Landscape drives zoonotic malaria prevalence in non-human primates

  1. Emilia Johnson
  2. Reuben Sunil Kumar Sharma
  3. Pablo Ruiz Cuenca
  4. Isabel Byrne
  5. Milena Salgado-Lynn
  6. Zarith Suraya Shahar
  7. Lee Col Lin
  8. Norhadila Zulkifli
  9. Nor Dilaila Mohd Saidi
  10. Chris Drakeley
  11. Jason Matthiopoulos
  12. Luca Nelli
  13. Kimberly Fornace

  • Curated by eLife

    eLife logo

    eLife assessment

    This useful study presents findings regarding the impact of forest cover and fragmentation on the prevalence of malaria in non-human primates. The evidence supporting the claims of the authors is solid.

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Zoonotic disease dynamics in wildlife hosts are rarely quantified at macroecological scales due to the lack of systematic surveys. Non-human primates (NHPs) host Plasmodium knowlesi, a zoonotic malaria of public health concern and the main barrier to malaria elimination in Southeast Asia. Understanding of regional P. knowlesi infection dynamics in wildlife is limited. Here, we systematically assemble reports of NHP P. knowlesi and investigate geographic determinants of prevalence in reservoir species. Meta-analysis of 6322 NHPs from 148 sites reveals that prevalence is heterogeneous across Southeast Asia, with low overall prevalence and high estimates for Malaysian Borneo. We find that regions exhibiting higher prevalence in NHPs overlap with human infection hotspots. In wildlife and humans, parasite transmission is linked to land conversion and fragmentation. By assembling remote sensing data and fitting statistical models to prevalence at multiple spatial scales, we identify novel relationships between P. knowlesi in NHPs and forest fragmentation. This suggests that higher prevalence may be contingent on habitat complexity, which would begin to explain observed geographic variation in parasite burden. These findings address critical gaps in understanding regional P. knowlesi epidemiology and indicate that prevalence in simian reservoirs may be a key spatial driver of human spillover risk.

Article activity feed

  1. Version published to 10.7554/elife.88616.3 on eLife
  2. Version published to 10.7554/elife.88616 on eLife
  3. eLife

    Author Response

    The following is the authors’ response to the previous reviews.

    eLife assessment

    This study presents useful findings regarding the impact of forest cover and fragmentation on the prevalence of malaria in non-human primates. The evidence supporting the claims of the authors is, however, incomplete, as the sampling design cannot adequately address the geospatial issues that this study focuses on.

    Public Reviews:

    Reviewer #1 (Public Review):

    The study as a concept is well designed, although there is still one issue I see in the methodology.

    I still have concerns with their attempts to combine the different scales of data. While the use of point data is great, it limits the sample size, and they have included the district to country level data to try and increase the sample size. The problem is that although they try to get an overall estimate at the district/state/country by taking 10 random sample points, which could be a method to get an estimate for the district/state/country. It would be a suitable method if the primates were evenly distributed across the district/state/country. The reality is that the primates are not evenly distributed across the district/state/country therefore the random point sampling is not a reasonable method to get an estimate of the environmental variables in relation to the macaques. For example if you had a mountainous country and you took 10 random points to estimate altitude, you would end up with a large number, but if all the animals of interest lived on the coast, your average altitude is meaningless in relation to the animals of interest as they are all living at low altitude. The fact that the model relies less on highly variable components and places more reliance on less variable components, is really not relevant as the district/state/country measurements have no real meaning in relation to the distribution of masques.

    A simple possible way forward could be to run the model without the district/state/country samples and see what the outcome is. If the outcome is similar then the random point method may be viable (but if it gives the same outcome as ignoring those samples then you don't need the district/state/country samples). If you get a totally different outcome then it should raise concerns about using the district/state/country samples.

    This paper is a really nice piece of work and is a valuable contribution but the district/state/country sample issue really needs to be addressed.

    Recommendations for the authors:

    Reviewer #1 (Recommendations For The Authors):

    A simple possible way forward could be to run the model without the district/state/country samples and see what the outcome is. If the outcome is similar then the random point method may be viable (but if it gives the same outcome as ignoring those samples then you don't need the district/state/country samples). If you get a totally different outcome then it should raise concerns about using the district/state/country samples.

