immunisation ^23–25.

Unmodified serological assays including enzyme-linked immunosorbent assays (ELISAs), immunofluorescence assays (IFA), and neutralization tests (NTs) struggle to distinguish the humoral immune response against different orthopoxvirus species due to the complex antibody response which targets various immunogenic viral proteins with high sequence homology ^20,26,27. However, assays capable of differentiating between mpox infections and VACV vaccinations have been developed, using techniques like cross-absorption of antibodies, identification of mpox-specific short peptides or proteins, analysis of T-cell immunity, or combinations thereof ^28–32. Just recently, differential binding profiles of antibodies to recombinant proteins derived from MPXV or VACV facilitated the differentiation between mpox infection and MVA vaccination ^17,32.

Here we show how machine learning (ML)-aided analysis of a novel serological multiplex assay enables the differentiation of sera after mpox infection from MVA vaccination or pre-immune sera. To this end, we implemented a total of 19 proteins including 15 antigens derived from MPXV, VACV and CPXV in a bead-based multiplex assay based on the Luminex® technique for IgG and IgM detection ³³. Our assay demonstrated a strong correlation with three established reference assays (IFA, ELISA, NT), hereby accurately reflecting results from traditional methods. For ML analysis we used six algorithms. These were trained and tested on two extensive serum panels: one containing acute-phase sera and another from an epidemiological study of an at-risk MSM population, as well as both panels combined. For most sera, the six tested algorithms returned congruent results with Gradient Boost Classifier and Random Forest showing higher sensitivity, while Linear Discriminant Analysis was more specific and robust. The algorithms performed better in the acute phase of infection and in patients below 40 years of age at the timepoint of sampling. Despite a higher rate of misclassifications, the algorithms remained useful at later timepoints post-infection and showed some effectiveness even in patients likely vaccinated against smallpox in early childhood based on their age. Crucially, by only considering predictions with a confidence degree above a threshold for the classification confidence from the LDA algorithm, we were able to exclude low confidence predictions, which were also more frequently misclassified. By harnessing the potential of ML algorithms in combination with multiplex serology, we were able to develop an assay capable of distinguishing vaccinated from mpox infected individuals, thereby creating a powerful tool for epidemiological studies of mpox.

Our goal was to develop an assay capable of distinguishing between sera generated due to mpox infection, MVA immunization, or pre-immune sera without a known history of either infection or MVA immunization. To this end, we created a panel of recombinant poxviral proteins, previously described as immunogenic and/or targets of neutralizing antibodies (see Table 1) ^34–46. Among these proteins, we included five pairs of homologous recombinant proteins derived from either VACV or MPXV: A27L/A29L, L1R/M1R, D8L/E8L, A33R/A35R, B5R/B6R. Despite the high sequence similarity between these homologues (ranging between 93.6% and 98.8% sequence homology based on the comparison of the amino acid sequences of MPXV Zaire and VACV; see supporting figures S1 to S8 for pairwise sequence alignments of all proteins from representative orthopoxvirus strains and supporting tables S1 to S8 for sequence homology based on the amino acid sequences), it has been demonstrated that polyclonal antibodies produced post-infection or immunization exhibit differential binding to some of these antigens, namely A27L and L1R. An observation which has been utilized for serological differentiation by ELISA ¹⁷. As another potential marker for differentiation, we included the recombinant ATI protein, previously used to distinguish between Dryvax vaccination (employed during the smallpox eradication campaign) and MVA immunization ⁴⁷. Additionally, H3L (VACV) and A5L (CPXV) were included as important immune dominant antigens ^39,48. Finally, we included a complex antigen mixture from VACV infected cells (and the corresponding uninfected cells) to encompass the entire complexity of the antibody immune response, characterized by high redundancy and plasticity ^20,27.

Table 1

Overview of poxviral antigens implemented in the multiplex assay.
Name	Source strain	Virion location	Function	Remarks
A27L/A29L	VACV/MPXV	IMV membrane	Attachment⁸⁰	Neutralizing mAbs^38,81,82
L1R/M1R	VACV/MPXV	IMV membrane	Entry/Fusion^83,84	Neutralizing mAbs³⁶
D8L/E8L	VACV/MPXV	IMV membrane	Attachment⁸⁵	Immunodominant, Neutralizing mAbs⁴⁰
H3L	VACV	IMV membrane	Attachment⁸⁶	Immunodominant, Neutralizing mAbs³⁹
A33R/A35R	VACV/MPXV	EEV membrane	EEV-spread, induction of actin-tails^87,88	EEV-neutralising mAbs^35,44
B5R/B6R	VACV/MPXV	EEV membrane	EEV-formation and transport^89,90	EEV-neutralising mAbs⁴⁶
A5L	CPXV	Core protein	Morphogenesis⁹¹	Immunodominant, cross reactive antigen⁴⁸
ATI-C/-N	CPXV	A-type inclusion bodies	Protects virion infectivity after cell lysis⁹²	Differentiation of MVA/Dryvax⁴⁷
Delta	VACV	Whole proteome	-	Normalised to uninfected cell lysate²⁰

Initially, we evaluated our novel multiplex assay using serum panels previously characterised by well-established reference methods, including an in-house ELISA, IFA, or virus NT ^19,20. This was done to ensure that the results from our new method agreed with those from established assays. Additionally, our goal was to verify the immunogenicity of the included antigens, and to assess the contribution of antibodies targeting individual recombinant antigens to the overall antibody immune response that is detected in the reference assays. We observed a high degree of agreement in both IgG and IgM measurements between our multiplex assay and the tested reference assays (Fig. 1). Notably, we found strong correlations between the binding to the complex VACV lysate antigen, normalized against the uninfected cell lysate antigen (resulting in the “Delta” response), and the results from the reference assays for both IgG and IgM (Fig. 1a). This indicates that complex antigen mixtures can be effectively adapted to bead-based multiplex assays. The high level of agreement with IFA and NT results suggests that the multiplex assay yields results consistent with reference assays, enabling reliable sero-diagnostics with higher throughput and automation. We determined cut-off values for each of the orthopoxvirus-specific antigens using IFA results as ground truth (Titer ≤ 1:80 negative, above positive) in a ROC analysis for both IgG and IgM results (supporting figure S9, supporting tables S9 and S10). Here, the highest level of agreement was found between the complex Delta antigen and the IFA results. From the recombinant proteins, D8L/E8L, H3L, A33R/A35R, and B5R/B6R agreed best with the IFA results (performance parameters shown in supporting tables S11 and S12). Next, we compared the binding to recombinant antigens in the multiplex assay with binding to the complex Delta antigen using the multiplex assay or ELISA. Here, we found the highest antibodies levels were directed against D8L/E8L, H3L, A33R/A35R, and B5R/B6R both for IgG and IgM detection (Fig. 1b and supporting figure S10). This also held true for the comparison of the IgG responses against the IFA or NT results (Fig. 1c and d, IgM results see supporting figure S11). These observations were strongly supported by data analysis using pearson’s or spearman correlation coefficients (supporting figure S12 and supporting tables S13 to S15) where those antigens showed the closest correlation to both the Delta antigen as well as ELISA, IFA, and NT results. Other antigens tested (A27L/A29L, L1R/M1R, ATI-C, ATI-N, and A5L) also showed significant correlations with the IFA and NT binding assays, albeit to a lesser extent. These findings indicate that all tested orthopoxvirus-specific antigens are immunogenic, with certain antigens contributing more significantly to the overall immune response than others.

Differential immune response after mpox infection and MVA immunisation revealed by the novel multiplex assay

After demonstrating that our multiplex assay closely mirrors the results of other serological assays, the next step was to determine, if our assay could differentiate between the antibody immune responses induced by mpox infection and MVA immunization by utilizing supervised machine learning algorithms. To achieve this, we analysed both IgG and IgM immune responses in two distinct serum panels, which both comprised sera post-mpox infection, post-MVA immunization, and pre-immunisation or infection, but differed with regards to their timepoint of collection (Fig. 2).

The first panel, labelled “acute”, included sera from PCR-confirmed mpox cases collected during the peak of Germany's mpox outbreak from May to August 2022, in the acute or early convalescent phase of the infection (n = 307). It also comprised sera collected two weeks after prime or boost immunization with the Imvanex vaccine (n = 48). The pre-immune sera in this panel were obtained prior to MVA immunization (n = 16) and before the mpox outbreak for diagnostics of measles, mumps, or rubella virus infections (n = 176). The second panel, termed “epi” (short for epidemiological), consisted of 1,120 sera collected from April to June 2023, after which the mpox outbreak had largely subsided in Germany. Those samples were collected from MSM visiting MSM-friendly practices for general STI screening, hence embracing a study population eligible for mpox vaccination and at risk for mpox infection. Metadata with regards to self-reported mpox infection (n = 59), Imvanex immunization (n = 476) or neither mpox infection nor Imvanex immunization (n = 324) was available for a total of 859 sera. Sera in this panel were gathered to conduct a sero-epidemiological study on the prevalence of antibodies post-mpox infection and/or MVA immunization following the outbreak and vaccination campaign. We deliberately chose these distinct panels to replicate scenarios commonly encountered in serological assay development during acute outbreak phases and for subsequent epidemiological studies. We evaluated the performance of various machine learning algorithms, namely Random Forrest (RF), Fuzzy Rule-based Classification (FRBC), Gradient Boosting Classifier (GBC), Linear Discriminant Analysis (LDA), and combinations of LDA and RF or LDA and FRBC, which were trained and tested on the same panel, or on both panels combined. Additionally, algorithms trained on the “acute” panel were applied to predict outcomes in the “epi” panel, and vice versa.

