The development and spread of Plasmodium falciparum resistance to the most commonly used anti-malarial drugs is a major challenge in the control of malaria 1 . Quantifying the level of drug resistance through regular monitoring provides essential information to the health authorities responsible for ensuring ready access to effective drugs. Methods for assessing drug resistance in malaria include in vivo drug efficacy studies (the gold standard), in vitro sensitivity testing of patient isolates, and surveys designed to detect molecular markers of drug resistance. Each approach has advantages and disadvantages (discussed in 2 3 4 ) and herein field data evaluating the usefulness of the molecular marker approach are presented.

Molecular screening to detect drug resistance was made possible by studies identifying the genetic basis of drug resistance in P. falciparum and this approach has subsequently become an integral part for the evaluation of resistance to treatment 5 6 . Chloroquine resistance (CQR) is attributable to single nucleotide polymorphisms (SNPs) in pfcrt and pfmdr1 (reviewed in 7 8 9 ). Resistance to sulphadoxine-pyrimethamine (SP) is associated with a stepwise accumulation of mutations in pfdhfr and pfdhps (reviewed in 10 ). Artemisinin derivates have been shown to inhibit pfATPase6 11 and SNPs in pfATPase6 may become associated with resistance in vitro 12 13 . It has also become apparent that these mutations do not act in isolation, but often act synergistically to encode or enhance resistance. Resistance may arise through sequential accumulation of mutations in a single gene, such as anti-folate resistance in the pfdhfr gene 4 14 . Mutations in different genes may also act synergistically. Linkage disequilibrium is often observed between pfcrtK76T and pfmdr1N86Y in vivo 15 16 17 and the commonly accepted hypothesis is that CQR is mediated by multigenic processes 18 , primary resistance being encoded by pfcrtK76T with SNPs in pfmdr1 playing an important role in modulating levels of CQR. The requirement to consider intra- and inter-genic genotypes impose limitations on the use of molecular methods such as PCR-RFLP which can identify the presence of mutations at individual markers, but not individual malaria genotypes defined at multiple loci in large sample sets.

The link between molecular markers and resistance can be clearly demonstrated in the laboratory, but the usefulness of these markers in the field has been controversial, primarily because the strength of their associations with in vivo treatment outcome is not consistent in different epidemiological settings. There are two plausible reasons for this 3 . Firstly, drug failure may occur for reasons other than parasite genetics. Patients may take incomplete drug courses, may have poor drug absorption or metabolism and so on. When marker frequencies are low, these human sources of drug failure may obscure the relationship between marker and resistance. Secondly, most studies have primarily focused on single genes and markers 19 , rather than multiple markers which may act synergistically. A further limitation is that most analyses of molecular markers have been conducted as part of clinical studies and have, therefore, been limited to sentinel sites with good access to health care facilities and mostly examined clinical cases within a restricted age group 20 . Moreover, most such studies used parasite genetic measures based on the proportions of patients carrying a given genetic marker, or composite measures formed from ad hoc functions of such proportions 5 . This ignores the fact that patients often carry multiple clones of parasites and so may carry a mutation even if most of their parasites are wild-type. In principle, the evolution of drug resistance should be assessed by genetic profiling of representative samples of the circulating parasite population, which includes the reservoir in asymptomatic carriers. This should be summarized using population-genetic measures of allele frequencies (for single alleles) and haplotype frequencies (for multiple linked alleles). However, genetic data from community samples have rarely been compared with in vivo treatment responses 21 22 . In addition, it is not clear whether the molecular profile of parasites in symptomatic patients seen at health facilities matches that of the large circulating parasite reservoir.

The current study has investigated the role and applicability of the molecular drug resistance profiles in community samples for the monitoring of drug resistant malaria in Papua New Guinea (PNG). For this purpose, in vivo drug efficacy studies with the first-line regimen of amodiaquine or chloroquine plus sulphadoxine-pyrimethamine (AQ+SP or CQ+SP) were conducted at three different sites. A DNA microarray-based technology was used to compare the molecular drug resistance profile in clinical samples and blood samples collected in the communities from the catchment areas of the corresponding health facilities. Finally, the relationships of the allele and haplotype frequencies of the molecular markers with the rates and time-trends in clinical failure at the different sites were investigated.

The present study was specifically designed to analyse the relationship between parasite genetics in the community and in vivo failure rates, but the complexity of this relationship was not initially appreciated. There is substantial evidence linking specific genetic markers to drug resistance in vitro and the presence of genetic resistance mutations often predicts treatment failure at the level of the individual patient, but it was observed that these individual level relationships are not reflected in straightforward relationships at the population level between molecular typing results and in vivo tests of resistance. A number of possible reasons for this discrepancy were investigated.