    Thank you for your comments, and for the suggestions to address the issues identified in your main commentary by running an analysis on exclusively GPS geolocated data points. This was the original plan for analysis, but the available data identified in the literature review includes only 14 data points (macaque P. knowlesi prevalence surveys) with associated GPS coordinates. This was found to be too limited to obtain meaningful results from a regression analysis, and hence we then explored methods for utilising all available data to identify trends whilst accounting for spatial uncertainty in the analysis. As the point location only represents the location of capture and not the extent of the home range of the NHPs, we additionally feel there is value in exploring methods to encompass the wider surrounding habitat.

    We do appreciate the concerns you raise with the random point method being used to represent macaque survey sites when species of interest are not necessarily evenly distributed across an area. To investigate this, we ran sensitivity analysis on a subset of the dataset according to whether the points fall in areas of >50%, >75% or >90% predicted probability of macaque occurrence, with maps derived from published models of macaque suitability in Southeast Asia. For each of these thresholds, points that fall outside these areas were removed – such that, if a random point is located on a mountain range where there is 0 likelihood of macaque occurrence, it is excluded from the analysis. We found that restricting analysis to areas with highly probably macaque habitat still shows a robust effect of forest cover on NHP prevalence, and additionally that for the most conservative (>90%) habitat threshold there remains an effect of forest fragmentation on prevalence (SI Table S17c, Figure S15c). Given that using the full data set increases the uncertainty, as there is more variation in covariates between the replicates, this can be considered a more conservative approach to detecting an effect of environment as reported in the main findings.

  4. eLife

    eLife assessment

    This useful study presents findings regarding the impact of forest cover and fragmentation on the prevalence of malaria in non-human primates. The evidence supporting the claims of the authors is solid.

  5. eLife

    Reviewer #1 (Public Review):

    The paper aims to determine the impact of forest cover and fragmentation on the prevalence of malaria in non-human primates. The paper uses existing spatial datasets, as well as data obtained through published studies on zoonotic malaria. The findings of this study are important, as forest loss is still occurring in the tropics which will impact human infections of zoonotic malaria.

  6. eLife

    Reviewer #2 (Public Review):

    This is the first comprehensive study aimed at assessing the impact of landscape modification on the prevalence of P. knowlesi malaria in non-human primates in Southeast Asia. This is a very important and timely topic both in terms of developing a better understanding of zoonotic disease spillover and the impact of human modification of landscape on disease prevalence.

    This study uses the meta-analysis approach to incorporate the existing data sources into a new and completely independent study that answers novel research questions linked to geospatial data analysis. The challenge, however, is that neither the sampling design of previous studies nor their geospatial accuracy are intended for spatially-explicit assessments of landscape impact. On the one hand, the data collection scheme in existing studies was intentionally opportunistic and does not represent a full range of landscape conditions that would allow for inferring the linkages between landscape parameters and P. knowlesi prevalence in NHP across the region as a whole. On the other hand, the absolute majority of existing studies did not have locational precision in reporting results and thus sweeping assumptions about the landscape representation had to be made for the modeling experiment. Finally, the landscape characterization was oversimplified in this study, making it difficult to extract meaningful relationships between the NHP/human intersection on the landscape and the consequences for P. knowlesi malaria transmission and prevalence.

    Despite study limitations, the authors point to the critical importance of understanding vector dynamics in fragmented forested landscapes as the likely primary driver in enhanced malaria transmission. This is an important conclusion particularly when taken together with the emerging evidence of substantially different mosquito biting behaviors than previously reported across various geographic regions.

    Another important component of this study is its recognition and focus on the value of geospatial analysis and the availability of geospatial data for understanding complex human/environment interactions to enable monitoring and forecasting potential for zoonotic disease spillover into human populations. More multi-disciplinary focus on disease modeling is of crucial importance for current and future goals of eliminating existing and preventing novel disease outbreaks.

  7. Version published to 10.1101/2022.06.17.496284v4 on bioRxiv
  8. Version published to 10.7554/elife.88616.2 on eLife
  9. eLife

    Author Response

    The following is the authors’ response to the original reviews.

    Reviewer #1 (Public Review):

    The study as a concept is well designed, although there are two issues I see in the methodology (these may be just needing further explanation or if I am correct in my interpretation of what was done, may need reanalysis to take into account). Both issues relate to the data that was extracted from the published literature on zoonotic malaria prevalence in the study area.

