A retrospective analasysis was carried out on blood samples collected from patients during several clinical trials of anti-malarial drugs conducted in Kilifi from 1993 to 2006. The 1993–2003 samples originated from several clinical trials of SP, chlorproguanil/dapsone (Lapdap™), and SP-artesunate 12 13 14 15 and were collected before drug treatment. Others were collected in 2006 from infected patients both before and after treatment in clinical trials of artemether/lumefantrine (AL)(Coartem™) versus piperaquine-dihydroartemisinin (Artekin^®) [unpublished data], and amodiaquine (AQ) 16 . No samples were collected in 1996, 2004 and 2005. Blood samples (~50 μl) were spotted onto filter paper, air-dried and stored in plastic bags with silica gel at ambient temperature. Around thirty samples were selected per year and analysed for P. falciparum mutation at codon 76 (Lys to Thr) of pfcrt (pfcrt-76) and at codon 86 (Asn to Tyr) of pfmdr1 (pfmdr1–86), and at codons 51,59 and 164 of dhfr.

A subset of the 2006 samples were adapted in vitro for long-term culture and their chemosensitivity profiles to CQ were analysed. Briefly, parasite adaptation was carried out as follows: 1 ml of patient venous blood was collected and added to 5 ml of transport medium (RPMI-1640 containing 10% albumin). Parasites in this medium could be kept at 4°C for 4 days before the in vitro adaptation or cryopreserved in liquid nitrogen and adapted later. Fresh parasites were put in culture after washing the suspension with normal RPMI-1640 (centrifugation at 2500 g for 15 min); cryopreserved parasites were thawed, washed and put into in vitro culture according to standard protocols of cryopreservation 16 . Parasitaemias were monitored until a per cycle increase in parasitaemia greater than two-fold was reached, at which point the culture was diluted by a factor >2 and if parasitaemia increased again (greater than two-fold), the culture was then considered "adapted": this process lasted two to eight weeks. Once parasites were adapted, their chemosensitivity to CQ was determined by the radioisotope incorporation method 17 . Results were expressed as the drug concentration required for 50% inhibition of [³H] hypoxanthine incorporation into parasite nucleic acid (IC₅₀), using regression analysis of the dose-response curve. Fifty microlitres of culture from these adapted parasites was spotted onto filter paper for DNA extraction and genotyping.

Parasite genomic material from filter paper was prepared using the methanol procedure described elsewhere 18 . The detection of single-base changes at pfcrt-76 and pfmdr1–86 was carried out using the PCR-restriction enzyme protocol described elsewhere 3 . Point mutations in codons 51, 59 and 164 of dhfr were analysed as described elsewhere 18 19 . The analysis of mutations at codon 164 of dhfr of samples was not carried out routinely, since previously reports showed that this mutation was not present at detectable levels in our study site and other African sites where SP resistance is high 20 21 . However, in light of reports on the emergence of 164 dhfr mutant in Africa 22 23 24 , the 164 codon of dhfr in isolates collected in 2006 was analysed. Mutation at codon 108 was not analysed since the presence of a mutation at codon 51 and/or 59 indicates that the parasite carries the resistant mutation at codon 108 10 11 .

Frequencies of mutations at pfcrt-76, pfmdr1–86 and dhfr (codons 51 and 59) alleles were calculated as the proportion of samples carrying the mutant form out of the total of all samples carrying either only the mutant form or only the wild-type form: thus samples carrying both wild-type and mutant forms, in which relative frequencies could not be determined, were excluded from the denominator and numerator. In the case of dhfr, those samples with either a mutation at dhfr-51 (from the amino acid Asn to Ile) or dhfr-59 (from Cys to Arg) were denoted as 'double mutants'. Those with mutations at both dhfr-51 and dhfr-59 were denoted as 'triple mutants' and those with or without a resistant allele mutation at dhfr-108 (Ser to Asn) were pooled and denoted as wild-type/single mutants ('WT/single mutants'). The data were first analysed by treating these as three haplotypes, and then by treating the double mutant forms as two separate alleles to give four haplotypes.

As some of the samples were collected from patients enrolled in clinical trials, an analysis was performed to determine the potential bias to population frequencies caused by drug treatment of the patients from which the parasites were sampled. This was done by comparing the frequencies of each mutant before versus after drug treatment using a Fisher's Exact test for each trial separately. Thereafter, an analysis was carried out for trends in the frequencies of mutants through time with and then without the post-treatment samples included in the data, by fitting a logistic regression model to the frequency data with generation as a linear covariate and assuming three parasite generations per year (an average of previously used estimates 25 26 and consistent with estimates of durations of infection in children 27 ). The selection coefficient, s, which describes how much more (or less) fit the resistant mutant is relative to the wild-type, and is a function of the amount of drug pressure and also the cost of resistance in the absence of drug pressure 28 , was calculated from the estimated slope of the logistic regression curve fitted to the observed allele frequencies over the generations, viz:

where the slope of this line is 29 , w _R and w _S are the fitnesses of the resistant and wild-type alleles, respectively, 'log' denotes the natural logarithm, p _R and p _S are the relative population frequencies of the resistant and wild-type allele with p _R = 1-p _S, and the superscripts denote the generation number. Solving for s is straightforward in the case of two alleles, but in the case of more alleles, as for dhfr, this equation cannot be directly used to estimate the selection coefficient because it gives the fitness of one allele relative to all other alleles combined. In this case, the expected slope for allele i (slope _i ) becomes giving rise to allelic fitnesses of:

with i = 1,2,3 for single/WT, double and triple mutants, respectively, and where j denotes all alleles other than i. After setting w ₁ = 1, Equation (2) can be solved for w ₂ and w ₃ by substituting in the observed values for p _i and p _j and minimizing the sum of squared differences between the observed slope _i 's (i.e., those obtained by logistic regression) and their expected values, (i.e., those in Equation (2)). Note, however, that w _i now depends on the frequencies of the other alleles: since these change with time, Equation (2) is an approximation 30 . In the case of dhfr in this study, the single and double mutants changed relatively little with time and so the approximation is likely to be accurate.

