Artificial intelligence–derived 3-Way Concentration-dependent Antagonism of Gatifloxacin, Pyrazinamide, and Rifampicin During Treatment of Pulmonary Tuberculosis

Abstract

Background

In the experimental arm of the OFLOTUB trial, gatifloxacin replaced ethambutol in the standard 4-month regimen for drug-susceptible pulmonary tuberculosis. The study included a nested pharmacokinetic (PK) study. We sought to determine if PK variability played a role in patient outcomes.

Methods

Patients recruited in the trial were followed for 24 months, and relapse ascertained using spoligotyping. Blood was drawn for drug concentrations on 2 separate days during the first 2 months of therapy, and compartmental PK analyses was performed. Failure to attain sustained sputum culture conversion at the end of treatment, relapse, or death during follow-up defined therapy failure. In addition to standard statistical analyses, we utilized an ensemble of machine-learning methods to identify patterns and predictors of therapy failure from among 27 clinical and laboratory features.

Results

Of 126 patients, 95 (75%) had favorable outcomes and 19 (15%) failed therapy, relapsed, or died. Pyrazinamide and rifampicin peak concentrations and area under the concentration-time curves (AUCs) were ranked higher (more important) than gatifloxacin AUCs. The distribution of individual drug concentrations and their ranking varied significantly between South African and West African trial sites; however, drug concentrations still accounted for 31% and 75% of variance of outcomes, respectively. We identified a 3-way antagonistic interaction of pyrazinamide, gatifloxacin, and rifampicin concentrations. These negative interactions disappeared if rifampicin peak concentration was above 7 mg/L.

Conclusions

Concentration-dependent antagonism contributed to death, relapse, and therapy failure but was abrogated by high rifampicin concentrations. Therefore, increasing both rifampin and gatifloxacin doses could improve outcomes.

Clinical Trials Registration

NCT002216385.

Fluoroquinolones are an essential component of the treatment for multidrug-resistant (MDR) tuberculosis. Moxifloxacin and gatifloxacin have also been examined for shortening treatment duration for drug-susceptible tuberculosis in large phase 3 clinical trials [1–6]. One of these, the OFLOTUB trial, which compared a 4-month gatifloxacin-containing regimen plus rifampicin, pyrazinamide, and isoniazid to the standard 6-month regimen with ethambutol, included a nested pharmacokinetic (PK) study [6, 7]. Several studies have now revealed that PK variability, which together with the dose administered determines the drug concentrations achieved in patients, is a major driver of therapy outcomes such as cure, relapse, and acquired drug resistance for both drug-resistant and drug-susceptible disease [8–10]. Here, we utilized the PK study to determine if PK variability also had an impact on the outcomes in the OFLOTUB experimental arm.

In a single-dose PK study of a 4-drug fixed-dose combination (FDC) containing rifampicin, isoniazid, pyrazinamide, and ethambutol in healthy volunteers, the addition of gatifloxacin was associated with a significant reduction of rifampicin’s 0–24 hour area under the concentration-time curve (AUC_0-24) [11]. The importance of these particular drug–drug PK interactions in clinical outcomes of patients with tuberculosis is unclear. Recently, several preclinical and clinical studies have demonstrated that concentration-dependent antagonism is a surprisingly common problem with anti-tuberculosis drugs, including quinolones [10, 12–17]. The same drugs can show antagonism at some concentrations and then synergy or additivity at other concentrations, in other words, they are best described by hinge functions and analyzed using nonlinear methods (such as nonparametric regression). Therefore, it is important to identify pharmacodynamic (PD) interactions whenever another drug is added to current combination therapy regimens or whenever a new regimen is tested as first- or second-line therapy.

We have previously used machine-learning (ML) algorithms to identify patterns of clinical factors and drug concentrations predictive of clinical outcomes, and to identify interactions between them, and the contribution of single drugs in combinations to therapy outcomes [12, 15, 18]. ML, coined by Arthur Samuel in 1959, has been a cornerstone concept of artificial intelligence (AI) since the field’s inception. ML algorithms are adept at disentangling nonlinear and higher-order interactions between predictors and outcomes and between potential predictors. The algorithms, designed to analyze hinge functions, are also effective at ranking these predictors by order of importance [19]. Here, we used an ensemble of ML methods, which we further supplemented with probit regression analysis, to identify predictors of clinical outcomes in the experimental OFLOTUB arm.