    1. No limit was set on the temporal range

    With no temporal limit on the range of studies, the landscape in many cases will have changes between the study being conducted and the spatial data. This will be particularly marked in areas where there has been clearing since the zoonotic malaria prevalence study. Also, population changes (either through population growth, decline or movement) will have occurred. All research is limited in what it can do with the available data, so I realise that there may not be much the authors can do to correct this. One possible solution would be to look at the land use change at each site between the prevalence study and the remote sensing data. I'm not sure if this is feasible, but if it is I would recommend the authors attempt this as it will make their results stronger.

    Thank you for the comments. We agree that matching the date of remote sensing data to samples is particularly important for environmental variables that change rapidly (such as forest loss). To clarify, no limit was set on the date range of the studies identified from the literature to ensure no articles were excluded due to arbitrary date restrictions. We have edited the manuscript to clarify this (line 422). Regarding landscape and environmental features, remote sensing data was extracted annually for every year for the full date range of the data (see Table 1 and S11, annual temporal resolution from 2006 to 2020). Forest was then matched contemporaneously (see lines 467–473) meaning that, insofar as it was possible, forest data was extracted for the same year as the data was collected. Where a date range was given for the primate data, the mean year was used. For human population density, covariate data were extracted for multiple years but were found to be relatively stable over the time period for the sites covered, so median year was used (see Supplementary Information, Appendix E and Table S11). Elevation is stable and typically only one time point is used as reference (in this instance the SRTM 90m Digital Elevation model, 2003).

    1. Most studies only gave a geographic area or descriptive location.

    The spatial analysis was based on a 5km and 20km radius of the 'study site' location, but for many of the studies the exact site is not known. Therefore the 'study site' was artificially generated using a polygon centroid. Considering that the polygon could be an administrative boundary (i.e., district/state/country), this is an extremely large area for which a 5km radius circle in the middle of the polygon is being taken as representative of the 'study site'. This doesn't make sense as it assumes that the landscape is uniform across the district, which in most cases it will not be (in rural areas it is going to be a mixture of villages, forest, plantation, crops etc which will vary across the landscape). This might just be a case of misunderstanding what was done (in which case the text needs rewording to make it clearer) or if I have interpreted it correctly the selection of the centroid to represent the study area does not make sense. I am not sure how to overcome this as it probably not possible to get exact locations for the study sites. One possibility could be to make the remote sensing data the same scale as the prevalence data ie if the study site is only identifiable at the polygon level, then the remote sensing data (fragmentation, cover and population) is used at the polygon level.

    Both these issues could have an impact on the study's findings. I would think that in both cases it might make the relationship between the environmental variables and prevalence even clearer.

    We would like to thank the reviewer for their concerns and provide some clarification on the methods used to extract environmental variables:

    • Centroid was initially explored, but not pursued for the same concerns raised by the reviewer. Taking the centroid would be arbitrary and the central point of a large polygon is not likely to be representative of habitat across the entire sampling area and introduces error so this was not pursued(Cheng et al., 2021). We have clarified the wording in the manuscript with reference to centroids to avoid confusion on this point (line 491).

    • We demonstrate a method to account for the lack of precise geolocation by taking 10 ‘pseudo-sampling’ points instead of a single random location, with environmental variables extracted at 5, 10 and 20km for each site (lines 487-500). By including 10 environmental realisations, surveys conducted in smaller or more uniform landscapes will have more consistent covariates and this will lend more weight to the model. Conversely, samples taken from large administrative polygons are likely to be highly variable, and these associations will have less representation in the final model. This approach was used to demonstrate an alternative to using a single arbitrary site to represent the area.

    To further support the validity of this technique:

    • Figures illustrating the variance of the environmental variables across the 10 sampling sites at 5, 10 and 15km for GADM administrative classifications at country level (GID0), state (GID1), district (GID2) and exact coordinates (GPS) are now included in the SI (Figure S12).

    • Sensitivity analyses were conducted, in which final GLMM models were fit again but using only acceptable levels of variance in environmental variables and/or acceptable size of administrative boundary (Table S15 and S16). In sensitivity analyses, forest cover and fragmentation retained a significant effect on prevalence of P. knowlesi in macaques, suggesting this effect is robust to spatial uncertainty.