Confidence intervals (95%) on estimates of s were calculated by taking the upper and lower CI values of the regression estimates of the slope and then transforming them using Equation (1). Significance levels of differences between rates of decline of different alleles between different time periods, and when including vs. excluding post-treatment samples, were tested by t-tests. data from a previously published study in Malawi 3 were analysed in the same manner. However, in that study, the frequencies of resistant mutants were calculated as the proportion of all samples with both single and mixed genotypes that carried the resistant allele, either alone or with wild-type alleles. Thus estimates of population allele frequency of the mutants in that study are over-estimates.

In Kilifi, the frequency of the pfcrt-76 mutant significantly decreased (P < 0.001) from around 94% to 63% over a period of 13 years (Figure 1a). This decline was occurring before the official withdrawal of CQ in 1999 (P < 0.05). There was no detectable difference in the rate of decline before and after this time (P > 0.9 by test for an interaction between the slope and time period). The estimated fitness of this resistant mutant was 5% lower than that of the wild-type (s = -0.05). In Malawi, the rate of decline after CQ was withdrawn was much faster than in Kilifi (from 85% to 13% in 9 years). The estimated fitness of the resistant mutant in the Malawi study was 12% lower than that of the wild-type (s = -0.12): this was significantly greater in magnitude than that in the Kilifi study (P < 0.001, by Student's t-test of the logistic regression slopes).

pfmdr-86 did not decline significantly (P = 0.6) in the Kilifi population (Figure 1b), and had a corresponding small selection coefficient of 1%. On the other hand, in the Malawi population, the frequency of the pfmdr1 mutant allele decreased slightly (P = 0.05) with an estimated selection coefficient of -3%.

In Kilifi, the frequency of the dhfr double and triple mutants combined was already high (around 80%) at the start of the study (five years before SP became the first-line treatment for uncomplicated malaria in Kenya). Nonetheless, this frequency increased significantly (P < 0.001) to above 95% during the study period (Figure 2a). Relative to the wild-type/single mutants, the triple and double mutants combined had an estimated selection coefficient of 7%. In contrast, the change in frequency in Malawi was faster (from 15% to above 80% within nine years) with a selection coefficient of 23% for the double and triple mutants combined. This was significantly higher than that for Kilifi (P < 0.001 by Student's t-test of the regression slopes). No isolate had a mutation at 164 codon of dhfr.

Analysis of the mutant alleles separately in the Kilifi data showed that the selection coefficients of the dhfr Asn51Ile-Cyst59Arg double mutants were 5% and 7%, respectively, and 8% for the triple mutant. The frequencies of the double mutants stayed approximately the same through time because, at the same time as they were gaining ground to the wild-type/single mutant, they were losing ground to the triple mutant (Figure 2b). The values of s found here span the range of estimates from other sites in Africa and South-East Asia 31 . The effect of including mixed infections on estimates of s was negligible. Thus the inability to detect alleles at low frequencies within a patient sample by the genotyping methods used here is unlikely to influence the results.

Based on the above estimates and Equation (1), it is predicted that, if the current drug pressure were to be maintained, Kilifi isolates will have largely reverted to CQ-sensitive phenotypes (frequency of the pfcrt-76 mutant less than 10%) by the year 2023, over 20 years after the official removal of the drug in Kenya. This contrasts with the case in Malawi where this level of sensitivity was reached within nine years following CQ withdrawal. In Kilifi, virtually no change is expected in pfmdr1–86 frequency while in Malawi by 2009, the frequency of pfmdr1–86 mutant is expected to fall below 10%.

CQ IC₅₀ against laboratory reference strains V1S, W2 (pfcrt-76 mutant) and 3D7 and M24 (wild type) were analysed and and these IC₅₀ values are 158 ± 75, 94.15 ± 16, 6.5 ± 2.3, 15.6 ± 8.6 nM respectively. The results of field isolates are summarized in Figure 3. Depending on their pfcrt-76 genotypes, parasites can be classified into three distinct groups. The first group, the wild-type group, has IC₅₀ values ranging between 1 and 25 nM, with an average of 13 ± 12 nM. The second group is composed of mixed genotype infections (wild-type and mutant), with a mean of 24 ± 10 nM; and the third group of pfcrt-76 mutants, which forms the majority of isolates, has a mean of 63 ± 90 nM (ranging from 5 to 150 nM). To distinguish CQ-sensitive and CQ-resistant isolates, a cut-off point of 100 nM is generally used 32 33 . However, using this cut-off point, most isolates would be classified as CQ-sensitive, yet they are pfcrt-76 mutant. Thus, the most accurate cut-off point to use in our setting would be 25 nM (Figure 3). The analysis of pfmdr1–86 genotype showed that wild-type isolates had lower CQ IC₅₀ than mutant (57 ± 50 nM versus 68 nM+87 nM), but this difference was not statistically significant (p > 0.05).