Prediction of drug permeation through microneedled skin by machine learning

Abstract

Stratum corneum is the outermost layer of the skin preventing external substances from entering human body. Microneedles (MNs) are sharp protrusions of a few hundred microns in length, which can penetrate the stratum corneum to facilitate drug permeation through skin. To determine the amount of drug delivered through skin, in vitro drug permeation testing is commonly used, but the testing is costly and time-consuming. To address this issue, machine learning methods were employed to predict drug permeation through the skin, circumventing the need of conducting skin permeation experiments. By comparing the experimental data and simulated results, it was found extreme gradient boosting (XGBoost) was the best among the four simulation methods. It was also found that drug loading, permeation time, and MN surface area were critical parameters in the models. In conclusion, machine learning is useful to predict drug permeation profiles for MN-facilitated transdermal drug delivery.

Keywords: XGBoost; machine learning; microneedle; multiple linear regression; random forest; transdermal.

FIGURE 1

Four methods were used to predict drug permeation through skin: (a) Fick's law; (b) multiple linear regression; (c) random forest; and (d) XGBoost.

FIGURE 2

A flowchart showing the model training and prediction process of all four models: Fick's law, MLR, RF, and XGBoost. MLR, multiple linear regression; RF, random forest.

FIGURE 3

The data distribution and molecular structure of the drugs delivered using MNs in the skin permeation study., , , MNs, microneedles.

FIGURE 4

(a) An illustration of the Fick's law prediction curve compared with the experimental data. The drug distribution in the MN‐skin system at (b) 15 min, (c) 1 h, (d) 3 h, (e) 6 h, and (f) 24 h. MN, microneedle.

FIGURE 5

Effect of different parameters in Fick's law simulation: (a) diffusion coefficient; (b) number of MN; (c) MN length, and (d) mass of loaded drugs. MN, microneedle.

FIGURE 6

The comparison of the predicted permeation results with experimental data: (a) BSA(R); (b) GHK(H); (c) GHK(R); (d) rhodamine B(R); (e) lidocaine(H); (f) lidocaine(R); (g) caffeine(H); (h) Cu(R); and (i) Cu(H) (R: rat skin; H: human skin). BSA, bovine serum albumin; GHK, glycyl‐l‐histidyl‐l‐lysine.

FIGURE 7

The comparison of the predicted drug permeation percentage with experimental data: (a) BSA; (b) GHK(H); (c) GHK(R); (d) rhodamine B(R); (e) lidocaine(H); (f) lidocaine(R); (g) caffeine(H); (h) Cu(R); and (i) Cu(H) (R: rat skin and H: human skin). BSA, bovine serum albumin; GHK, glycyl‐l‐histidyl‐l‐lysine.

FIGURE 8

The feature significance in the RF and XGBoost models: (a) permeation percentage (RF); (b) permeation amount (RF); (c) permeation percentage (XGBoost); (d) permeation amount (XGBoost). RF, random forest.