    We would also like to highlight that the main finding of this research is the novel synthesis of regional prevalence of P. knowlesi in simian reservoirs across Southeast Asia, which was formerly assumed to be ubiquitous high prevalence, and which can now be used to inform regionally specific transmission modelling, better estimate spatial risk and parameterise early warning systems for P. knowlesi malaria in countries approaching elimination of human malarias. The risk factor analysis here is provided to begin to understand what may be driving this geographic heterogeneity in P. knowlesi prevalence at finer scales and demonstrate methods that could be used to accommodate spatial uncertainty in secondary data. We appreciate that this may not have been clear and have edited the manuscript accordingly.

    Reviewer #2 (Public Review):

    This is the first comprehensive study aimed at assessing the impact of landscape modification on the prevalence of P. knowlesi malaria in non-human primates in Southeast Asia. This is a very important and timely topic both in terms of developing a better understanding of zoonotic disease spillover and the impact of human modification of landscape on disease prevalence.

    This study uses the meta-analysis approach to incorporate the existing data sources into a new and completely independent study that answers novel research questions linked to geospatial data analysis. The challenge, however, is that neither the sampling design of previous studies nor their geospatial accuracy are intended for spatially-explicit assessments of landscape impact. On the one hand, the data collection scheme in existing studies was intentionally opportunistic and does not represent a full range of landscape conditions that would allow for inferring the linkages between landscape parameters and P. knowlesi prevalence in NHP across the region as a whole. On the other hand, the absolute majority of existing studies did not have locational precision in reporting results and thus sweeping assumptions about the landscape representation had to be made for the modeling experiment. Finally, the landscape characterization was oversimplified in this study, making it difficult to extract meaningful relationships between the NHP/human intersection on the landscape and the consequences for P. knowlesi malaria transmission and prevalence.

    Thank you for the feedback on the manuscript. We agree that the data was not originally intended for spatial assessment of landscape impact nor represents a full range of landscape conditions across the region. However, we would like to highlight the first set of results from the meta-analysis. Here, the synthesis of all available data allows for the detection of regional disparities and geographic heterogeneity of prevalence in host species, which individual small-scale opportunistic studies are not powered to do, and which had not been identified before this investigation.

    In this context, the risk factor analysis is an exploratory analysis to understand what may be driving the observed geographic variation at broad scales as well as provide a framework for dealing with spatial uncertainty. Landscape data was extracted at a level deemed appropriate given the limitations of the data. The majority were geolocated to district level and sensitivity analysis showed a reasonable consistency of landscape features at our chosen scales (Table S8, Figure S12A). To address some of these concerns, we conducted further analysis to explore the deviation of environmental covariates in each sampling area and ran sensitivity analysis by removing extremely variable datapoints (Table S15 and Table S16). When removing highly uncertain data and/or countrylevel data, effects of canopy cover on non-human primate malaria prevalence is retained, supporting the original findings.

    Despite many study limitations, the authors point to the critical importance of understanding vector dynamics in fragmented forested landscapes as the likely primary driver in enhanced malaria transmission. This is an important conclusion particularly when taken together with the emerging evidence of substantially different mosquito biting behaviors than previously reported across various geographic regions.

    Another important component of this study is its recognition and focus on the value of geospatial analysis and the availability of geospatial data for understanding complex human/environment interactions to enable monitoring and forecasting potential for zoonotic disease spillover into human populations. More multi-disciplinary focus on disease modeling is of crucial importance for current and future goals of eliminating existing and preventing novel disease outbreaks.

    Reviewer #1 (Recommendations For The Authors):

    A couple of minor points

    1. Was the human density and forest cover correlated? If so was this taken into account

    Human density and forest cover at selected scales were not found to be strongly correlated (Spearman’s rank values -0.38 and -0.45 within 5km and 20km buffer radii for human population density respectively).

    In selecting variables for inclusion in the final model, we examined variance inflation factors (VIF) to detect and minimise multicollinearity in the model. VIF measures the correlation and strength of correlation between independent predictors. VIF of each predictor variable was examined starting with a saturated model and sequentially excluding the variable with the highest VIF score from the model. Stepwise selection continued until the entire subset of explanatory variables in the global model satisfied a conservative threshold of VIF ≤6 (Rogerson, 2001), which ensures that the remaining variables included in the final model have minimal correlation. Spearman’s correlation matrices for all variables at all scales and final selected variables (below VIF threshold) are included in the Supplementary Information (Figure S13 and Figure S14).