Before training the different ML algorithms, we analysed the IgG and IgM immune response targeting the different antigens stratified by the different panels (“acute” or “epi”) or infection/immunization status (pre, MVA, mpox) (Fig. 3a and b). Additionally, we analysed differences with regards to possible childhood immunizations against smallpox by using an age-based cut-off to classify sera with high probabilities for having received a vaccination (age ≥ 60 years as of 2023), low probabilities (age below 40 years) or ambiguous serostatus (between). When we compared the IgG and IgM immune response over all analytes, we found that the IgG response was higher in patients with presumed childhood vaccination against smallpox (linear mixed model: normalized data ~ childhood immunisation (“Yes” > “No”); intercept = 0.49, slope = -0.05; Wilcoxon p < 2.2e-16) whereas the IgM response was higher in patients without presumed childhood vaccination (intercept = 0.39, slope = 0.02; Wilcoxon p < 2.2e-16, supporting figure S13) irrespective of the serostatus (Pre, MVA, mpox) or serum panel tested (“acute”, “epi”). To quantify the impact of the three categorial variables childhood immunisation, serostatus, or analyte, on the antibody immune profile we performed four three-way ANOVAs, stratified by antibody isotype (IgG or IgM) and serum panel (“acute” or “epi”) (supporting tables S16 to S19) with subsequent determination of the simple main effects (supporting tables S20 to S23) and pairwise comparisons (supporting tables S24 to S27). When we analysed the simple main effects, we found significant contributions from almost all antigens to the IgG response (supporting tables S20, S22). Interestingly, we found antigens contributing differently, depending on the serum panel. In the “acute” panel A27L, ATI-C, D8L, ATI-N, and A29L contributed most in patients with assumed childhood vaccination, while in patients without assumed childhood vaccination, A29L, A35R, B6R, A27L, and E8L contributed most. In the “epi” panel, E8L, D8L, Delta, B5R/B6R, A33R/A35R contributed most in patients irrespective of assumed childhood vaccination. As the IgM response was weaker in patients with previous childhood immunization, fewer antigens contributed significantly to differences between the three serostatuses, with D8L/E8L being the most prominent. Besides that, B6R and A33R/A35R also contributed significantly in both smallpox-vaccinated and non-vaccinated patients. As expected, the overall contribution of the IgM response was much more pronounced in the “acute” panel as compared to the “epi” panel (supporting tables S21, S23). When we performed pairwise comparisons, we found statistically significant differences between pre-immune sera and sera after either MVA immunisation or mpox infections for most antigens in the “acute” panel, when the panels were stratified by presumed childhood immunisation (Fig. 3c and supporting table S24). However, differentiation between MVA immunisation and mpox infection was not possible based on binding to most antigens when analysed in isolation. Additionally, as the immune response in younger patients without previous childhood immunisation against smallpox was generally lower, the immune response after MVA vaccination in those patients was similar to pre-immune sera of older patients with a high likelihood of a previous childhood immunization against smallpox. Remarkably, the immune response directed at the N-terminal fragment of the A-type inclusion protein (ATI-N) was significantly elevated only after mpox infection, in comparison to both pre-immune sera and sera after MVA vaccination. This indicated that reactivity against ATI-N could be a valuable marker to discern acute infection from previous vaccination. As expected, the IgM response in the “acute” panel was elevated significantly mostly in younger patients without previous childhood vaccination (Fig. 3d and supporting table S25). Interestingly, those differences were more pronounced in the MPXV derived proteins as compared to the VACV derived proteins. When we analysed the IgG immune profile in the “epi” panel, we found the highest immune response after mpox infection, followed by MVA vaccination (Fig. 3e and supporting table S26). Due to the later timepoints of sampling after infection, the IgM response was lower with fewer significant differences (Fig. 3f and supporting table S27) while the IgG response especially against the immune dominant antigens D8L/E8L, A33R/A35R, and B5R/B6R was fully matured with fewer low positive signals in the mpox infected patients as compared to the acute phase sera. Conversely, the difference in the IgG response against ATI-N after mpox infection as compared to the pre-immune sera or after MVA vaccination was still highly significant, yet less pronounced than in the acute panel, indicating that ATI-N might perform best in the early phase of mpox infection as a marker for differentiation.

Lastly, we examined whether binding to homologue antigens derived from MPXV or VACV differed among the three immune statuses (pre-immune, MVA, mpox) stratified by a presumed childhood vaccination against smallpox (supporting Figure S14). We compared the IgG and IgM binding responses to MPXV-derived proteins versus VACV-derived proteins across all five tested antigen pairs. We observed stronger binding to MPXV proteins post-mpox infection and to VACV proteins post-MVA immunization for the antigen pairs A33R/A35R, B5R/B6R, D8L/E8L, and to a lesser extent, A27L/A29L. However, no difference in binding to MPXV vs. VACV derived proteins was observed for the L1R/M1R antigen pair. The effect was more pronounced in younger patients without assumed childhood immunization against smallpox for both IgG and IgM binding.

Discrimination of sera post-MVA immunization and post-mpox infection using the novel multiplex assay and machine learning-guided analysis

As we observed notable differences between pre-immune sera and those collected post-mpox infection or post-MVA vaccination, particularly with regard to certain antigens, we trained supervised ML algorithms, to distinguish between the three different serostatuses: pre-immunization, post-MVA vaccination, and post-mpox infection (Fig. 2b). Additionally, we tested combinations of both LDA and RF or FRBC where LDA was employed to distinguish the most relevant features or characteristics for the collected data and, thus, reduce the number of dimensions while RF or FRBC were trained on the updated data. Regarding orthopoxvirus-specific antigens, we included all except L1R and M1R. These were excluded due to batch-to-batch variations observed during the antigen-coupling process for L1R (supporting Figure S15), their relatively weaker correlation with the reference assays compared to other tested antigens (Fig. 1), as well as their weaker contribution to the mpox-specific immune response (Fig. 3). In addition to comparing different ML algorithms, we investigated the influence of different training datasets, hyperparameters, and features on the prediction performance. Firstly, we used data from either the “acute”, the “epi” panel, or a combination of both for training to assess how the algorithms perform across different datasets. As the “epi” panel comprised sera from an at-risk population of MSM, we excluded sera from younger patients (< 40 years at the time of sampling) labelled as pre-immune based on self-reported exposure to either MVA immunisation or mpox infection with clear indicators for exposure to orthopoxviruses, or labelled as MVA with elevated ATI-N binding also hinting towards recent MPXV exposure, from the training dataset (n = 132 sera, supporting Figures S16 and S17). Secondly, we trained the models using either IgG data alone or a combination of IgG and IgM data to determine if including IgM results improved model performance or if IgG data alone was sufficient for discrimination. Although we did see that the childhood vaccination status has an impact on the immune response, we did not further stratify our training datasets for this hyperparameter. This was done to be able to include a more complete dataset as we did not have sufficient data on childhood vaccination status for all samples.

We then applied these models to predict serostatus outcomes in various panels. This included not only testing models on the same serum panels used for training, but also applying models trained on one panel (“acute” or “epi”) to make predictions on the other. Furthermore, all models were trained on a combined dataset containing results from both panels (“all”). To validate the models trained and tested on the same serum panels, we employed a 5-fold cross-validation approach, repeated thrice. For assessing the transferability of the models between different panels, we used the complete “acute” dataset for training when testing on the “epi” dataset, and vice versa.

When we compared accuracy, precision, recall, and F1 scores for the test datasets of the 5-fold cross-validation approach, we found marked differences between the tested panels and ML algorithms (Fig. 4a and supporting table S28). With regards to the tested ML algorithms, GBC showed the best performance, followed by LDA and LDA RF, which performed approximately equally well, while both RF and FRBC performed slightly worse with LDA FRBC showing the overall worst performance. The inferior performance in the latter cases is caused by the dimensional reduction by LDA, which leads to information loss, that affects FRBC more than RF. When we compared the different panels used for training and testing of the ML algorithms, we found good and largely comparable performances when homologue panels were used for training and testing (“acute acute”, “epi epi”, and “all all”). Of those, the performance was slightly better in the “acute acute” panel as compared to both the “epi epi” and “all all” panels. Additionally, the inclusion of IgM data improved the performance, which was more pronounced in the “acute acute” panel. Conversely, the overall performance decreased dramatically, when the “acute” panel was used for training of the algorithms and the models were tested on the “epi” panel and vice versa, highlighting the observed differences in the immune signature between both panels. The performance was better, when the “epi” panel was used as the training dataset and testing was done on the “acute” panel, and LDA-based algorithms seemed to be more robust, however the performance was still not satisfactory in either of these cases (mean F1 scores of ~ 0.6 as compared to ~ 0.8 in the homologue panels).