METHODS

Description of Clinical Dataset

The OFLOTUB study was a randomized, controlled trial of a 4-months gatifloxacin-containing regimen vs the standard 6-month first-line therapy. The study was conducted at tertiary facilities in 5 cities: Cotonou in Benin, Conakry in Guinea, Nairobi in Kenya, Dakar in Senegal, and Durban in South Africa in 1836 patients with pulmonary tuberculosis. Details of the OFLOTUB trial are provided at the Clinical Trials website (ClinicalTrial.gov), and complete demographic and clinical information of study participants, gatifloxacin population PKs, and clinical outcomes have been published [6, 7, 20, 21]. All patients were initially smear positive and rifampicin susceptible based on rapid detection of drug-resistant tuberculosis on at least 2 sputum samples. A total of 291 (6%) patients underwent blood sampling for the nested PKs study, of which 128 (44%) patients were in the gatifloxacin arm.

All patients received supervised doses. Patients in the intervention arm received 400 mg of gatifloxacin regardless of weight and FDC tablets of 150 mg rifampicin and 75 mg isoniazid in addition to 400 mg pyrazinamide during the first 2 months of therapy and FDC tablets of rifampicin 150 mg and isoniazid 75 mg during the last 2 months. Patients weighing <50 kg received 3 FDC tablets, while those weighing ≥50 kg received 4 FDC tablets. This means that, depending on the baseline weight, each patient essentially received different drug dosages in milligrams per kilogram for each of the 4 drugs in the intervention arm.

Definitions of Outcomes

Patients who failed at the end of therapy, relapsed during the 24-month follow-up, or died while on therapy and during follow-up were defined as having unfavorable outcomes [6]. Patients who were lost to follow-up were also initially considered to have unfavorable outcomes and were analyzed for effect. Therapy failure and relapse results were based on sputum smear or culture results. Relapse was confirmed with mycobacterial interspersed repetitive unit-variable number tandem repeat spoligotyping.

Machine-learning Methods

We used an ensemble of ML algorithms; steps and softwares are detailed in the Supplementary Methods [19, 22–25]. Random forest and stochastic gradient boosted classification and regression trees (CARTs) were used to generate variable importance ranking scores (VIRS), as we have done in the past [12, 14, 15, 18, 26]. Gradient boosting of algorithms in ML helps in identifying weak but otherwise important signals (that would ordinarily be missed) in data and combines these with the other variables that they interact with to improve model prediction [24, 27]. CART, which is nonparametric, was used to detect high-order interactions and produce simpler decision trees for clinical decision-making [28].

Statistical Analyses

Comparisons of proportions between groups were made using χ² or Fisher exact tests. Probit regression and multivariate logistic regression models were fitted using the thresholds obtained from the ML outputs to compare favorable outcomes between groups. In building the multivariate logistic regression models, iterative stepwise backward elimination of 1 variable at a time was performed manually to produce models that could be used in the clinic.

RESULTS

Patient Characteristics

Of the 128 patients who underwent PK sampling in the intervention arm, 2 were excluded from further analyses because they had been administratively withdrawn. The demographic characteristics in Table 1 show that 71 (56%) enrolled patients were from South Africa; patients from South Africa had higher body mass indices (BMIs) and higher human immunodeficiency virus infection rates. Moreover, patients from Benin and Guinea had very low rates of cavitary disease compared to those from other countries, and patients from Senegal had significantly higher rifampicin concentrations. Furthermore, the dose in milligrams per kilogram that patients received ranged from 5.0 to 11.43 mg/kg, and the mean milligrams per kilogram dose varied between sites. The distribution of gatifloxacin dose observed in patients is shown in Figure 1A.

Table 1.

