Figure 1.
Diagram of the proposed machine learning procedure.
The top part of the diagram shows the training procedure, which contains the clustering module and temporal analysis module. The bottom part shows the testing procedure.
Figure 2.
Example of second order PSA built from the state sequence, .
This state transition diagram shows the probability of transitioning between states. The sum of all possible states transitions equals one, indicating that a state transition must occur.
Table 1.
Performance metrics.
Figure 3.
Representative Giardia killing curves.
Killing curves generated from the average of three replicates per dose. The cell counts were compared to the control, which is typical of killing curves. Note that the general trend is towards an exponential decrease over the time interval of the study, 18 hours. In the lowest dose tested, 1.9 µg/ml, an obvious plateau is reached at 18 hours indicating an ineffective dosing, while at the other doses, the trend indicates a decline.
Table 2.
Cell number over time for each dose prior to pre-processing.
Figure 4.
An example of 50 µg/ml trial and its states.
The separated boxes indicate different clusters/states. The width of the boxes indicates the time duration of each state. The dotted curve represents the drug concentration over time, while the solid curve represents the percent change of the pathogen over time.
Figure 5.
The Markov model state transition diagram built from the 15 effective drug delivery trials.
The “start” and “end” states are added for illustrative purposes. In the effective delivery strategy, it is possible to transition between three states. From the high drug state it is only possible to remain in that state, or transition to the medium drug state. Similarly, once in the medium drug state it is not possible to transition back to the high drug state, it is only possible to remain in that state or transition to the low drug state. Once in the low drug state, the system will remain in the state for various iterations before finally ending.
Figure 6.
The Markov model state transition diagram built from an ineffective drug delivery.
The “start” and “end” states are added for illustrative purpose. In this example of ineffective delivery, the model has only two transition states. When the concentration of drug is low and the current population is below the starting population, the system is more likely to remain in this state for several iterations. Eventually, however, a transition out of this state will occur resulting in a low drug concentration and a larger current population. Once in this state it is impossible to leave this state, and eventually an end state will be reached.
Table 3.
The prediction performance of 3-fold cross-validation for all doses.
Table 4.
The average prediction accuracies of 3-fold cross-validation for each dose.
Table 5.
The prediction performance for in silico dosing.
Figure 7.
The ROC curve for determining prediction performance.
The ROC curve shows the tradeoff between sensitivity and specificity (any increase in sensitivity will be accompanied by a decrease in specificity). The closer the curve is to the minimum false alarm rate (x-axis) and the maximum sensitivity (y-axis), the more accurate the test. As the ROC curve approaches y = x, the less accurate the test becomes. The intersection point of the ROC curve with the line y = −x is defined as the optimum operation point. In this ROC curve, the optimum operation point had an 80% true positive rate, with a 20% false positive rate.
Table 6.
Thresholds for different trade-offs.