Next, we analysed the feature importance for those algorithms, where this information was available (LDA, GBC, RF) on the different panels and antibody isotypes used for training. When the algorithms were trained on IgG data from the “acute” panel, ATI-N, A27L/A29L, and ATI-C contributed most, i.e. the algorithms relied on those antigens more often for differentiation as compared to the other antigens for GBC and RF. For LDA, ATI-N and A29L contributed most to differentiation in most cases of the 5-fold cross validation, followed by D8L, B5R, and Delta. When the IgM response was included, the IgM response of E8L and A33R contributed most for the LDA, followed by the ATI-N IgM response. Similar results were obtained for GBC and RF, but with the IgM response against A33R contributing most, followed by E8L and ATI-N. When the algorithms were trained on the “epi” panel, E8L and D8L were contributing most with ATI-N still contributing, but to a lesser degree as compared to the “acute” panel, while IgM contribution was also much less important for the differentiation between the different serostatuses as compared to the “acute” panel (supporting figures S18 to S20).

To assess the specificity of the different ML algorithms trained and tested on the combined dataset (“all all”) in dependence of a presumable childhood vaccination against smallpox, we compared the predictions on a panel of 88 negative sera, which have been sampled before the mpox outbreak, stratified by age groups (Fig. 4b). Here we found that, depending on the algorithm and antibody isotypes used for the predictions, RF and GBC tended to classify sera in older subjects as “mpox”, indicating that those algorithms have a lower specificity in subjects with presumed childhood vaccination. Conversely, LDA was more robust, leading to fewer false misclassifications in older subjects. Classifications as “MVA” also occurred with higher frequencies, indicating that childhood vaccination might also be confused with MVA vaccination to a degree. In agreement with the overall performance, FRBC based algorithms also showed higher rates of misclassification in comparison with the other algorithms tested.

Finally, as we observed a higher rate of misclassification in older patients indicating a more challenging differentiation in patients with previous childhood vaccination, we compared the prediction of the GBC and LDA algorithms on the “all all” dataset in comparison to the ground truth either on all data, or, where available, on a dataset comprising patients with a high likelihood of childhood vaccinations (age 60+), on a data set with patients without childhood vaccinations (age < 40) or with a more ambiguous vaccination status (age between) (Fig. 4c; results for RF and FRBC in supporting figure S20, full results supporting table S29). Here we found that classifications of pre-immune or MVA sera as mpox sera were increased in patients with a high likelihood of childhood smallpox vaccination. This effect was more pronounced when the GBC algorithm was used as compared to the LDA algorithm (see supporting table S30 for assay performance of the LDA algorithm stratified by childhood vaccination). Conversely, a much smaller fraction of pre-sera from patients without a history of childhood vaccination were classified as mpox whereas more MVA and mpox sera were misclassified as pre-sera, especially when LDA was used for classification. This indicated that LDA was more specific, but less sensitive as compared to the GBC algorithm. Lastly, another advantage of the LDA algorithm is, that the classification probability, which the algorithm uses to predict different classes, can be used to further increase the stringency of the predictions (Fig. 5). By excluding low confidence predictions, which also have a higher rate of misclassification (Fig. 5a), the accuracy of predictions for the remaining samples could be increased. The cut-off values were determined by ROC analysis (Fig. 5b, supporting table S31) to optimize the separation between correctly classified and incorrectly classified samples while simultaneously minimizing the number of excluded samples. When we applied this approach to the different datasets used for training and testing, we were able to improve the performance parameters in the remaining samples above the cut-off (Fig. 5c) while the performance parameters decreased below the cut-off value (Fig. 5d). For example, the overall accuracy of predictions based on the IgG data above the cut-off value applied to the “all all” seropanel improved to 0.89 (95% CI: 0.87 to 0.91) while including 70% of all serum samples. However, the accuracy decreased to 0.63 (95% CI: 0.59 to 0.66) when including 52% of all serum samples, which fell below the cut-off value (see full results in supporting table S30). For samples close the cut-off values, classification scores fell both above and below the cut-off, depending on the run and replication of the 5-fold cross validation approach hence leading to sums above 100% for individual sera.

Differentiation between orthopoxviruses can be challenging due to the presence of cross-reactive antibodies ^49,50. Hence, technical measures to enhance discernibility like the absorption of these cross-reactive antibodies through preincubation with heterologous antigens ^28,51,52, differential binding to heterologous antigens ^17,53 or MPXV-specific peptides ^29–32,54 have been developed to differentiate between mpox infections and vaccination. Clearly, all the assays described to date have strengths as well as shortcomings. In the case of the cross-absorption ELISA, while antigen preparation and assay setup are relatively straightforward, they are feasible only in laboratories equipped to culture and which can safely handle or inactivate virus preparations ²⁰. Furthermore, this assay struggles to differentiate between individuals vaccinated against smallpox who also had clinically inapparent mpox infections and those only vaccinated against smallpox. This highlights an area of ambiguity, where differentiation is challenging for serological assays. Regarding assays performed with specific peptides, the identification of specific and immunogenic peptides for differentiation necessitates evaluating several peptides against a variety of antigens to pinpoint those that exhibit high sensitivity and specificity. Once such peptides are identified, they can be extremely valuable for the precise identification of mpox-specific antibodies ³¹. These peptides can also be integrated into serological multiplex assays in combination with recombinant proteins, as recently demonstrated ³². This particular assay showed, how the sensitivity and specificity of an assay can be improved by simultaneously targeting two MPXV-specific peptides and five cross-reactive recombinant proteins derived from either VACV or MPXV. It further highlighted the presence of unique epitopes that lead to differences in binding to homologous antigen pairs, despite a largely cross-reactive polyclonal immune response. This suggests that differentiation can be achieved without the need for cross-adsorption when peptides or recombinant proteins are targeted.

Building on this body of work, our assay not only confirms those findings but also introduces novel elements that can enhance the ability to distinguish between mpox infection and MVA immunization. By targeting a broader panel of five homologous pairs, four of which are included for differentiation purposes, we incorporated the distinct binding characteristics to the VACV/MPXV antigen pairs following either vaccination or infection, thereby harnessing the strengths of the individual antigens. In our study, antigens known for their immunogenicity (H3L, D8L, A33R, B6R) elicited strong reactions, while others exhibited reduced immunogenicity in humans (especially L1R), as previously described ^27,55–59. A key antigen for differentiation between MVA vaccination and mpox infection is ATI ⁴⁷. Although ATI was originally identified for differentiating between Dryvax and MVA immunization, it is also reliably triggered by mpox infection, as our work demonstrates.

The most significant aspect of our assay — compared to previous studies differentiating between infection and immunization for orthopoxviruses — is the use of machine learning algorithms for differentiation. This approach allowed us to integrate data from the combined immune response against multiple antigens simultaneously, avoiding the need for multi-step algorithms that were employed successfully in earlier research ^31,32. The use of machine learning algorithms, however, requires access to large datasets with numerous observations (patient samples) across different groups and numerous features (antigens). Therefore, this kind of assay became feasible only after the 2022 mpox outbreak provided significantly higher numbers of cases suitable for training and testing as well as the substantial amount of work that was done on immunogenic proteins over the last decades ^27,47,55,60. While the extensive data gathered by multiplex assays make the use of machine learning algorithms a logical choice for cutting-edge data analysis, there are still relatively few studies that employ this approach for differentiating between various states. Recent studies applied machine learning to serological multiplex data to distinguish recent from historical malaria transmission ^61,62, to enhance serodiagnostics for SARS-CoV-2 ⁶³, or to identify cases of active tuberculosis ⁶⁴.

Our work is also significant because it addresses the impact of various parameters on prediction outcomes, thereby highlighting certain caveats when working with different machine learning (ML) algorithms or datasets for classification tasks. When we trained our models on the acute dataset and applied them to the epidemiological panel, or vice versa, we observed a dramatic drop in prediction accuracy indicating an inherent difference between the two datasets. This underscores the need to train and test machine learning models on appropriate datasets to ensure the general applicability of a trained model. Otherwise, longitudinal differences between the acute and epidemiological phases of infection, as well as class imbalances, can lead to suboptimal assay performance. This consideration is particularly crucial in serological settings, where assay development often involves acute phase sera post-infection, and epidemiological studies utilize different study panels ¹⁸. Thus, careful selection of panels for training and testing of ML algorithms is essential, as they may be more sensitive to such variations compared to traditional data analysis. However, traditional analysis can also suffer from reduced sensitivity or specificity when different time points or panels are analysed ³². Consequently, the use of epidemiologically diverse panels during assay development is encouraged to account for these effects ⁶³.

Another strength of our assay is its validation against three reference assays for both IgG and IgM detection. This means that, in addition to ML-based differentiation, our assay serves as a valuable tool for comparatively quantifying the immune response following both infection and immunization. Our assay demonstrates a high level of agreement with all three methods. By utilizing both viral lysate and distinct recombinant proteins as antigens, it effectively captures both the breadth and the specificities of individual immune responses against the targeted antigens.

