Machine Learning-Led Optimization of Combination Therapy: Confronting the Public Health Threat of Extensively Drug Resistant Gram-Negative Bacteria.

Abstract

Developing optimized regimens for combination antibiotic therapy is challenging and often performed empirically over many clinical studies. Novel implementation of a hybrid machine-learning pharmacokinetic/pharmacodynamic/toxicodynamic (ML-PK/PD/TD) approach optimizes combination therapy using human PK/TD data along with in vitro PD data. This study utilized human population PK (PopPK) of aztreonam, ceftazidime/avibactam, and polymyxin B along with in vitro PDs from the Hollow Fiber Infection Model (HFIM) to derive optimal multi-drug regimens de novo through implementation of a genetic algorithm (GA). The mechanism-based PD model was constructed based on 7-day HFIM experiments across 4 clinical, extensively drug resistant Klebsiella pneumoniae isolates. GA-led optimization was performed using 13 different fitness functions to compare the effects of different efficacy (60%, 70%, 80%, or 90% of simulated subjects achieving bacterial counts of 102 CFU/mL) and toxicity (66% of simulated subjects having a target polymyxin B area under the concentration-time curve [AUC] of 100 mg·h/L and aztreonam AUC of 1,332 mg·h/L) on the optimized regimen. All regimens, except those most heavily weighted for toxicity prevention, were able to achieve the target efficacy threshold (102 CFU/mL). Overall, GA-based regimen optimization using preclinical data from animal-sparing in vitro studies and human PopPK produced clinically relevant dosage regimens similar to those developed empirically over many years for all three antibiotics. Taken together, these data provide significant insight into new therapeutic approaches incorporating ML to regimen design and treatment of resistant bacterial infections.