    1. Reference (Speldewinde et al., 2019) is down as Davidson et al. in the reference list

    Thank you for the thoroughness in this review. There are two similar but separate references, both published in 2019 with the same co-authors, and the (Speldewinde et al, 2019) was incorrectly referenced. They should be (Davidson et al., 2019a) and Davidson et al., 2019b) respectively. This has now been corrected in the manuscript.

    Davidson, G., Chua, T.H., Cook, A. et al. Defining the ecological and evolutionary drivers of Plasmodium knowlesi transmission within a multi-scale framework. Malar J 18, 66 (2019). https://doi.org/10.1186/s12936-019-2693-2

    Davidson G, Chua TH, Cook A, Speldewinde P, Weinstein P. The Role of Ecological Linkage Mechanisms in Plasmodium knowlesi Transmission and Spread. Ecohealth. 2019;16(4):594-610. https://doi:10.1007/s10393-019-01395-6

    Reviewer #2 (Recommendations For The Authors):

    Line 143: "We hypothesise that higher prevalence of P. knowlesi in primate host species is driven by landscape change..." without specifying here the kind of landscape change (e.g. "forest degradation and fragmentation") it is virtually impossible to confirm or reject this hypothesis.

    We agree that the wording of the hypotheses needed to be more specific. We have edited lines 142 – 145 to specify forest fragmentation as our landscape variable of interest, and to more explicitly include the regional meta-analysis of P. knowlesi prevalence.

    Table 1 vs Table S11 discrepancy regarding spatial resolution of Forest cover and fragmentation variables. The original dataset resolution is 30m but I don't think one can compute a PARA index at 30 m since it really requires a polygon that is larger than the single value pixel. Table S11 indicates a 30 km gridcell with some postprocessing of the original datasets.

    We appreciate this being identified. The resolution refers to the input layer (tree canopy cover, 30m). PARA was calculated from the binary forest cover layer (30m resolution) within each buffer radii 5, 10 and 20km. We have edited both Table 1 and Table S11 to help clarify this.

    It would be very helpful if you provided justification for selecting specific metrics to represent the key landscape variables. How are these particular landscape variables relevant? Why not other land cover/land use components?

    We have now included a paragraph in the Supplementary Information (Appendix D) to explain the choice of environmental covariates. Elevation was chosen as an important proxy for vector distribution (but was not retained in model selection). Human population density was chosen as a measure of proximity to human settlement, rather than relying on qualitative assessment of rural/peri-urban/urban. Tree canopy cover and fragmentation indices are key determinants of primate habitat selection and of vector breeding habitat, and justification for the use of perimeter: area ratio is included in the methods section (section beginning line 462).

    I think the other issues present substantial weaknesses that you cannot address without redoing the study. I will list those below just for reference.

    1. If the forest is so dominant (which I would agree with based on my understanding of macaque ecology), how does it make sense to select completely random points (especially at the country or even state level) to represent landscape covariates? At a minimum, I would suggest getting random points within the forest or better yet forest edge habitat. But even then, I doubt that these points would be at all representative of the conditions of a specific study. The geospatial uncertainty is just too large. The dataset simply doesn't support the analysis that is attempted here.

    On the point of selecting from only within forest: forest is a dominant habitat, but Long-tailed macaques are anthropophilic and not exclusively found in forest (Stark et al., 2019), and a proportion of the more opportunistic and nuisance samples caught were found in areas more associated with human activity (Li et al., 2021). As such, random points only within forested areas is also unlikely to capture the true habitat of the primates sampled and selecting only from forested areas would bias the results.

    Whilst fully georeferenced samples would be the ideal scenario, the idea behind selecting random points from the sampling polygon is that for smaller areas (with higher spatial certainty), habitat would be more consistent between random points and lend more weight to the final model, whereas large polygons with high uncertainty are likely to vary and lend less weight to the final model. In response to these comments, we have further supported this by running regression models only on samples within a reasonable administrative boundary size and on samples within reasonable threshold of uncertainty (i.e., data points are removed if the deviation of environmental covariates across the 10 random points is so high that the sample is uninformative, or if datapoints can only be geolocated to country-level). In these sensitivity analyses, forest cover and species are retained as factors associated with higher malarial prevalence in non-human primates (Table S15S16).