One limitation of many studies is that they have included only small numbers of SNPs. The chosen SNPs might prove to be the wrong mutations in the given population or secondary mutations with indirect relationships to resistance might be important. This was addressed by using a DNA microarray-based technology that allowed parallel assessment of many SNPs in several genes 33 . The analysis of a total of 12 polymorphic SNPs, as well as a further 21 SNPs that proved to be uniformly wild-type, enabled to consider all candidate SNPs so far identified. The study populations had highly mutated genes involved in CQR resistance (pfcrt and mdr1) combined with pfdhfr mutations consistent with a moderately pyrimethamine resistant phenotype and the emergence of key mutations in pfdhps. The key CQR markers pfcrt K76T, N326D and N326D were almost fixed at all three sites, in keeping with other recent findings 38 39 40 . Similarly, the high allele frequencies for pfdhfr S108N and C59R and the resulting high frequency of the double-mutant haplotype are consistent with a recent report from East Sepik Province 41 . Despite this substantial amount of information about each sample, the allele and single-locus haplotype frequencies could still not easily be related to in vivo failure rates.

The choice of strategy for sampling the parasite population does not seem to be an important determinant of whether a relationship is seen with treatment failure. The differences between corresponding genetic profiles in health facility and community samples were small in comparison with the differences between sites. In Karimui, the profiles were near identical, and the small differences between the two profiles in the Wosera, which may reflect heterogeneities in access and hence drug pressure for different parts of the catchment area, are too small to influence the choice of sampling strategy. The decision as to whether to sample the community or health centres should therefore be made on logistic grounds. Community sampling can be carried out during a single short visit to a site, while health facility sampling is likely to take longer, but may be more convenient especially if blood sampling for drug resistance can be combined with other procedure(s).

A more important factor in being able to predict treatment failure from parasite genetic data may be the appropriate analysis of the data. Only by analysing two-locus pfhfr/pfdhps haplotypes coherent relationships between the genetic data and TFRs could be seen, corroborating the previous finding that the pfdhfrC59R+S108N/pfdhpsA437G triple mutant predicts treatment failure best on individual level in PNG 34 . While this haplotype was absent in the North Coast and so cannot account for the treatment failures observed there, increases in haplotype frequency in the other sample sets matched remarkably well the increased TFRs. This is in agreement with studies from Uganda 42 and Laos 43 , that also report pfdhps mutations to be important markers for unsuccessful treatment response to combination therapy with CQ plus SP, although the pfdhps genotype was not indicative for treatment failure with SP monotherapy in Tanzania 44 45 or Ghana 46 .

It has been proposed 47 48 49 that genetic resistance can be measured by the ratio between the prevalence of mutated genotypes to the TFR (the genetic failure index, GFI). A low, reasonably stable GFI would indicate a good marker of resistance. The theoretical properties of GFI are discussed in detail elsewhere 37 where it was pointed out that the GFI for any given SNP could vary widely depending on genotypes at other SNPs modulating the parasites ability to survive therapy. This study quantified this effect, showing that the relationship between SNPs and clinical outcome needs to be based on haplotype frequencies rather than crude prevalence of a single SNP. The estimation of further multi-locus haplotype frequencies may be needed before the full potential of such analysis becomes clear.

Currently, there are technical limitations in the ability to estimate multi-locus haplotype frequencies. Where there are no polyclonal infections, haplotype frequencies can be estimated by simple gene counting. However, 68% of the PCR-positive community samples had multiple msp2 types, implying that only 17% of the parasite clones could validly be used to estimate allele frequencies by simple gene counting, making this a very inefficient and imprecise procedure. The estimation of haplotype frequencies from multi-clone samples using the MalHaploFreq programme makes use of substantially more information, but with the current implementation it is not feasible to identify the 3- and 4-locus haplotype frequencies.

Fortunately, there are two ways of avoiding this limitation. Firstly, to date, most resistance appears to have arisen through a few single key SNPs such as pfcrtK76T, pfcrtA220S and pfmdr1N86Y, or by a series of mutations such as the pfdhfr108-51-59-164 sequence. Thus, most key multi-locus haplotypes can be characterized at a few numbers of SNPs (e.g. pfcrtK76T+pfmdr1N86Y, pfcrtK76T+dhfr I164L, and so on). Secondly, in areas of moderate to high transmission there is likely to be large numbers of mixed infections and hence large amounts of sexual recombination which tends to bring separate genes into linkage disequilibrium (LD). If the genes are in LD then the frequency of multi-gene haplotypes can be obtained simply by multiplying their individual frequencies. Alternately, if mixed infections are rare then haplotypes can be directly observed in single-clone infections and MalHaploFreq is not required.

The inter-site differences in TFR need to be understood in the context of the long-term histories of malaria transmission and drug use at each site, possibly including locality-specific factors, such as clinical practice, genetic background of human population and degree of isolation. After a long history of 4-aminoquinoline use, treatment policy against uncomplicated malaria in PNG changed to combination therapy with AQ or CQ plus SP only in 2000. Both, the TFR and allele/haplotype frequencies, including high frequencies of important markers in pfdhfr and pfdhps, are therefore the outcomes after only two years of implementation of combination therapy and are probably highly dependent on the baseline frequency of resistant mutations at the time of the policy change.

