Predicting permeability

HANGZHOU, China—Caco-2 assays have become a gold standard for assessing oral absorption of potential drugs. Permeability prediction algorithms, however, offer scientists an opportunity to screen many more compounds without expending a lot of resources. With this in mind, researchers at Zhejiang University developed prediction models based on open-source software, looking to correlate permeability with molecular descriptors.

As they presented in the Journal of Pharmacy and Pharmaceutical Sciences, the researchers developed two models—one based on multiple linear regression (MLR), the other on support vector machine (SVM)—using molecular descriptors developed from the open-source system Chemistry Development Kit (CDK). For both models, they found that the hydrogen-bond donors and charged polar surface area descriptors were keys to determining permeability of a compound.

The researchers determined that when they tested the MLR model, it did not perform as well as they had hoped, exhibiting a correlation coefficient (r) of 0.70 for the test set. The SVM model performed much better, however, offering an r value of 0.85 for the test set. They were quick to point out that their models were limited by a small data set and that improvement is possible, but by developing an open-source system, they are confident that others can build on their initial efforts.