Optimizing tuberculosis treatment efficacy: comparing the standard regimen with Moxifloxacin-containing regimens

Tuberculosis (TB) continues to be one of the deadliest infectious diseases in the world, causing ~1.5 million deaths every year. The World Health Organization initiated an End TB Strategy that aims to reduce TB-related deaths in 2035 by 95%. Recent research goals have focused on discovering more effective and more patient-friendly antibiotic drug regimens to increase patient compliance and decrease emergence of resistant TB. Moxifloxacin is one promising antibiotic that may improve the current standard regimen by shortening treatment time. Clinical trials and in vivo mouse studies suggest that regimens containing moxifloxacin have better bactericidal activity. However, testing every possible combination regimen with moxifloxacin either in vivo or clinically is not feasible due to experimental and clinical limitations. To identify better regimens more systematically, we simulated pharmacokinetics/pharmacodynamics of various regimens (with and without moxifloxacin) to evaluate efficacies, and then compared our predictions to both clinical trials and nonhuman primate studies performed herein. We used GranSim, our well-established hybrid agent-based model that simulates granuloma formation and antibiotic treatment, for this task. In addition, we established a multiple-objective optimization pipeline using GranSim to discover optimized regimens based on treatment objectives of interest, i.e., minimizing total drug dosage and lowering time needed to sterilize granulomas. Our approach can efficiently test many regimens and successfully identify optimal regimens to inform pre-clinical studies or clinical trials and ultimately accelerate the TB regimen discovery process. Author summary Tuberculosis (TB) is a top global health concern and treatment for TB requires multiple antibiotics taken for long periods of time, which is challenging for TB patients. Therefore, identifying regimens that are more effective and more patient-friendly than the standard treatment is urgently needed. It is also known that non-compliance leads to the development of drug resistant TB. In this work, we pair computational and experimental models to predict new regimens for the treatment of TB that optimize how fast bacteria are cleared using minimal dosage. We apply novel approaches to this goal and validate our predictions using a non-human primate model. Our findings suggest that systems pharmacological modeling should be employed as a method to narrow the design space for drug regimens for tuberculosis and other diseases as well.

114 that are more effective in treating TB granulomas and that require shorter treatment times 115 compared to the standard regimen. We used our computational model GranSim to create 116 an in silico biorepository of hundreds of granulomas, combined with in vivo data 117 generated from a NHP model and applied a surrogate-assisted optimization algorithm to 118 identify regimen success and failure. We first simulated moxifloxacin-containing regimens 119 using GranSim and identified regimens that are superior to the standard treatment based 120 on sterilization times. Informed by our simulation results, we performed an in vivo study 121 in NHPs to test our predicted regimens that haven't been studied before, validating our 122 simulation predictions. Thus, our study identifies new regimens that can inform pre-clinical 123 trials to shorten treatment times and minimize dosages. This highlights the importance of 124 using modeling prior to pre-clinical trials as a step towards a more efficient and directed 125 regimen design for TB.