The performance of our assay was comparable to previously published results. Depending on the assay and setup, previously reported sensitivities and specificities in the literature vary, ranging from 93%/98% for a combination of mpox-specific peptides and recombinant proteins ³², 86%/90% for an ELISA with MPXV-specific peptides ³⁰, 86%/100% for a post-absorption ELISA²⁸, 100%/90% for another peptide-specific ELISA and 92–100%/100% for a combination between titres to VACV/MPXV binding and titre decline³¹. As the assay performance of our assay was highly dependent on the chosen algorithm and the seropanels used for training and testing, a direct comparison with regards to sensitivity and specificity is not easily possible. As representative data, we could have a look at the prediction of the LDA when sera from both the “acute” and “epi” panel were used for training and testing using the IgG data only. Here, sensitivity/specificity values of 86%/88% (Mpox), 80%/87% (MVA), and 68%/90% (Pre) were reached for our ML-based approach (supporting table S30). This was despite the fact that a large number of challenging sera were included in our assay's training and testing procedures. Sera from PCR-confirmed MPXV-positive patients in the acute panel contained both seropositive and seronegative samples in the early phase of seroconversion, which impacts the results due to lowered sensitivity. Additionally, we included sera regardless of childhood immunization status in the training and testing, acknowledging that previous childhood immunization poses an additional challenge for differentiation. Our assay performs slightly better in differentiating between pre-immunisation, post-MVA, and post-mpox immune responses in patients without previous smallpox vaccination in childhood (e.g. overall accuracy was 0.79 in sera of subjects without childhood vaccination as compared to 0.68 in sera of subjects with presumed childhood vaccination for the “all all” seropanel, trained on the IgG dataset, supporting Table S30). Due to its higher specificity, the LDA algorithms especially were also useful for predictions in patients with presumable smallpox vaccination. However, a certain rate of false positives in older patients should be taken into consideration when interpreting epidemiological findings using our novel assay. Finally, the epidemiological panel we used, was sampled from an at risk MSM population with a high likelihood of breakthrough infections or at least both mpox contact and immunization, hereby further adding to the challenge of differentiation. Another strength of our assay lies in its comprehensive development and subsequent testing, which included the thorough comparison between different serum panels and ML algorithms, the impact of using only IgG or both IgG and IgM to develop the assay, as well as multiple different antigens and a large panel of different sera. The robustness of LDA and the higher sensitivity of GBC and RF, but also the weaker performance of FRBC became obvious and guided the selection of appropriate algorithms for different tasks. Finally, the ability of the assay to utilize the confidence score of the LDA algorithms to effectively exclude uncertain samples, which are also more often misclassified, is another strongpoint (e.g. the overall accuracy increased to 0.89 for the “all all” seropanel, trained on the IgG dataset, leading to sensitivities/specificities of 95%/94% for mpox, 91%/94% for MVA and 84%/96% for Pre, see supporting table S30). Although this reduces the number of sera for which predictions are feasible by ~ 30%, the predictions for the remaining sera are highly reliable and accurate. This is particularly crucial for epidemiological studies in populations with a low expected prevalence of mpox, where maintaining a low number of false positives is essential due to the assay's high specificity. Additionally, this approach also agrees well with the biology of serology where samples can fall into an area of uncertainty, when titers after an infection fall back to baseline levels, the immune response was sampled to late or too early, or when the immune profile is not clearly distinguishable between infection, vaccination, or childhood vaccination. Hence, having an additional metrics to judge the reliability of predictions is a clear bonus over other assays.

In conclusion, our novel serological multiplex assay, enhanced by machine learning analysis, enables highly accurate differentiation between mpox infection and MVA vaccination. This novel assay eliminates the need for cross-absorption, effectively detecting subtle yet distinct differences across a broad spectrum of homologous antigen pairs and other relevant antigens thereby facilitating high-throughput measurements with minimal sample processing. Due to its modular setup, the assay performance could be improved by including more specific protein antigens, for example by adding MPXV-specific peptides to the multiplex approach. Moreover, the study's methodology, blending multiplexed serological assays with machine learning algorithms, sets a new standard in the field. It demonstrates how integrating cutting-edge technology can enhance the accuracy and efficiency of diagnostic assays. This approach is also readily applied to serological assays for other infectious diseases, hereby paving the way for more sophisticated and reliable public health surveillance tools.

Serum panels and ethics statements

For all serum panels used in this study, written consent by the patients and/or ethics or data security committees was obtained. Ethical clearance was obtained from the Berliner Ärztekammer (BÄK Eth-44/22). Data was only stored and handled pseudonymised without links to patient data apart from the data mentioned in this manuscript. All research was performed in accordance with relevant guidelines/regulations.

Where available and applicable, the following metadata were linked to the experimentally collected data: year of birth, previous infection with MPXV and timepoint since sample receipt, previous vaccination with MVA and dosing schedule (pre, prime, boost), previous vaccination against smallpox, results of reference assays as described (IFA titre, ELISA results, NT titre) ²⁰.

To establish the multiplex assay and compare it with IFA, ELISA, and NT, we selected a diverse panel of sera. This method comparison panel included known sera of patients after PCR-confirmed CPXV or mpox infections, as well as sera collected before and after MVA immunization, to encompass a wide range of IgG and/or IgM reactivities. For the IFA method comparison, we used a panel of 25 sera tested for IgG and IgM antibodies against orthopoxviruses (CPXV and/or VACV). This panel included sera from 20 CPXV infections, one PPV infection, and four post-MVA vaccinations. In the comparison with our in-house ELISA, we included 14 sera from 5 subjects at different stages: pre-MVA immunization (n = 5), two weeks post-prime immunization (n = 5), and two weeks post-boost immunization (n = 4). We included a total of 36 sera after PCR-confirmed mpox infection from 18 patients in the acute phase (< 33 days post first sample receipt). This included 11 patients with samples from one time point post-infection, two patients with samples from two time points, one patient with samples from three, and two patients each with samples from four and five time points. Additionally, 21 sera from patients tested by IFA for anti-orthopoxvirus antibodies (CPXV and/or VACV) were also included for the ELISA method comparison. Finally, for the NT assay comparison, we utilized a panel of 25 sera from 10 patients with PCR-confirmed mpox infections, specifically tested for VACV-neutralization capability as described ²⁰.

For training and evaluating the ML models, we utilized the following panels: First, the “acute” panel included 307 sera from 173 patients with PCR-confirmed MPXV infections, sampled between 2022/05/21 and 2022/08/11. These serum samples, sent along with swabs from various locations (lesions, rectum, throat, etc.), were exclusively from patients with PCR-confirmed mpox and were collected during the acute phase of infection or early convalescence (with the latest follow-up samples taken 33 days after receipt of the primary sample). Additionally, this panel included 48 sera from 18 subjects who voluntarily received MVA vaccinations. Of these, 17 sera were collected pre-immunization, 15 post-prime immunization, 13 after the first boost (four weeks post-prime), 5 before a second boost, and 4 after a second boost (about one year after the first). Two pre-pandemic control panels, each containing 88 sera from individuals without prior mpox infection or MVA vaccination, were also included. These panel were sent to the Robert Koch Institute for diagnostics of vaccine-preventable diseases before the mpox outbreak in Germany. The first panel has been previously described ³³, and the second panel was included to provide year of birth data not available in the first. Second, an epidemiological (“epi”) panel was collected for a seroprevalence study in a high-risk MSM population. Sampling occurred in MSM friendly practices specialised in HIV/STI care in Berlin from April to June 2023. Accompanying questionnaires gathered data on diagnosed mpox infections, MVA immunizations, year of birth, and childhood smallpox vaccinations (if known). Participants also reported any recent mpox-specific symptoms, despite not having a confirmed diagnosis. For ML algorithm training, 59 study subjects reported previous mpox infections, 476 reported MVA immunization (one or two doses), and 324 reported neither mpox infection nor MVA immunization. Subjects with both MVA immunization and mpox infection were categorized as MPXV infected.

Reagents and antibodies

The following antibodies were used during the study: R-phycoerythrin (PE) labelled (109-115-098) as well as unlabelled (109-005-008) goat anti-human IgG antibody (Fc-γ specific), or unlabelled goat anti-human IgM (109-005-129) or PE-labelled (709-116-073) donkey anti-human IgM antibody (Fc-µ-specific) as well as PE-labelled goat anti-mouse IgG antibody (Fc-γ specific), obtained from Dianova (Hamburg, Germany). Monoclonal mouse anti-A27 antibody A1/40 and biotinylated polyclonal antibodies generated in goats (anti D8L, anti H3L) or in rabbits (anti L1R) were prepared as previously described ⁶⁵. Rabbit anti-A33R antibodies were generated analogous to the anti L1R by Open Biosystems (Thermo Fisher Scientific, Dreieich, Germany) using the standard 70-day protocol as described by the manufacturer. Monoclonal anti-6×His antibody (TD264974) was obtained from Invitrogen (Thermo Fisher Scientific, Dreieich, Germany), Streptavidin-R-Phycoerythrin (SA-PE) PJRS27 was purchased from Agilent (Santa Clara, CA, USA). The following reagents were obtained through the NIH Biodefense and Emerging Infections Research Resources Repository, NIAID, NIH: polyclonal anti-Vaccinia virus (immune globulin G, Human), NR-2632 (termed VIG), and monoclonal anti-Vaccinia virus (WR) B5R protein, residues 20 to 275 (ectodomain), (similar to VMC-14 and produced in vitro), NR-429.