    1. Hansen et al. dataset reflects "tree cover" - which is not the same as "forest cover" since it would also include plantations that are very widely distributed across Southeast Asia. If the animal use of plantations differs from that of natural forests, it will present a large issue for the study.

    In this analysis the feature of interest was habitat configuration (fragmentation) and deforestation (forest loss) rather than specific land class. We have defined forest as >50% canopy cover, which considers canopy density given historical forest loss and has precedence in other work (Fornace et al.,, 2016). In addition to importance to macaque ecology, forest (canopy) cover, forest loss and forest edge are noted to be key determinants of vector breeding and vector habitat (Byrne et al., 2021, Chua et al., 2019). For this reason, these are important variables to include in analyses. More specific landscape variables were explored, but the temporal and spatial range of the data precluded fine-scale land classification data. To investigate preliminary links to landscape configuration and habitat fragmentation at broad scales this is felt to be sufficient. We have also amended the manuscript to be more discerning with the use of ‘forest’ to avoid confusion throughout.

    1. Tree regrowth in the ecosystems of monsoonal Asia is very rapid. Based on the study description, tree regrowth was not accounted for in the study which could potentially lead to a very large underestimation of tree cover if only tree loss since 2000 was monitored. Again unless there is a reason to assume that macaques do not use young successional forests or use it at a highly reduced rate. Both of these points are acknowledged as limitations at the end of the discussion section but in my opinion they have a very strong impact on the study, making the results non-significant.

    This is an interesting suggestion. Macaques do forage in plantations and cultivated landscapes to supplement food, but preferentially roost and range in forest edges and interior forest, though ranging behaviour will be complex and vary across Southeast Asia. In this study the primary interest was in deforestation (forest loss) and fragmentation of old growth forested landscapes, which are key variables both for macaque ecology and for vector breeding sites. Therefore, it was felt that forest loss (transition from >50% canopy cover to <50% canopy cover since 2000) was sufficient to capture this. Ranging behaviour of individual animals and macaque troops would not be captured at this scale, and higher spatial and temporal resolution would be required to characterise relationships with tree regrowth and young plantations which is outside the scope of this study. In all regions, purposeful fine scale follow-up studies would be required to unpick fine scale relationships across a habitat gradient.

    I am not 100% sure I understand the geospatial design fully. The pieces are distributed between different subsections and it was challenging to string together the processing chain between subsections of the manuscript and the supplemental information. I would help to add a figure (a flowchart, perhaps?) to the supplemental section that walks through the entire geospatial covariates assembly. E.g.

    • GPS location create 5, 10, and 20 km buffers mean elevation, mean population, %(?) Forest, PARA(?) for each buffer - I still don't understand the 30m or 30 km spatial resolution reference for forest and PARA in this context.

    This was an error in the table in the Supplementary Information and has been corrected – the forest cover raster has a resolution of 30m, and the perimeter: area ratio is calculated within 5, 10 and 20km buffers.

    • landscape covariates receive the full weight (1) in the model. - This is defensible even though not ideal

    This is equivalent, but we felt more intuitive, to sampling GPS points x10 and inputting with equal weights to the areal data.

    • No GPS location assign to the best identifiable administrative unit (country, state, or district) generate 10 random points within the administrative unit create 5, 10, and 20 km buffers mean elevation, mean population, %(?) Forest, PARA(?) for each buffer landscape covariates from each point receive the proportional weight (0.1) in the model. I do not believe that this approach is representative of macaque habitat/macaque human interaction characterization.

    In other examples dealing with spatial uncertainty, the centroid is taken to be representative of an area. This method generates considerable bias and uncertainty – particularly if the uncertainty is not then accounted for by weighting subsequent models (Cheng, 2021). In this exploratory analysis, pseudo-sampling from 10 random sites generates a more realistic generalised environmental realisation than taking a centroid/random point. This was used as an exploratory analysis to explain broad regional trends in prevalence between, which can be used to guide further investigation on fine scale studies which are required to completely describe disease dynamics in specific macaque habitats.

    Thank you for this useful suggestion – we have taken this advise and added a flowchart of data processing to the Supplementary Information (Appendix D, Figure S8).