127 Results
128 In silico library of granulomas for treatment simulations and dose optimization 129 We first generated an in silico library of 750 granulomas over 300 days that matches NHP 130 dataset of 600 granulomas [46,47]. To do that, we sampled 250 granuloma parameter 131 sets within biological feasible ranges using the LHS method and simulating three 132 replications with each parameter set to capture both types of uncertainty present [48]. We 133 then classified granulomas that have nonzero bacterial loads (those that did not sterilize) 134 by measuring their colony forming units (CFUs) as either low-CFU or high-CFU 135 granulomas, depending on their CFU trends (Fig 1) In this work, we simulated different 136 treatments on subsets of granulomas from this library of both high-and low-CFU 137 granulomas as well as combined. This follows as humans and NHPs have multiple 138 granulomas within their lungs, and ensures that we test each regimen on a variety of 139 granuloma types and multiple granulomas, making it relevant to both experimental data 140 and clinical TB outcomes. Here, low-CFU granulomas represent the state where the 141 immune system controls bacterial growth within a granuloma, whereas within high-CFU 142 granulomas, bacteria grow to large numbers and can disseminate [8,49,50]. Specifically, 143 if the number of CFUs within a granuloma is less than 10 4 at the end of the simulation 144 and has not increased more than 50 CFUs in the last 20 days of simulation, we label it as 145 a low-CFU granuloma (Fig 1, blue curves). If the number of CFUs in a granuloma is 146 between 10 4 and 10 7 at the end of the simulation or it has increased by more than 50 147 CFUs in the last 20 days of simulation, we label it as a high-CFU granuloma (Fig 1, red 148 curves). We proposed 10 4 CFUs/granuloma as a threshold for low-CFU granulomas, 149 based on the observed CFU trends of the 750 granulomas we simulated: granulomas 150 with CFUs lower than this threshold tend to stabilize in our simulations (Fig 1, blue 151 curves), representing controlled growth. However, granulomas with CFUs higher than this 152 threshold tend to grow uncontrollably (Fig 1, red curves). We can alter this threshold 153 without loss of generality.
154 155 Fig 1. CFU trends within the in silico repository of simulated granuloma generated by GranSim after 156 the start of infection. Each curve represents a single granuloma simulation with a single parameter set 157 using GranSim, and black dots are data from NHP studies [46,47]. Based on their CFU trajectories, we 158 categorize granulomas into low-CFU (blue curves) and high-CFU (red curves) granulomas. Low-CFU 159 granulomas represent granulomas that have controlled bacterial burden; high-CFU granulomas are those 160 where bacterial growth is uncontrollable by the immune system, respectively [8,49,50]. 161 162 Simulations capture the rapid rate of sterilization with moxifloxacin-containing 163 regimens that is observed in clinical trials 164 We first compare the standard regimen for TB, i.e., HRZE, with various moxifloxacin-165 containing regimens. A recent clinical trial (REMoxTB) compared the 6-month standard 166 regimen HRZE treatment (control group) to 4-month treatment with two moxifloxacin-167 containing regimens, HRZM (termed the "isoniazid group" in the original study) and RMZE 168 (termed the "ethambutol group" in the original study) (see Table 1 for the protocol) [33].
169 Regimens with moxifloxacin were not found to be suitable replacements for the standard 170 regimen, as they had a higher rate of relapse in patients after the end of treatment, even 9 171 though they decreased the bacterial load in patients' sputum more rapidly at the beginning 172 of the treatment (Fig 2A). 203 Therefore, our simulations suggest that the HRZM is the most effective regimen in terms  Our results demonstrate that regimens that are more 308 effective in sterilizing granulomas than HRZE each contain moxifloxacin (colored curves 309 in Fig 6). For high-CFU granulomas, a moxifloxacin-containing regimen with at least 3 310 antibiotics is needed to achieve a better performance than HRZE ( Fig 6A). However, 311 sterilizing low-CFU granulomas faster than HRZE is possible even with regimens 312 containing two antibiotics (HM, RM and ZM in Fig 6B). These comparisons are based 313 only on the standard doses of regimens; optimization of doses is also possible.  339 panels A-E) compared to the standard regimen HRZE with CDC-recommended doses (X in Panel F). Dots 340 in the dashed gray rectangle indicate the regimens that have lower total drug dose and lower average 341 sterilization times (see Table 2 for the doses of each antibiotic in these regimens). Triangles indicate 342 optimized regimens with 3-way combinations, as the optimal doses of one antibiotic (E or Z) in these 343 regimens are predicted as 0. 344 345 In general, regimens that simulations identify as optimal (i.e., regimens in the  (Table 2), resulting in shorter sterilization times, which is in 356 line with clinical trials that showed a reduction in time to culture conversion using higher 357 doses of rifampicin [54,55]. Based on our earlier results, it is not surprising that these 358 optimized regimens mostly contain moxifloxacin (Table 2). This is also expected based 359 on clinical studies where moxifloxacin-containing regimens sterilize granulomas more 360 efficiently (c.f. Fig 2). Further, although most of these optimal regimens contain four 361 antibiotics, our pipeline also predicted a few optimal combinations with less than four 362 antibiotics (see triangles in the rectangle region of Fig 7F; rows labeled with triangles in 363 Table 2). (Our pipeline predicts the ethambutol optimal dose as 0 for HRME and RMZE 364 regimens and the pyrazinamide optimal dose as 0 for RMZE, thus identifying HRM, RMZ 365 and REM as additional optimal regimens). This agrees with our systematic study of all 366 possible combinations that determined HRM, RZM and REM as more efficient regimens 367 than the standard regimen HRZE (c.f. Fig 6). Our optimization approach provides a more 368 efficient way to identify regimens with different combinations of antibiotics than is possible 369 in clinical or experimental studies.
370 Table 2. Simulated Doses of Antibiotics that optimize treatment objectives (compare with Fig 7). The 371 doses for each antibiotic in the regimens that have lower average sterilization time and lower dosage than 372 the standard regimen (black row) as shown in Fig 7F (i.e., all regimens in dashed gray rectangle). Optimized 373 doses for HMZE, RMZE, HRME, HRZM and HRZE are color-coded as blue, brown, yellow, purple and 374 green, respectively. The rows labeled with a triangle indicate optimal 3-way combinations, where the 375 optimal dose of E or Z is predicted as 0.   (Fig 2). We also perform NHP studies with promising 392 regimens predicted by GranSim. (Figs 3 and 4). Different from our previous studies, we 393 systematically analyzed all possible combinations with or without moxifloxacin and 394 employed a new optimization pipeline to identify optimal regimens that sterilize 395 granulomas more efficiently than HRZE.

396
Previous clinical trials concluded that both 4 months of HRZM treatment and 4 397 months of RMZE had better bactericidal activity than the control group (HRZE) based on 398 the conversion to culture negativity status of the patients (Fig 2A) [33]. In our simulations,

405
NHP experiments with standard and moxifloxacin-containing regimens indicate 406 that all regimens reduce the total CFU of NHPs (Fig 3D) by sterilizing the majority of NHP 407 granulomas (Fig 3E and S5 Fig)

421
To test the efficacy of moxifloxacin-containing regimens more systematically and In this study, we conclude that any 4-way, 3-way 425 or 2-way (except EM) combinations that include moxifloxacin are more efficacious in 426 eliminating bacteria within low-CFU granulomas than HRZE (Fig 6B). However, only 4-427 way combinations and some of the 3-way combinations work better than HRZE for 428 treating high-CFU granulomas (Fig 6A). This suggests that decreasing the number of

559
We modeled the PD by using a Hill function that determines the concentration (C) 560 dependent antibiotic killing rate constant (k), which is the rate of bacterial death per time  675 Minimizing drug dose will decrease potential side effects. In our optimization pipeline, we 676 aim to find the regimens that minimize both objective functions.

677
The sampling ranges for each dose variable were set to range from 0 mg/kg to 678 double the standard CDC dose [4]. Maximum safe doses for each antibiotic were set to 702 ( ) = +∈ ( ) (Eq.8) 703 We also assume that the error term (x) is normally distributed with a mean of 0 and a 704 standard deviation of 2 . To provide an estimate for the error at any given x, we assume 705 the errors at two points are correlated based on the distance between those two points.
706 This means points that are closer in the variable space tend to be more related and have 707 smaller variance than those that are farther. Hence, the correlation in error between points 708 i and j, equal to component R ij in the correlation matrix R, exponentially decays with 709 respect to the weighted distance between them: