Fig 1.
Schematic diagram of the dynamical model given by Eqs 1–5.
The variables X, S, R1, R2, and R3 represent uninfected patients, and patients infected with a sensitive strain, a strain resistant to drug 1, a strain resistant to drug 2, and a strain resistant to both drugs, respectively.
Table 1.
Model parameters with their corresponding description, unit, and process.
Fig 2.
Illustration of the efficacy score.
The score is calculated by integrating over the difference in the number of infecteds in the absence and presence of treatment, and the area representing the efficacy score is marked by the dashed lines.
Table 2.
Model parameters given with their sampling range and sampling schemes.
Fig 3.
Chance of being the best strategy according to the random sampling results.
500,000 randomly sampled parameter sets are used. For each parameter set, efficacy scores are calculated for the five treatment strategies, and the winning strategy for that particular parameter set is determined. Based on these results, the probability of being the best strategy for each treatment is calculated.
Fig 4.
LDA of random sampling results.
LDA classifying parameter sets according to the strategy with the highest efficacy score. Shaded areas represent the density of the parameter sets colored according to which strategy wins. Each treatment strategy is represented by a different color, and the opacity of each color is proportional to the number of parameter sets that fall into that corresponding region. 500,000 randomly sampled parameter sets are used. LD1 and LD2 are the two principal axes of the LDA. The parameter vectors are given with their relative magnitudes and counterclockwise angles. Parameter vectors are amplified by a factor of 3 in order to make them better visible. The most important parameters in terms of class separation are q (rate of de novo emergence of double resistance), c1, c2, and c3 (fitness costs of resistance to drug 1, drug 2, and both drugs).
Fig 5.
Marginal benefit comparisons for random sampling results.
Results for the marginal benefit comparisons where 2000 samples are used. Disks are colored according to the winning strategy, and both the disk size and the transparency are proportional to the efficacy score difference between the best (worst) and the second best (second worst) strategy. Transparency values are normalized for each panel separately. Three examples of disks are provided with the values they represent on the right legend of each panel. LD1 and LD2 are the two principal axes of the LDA, and the most important parameters in terms of class separation are q (Rate of de novo emergence of double resistance), c1, c2, and c3 (Fitness costs of resistance to drug 1, drug 2, and both drugs). (A) Comparison of the best vs. the second best strategy. Disks are colored according to the best strategy. (B) Comparison of the worst vs. the second worst strategy. Disks are colored according to the worst strategy.
Fig 6.
Analysis of the important parameters.
Analysis of the role of de novo emergence of resistance for RS results. (A) Semi logarithmic plot of fraction of combination therapy being the best or the worst within all strategies vs. q/ν, (q: rate of de novo emergence of double resistance, ν: rate of denovo emergence of single resistance) for 500,000 RS results. (B) Plot of fraction of combination therapy being the best or the worst within all strategies vs. c3 − (c1 + c2) (c1,c2,c3: fitness costs of resistance to drug 1, drug 2, and both drugs), for 500,000 RS results.
Fig 7.
PSO results, where 1000 realizations of the PSO algorithm are used for each optimization target. (A) PSO results projected on the same LDA space as the RS results. (B) Histograms of the most important parameters for each optimization target.