    Discussion:

    Based on information in Table S4, sampled NHPs were predominantly from human-dominated (peridomestic, agricultural, and urban) landscapes. In forested landscapes, only macaques that live in forest edge habitats were likely sampled in the first place just simply due to extreme challenges in getting to macaques in remote inaccessible areas. There is a very substantial spatial bias in sampling will undoubtedly reflect that fragmented habitat is a key landscape component impacting the prevalence of Pk in NHP, especially as the authors point out in the later part of the discussion, the critical vectors for transmission are also associated with forest edge habitats. High forest fragmentation is also linked to the presence/ increase in migrant human workers (logging or plantation activities) - a population also strongly associated with higher malaria prevalence for a variety of P spp (although I am not aware of studies that are specific to Pk malaria). However, the living conditions for migrant workers have frequently been implicated in higher rates of malaria transmission and thus those could, hypothetically, also contribute to Pk infection rates in NHP. Ultimately, the discussion appears to suggest that the biggest gap in our understanding is within vector ecology and understanding the NHP-vector-human dynamics within local landscape settings. It is an interesting finding. However, my overall conclusion would be that the sampling strategy (both for NHP and geospatial covariates) renders this study as "exploratory" at maximum and that all findings would need to be tested and verified through independent and more rigorously designed studies.

    Thank you to the reviewer for a comprehensive assessment. We would first like to highlight the regional meta-analysis, which was one of the main findings. This is a novel result for P. knowlesi literature; being the first demonstration of regional differences in prevalence that correlate to regional hotspots of human incidence, the force of infection from NHP may drive hotspots of P. knowlesi in human populations.

    We include a risk factor analysis that suggests a method for dealing with high spatial uncertainty, and an exploratory analysis that finds landscape complexity may be a contributory factor to broad regional heterogeneity. These associations are robust to sensitivity analysis where data with extreme variability in environmental variables is removed (Table S15-S16).

    Habitat descriptions in original studies are qualitative, likely subjective, and whilst there is likely to be an important sampling bias there was also evident differences in prevalence between the NHP sampled in different environments from the available data that we have further characterised. Risk factors for human P. knowlesi do include forest loss (reduction in canopy cover) within 5 years and within 2km, as well as contact with macaques and occupations in plantations (Fornace et al., 2014; Fornace et al., 2016). Reverse spillover from humans to NHP is an interesting suggestion, but outside the scope and scale of the study. Given known links of deforestation (forest loss) with human incidence of P. knowlesi and also with increased vector breeding sites (Byrne et al., 2021), this analysis explores whether deforestation is linked to prevalence in reservoir species thus contributing to the force of infection at broad scales.

  10. eLife

    eLife assessment

    This study presents useful findings regarding the impact of forest cover and fragmentation on the prevalence of malaria in non-human primates. The evidence supporting the claims of the authors is, however, incomplete, as the sampling design cannot adequately address the geospatial issues that this study focuses on.

  11. eLife

    Joint Public Review:

    The study as a concept is well designed, although there is still one issue I see in the methodology.

    I still have concerns with their attempts to combine the different scales of data. While the use of point data is great, it limits the sample size, and they have included the district to country level data to try and increase the sample size. The problem is that although they try to get an overall estimate at the district/state/country by taking 10 random sample points, which could be a method to get an estimate for the district/state/country. It would be a suitable method if the primates were evenly distributed across the district/state/country. The reality is that the primates are not evenly distributed across the district/state/country therefore the random point sampling is not a reasonable method to get an estimate of the environmental variables in relation to the macaques. For example if you had a mountainous country and you took 10 random points to estimate altitude, you would end up with a large number, but if all the animals of interest lived on the coast, your average altitude is meaningless in relation to the animals of interest as they are all living at low altitude. The fact that the model relies less on highly variable components and places more reliance on less variable components, is really not relevant as the district/state/country measurements have no real meaning in relation to the distribution of masques.

    A simple possible way forward could be to run the model without the district/state/country samples and see what the outcome is. If the outcome is similar then the random point method may be viable (but if it gives the same outcome as ignoring those samples then you don't need the district/state/country samples). If you get a totally different outcome then it should raise concerns about using the district/state/country samples.

    This paper is a really nice piece of work and is a valuable contribution but the district/state/country sample issue really needs to be addressed.