Recombinant proteins and antigens

An overview of all antigens used in this work is shown in supporting table S32. The following reagents were obtained through the NIH Biodefense and Emerging Infections Research Resources Repository, NIAID, NIH: Vaccinia virus (WR) A27L protein with C-terminal histidine tag, recombinant from baculovirus, NR-22133, Vaccinia virus (WR) L1R protein with C-terminal histidine tag, recombinant from baculovirus, NR-21986, Vaccinia virus, Western Reserve, A33R protein with C-terminal histidine Tag, recombinant from baculovirus, NR-2623, and Vaccinia virus (WR) B5R protein with N-terminal histidine tag, recombinant from baculovirus, NR-22132. Recombinant MPXV-derived proteins A29L, M1R, E8L, A35R, and B6R were obtained from ProteoGenix (Schiltigheim, France). The following proteins were custom expressed in E. coli (GenExpress, Berlin, Germany): D8L, H3L, A5L, ATI-N, and ATI-C. Lysates of VACV-infected Hep-2 cells or non-infected cells were produced as described ¹⁹. The purity and identity of all recombinant orthopoxvirus-specific antigens was checked by LC-MS/MS (see supporting information). Due to the high sensitivity of the method, contaminations deriving mostly from the host organisms used for recombinant expression (baculovirus, E. coli, HEK293 cells) could also be detected. However, the identity of all antigens was confirmed as antigen-specific peptides which generated the highest signal intensities, indicating sufficient purity.

Serological bead-based multiplex assay

Reagents and preparations of buffers used for coupling MagPlex® Microspheres and performing the multiplex assay have been described previously ³³. MagPlex® Microspheres covering bead regions 7, 8, 15, 20, 26, 33, 36, 38, 42, 48, 52, 54, 55, 59, 67, 76, 81, 82, and 89 were obtained from Luminex® Corporation (Austin, TX, USA) and coupled with proteins as outlined in supporting table S32. The serological multiplex assay was performed as described with minor modifications ³³. Briefly, coupling of proteins to MagPlex® microspheres was done using EDC/NHS-coupling chemistry following the suggestions published in the xMAP® Cookbook (4th Edition, Luminex® Corporation) and the antigen-specific amounts and bead-regions specified in supporting table S32. After coupling and blocking, the microspheres were quantified using a Neubauer counting chamber and the concentration adjusted to 1000 microspheres per µL. Successful coupling and retained antigenicity were proven by specific binding of antibodies targeting one of the recombinant proteins ⁶⁵ (supporting information and supporting figure S15 a) as well as broad recognition of all specific antigens by vaccinia immune globulin (VIG, BEI Resources, NR-2632) as described below.

All sera were inactivated by incubation with 0.5% Triton X-100/ 0.5% Tween-20 in PBS for 30 min at 56°C. This protocol was shown to be highly effective for MPXV inactivation from serum ²⁰. To determine IgG and IgM antibodies against the orthopoxviral proteins in the bead-based multiplex assay, each serum sample was further diluted 1:100 and 1:1000 in PBS/B (PBS + 1% BSA) assay buffer, before addition of the multiplex bead mix containing all antigen-coupled MagPlex™ beads to be tested. On each plate, five dilutions of VIG in a 1:4 dilution series starting at a 1:500 dilution was included as well as a blank control containing buffer only, and one negative serum measured in a 1:100 and 1:1000 dilution. A layout using only the first 6 columns of a 96 well plate of all sera and standards was mirrored in the second 6 columns so that both IgG and IgM detection could be performed on the same plate. The multiplex bead mix containing 1000 beads per region for each antigen tested was freshly prepared for each measurement. 50 µL of the mix was added per well of a polystyrene 96 well flat bottom microplate (Greiner Bio-One, Frickenhausen, Germany) to which 50 µL of diluted serum or standard were added per well, then incubated for 1 hour at room temperature on a plate shaker (IKA MTS 2/4 digital, IKA-Werke, Staufen im Breisgau, Germany) at 750 rpm. Using a magnetic plate washer (HydroSpeed™, Tecan, Männedorf, CH), plates were washed 3-times with 200 µL PBS/T per well, using soak times of 1 minute for each washing cycle. Subsequently, 100 µL PE-labelled goat anti-human IgG antibody was added at a concentration of 1 µg/mL in PBS/B assay buffer to each well in the first six columns of each 96 well plate, and 100 µL PE-labelled donkey anti-human IgM antibody was added to the second six columns. Plates were incubated and washed as described before. Thereafter the washed beads were resuspended in 100 µL sheath fluid (PBS) per well by shaking for 1 min at 750 rpm on the plate shaker. Finally, fluorescent signals were analysed on a Bio-Plex 200 instrument (Bio-Rad, Munich, Germany) using normal detector gain (RP1 target) collecting data from at least 50 beads per region. Raw data was exported into a Microsoft Excel® spreadsheet and further processed and analysed as described below.

Test of coupling efficiency with monospecific poly- or monoclonal antibodies

To test the coupling efficiency for the different orthopoxvirus-specific antigens, polyclonal, but monospecific antibodies directed against a single protein (anti D8L, anti H3L, anti A33R, and anti L1R) were biotinylated using Biotinamidohexanoyl-6-aminohexanoic acid-N-hydroxysuccinimide ester (Sigma-Aldrich) at a molar coupling ratio of 20. For the detection of A27L monoclonal antibody A1/40 and for B5R monoclonal antibody NR-429 were used. Polyclonal antibodies were tested at final concentrations of 10, 2, and 0.4 µg/mL, monoclonal antibodies at final concentrations of 1, 0.2 and 0.04 µg/mL while VIG was tested at 1:100, 1:500 and 1:2500 dilutions diluted in assay buffer (PBS, 1% BSA) as described above. Biotinylated polyclonal antibodies were detected using SA-PE at 2 µg/mL, mouse monoclonal antibodies were detected using goat anti-mouse IgG (Fc-γ specific) PE-labelled antibodies, and VIG was detected using goat anti-human IgG (Fc-γ specific) PE-labelled antibodies at a final concentration of 1 µg/mL each. Batch-to-batch variability for each new batch of coupled beads was tested as described with a 1:4 dilution series of VIG starting at a 1:100 dilution in a total of 7 dilutions together with a buffer only blank control in technical duplicates.

Reference assays

The three reference methods for validation of the multiplex assay, the immunofluorescence assay (IFA), an in-house indirect enzyme-linked immunoassay (ELISA), as well as a virus neutralization assay (NT) were performed as published ²⁰.

Data analysis for multiplex assay quantification and validation

All data analysis was done using the statistical programming language R (version 4.1.2) ⁶⁶. For data handling and wrangling, the packages rio, tidyverse⁶⁷, lubridate⁶⁸, purrr, and reshape⁶⁹ were used, for plotting, the packages corrplot, ggthemes, ggforce, ggvenn, and drc⁷⁰ were used. Calculation of cut-off values and determination of the assay performance of our novel multiplex assay was performed with the packages pROC⁷¹, rstatix, mrc, and datawizzard⁷². Quantification of the binding of sera to individual antigens was done using the package drLumi⁷³. For analysis of the epidemiological panel, a Shiny app was developed to read experimental data and metadata using the packages docxtractr and officer before the results were analysed and quantified as described below. To analyse the ML-based prediction, the package caret⁷⁴ was used.

To reference and normalize the binding responses of sera to the orthopoxvirus-specific antigens tested in the multiplex assay, we measured a standard curve using five VIG-dilutions on each plate. We set IgG concentrations to a starting concentration of 10 mg/mL, as starting dilutions for the standard were fivefold higher as compared to the first serum dilution tested (1:500 vs. 1:100). Using IgG concentrations ranging from 10,000 µg/mL to 39 µg/mL (five 1:4 dilutions), a 4-parameter logistic regression curve was fitted through log10-transformed MFI-values over log10-transformed IgG concentrations using the drLumi package for each orthopoxvirus-specific antigen. A background measurement using buffer only was used to set a bottom constraint for the curve-fitting. For each serum both dilutions (1:100 and 1:1000) were quantified by interpolation over the standard curve for each antigen. The upper and lower limits of quantification (LOQ) were determined based on a coefficient of variation of less than 30% for each standard curve. Measurements above or below were disregarded from quantification results. The final result was either the mean of both measurements with the dilution factors considered, only the measurement for the 1:1000 dilution if the measurement for the 1:100 was above the LOQ, or only the measurement for the 1:100 dilution if the measurement for the 1:1000 was below the LOQ. If all values were either below or above the LOQs in both dilutions, the antigen-specific minimum or maximum was used for the respective sera as a lower or upper limit. In the absence of a suitable IgM standard, IgM measurements were also quantified using the IgG standards to minimize plate-to-plate variability.