Clinical Factors and Outcomes After 24 Months Follow-up of 126 Patients Enrolled at the Different Country Sites

	Country
Variable	Benin, n = 23 (%)	Guinea, n = 16 (%)	South Africa, n = 71 (%)	Senegal, n = 16 (%)	P Value
Demographics
Sex, male	19 (83%)	8 (50%)	46 (65%)	15 (94%)	.019
Age, years*	31 (21–55)	24 (18–42)	30 (19–57)	29 (20–47)	.058
Weight, kg*	53 (35–64)	52 (40–76)	56 (39–80)	56 (43–67)	.074
Body mass index, kg/m2*	17 (14–22)	20 (14–23)	20 (16–29)	18 (14–21)	<.001
Clinical
Gatifloxacin dose, mg/kg*	7.55 (6.25–11.43)	7.69 (5.26–10)	7.19 (5–10.27)	7.11 (5.97–9.30)	.074
Human immunodeficiency virus infection	3 (13%)	0	32 (45%)	0	<.001
Cavitary disease	7 (30%)	2 (13%)	68 (96%)	15 (94%)	<.001
Sputum microscopy results, ≥2+	23 (100)	13 (81)	64 (90)	13 (81)	.163
Chest X-ray, >3+	10 (43)	6 (38)	33 (46)	10 (63)	.523
Drug exposures
Gatifloxacin 24-hour area under the concentration-time curve	37.76 (10.92)	38.19 (20.09)	32.33 (12.57)	37.66 (10.57)	.214
Gatifloxacin Cmax	3.79 (0.71)	4.07 (1.0)	3.87 (1.03)	4.31 (0.76)	.203
Rifampicin Cmax	6.24 (1.53)	7.01 (2.29)	6.04 (2.00)	8.93 (2.23)	<.003
Pyrazinamide Cmax	33.31 (3.86)	39.92 (7.30)	37. 55 (8.01)	35.32 (3.31)	.033
Outcomes
Favorable	10 (44)	13 (81)	56 (79)	16 (100)	.001
Dropped out	7 (30)	1 (6)	4 (6)	0
Relapsed	6 (26)	1 (6)	8 (11)	0
Failed	0	0	1 (1)	0
Died	0	1 (6)	2 (3)	0

Bold indicates statistical significance; * indicates median (range).

Figure 1.

Distribution of drug doses and concentrations in the 126 study patients. (A) The distribution of gatifloxacin dose in milligrams per kilogram received by each patient is shown. (B) Distribution of the gatifloxacin 24-hour area under the concentration-time curve (AUC_0-24) has a different shape from that of the dose since each patient has a different clearance of the drug due to pharmacokinetic (PK) variability. The median AUC_0-24 (range) was 33.95 (5.27–94.92) mg*h/L. (C) Distribution of gatifloxacin peak (C_max) concentration differs from the dose, indicating influence of between-patient variability in volume of distribution. (D) The distribution of rifampicin peak concentration; the ratio of the lowest-to-highest peak concentration was 9.0. (E) Distribution of pyrazinamide peak concentration was much narrower than for rifampin. (F) Distribution of isoniazid AUC was wide, with a highest-to-lowest AUC ratio of 13.5, to be expected given the acetylation rates in patients. Detailed description of the PKs of drugs used in the OFLOTUB trial, including the clearance and volume of distribution of each of these drugs in the study population, have been described y Smythe et al [7]. Abbreviations: AUC_0-24, 24-hour area under the concentration-time curve; MIC, minimum inhibitory concentration.

Pharmacokinetic Characteristics

The distributions of AUC_0-24 associated with the gatifloxacin doses are shown in Figure 1B. The distributions of peak concentrations of rifampicin and pyrazinamide and isoniazid AUC₂₄ are shown in Figure 1D–1F. These concentrations are consistent with prior observations in the same patient populations in the past [7, 29].

Boosted CART and Random Forests Output

We ran paired random forests and boosted CART of all 27 potential predictors for the binary outcomes of favorable vs unfavorable outcomes after 24 months of follow-up in 126 enrolled patients. The initial random forests results, which had a receiver operating characteristic (ROC) of 73% in the out-of-bag sample, identified 4 interesting patterns, which informed further analyses. First, the study site was an important predictor and was ranked second in both random forests and boosted CART. Given this and the results listed in Table 1, subsequent analyses separated South African from West African patients. Second, for all drugs, steady-state drug concentration was ranked higher than initial concentration in predicting patients’ outcomes; thus, subsequent analyses utilized steady-state concentrations. Third, gatifloxacin AUC_0-24 outranked peak concentration that in turn outranked isoniazid AUC_0-24 in predicting outcomes. Fourth, the “lost to follow-up” category did not inform us as to who got negative sputum or relapsed (ie, outcomes); therefore, we excluded the 12 patients who were lost to follow-up in subsequent analyses.

