Addressing antimicrobial resistance with the IDentif.AI platform: Rapidly optimizing clinically actionable combination therapy regimens against nontuberculous mycobacteria

Abstract

Background: Current standard of care (SOC) regimens against nontuberculous mycobacteria (NTM) usually result in unsatisfactory therapeutic responses, primarily due to multi-drug resistance and antibiotic susceptibility-guided therapies. In the midst of rising incidences in NTM infections, strategies to develop NTM-specific treatments have been explored and validated. Methods: To provide an alternative approach to address NTM-specific treatment, IDentif.AI was harnessed to rapidly optimize and design effective combination therapy regimens against Mycobacterium abscessus (M. abscessus), the highly resistant and rapid growth species of NTM. IDentif.AI interrogated the drug interaction space from a pool of 6 antibiotics, and pinpointed multiple clinically actionable drug combinations. IDentif.AI-pinpointed actionable combinations were experimentally validated and their interactions were assessed using Bliss independence model and diagonal measurement of n-way drug interactions. Results: Notably, IDentfi.AI-designed 3- and 4-drug combinations demonstrated greater %Inhibition efficacy than the SOC regimens. The platform also pinpointed two unique drug interactions (Levofloxacin (LVX)/Rifabutin (RFB) and LVX/Meropenem (MEM)) that may serve as the backbone of potential 3- and 4-drug combinations like LVX/MEM/RFB, which exhibited 58.33±4.99 %Inhibition efficacy against M. abscessus. Further analysis of LVX/RFB via Bliss independence model pointed to dose-dependent synergistic interactions in clinically actionable concentrations. Conclusions: IDentif.AI-designed combinations may provide alternative regimen options to current SOC combinations that are often administered with Amikacin, which has been known to induce ototoxicity in patients. Furthermore, IDentif.AI pinpointed 2-drug interactions may also serve as the backbone for the development of other effective 3- and 4-drug combination therapies. The findings in this study suggest that this platform may contribute to NTM-specific drug development.

Keywords: Mycobacterium abscessus; Artificial Intelligence; Combination Therapy; Infectious Disease; Nontuberculous Mycobacteria.

Figure 1

IDentif.AI workflow to optimize and design combination therapy regimens against M. abscessus. The workflow begins by selecting repurposed drug candidates that are currently used against other NTM species and those that demonstrated in vitro efficacy against M. abscessus in previous studies. The dose response curves of the selected drugs are then generated, and two clinically actionable concentrations are determined. Subsequently, OACD-designed combinations are experimentally validated, and IDentif.AI harnesses the data to interrogate the drug interaction space. Efficacious IDentif.AI-designed combinations are rapidly pinpointed and subsequently, experimentally validated and compared to standard of care regimens. In the final step, IDentif.AI-pinpointed unique 2-drug combinations are included for further synergy analyses.

Figure 2

Efficacy of monotherapies and OACD-designed combinations. (A) Experimentally measured %Inhibitions for all 6 selected drugs at level 1 (green) and level 2 (yellow) concentrations. The error bars represent the propagated SD, arising from the spread of controls, and each individual replicate is represented in black dots. (B) All 50 OACD-designed combinations were experimentally validated, and their corresponding average %Inhibitions are plotted. The combinations are in order in accordance to the design in Table S1, and each replicate is represented in red dots. The monotherapy and combinatorial experiments were performed in the same experiment, and data points are presented as mean ± propagated SD (N = 3). Experimental data are summarized in Table S3 and S4. Level 1 Conc.: level 1 concentration, Level 2 Conc.: level 2 concentration, AMK: amikacin, CLR: clarithromycin, LVX: levofloxacin, LZD: linezolid, MEM: meropenem, and RFB: rifabutin.

Figure 3

IDentif.AI drug interaction analysis. The IDentif.AI analysis identified two 2-drug combinations that may have unique interactions. (A) LVX/MEM surface indicated that highest %Inhibition may be achieved when both drugs are at L2, suggesting a synergistic interaction. (B) However, LVX/RFB surface suggested that the interaction is mildly synergistic and is mostly driven by LVX. L0, L1, and L2 correspond to the OACD concentration levels: level 0, level 1, and level 2. LVX: levofloxacin, MEM: meropenem, and RFB: rifabutin.