Cut-off values and assay performances (sensitivity, specificity, accuracy) based on method comparisons were calculated using the pROC package. To this aim, quantified binding signals for the orthopoxvirus-specific antigens of the multiplex assay were used as predictors to calculate receiver operator characteristics using classifications in positive and negative based on IFA titres below 1:320 or NT titres ≤ 14.

Three-way ANOVA

Three-way ANOVA was performed as described ⁷⁵ on the dataset with complete cases with the following grouping variables: childhood smallpox vaccination likely (“No”, “Yes”, based on age < 40 years or age > = 60 years), orthopoxvirus-specific antigen (A27L, A29L, L1R, M1R, D8L, E8L, H3L, A33R, A35R, B5R, B6R, A5L, ATI-C, ATI-N, Delta, i.e. VACV minus Hep-2 lysate), or serostatus with regards to infection or immunization (“Pre”, “MVA”, “Mpox”), which has also been used to train the ML-algorithms. A total of four ANOVA analysis were performed on datasets filtered for IgG or IgM data, on the “acute” or the “epi” panel. To this aim, outliers in the dataset were identified and excluded using the boxplot method implemented in the rstatix package ⁷⁶. Next, Shapiro-Wilk tests of normality were performed to test if the data is normally distributed (supporting figures S22 and S23). Although not all data for all antigens, serostatus, or childhood vaccination the prerequisite of normal distribution was met, we continued with the ANOVA analysis as the distributions were largely normally distributed or close to normal distributions.

Machine learning based differentiation

The machine learning based differentiation was implemented in Python using the scikit-learn⁷⁷ module. FRBC was implemented in Python but using skmoefs package ⁷⁸. To train and test the different ML-models, quantified IgG and IgM measurements from the “acute” and “epi” panels as described above were used as input. Input data were normalized using a min-max approach. As antigens, all orthopoxvirus-specific antigens were included with the exception of L1R and M1R. These antigens were excluded based on the overall poor IgG and IgM antibody immune response, as well as some difficulties with reproducibility during bead coupling of L1R, which lead to a largely inflated impact on model performance which could be traced back to inconsistent coupling efficacy during preliminary trainings. Each serum was classified as belonging to either the “acute” or “epi” panel, with three potential statuses with regard to immunization or infection: “pre” comprising sera without any known history of MVA immunization or mpox infection, “MVA” comprising sera with a known history of MVA vaccination, and “MPXV” comprising sera with PCR-confirmed (“acute” panel) or self-reported (“epi” panel) previous mpox infections. To keep our models simple and the stratified groups sufficiently large for training of the models, we did not further stratify the groups with regards to previous childhood immunization against smallpox, or number of MVA vaccinations (one or two) as these data were not available for all sera. The models were trained in a supervised learning approach to classify each serum as belonging to either one of the three groups “Pre”, “MVA”, or “mpox” based on the normalized data input. We trained a total of six models on different serum panels using different data as input (see Fig. 3), using the following algorithms: Random Forest, Gradient Boosting Classifier, and Linear Discriminant Analysis. This was done in order to only make predictions for samples with high reliability. The decision was based on the rationale that in an epidemiological setting it is more desirable to separate highly reliable predictions from uncertain results, in order to increase the overall validity of a study by avoiding misclassified data points. To determine the model performance, a 5-fold cross validation approach was applied, which was repeated three times to include 15 runs for each model and dataset. Each model was trained on either IgG and IgM data or IgG data alone, from either the “acute” panel, the “epi” panel or both panels combined (“all”). Finally, to test the transferability of models generated on one panel for prediction on another panel, we used either the “epi” panel for training and the “acute” panel for testing or vice versa. For the linear discriminant analysis, we also employed thresholds for confidence levels to separate confident from uncertain predictions (supporting table S23). Those thresholds were calculated on the datasets from the 5-fold cross validation approach with the prediction confidence of the LDA as variable and the correctness of the prediction as the grouping variable. ROC analysis was performed as described above. Metrices to evaluate and to compare different ML algorithms and parameters (accuracy, precision, recall, F1 score) wered calculated according to standard equations ⁷⁹.

Technical use disclosure

The use of ChatGPT 4.0 was focused on language editing tasks such as typesetting, rectifying spelling and grammatical inaccuracies, and rephrasing to enhance the readability and comprehensibility of existing text.

Tables

Figures

Additional Information

Supplementary information accompanies this paper at doi:

Competing Interests

The authors declare that they have no competing interests.

Author contributions

R.S., F.T., F.B., A.S., M.L.N., T.R., M.G. M.S., J.M., and D.S. designed and/or performed the experiments. S.A., D.B. and N.K. performed and/or supervised the ML analysis. S.A., D.B., M.S., H.W.M. and D.S. analysed the data. K.J., U.K., U.M., N.K., D.S. and A.N. conceived the study or parts of the study, B.G.D. and K.L obtained funding and supervised the project. D.S. and R.S. wrote the paper with input from all other authors. All authors read and approved the final manuscript.

Acknowledgements

This work was supported by grants from the German Federal Ministry of Health. The authors thank Josephina Lambrecht and Silvia Muschter for excellent technical assistance.