Next, our analyses focused on the comparison of 114 patients: 95 with favorable outcomes (ie, microbiologic cure) vs 19 who died, failed therapy, or relapsed. Random forests output for the 114 patients is shown in Figure 2A. The ROC was 80.8%, and the misclassification rate was 0.333. Pyrazinamide peak (C_max) concentration ranked first (VIRS = 100%) followed by rifampicin C_max (95%), isoniazid AUC_0-24 (71%), and gatifloxacin AUC_0-24 (64%). Figure 2B depicts the ranking for West African patients, which had an ROC of 93% and misclassification rate of 0.213. Rifampin C_max was ranked at the apex, followed by age (VIRS = 58%), gatifloxacin AUC_0-24 (51%), and isoniazid AUC_0-24 (46%). Figure 2C shows the ranking from the South African patients, which had an ROC of 76% and misclassification rate of 0.298. Pyrazinamide C_max was the primary predictor followed by BMI (VIRS = 44%), age (39%), and gatifloxacin AUC_0-24 (35%).

Figure 2.

Artificial intelligence approach to identifying predictors of therapy failure. (A) The variable importance ranking from a random forests model, which started with 27 clinical, radiologic, and laboratory variables from 114 study patients. The ranking is produced by several classification and regression trees (CART) generated from the data “voting” for the variable most used as a surrogate or as a main variable for splitting data into homogenous groups. Here, the target variable was binary favorable outcome (ie, microbiologic cure) vs unfavorable (ie, microbiologic failure, relapse, or death) outcome. (B) The variable importance ranking from 47 patients from the West African sites. (C) Variable importance ranking in 67 patients from the South African sites. (D) Trees generated for predictors of patients from West Africa. All 8 patients who failed therapy had low rifampicin peak concentrations (≤7.01 mg/L) but pyrazinamide peak concentrations higher than 29.08 mg/L. (E) Boosted CART from South African patients; all 11 patients who failed therapy weighed ≥48 kg and also had low rifampicin peak concentration, although the threshold was slightly higher at 8.86 mg/L. (F) The impact of the interactions between gatifloxacin area under the concentration-time curve and rifampicin peak concentrations on outcome. Blue/green color zones indicate either drug concentrations with negative effects contributing to therapy failure, while brown/red color indicates drug concentration with positive effects associated with cure. Abbreviations: AUC_0-24, 24-hour area under the concentration-time curve; BMI, body mass index; Fav, favorable; GATI, gatifloxacin, HIV, human immunodeficiency virus; PZA, pyrazinamide; RIF, rifampin; ROC, receiver operating characteristic; SS, steady state; Unfav, unfavorable.

The representative decision trees from random forests are shown in Figure 2D for the West African patients. The following thresholds were identified for West Africa: rifampicin C_max of 7.01 mg/L and pyrazinamide C_max of 33.29 mg/L. When rifampicin C_max was >7.01 mg/L, all (100%) patients had favorable outcomes. However, when rifampicin C_max was <7.01, the gatifloxacin AUC_0-24 of 28.51 mg*h/L was the next predictor. Surprisingly, all patients with a gatifloxacin AUC_0-24 less than 28.51 mg*h/L had a favorable outcome, suggesting concentration-dependent antagonism. All 8 patients with unfavorable outcomes in West Africa had a low rifampicin C_max and a high gatifloxacin AUC_0-24. The third ranked factor was pyrazinamide C_max of 33.29 mg/L, which also showed concentration-dependent antagonism and accounted for all 8 patients with unfavorable outcomes.

Figure 2E shows the representative tree for South African patients. The highest ranked predictor was pyrazinamide C_max >37.45 mg/L; all patients with unfavorable outcomes achieved concentrations above this. In patients with pyrazinamide C_max >37.45 mg/L, a gatifloxacin AUC_0-24 >47.17 mg*h/L resulted in cure of 100% of patients. With gatifloxacin AUC_0-24 below this threshold, a low rifampin C_max was associated with poor outcome, as were patients with weight >48 kg. Indeed, 11/11 (100%) South African patients with unfavorable outcomes had weight >48 kg.