Figure 4

Validation of standard of care and IDentif.AI-designed combinations against M. abscessus. Two SOC combinations (gray) were validated and compared to IDentif.AI-designed combinations (blue). Furthermore, IDentif.AI-pinpointed non-effective combinations (red) were also experimentally validated. The concentrations of each drug in each combination are listed in the table below. All combinations were experimented in triplicates (N = 3). Each replicate is represented in black dots. Data points are presented as mean ± propagated SD. The error bars represent the propagated SD, which is the measure of the plate-plate variation, instead of the spread of the triplicates. Kruskal-Wallis test detected statistically significant differences at P < 0.01 for the %Inhibitions among all validated combinations. Subsequently, pairwise comparisons via Dunn's post hoc test identified statistically significant differences in two pairs of combinations: (1) LVX/MEM and RFB/CLR (2) LVX/MEM/RFB and RFB/CLR (*P < 0.05). Experimental data are summarized in Table S5 and S6. AMK: amikacin, CLR: clarithromycin, LVX: levofloxacin, LZD: linezolid, MEM: meropenem, and RFB: rifabutin.

Figure 5

Validation of LVX/RFB drug interaction space. (A) Response surface of LVX/RFB in the validation interaction space (0% - 20% C_max). All replicates (N = 3) were included for the construction of the response surface. The clinically actionable interaction space (< 10% C_max) is within the dotted black box. The adjusted R² indicates the goodness of the fit for the response surface. (B) Clinically actionable interaction space is the magnification of the dotted black box in Figure 5A. (C) The heatmap represents the 2-dimensional view of the LVX/RFB response surface in the validation interaction space (0% - 20% C_max). The clinically actionable interaction space (< 10% C_max) is within the dotted black box. Experimental data are summarized in Figure S5 and Table S7. Statistics for the response surface are detailed in Table S8. LVX: levofloxacin and RFB: rifabutin.

Figure 6

Bliss independence model analysis of LVX/RFB. (A) Interaction map of LVX/RFB with measured %Inhibitions at each corresponding dose ratio. The %Inhibitions of LVX/RFB combinations and monotherapies were tested from 0% to 20% C_max, and the clinically actionable interaction space (< 10% C_max) for LVX/RFB is within dotted black box, which was also the interaction space analyzed by IDentif.AI (N = 3). (B) Synergy map of LVX/RFB with corresponding Bliss Synergy δ-Scores. Scores greater than 10, between -10 and 10, and less than -10 are considered synergistic, additive, and antagonistic, respectively. Bliss independence analysis revealed a mild synergistic dose region (dotted pink box). Clinically actionable interaction space (< 10% C_max) is within the dotted black box. Statistical significance of the Bliss Synergy δ-Scores was determined by one-sample t-test (*P < 0.05). (C-D) Dose response curves of LVX and RFB in monotherapies. The dotted line represents absolute IC₅₀. Data points are presented as mean ± propagated SD (N = 3). (E) Dose response curve of LVX in monotherapy and in combination with RFB. The IC₅₀ values for LVX and LVX/RFB are summarized in the legend. The dotted line represents absolute IC₅₀. Data points are presented as mean ± propagated SD (N = 3). Experimental data are summarized in Table S7. LVX: levofloxacin and RFB: rifabutin.

Figure 7

DiaMOND synergy analysis of LVX/MEM. (A) Interaction map of LVX/MEM with measured %Inhibitions at each corresponding dose ratios, and the clinically actionable interaction space (< 10% C_max) is within the dotted black box. The dotted red box represents the combination in original IDentif.AI concentrations (level 2), and the dotted pink box represents the 4-fold MIC shift of the originally identified synergy (N = 3). (B-C) Dose response curves of LVX and MEM in monotherapies, and the dotted line represents absolute IC₅₀. Data points are presented as mean ± propagated SD (N = 3). (D) Dose response curve of LVX in monotherapy and in combination with MEM. The IC₅₀ values for LVX and LVX/MEM are summarized in the legend. The dotted line represents absolute IC₅₀. Data points are presented as mean ± propagated SD (N = 3). Experimental data are summarized in Table S9. LVX: levofloxacin and MEM: meropenem.