Silva NIO, de Oliveira JS, Kroon EG, Trindade GS, Drumond BP (2020) Here, There, and Everywhere: The Wide Host Range and Geographic Distribution of Zoonotic Orthopoxviruses. Viruses 13. https://doi.org/10.3390/v13010043
Adetifa I, Muyembe JJ, Bausch DG, Heymann DL (2023) Mpox neglect and the smallpox niche: a problem for Africa, a problem for the world. Lancet 401:1822–1824. https://doi.org/10.1016/S0140-6736(23)00588-3
Lum FM et al (2022) Monkeypox: disease epidemiology, host immunity and clinical interventions. Nat Rev Immunol 22:597–613. https://doi.org/10.1038/s41577-022-00775-4
Beer EM, Rao VB (2019) A systematic review of the epidemiology of human monkeypox outbreaks and implications for outbreak strategy. PLoS Negl Trop Dis 13:e0007791. https://doi.org/10.1371/journal.pntd.0007791
Centers for Disease Control and Prevention (2024) 2022–2023 Mpox Outbreak Global Map, <https://www.cdc.gov/poxvirus/mpox/response/2022/world-map.html
Thornhill JP et al (2022) Monkeypox Virus Infection in Humans across 16 Countries - April-June 2022. N Engl J Med 387:679–691. https://doi.org/10.1056/NEJMoa2207323
European Centre for Disease Prevention and Control/WHO Regional Office for Europe (2024) Mpox, Joint Epidemiological overview, https://monkeypoxreport.ecdc.europa.eu/
Selb R et al (2022) A shift from travel-associated cases to autochthonous transmission with Berlin as epicentre of the monkeypox outbreak in Germany, May to June 2022. Euro Surveill 27. https://doi.org/10.2807/1560-7917.ES.2022.27.27.2200499
European Centre for Disease Prevention and Control (ECDC). Communicable Disease Threats Report Week 47, 20-26 November (2022) https://www.ecdc.europa.eu/en/publications-data/communicable-disease-threats-report-20-26-november-2022-week-47 (2023)
Zhang XS et al (2024) Transmission dynamics and effect of control measures on the 2022 outbreak of mpox among gay, bisexual, and other men who have sex with men in England: a mathematical modelling study. Lancet Infect Dis 24:65–74. https://doi.org/10.1016/S1473-3099(23)00451-6
Copen CE et al (2022) Modifications to Sexual Behaviors Associated With Mpox (Monkeypox) Virus Transmission Among Persons Presenting for mpox Vaccination, Washington, DC, August-October, Sex Transm Dis 51, 54–60 (2024). https://doi.org/10.1097/OLQ.0000000000001889
World Health Organization (WHO) (2023) Fifth Meeting of the International Health Regulations (2005) (IHR) Emergency Committee on the Multi-Country Outbreak of mpox (monkeypox), https://www.who.int/news/item/11-05-2023-fifth-meeting-of-the-international-health-regulations-(2005)-(ihr)-emergency-committee-on-the-multi-country-outbreak-of-monkeypox-(mpox)
World Health Organization (WHO) (2024) Mpox - Democratic Republic of the Congo, <https://www.who.int/emergencies/disease-outbreak-news/item/2024-DON522
Cheng SMS et al (2023) Serological assays for differentiating natural COVID-19 infection from vaccine induced immunity. J Clin Virol 170:105621. https://doi.org/10.1016/j.jcv.2023.105621
Bohm K, Filomena A, Schneiderhan-Marra N, Krause G, Sievers C (2017) Validation of HAV biomarker 2A for differential diagnostic of hepatitis A infected and vaccinated individuals using multiplex serology. Vaccine 35:5883–5889. https://doi.org/10.1016/j.vaccine.2017.08.089
Albinsson B et al (2018) Distinction between serological responses following tick-borne encephalitis virus (TBEV) infection vs vaccination, Sweden 2017. Euro Surveill 23 https://doi.org/10.2807/1560-7917.ES.2018.23.3.17-00838
Otter AD et al (2023) Monkeypox virus-infected individuals mount comparable humoral immune responses as Smallpox-vaccinated individuals. Nat Commun 14:5948. https://doi.org/10.1038/s41467-023-41587-x
Reynolds MG et al (2010) A silent enzootic of an orthopoxvirus in Ghana, West Africa: evidence for multi-species involvement in the absence of widespread human disease. Am J Trop Med Hyg 82:746–754. https://doi.org/10.4269/ajtmh.2010.09-0716
Leendertz SAJ et al (2017) A Cross-Sectional Serosurvey of Anti-Orthopoxvirus Antibodies in Central and Western Africa. Viruses 9. https://doi.org/10.3390/v9100278
Grossegesse M et al (2023) Serological methods for the detection of antibodies against monkeypox virus applicable for laboratories with different biosafety levels. J Med Virol 95:e29261. https://doi.org/10.1002/jmv.29261
Hammarlund E et al (2003) Duration of antiviral immunity after smallpox vaccination. Nat Med 9:1131–1137. https://doi.org/10.1038/nm917
Li E et al (2023) Duration of humoral immunity from smallpox vaccination and its cross-reaction with Mpox virus. Signal Transduct Target Ther 8:350. https://doi.org/10.1038/s41392-023-01574-6
Kennedy RB, Ovsyannikova IG, Jacobson RM, Poland GA (2009) The immunology of smallpox vaccines. Curr Opin Immunol 21:314–320. https://doi.org/10.1016/j.coi.2009.04.004
Kennedy JS, Greenberg RN (2009) IMVAMUNE: modified vaccinia Ankara strain as an attenuated smallpox vaccine. Expert Rev Vaccines 8:13–24. https://doi.org/10.1586/14760584.8.1.13
European Medicines Agency (2022) EMA recommends approval of Imvanex for the prevention of monkeypox disease, https://www.ema.europa.eu/en/news/ema-recommends-approval-imvanex-prevention-monkeypox-disease
Lim CK et al (2023) Mpox diagnostics: Review of current and emerging technologies. J Med Virol 95:e28429. https://doi.org/10.1002/jmv.28429
Benhnia MR et al (2008) Redundancy and plasticity of neutralizing antibody responses are cornerstone attributes of the human immune response to the smallpox vaccine. J Virol 82:3751–3768. https://doi.org/10.1128/JVI.02244-07
Dubois ME, Slifka MK (2008) Retrospective analysis of monkeypox infection. Emerg Infect Dis 14:592–599. https://doi.org/10.3201/eid1404.071044
Dubois ME, Hammarlund E, Slifka MK (2012) Optimization of peptide-based ELISA for serological diagnostics: a retrospective study of human monkeypox infection. Vector Borne Zoonotic Dis 12:400–409. https://doi.org/10.1089/vbz.2011.0779
Taha TY et al (2023) Design and Optimization of a Monkeypox virus Specific Serological Assay. Pathogens 12 https://doi.org/10.3390/pathogens12030396
Hammarlund E et al (2005) Multiple diagnostic techniques identify previously vaccinated individuals with protective immunity against monkeypox. Nat Med 11:1005–1011. https://doi.org/10.1038/nm1273
Yates JL et al (2023) Development of a Novel Serological Assay for the Detection of Mpox Infection in Vaccinated Populations. medRxiv https://doi.org/10.1101/2023.04.18.23288419
Stern D et al (2023) A bead-based multiplex assay covering all coronaviruses pathogenic for humans for sensitive and specific surveillance of SARS-CoV-2 humoral immunity. Sci Rep 13:21846. https://doi.org/10.1038/s41598-023-48581-9
Aldaz-Carroll L et al (2005) Physical and immunological characterization of a recombinant secreted form of the membrane protein encoded by the vaccinia virus L1R gene. Virology 341:59–71. https://doi.org/10.1016/j.virol.2005.07.006
Hooper JW, Custer DM, Schmaljohn CS, Schmaljohn AL (2000) DNA vaccination with vaccinia virus L1R and A33R genes protects mice against a lethal poxvirus challenge. Virology 266:329–339. https://doi.org/10.1006/viro.1999.0096
Wolffe EJ, Vijaya S, Moss B (1995) A myristylated membrane protein encoded by the vaccinia virus L1R open reading frame is the target of potent neutralizing monoclonal antibodies. Virology 211:53–63. https://doi.org/10.1006/viro.1995.1378
Berhanu A et al (2008) Vaccination of BALB/c mice with Escherichia coli-expressed vaccinia virus proteins A27L, B5R, and D8L protects mice from lethal vaccinia virus challenge. J Virol 82:3517–3529. https://doi.org/10.1128/JVI.01854-07
Rudraraju R, Ramsay AJ (2010) Single-shot immunization with recombinant adenovirus encoding vaccinia virus glycoprotein A27L is protective against a virulent respiratory poxvirus infection. Vaccine 28:4997–5004. https://doi.org/10.1016/j.vaccine.2010.05.023
Davies DH et al (2005) Vaccinia virus H3L envelope protein is a major target of neutralizing antibodies in humans and elicits protection against lethal challenge in mice. J Virol 79:11724–11733. https://doi.org/10.1128/JVI.79.18.11724-11733.2005
Diesterbeck US et al (2021) Characterization of an In Vivo Neutralizing Anti-Vaccinia Virus D8 Single-Chain Fragment Variable (scFv) from a Human Anti-Vaccinia Virus-Specific Recombinant Library. Vaccines (Basel) 9. https://doi.org/10.3390/vaccines9111308
Matho MH et al (2014) Murine anti-vaccinia virus D8 antibodies target different epitopes and differ in their ability to block D8 binding to CS-E. PLoS Pathog 10:e1004495. https://doi.org/10.1371/journal.ppat.1004495
Matho MH et al (2012) Structural and biochemical characterization of the vaccinia virus envelope protein D8 and its recognition by the antibody LA5. J Virol 86:8050–8058. https://doi.org/10.1128/JVI.00836-12
Matho MH et al (2018) Structure-function characterization of three human antibodies targeting the vaccinia virus adhesion molecule D8. J Biol Chem 293:390–401. https://doi.org/10.1074/jbc.M117.814541
Law M, Hollinshead R, Smith GL (2002) Antibody-sensitive and antibody-resistant cell-to-cell spread by vaccinia virus: role of the A33R protein in antibody-resistant spread. J Gen Virol 83:209–222. https://doi.org/10.1099/0022-1317-83-1-209
Aldaz-Carroll L et al (2005) Epitope-mapping studies define two major neutralization sites on the vaccinia virus extracellular enveloped virus glycoprotein B5R. J Virol 79:6260–6271. https://doi.org/10.1128/JVI.79.10.6260-6271.2005
Bell E et al (2004) Antibodies against the extracellular enveloped virus B5R protein are mainly responsible for the EEV neutralizing capacity of vaccinia immune globulin. Virology 325:425–431. https://doi.org/10.1016/j.virol.2004.05.004