Finally, if the entire dataset was recombined (West Africa plus South Africa), 14/19 (74%) unfavorable outcomes were predicted by low rifampicin peak concentration, while 11/19 (58%) unfavorable outcomes were in patients who weighed >48 kg in boosted CART (Figure 2F). At higher gatifloxacin AUC_0-24, the outcomes were consistently unfavorable, including antagonism with pyrazinamide. Taken together, these data suggest multiway concentration-dependent antagonism at low rifampicin concentrations, as reflected by higher relapse rates and deaths.

Multivariate Adjusted Regression Splines

Clearly, while PK variability and patient weight were the main predictors of outcome, the picture was complex. In order to better understand this, we utilized multivariate adjusted regression splines (MARS) to identify the predictors of response rates (rates of favorable outcomes) as continuous variables. The probability of favorable outcome in West African patients was estimated with 6 basis functions (BFs) that explained 75% of the variance in outcomes (see Supplementary Table S1.) The relationships between outcome and BFs are depicted by equation 1 in the table.

Gatifloxacin AUC_0-24 below 41.523 mg/L*h were antagonistic with isoniazid AUC_0-24 above 21.42 mg/L*h. On the other hand, rifampicin C_max <5.487 mg/L was antagonistic with both gatifloxacin and pyrazinamide (Supplementary Table S1). Similarly, equation 2 (Supplementary Table S1) describes the probability of favorable outcomes in South African patients, which was best described with the 7 BFs shown in Supplementary Table S2, which explained 31% of the variance in outcomes.

It is noteworthy that at the lower rifampicin exposures, there was synergy with gatifloxacin; however, the overall contribution was miniscule, as shown by the small coefficients. Nonetheless, the multiway concentration-dependent antagonism that we picked using CART was also confirmed using the MARS method.

Probit and Multivariate Logistic Regression Analyses

In order to better inform clinical decision-making and to synthesize the findings into a simple decision-making table, we performed standard statistical tests using the thresholds and output from ML analyses, with results shown Table 2 and Figure 3. Table 2 shows the odds of therapy failure with rifampin C_max <7.01 mg/L in univariate analysis. The adjusted odds for favorable outcomes on multivariate analyses for rifampin C_max>7.01 mg/L was 10.40 (95% confidence interval, 2.20–49.31). Next, we examined the odds of failure in the 72 patients who had rifampin C_max <7.01 mg/L. Since the BFs pointed to an M-shaped relationship, with 3 hinges for gatifloxacin AUC that separated the data into 3 spaces, we examined the odds ratio of failure in each of the 3 gatifloxacin concentration zones: AUC_0-24 >50.29 mg*h/L, AUC_0-24 of 35.48–50.29 mg*h/L, and AUC_0-24 <35.48 mg*h/L. Table 2 shows that the odds of failure were significantly higher in the middle zone, which is where concentration-dependent antagonism was encountered. Figure 3A shows that the gatifloxacin AUCs were driven by weight of patients, with a threshold weight around 50 kg. Next, we performed the same analysis for pyrazinamide; however, we found that no patient had both a pyrazinamide C_max >54.92 mg/L and a rifampin C_max <7.01 mg/L. The antagonism zone of a pyrazinamide C_max of 33.29–54.92 mg/L was confirmed to have an odds ratio of success of only 0.3 when compared to that of patients with pyrazinamide C_max <33.29 mg/L who also had a low rifampin C_max. The contribution of pyrazinamide C_max is also shown in Figure 3B, which shows the BF graphically.

Table 2.

Odds Ratios Associated With Favorable Outcomes for Different Concentration Thresholds in All 114 Patients

Decision Points		Treatment Outcomes		Unadjusted Odds Ratios
Drugs	Thresholds Used for Decisions	Favorable, n (%)	Unfavorable, n (%)	(95% Confidence Intervals)
All patients (n = 114)
Rifampicin	Cmax ≥7.01 mg/L	40 (42)	2 (11)	Referent
	Cmax<7.01 mg/L	55 (58)	17 (89)	0.16 (0.04–0.67)
For rifampin Cmax<7.01 only (n = 72)
Gatifloxacin	AUC0-24>50.29 mg*h/L	5 (9)	2 (12)	Referent
	AUC0-24 35.48-50.29 mg*h/L	8 (15)	9 (53)	0.12 (0.04–0.46)
	AUC0-24≤35.48 mg*h/L	42 (76)	6 (35)	0.36 (0.06–2.27)
Pyrazinamide	Cmax>54.92	0	0	NA
	Cmax >33.29 mg/L	14 (25)	9 (53)	0.30 (0.11–0.99)
	Cmax≤33.29 mg/L	41 (75)	8 (47)	Referent

Abbreviation: AUC_0-24, 24-hour area under the concentration-time curve.