  12. Version published to 10.1101/2022.06.17.496284v3 on bioRxiv
  13. Version published to 10.7554/elife.88616.1 on eLife
  14. eLife

    eLife assessment

    This study presents useful findings regarding the impact of forest cover and fragmentation on the prevalence of malaria in non-human primates. The evidence supporting the claims of the authors is, however, incomplete, as the sampling design cannot adequately address the geospatial issues that this study focuses on.

  15. eLife

    Reviewer #1 (Public Review):

    The study as a concept is well designed, although there are two issues I see in the methodology (these may be just needing further explanation or if I am correct in my interpretation of what was done, may need reanalysis to take into account). Both issues relate to the data that was extracted from the published literature on zoonotic malaria prevalence in the study area.

    1. No limit was set on the temporal range
    With no temporal limit on the range of studies, the landscape in many cases will have changes between the study being conducted and the spatial data. This will be particularly marked in areas where there has been clearing since the zoonotic malaria prevalence study. Also, population changes (either through population growth, decline or movement) will have occurred. All research is limited in what it can do with the available data, so I realise that there may not be much the authors can do to correct this. One possible solution would be to look at the land use change at each site between the prevalence study and the remote sensing data. I'm not sure if this is feasible, but if it is I would recommend the authors attempt this as it will make their results stronger.

    2. Most studies only gave a geographic area or descriptive location.
    The spatial analysis was based on a 5km and 20km radius of the 'study site' location, but for many of the studies the exact site is not known. Therefore the 'study site' was artificially generated using a polygon centroid. Considering that the polygon could be an administrative boundary (ie district/state/country), this is an extremely large area for which a 5km radius circle in the middle of the polygon is being taken as representative of the 'study site'. This doesn't make sense as it assumes that the landscape is uniform across the district, which in most cases it will not be (in rural areas it is going to be a mixture of villages, forest, plantation, crops etc which will vary across the landscape). This might just be a case of misunderstanding what was done (in which case the text needs rewording to make it clearer) or if I have interpreted it correctly the selection of the centroid to represent the study area does not make sense. I am not sure how to overcome this as it probably not possible to get exact locations for the study sites. One possibility could be to make the remote sensing data the same scale as the prevalence data ie if the study site is only identifiable at the polygon level, then the remote sensing data (fragmentation, cover and population) is used at the polygon level.

    Both these issues could have an impact on the study's findings. I would think that in both cases it might make the relationship between the environmental variables and prevalence even clearer.

  16. eLife

    Reviewer #2 (Public Review):

    This is the first comprehensive study aimed at assessing the impact of landscape modification on the prevalence of P. knowlesi malaria in non-human primates in Southeast Asia. This is a very important and timely topic both in terms of developing a better understanding of zoonotic disease spillover and the impact of human modification of landscape on disease prevalence.

    This study uses the meta-analysis approach to incorporate the existing data sources into a new and completely independent study that answers novel research questions linked to geospatial data analysis. The challenge, however, is that neither the sampling design of previous studies nor their geospatial accuracy are intended for spatially-explicit assessments of landscape impact. On the one hand, the data collection scheme in existing studies was intentionally opportunistic and does not represent a full range of landscape conditions that would allow for inferring the linkages between landscape parameters and P. knowlesi prevalence in NHP across the region as a whole. On the other hand, the absolute majority of existing studies did not have locational precision in reporting results and thus sweeping assumptions about the landscape representation had to be made for the modeling experiment. Finally, the landscape characterization was oversimplified in this study, making it difficult to extract meaningful relationships between the NHP/human intersection on the landscape and the consequences for P. knowlesi malaria transmission and prevalence.

    Despite many study limitations, the authors point to the critical importance of understanding vector dynamics in fragmented forested landscapes as the likely primary driver in enhanced malaria transmission. This is an important conclusion particularly when taken together with the emerging evidence of substantially different mosquito biting behaviors than previously reported across various geographic regions.

    Another important component of this study is its recognition and focus on the value of geospatial analysis and the availability of geospatial data for understanding complex human/environment interactions to enable monitoring and forecasting potential for zoonotic disease spillover into human populations. More multi-disciplinary focus on disease modeling is of crucial importance for current and future goals of eliminating existing and preventing novel disease outbreaks.

  17. Version published to 10.1101/2022.06.17.496284v2 on bioRxiv
  18. Version published to 10.1101/2022.06.17.496284v1 on bioRxiv