Hermanson G et al (2012) Measurement of antibody responses to Modified Vaccinia virus Ankara (MVA) and Dryvax((R)) using proteome microarrays and development of recombinant protein ELISAs. Vaccine 30:614–625. https://doi.org/10.1016/j.vaccine.2011.11.021
Miller L, Richter M, Hapke C, Stern D, Nitsche A (2011) Genomic expression libraries for the identification of cross-reactive orthopoxvirus antigens. PLoS ONE 6:e21950. https://doi.org/10.1371/journal.pone.0021950
Diaz-Canova D, Moens UL, Brinkmann A, Nitsche A, Okeke MI (2022) Genomic Sequencing and Analysis of a Novel Human Cowpox Virus With Mosaic Sequences From North America and Old World Orthopoxvirus. Front Microbiol 13:868887. https://doi.org/10.3389/fmicb.2022.868887
Gruber CEM et al (2018) Whole Genome Characterization of Orthopoxvirus (OPV) Abatino, a Zoonotic Virus Representing a Putative Novel Clade of Old World Orthopoxviruses. Viruses 10 https://doi.org/10.3390/v10100546
Esposito JJ, Obijeski JF, Nakano JH (1977) Serological relatedness of monkeypox, variola, and vaccinia viruses. J Med Virol 1:35–47. https://doi.org/10.1002/jmv.1890010107
Hutchinson HD, Ziegler DW, Wells DE, Nakano JH (1977) Differentiation of variola, monkeypox, and vaccinia antisera by radioimmunoassay. Bull World Health Organ 55:613–623
Marennikova SS, Malceva NN, Habahpaseva NA (1981) ELISA–a simple test for detecting and differentiating antibodies to closely related orthopoxviruses. Bull World Health Organ 59:365–369
Kinganda-Lusamaki E et al (2023) Use of Mpox Multiplex Serology in the Identification of Cases and Outbreak Investigations in the Democratic Republic of the Congo (DRC). Pathogens 12 https://doi.org/10.3390/pathogens12070916
Townsend MB et al (2013) Humoral immunity to smallpox vaccines and monkeypox virus challenge: proteomic assessment and clinical correlations. J Virol 87:900–911. https://doi.org/10.1128/JVI.02089-12
Amanna IJ, Slifka MK, Crotty S (2006) Immunity and immunological memory following smallpox vaccination. Immunol Rev 211:320–337. https://doi.org/10.1111/j.0105-2896.2006.00392.x
Moss B (2011) Smallpox vaccines: targets of protective immunity. Immunol Rev 239:8–26. https://doi.org/10.1111/j.1600-065X.2010.00975.x
Yefet R et al (2023) Monkeypox infection elicits strong antibody and B cell response against A35R and H3L antigens. iScience 26:105957. https://doi.org/10.1016/j.isci.2023.105957
Lantto J et al (2011) Capturing the natural diversity of the human antibody response against vaccinia virus. J Virol 85:1820–1833. https://doi.org/10.1128/JVI.02127-10
Kennedy RB, Ovsyannikova IG, Haralambieva IH, Grill DE, Poland GA (2022) Proteomic assessment of humoral immune responses in smallpox vaccine recipients. Vaccine 40:789–797. https://doi.org/10.1016/j.vaccine.2021.12.033
Macalinao MLM et al (2023) Analytical approaches for antimalarial antibody responses to confirm historical and recent malaria transmission: an example from the Philippines. Lancet Reg Health West Pac 37:100792. https://doi.org/10.1016/j.lanwpc.2023.100792
Byrne I et al (2023) Serological evaluation of risk factors for exposure to malaria in a pre-elimination setting in Malaysian Borneo. Sci Rep 13:12998. https://doi.org/10.1038/s41598-023-39670-w
Rosado J et al (2021) Multiplex assays for the identification of serological signatures of SARS-CoV-2 infection: an antibody-based diagnostic and machine learning study. Lancet Microbe 2:e60–e69. https://doi.org/10.1016/S2666-5247(20)30197-X
Rashidi HH, Dang LT, Albahra S, Ravindran R, Khan IH (2021) Automated machine learning for endemic active tuberculosis prediction from multiplex serological data. Sci Rep 11:17900. https://doi.org/10.1038/s41598-021-97453-7
Stern D et al (2016) Development of a Genus-Specific Antigen Capture ELISA for Orthopoxviruses - Target Selection and Optimized Screening. PLoS ONE 11:e0150110. https://doi.org/10.1371/journal.pone.0150110
R: A Language and Environment for Statistical Computing (2021)
Wickham H et al (1999) Welcome to the tidyverse. J Open Source Softw 4. https://doi.org/10.21105/joss.01686
Grolemund G, Wickham H (2011) Dates and Times Made Easy with lubridate. J Stat Softw 40:1–25
Wickham H (2007) Reshaping data with the reshape package. J Stat Softw 21
Ritz C, Baty F, Streibig JC, Gerhard D (2015) Dose-Response Analysis Using R. PLoS ONE 10:e0146021. https://doi.org/10.1371/journal.pone.0146021
Robin X et al (2011) pROC: an open-source package for R and S + to analyze and compare ROC curves. BMC Bioinformatics 12:77. https://doi.org/10.1186/1471-2105-12-77
Patil I et al (2022) datawizard: An R Package for Easy Data Preparation and Statistical Transformations. Journal of Open Source Software 7, 4684 https://doi.org/10.21105/joss.04684
Sanz H et al (2017) An open-source package to manage data, calibrate, and conduct quality control of multiplex bead-based immunoassays data analysis. PLoS ONE 12:e0187901. https://doi.org/10.1371/journal.pone.0187901. drLumi
Kuhn M (2008) Building Predictive Models in R Using the caret Package. J Stat Softw 28:1–26. https://doi.org/10.18637/jss.v028.i05
Kassambara A (2020) ANOVA in R, https://www.datanovia.com/en/lessons/anova-in-r/
rstatix: Pipe-Friendly Framework for Basic Statistical Tests (2023)
Pedrogosa F et al (2011) Scikit-learn: Machine learning in Python. J Mach Learn Res 12:2825–2830
Gallo GF, Marcelloni V, Ducange F (2020) Information Processing and Management of Uncertainty in Knowledge-Based Systems. IPMU 2020. Communications in Computer and Information Science. Springer, ed MJ.; et al. Lesot
Rainio O, Teuho J, Klen R (2024) Evaluation metrics and statistical tests for machine learning. Sci Rep 14:6086. https://doi.org/10.1038/s41598-024-56706-x
Chung CS, Hsiao JC, Chang YS, Chang W (1998) A27L protein mediates vaccinia virus interaction with cell surface heparan sulfate. J Virol 72:1577–1585. https://doi.org/10.1128/JVI.72.2.1577-1585.1998
Meyer H, Osterrieder N, Czerny CP (1994) Identification of binding sites for neutralizing monoclonal antibodies on the 14-kDa fusion protein of orthopox viruses. Virology 200:778–783. https://doi.org/10.1006/viro.1994.1241
He Y et al (2007) Antibodies to the A27 protein of vaccinia virus neutralize and protect against infection but represent a minor component of Dryvax vaccine–induced immunity. J Infect Dis 196:1026–1032. https://doi.org/10.1086/520936
Bisht H, Weisberg AS, Moss B (2008) Vaccinia virus l1 protein is required for cell entry and membrane fusion. J Virol 82:8687–8694. https://doi.org/10.1128/JVI.00852-08
Foo CH et al (2009) Vaccinia virus L1 binds to cell surfaces and blocks virus entry independently of glycosaminoglycans. Virology 385:368–382. https://doi.org/10.1016/j.virol.2008.12.019
Hsiao JC, Chung CS, Chang W (1999) Vaccinia virus envelope D8L protein binds to cell surface chondroitin sulfate and mediates the adsorption of intracellular mature virions to cells. J Virol 73:8750–8761. https://doi.org/10.1128/JVI.73.10.8750-8761.1999
Lin CL, Chung CS, Heine HG, Chang W (2000) Vaccinia virus envelope H3L protein binds to cell surface heparan sulfate and is important for intracellular mature virion morphogenesis and virus infection in vitro and in vivo. J Virol 74:3353–3365. https://doi.org/10.1128/jvi.74.7.3353-3365.2000
Roper RL, Wolffe EJ, Weisberg A, Moss B (1998) The envelope protein encoded by the A33R gene is required for formation of actin-containing microvilli and efficient cell-to-cell spread of vaccinia virus. J Virol 72:4192–4204. https://doi.org/10.1128/JVI.72.5.4192-4204.1998
Katz E, Ward BM, Weisberg AS, Moss B (2003) Mutations in the vaccinia virus A33R and B5R envelope proteins that enhance release of extracellular virions and eliminate formation of actin-containing microvilli without preventing tyrosine phosphorylation of the A36R protein. J Virol 77:12266–12275. https://doi.org/10.1128/jvi.77.22.12266-12275.2003
Mathew E, Sanderson CM, Hollinshead M, Smith GL (1998) The extracellular domain of vaccinia virus protein B5R affects plaque phenotype, extracellular enveloped virus release, and intracellular actin tail formation. J Virol 72:2429–2438. https://doi.org/10.1128/JVI.72.3.2429-2438.1998
Wolffe EJ, Isaacs SN, Moss B (1993) Deletion of the vaccinia virus B5R gene encoding a 42-kilodalton membrane glycoprotein inhibits extracellular virus envelope formation and dissemination. J Virol 67:4732–4741. https://doi.org/10.1128/JVI.67.8.4732-4741.1993
Williams O, Wolffe EJ, Weisberg AS, Merchlinsky M (1999) Vaccinia virus WR gene A5L is required for morphogenesis of mature virions. J Virol 73:4590–4599. https://doi.org/10.1128/JVI.73.6.4590-4599.1999
Howard AR, Weisberg AS, Moss B (2010) Congregation of orthopoxvirus virions in cytoplasmic A-type inclusions is mediated by interactions of a bridging protein (A26p) with a matrix protein (ATIp) and a virion membrane-associated protein (A27p). J Virol 84:7592–7602. https://doi.org/10.1128/JVI.00704-10

There is NO Competing Interest.

SurteesetalSuppInformationsubmission.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Differentiation between mpox infection and MVA immunization by a novel machine learning-supported serological multiplex assay

Status:

Version 1

Abstract

Figures

Introduction

Results

Differential immune response after mpox infection and MVA immunisation revealed by the novel multiplex assay

Discussion

Online methods

Serum panels and ethics statements

Reagents and antibodies

Recombinant proteins and antigens

Serological bead-based multiplex assay

Test of coupling efficiency with monospecific poly- or monoclonal antibodies

Data analysis for multiplex assay quantification and validation

Three-way ANOVA

Machine learning based differentiation

Technical use disclosure

Figures

Declarations

Additional Information

Competing Interests

Author contributions

Acknowledgements

References

Additional Declarations

Supplementary Files

Status:

Version 1