Figure 3.

Statistical analyses approach to identifying predictors of therapy failure. (A) Gatifloxacin peak and 0–24 hour area under the concentration-time curve (AUC_0-24) concentrations were significantly higher in patients with weight <50 kg compared to those with weight >50 kg. Mean ± standard deviation peak concentration was 4.42 ± 0.98 mg/L in patients weighing <50 kg and 3.79 ± 0.90 mg/L in patients >50 kg (P = .003). Similarly, gatifloxacin AUC_0-24 concentrations were also significantly lower in heavier patients: 40.61 ± 17.57 mg*h/L in patients weighing <50 kg and 32.67 ± 11.23 mg/L*h in patients weighing >50 kg (P = .014). (B) Relative contribution of changes in pyrazinamide peak concentration basis function (BF) to probability of microbiologic cure, from Table 2 and equation 1. The BF was (0, 49.511 – pyrazinamide C_max). Each decrease in pyrazinamide concentration below 49.511 mg/L contributes to lower microbiologic cure up to 27 mg/L conditional on low rifampin concentration. (C) Regression of probability of microbiologic cure vs gatifloxacin dose in milligrams per kilogram stratified by the rifampin peak concentration threshold of 7.01 mg/L identified by the machine-learning methods. Based on extrapolation of the data from 72 patients with rifampin peak concentration below thresholds, a minimum of 800 mg gatifloxacin dose would be required to attain cure in 90% of patients. Abbreviations: AUC_0-24, 24-hour area under the concentration-time curve.

Given that the data suggest different gatifloxacin dose-response effects dependent on rifampin peak concentrations, we performed a probit analysis of gatifloxacin dose vs probability of outcome, shown in Figure 3C. Extension of the curve beyond the highest dose given revealed that increasing the gatifloxacin dose to 20 mg/kg (about 800 mg/day) could lead to optimal outcomes, even with rifampicin C_max <7.01 mg/L.

DISCUSSION

Our findings underline the critical importance of considering PK variability and drug–drug interactions when combining anti-tuberculosis drugs into regimens. First, in this gatifloxacin-substitution regimen meant to shorten therapy, PK variability was the most important driver of outcome; all of the highest ranked predictors were drug concentrations [28, 30, 31]. This means that we can use a PK/PD approach to further optimize regimens in order to shorten therapy duration [32–36]. As a prime example, high rifampicin C_max overrode the antagonism with pyrazinamide and gatifloxacin, which was mainly manifest as relapse. Antagonism conditional on low rifampicin C_max has also been observed in the antagonism of rifampicin with isoniazid in children and adults [12, 14, 15]. Antagonism disappeared when gatifloxacin AUC_0-24 exceeded 42 mg*h/L. Thus, if rifampin and gatifloxacin dosing were optimized to achieve this, shorter therapy duration could be effective, which is consistent with a high-dose rifampin and moxifloxacin hollow fiber study described in this supplement [37]. Another idea could be to replace pyrazinamide, perhaps with an optimized dose of an oxazolidinone to reduce antagonism, while tapping into the synergy of fluoroquinolones and oxazolidinones identified elsewhere [13, 38].

