Figure 1
Overview of our multiscale pipeline to predict in vivo drug interactions. (a) INDIGO-MTB uses Mtb transcriptomic responses to drugs and experimentally measured drug–drug interactions as inputs for training a machine-learning model, inferring synergistic and antagonistic interactions between new drug combinations as output8,27. (b) Components of the model integrating GranSim and INDIGO-MTB. From right to left, the plasma PK model determines the time-dependent concentration of all antibiotics following oral doses, which in turn determines the amount of antibiotic delivered onto the agent-based model grid. The computational grid is 200 × 200 square grid spaces, representing 4 mm × 4 mm of lung. Within the agent-based model, the tissue PK model describes antibiotic diffusion and binding as well as immune cell accumulation. Based on the local concentration of antibiotics, the PD model evaluates an antibiotic killing rate constant based on an effective concentration that is calculated from each individual antibiotic concentration. The corresponding FIC predicted from INDIGO-MTB either increases or decreases this effective concentration, depending on whether the combination is synergistic or antagonistic. (c) Different predictions and outcomes, with the gradient above corresponding to the relevant length scale for the model/prediction. From left to right, predictions made by integration of GranSim and INDIGO-MTB are shown, including FIC predictions from INDIGO-MTB, Mtb-specific killing rate and interactions, number of cells/Mtb overtime used to evaluate simulated EBA, spatial analysis of antibiotic concentration and interactions, and sterilization time distributions from a collection of granulomas.