The sites differed in the frequency of pfcrtA220S corresponding to the highly CQ resistant pfcrt quadruple mutant haplotype (K76T+N326D+N326D+A220S), which was close to fixation only at the North Coast. Similarly, the pfmdr1N86Y mutant was close to fixation in Karimui and at the North Coast (>92%), but was less frequent in the Wosera (71%), and the pfmdr1 polymorphisms Y184F and N1042D were absent in Karimui. These genetic differences might simply reflect stochastic differences in source parasite populations, but they could also reflect differences in historical drug pressure. In particular, while pfcrt mutations presumably reflect a history of CQ use, SNPs in pfmdr1 conferring both, resistance to CQ and altered sensitivity to other drug classes, including quinine (which was used for treatment of severe malaria in PNG), arylaminoalcohols and artemisinin derivates 37 50 , are presumably selected by exposure to these. The fixation of a highly CQ resistant haplotype at the North Coast may reflect a history of very high CQ pressure due to good health care provision as well as easy (and unregulated) access to drugs in the nearby town of Madang. In contrast, former 4-aminoquinoline use is assumed to be lower in the remote regions in Karimui and the Wosera, so it is possible that the North Coast pattern is one towards which the others are progressing as a result of drug pressure, yet the North Coast had the lowest TFR. However, former drug pressure with anti-folates in the course of previous mass drug administration campaigns was higher in the two more remote regions 51 52 , so a reservoir of anti-folate resistance may have been present there for a long time. Parasite CQR at all three sites may have reached a plateau, while resistance to SP evolved at a different speed depending on both, historical and current anti-folate drug pressure. The frequency of the pfcrtA220S mutation increased in the Wosera between 2003 and 2004, most likely as a consequence of inappropriate drug use exerted on a large and heterogeneous parasite population. This arose because of 1) health systems management problems in the area, which lead to an undersupply with SP (local health extension officer at the Kunjingini health centre, personal communication) and 2) drug consumption outside health facilities which was relatively frequent by local standards (10% and 12% of the survey populations in 2003 and 2004, respectively) with AQ or CQ monotherapy distributed by inadequately trained community health workers. These factors may have further enhanced CQR leading to a reduced capacity of CQ to protect SP. Within the same time period, decreasing trends of TFRs were accompanied by decreasing frequencies of pfcrtA220S and pfdhpsA437G in Karimui. Health care provision generally improved due to better medical supply to this remote highland area and the study increased awareness of malaria. The result appears to have been effective delivery of combination therapy to a higher proportion of malaria cases, acting on a relatively small parasite population, therefore preventing the further development and spread of parasite resistance in this moderate transmission area 53 54 .

The in vivo and molecular data suggest that in the current first-line regimen in PNG, SP, and more specifically sulphadoxine is the effective component. Hence, molecular monitoring of resistance to this component is important under constant treatment policy. However, significant levels of in vivo failure are a strong argument for the introduction of an artemisinin-based combination regimen. Once such a regimen is introduced, rapid, easy-to-use and affordable surveillance systems will be needed to monitor emergence and spread of resistance to the components of the new combination. At the same time, surveillance of resistance to withdrawn drugs should also be continued, with the prospect that decreases in resistance may make it possible to reuse one or more of these safe and cheap drugs as partner compounds in combination regimens 55 56 .

In summary, despite there being considerable interest in using molecular data to predict drug failure rates, and hence inform and guide policy choice 5 57 , the current analyses showed that the relationship was not as simple as might be hoped. Molecular drug resistance profiles were similar in parasites from community and clinical samples, but these did not closely reflect the longitudinal and geographical variation in in vivo drug efficacy observed at health centres. The data confirm that the genetic drug resistance background of the parasite is only one of many factors determining clinical outcome and one which has evolved differently according to epidemiological characteristics as well as history and patterns of drug use in a given area. In particular it was shown that (i) using genetic markers to predict failure rates is much better if based on haplotype frequencies rather than crude prevalence of individual SNPs, and (ii) SNP marker sets will have to be constantly updated and adapted to future monitoring purposes, such as the inclusion of markers for newly introduced or withdrawn drug classes. Finally, this community-based molecular monitoring approach will have to be further evaluated in geographical areas at both extremes of transmission intensity and different drug use patterns (e.g. rural versus urban areas) in order to test its validity as a complementary resistance monitoring tool.

pfmdr1: Plasmodium falciparum multidrug resistance gene 1; pfcrt: Plasmodium falciparum chloroquine resistance transporter; pfdhfr: Plasmodium falciparum dihydrofolate reductase; pfdhps: Plasmodium falciparum dihydropteroate synthase; pfATPase6: sarcoendoplasmic reticulum Ca²⁺-ATPase (SERCA) of P. falciparum; PCR-RFLP: polymerase chain reaction-restriction fragment length polymorphism.

The authors declare that they have no competing interests.

JM participated in the coordination of the field and laboratory studies, performed data acquisition and molecular and statistical analyses and drafted the manuscript. TAS, IMH, HPB and BG participated in the design of the study, the statistical analysis and the drafting of the manuscript. IM, AS, OO, MB and JCR participated in the coordination of the field studies. All authors read and approved the final manuscript.