Second, antagonism seems to be a rather common phenomenon with our current treatment regimens for tuberculosis [12, 14, 15, 39]. Indeed, we have identified PD antagonism in 4/5 studies that examined first-line regimens. The belief and often-heard mantra that this does not matter much since most patients with drug-susceptible tuberculosis do well is contradicted by the data. In fact, success rates in Europe, the United States, and Africa at 6 months are around 80% at best [6, 18, 40, 41]. In the current study and in past studies of 6-month regimens, we found that antagonism is associated with slower sterilizing effect rates, lower sputum conversion rates, higher relapse rates, and higher mortality [14, 39]. Indeed, the antagonism has a PK/PD basis; mouse and hollow fiber models of tuberculosis have identified these very antagonistic relationships [10, 12–17]. This cannot be ignored. Conversely, it should reenforce the need to increase doses of rifampicin, ethambutol, and quinolones in first-line regimens for children and adults; patient outcomes may be dependent on it. The likely reason why the field had not identified these antagonistic interactions in the clinic is likely just a statistical one. Comparisons of measures of central tendency are not geared to identify changes in direction on each side of a hinge in a V- or M-shaped relationship, for example, given that they average out concentrations. However, ML algorithms, which became available in recent decades, are optimized to identify such multiway higher-order interactions [19, 25].

Third, our findings have implications for the use of gatifloxacin in the treatment of MDR tuberculosis. Most of the World Health Organization (WHO)–recommended regimens that involve fluoroquinolones (usually moxifloxacin or gatifloxacin) include the use of pyrazinamide and isoniazid, which demonstrated antagonism with gatifloxacin in this study. While the context of gatifloxacin use in MDR tuberculosis is different (regimens also include clofazimine and aminoglycosides), several drugs such as isoniazid and pyrazinamide are also commonly used. These regimens have no rifampin, which has a high concentration that overrides the antagonism and protects the regimen. In these regimens, the central role of quinolones is undeniable, based on poor outcomes at higher gatifloxacin minimum inhibitory concentration (MICs) and in extremely drug-resistant tuberculosis [42]. Our probit analysis shows that gatifloxacin doses of 800 mg/day would be optimal, which is in agreement with findings from our PK/PD studies and systemic analyses [42]. In addition, given the pyrazinamide MIC distribution in MDR tuberculosis patients and our proposed clinical susceptibility breakpoints in places where we practice, the microbial kill derived from pyrazinamide in MDR tuberculosis is in any case questionable [43, 44]. As discussed for first-line regimens, replacing pyrazinamide with oxazolidinones or newer agents such as bedaquiline and delamanid may be more desirable and could improve patients’ outcomes as well as shorten therapy duration.

There are several strengths and limitations in this study. The ensemble of AI methods and standard inferential statistics all point to PK variability–derived concentration-dependent antagonism as one of the main reasons for therapy failure. While the effects differed between centers and different algorithms picked different concentration thresholds below that patients failed, the methods were very consistent in picking the antagonism and the role of high rifampin concentrations in achieving a favorable outcome. Another limitation is that we were unable to determine causality in the multiway antagonism and were unable to clearly ascertain the temporal sequences in these multiway PD interactions. A third limitation is that in other work with benzylpenicillin for treating tuberculosis, we found that in addition to concentration dependence of antagonism and synergy, duration of therapy may also be a determinant of antagonism and synergy [45]. Thus, it could be that what we identified with a 4-month regimen does not strictly apply to say a 9–12 month therapy duration regimen. Nonetheless, these limitations do not diminish the major finding that drug concentrations and the interactions are important drivers of outcomes in regimens for tuberculosis.

Supplementary Data

Supplementary materials are available at Clinical Infectious Diseases online. Consisting of data provided by the authors to benefit the reader, the posted materials are not copyedited and are the sole responsibility of the authors, so questions or comments should be addressed to the corresponding author.

Notes

Acknowledgments. We thank the participating patients, staff at the clinical trial sites, and the members of the Data Monitoring Committee (A. Nunn [Chair], A. Mwinga, and P. Godfrey-Fausset).

Disclaimer. P. O. and C. S. M. are currently staff members of the WHO. The authors alone are responsible for the views expressed in this publication and they do not necessarily represent the decisions, policy, or views of the WHO.

Financial support. This work was partly supported by the Wellcome Trust (grant 206379/Z/17/Z to H. M) and WHO Special Programme on Research and Training in Tropical Diseases fellowship grant to P. L. O.

Supplement sponsorship. This supplement is sponsored by the Baylor Institute of Immunology Research of the Baylor Research Institute.

Potential conflicts of interest. All authors: No reported conflicts of interest. All authors have submitted the ICMJE Form for Disclosure of Potential Conflicts of Interest. Conflicts that the editors consider relevant to the content of the manuscript have been disclosed.