Pulmonary infection after inhalation of Mycobacterium tuberculosis (Mtb) causes tuberculosis (TB). TB presents with lung granulomas - complex spheroidal structures composed of immune cells and bacteria. Granulomas often have centralized caseum (necrotic tissue) where mycobacteria are quarantined, complicating and prolonging multi-antibiotic regimens. Determining which antibiotic regimens are optimal for reducing treatment time and toxicity is a goal of recent TB eradication campaigns. Clinical trials are expensive and challenging, making it difficult to untangle which host-pathogen interactions drive heterogeneous infection and treatment outcomes observed both within and between hosts. To determine responses to antibiotic regimens, we simulate treatments in HostSim, our whole-host mechanistic, multi-scale computational model of Mtb-infection. HostSim tracks dynamics of pulmonary Mtb-infection over molecular, cellular, tissue, organ, and whole-host scales. We create a heterogenous virtual cohort, comprising distinct hosts, for virtual clinical trials. We represent drug treatments by newly-integrating pharmacokinetics / pharmacodynamics into HostSim, simulating treatment with commonly-prescribed TB antibiotic regimens (e.g., HRZE or BPaL). Our approach allows us to identify both (1) which hosts/granulomas improve with treatment, and (2) which mechanisms influence outcome heterogeneity. By tracking experimental and clinical measurements, we virtually recreate several drug rankings from literature. We find that many methods of ranking treatment efficacy are strongly influenced by the 'definition of improvement' used and, in some cases, the detection threshold of CFU. Our work suggests that a study's reported optimal treatment may depend on its experimental design, including initial disease state and bacterial burden measures, possibly explaining seemingly-contradictory findings from prior studies.