The target, Janus Kinase (JAK), is known for four types of isozyme: Janus kinase 1(JAK1), Janus kinase 2(JAK2), Janus kinase 3(JAK3), and Tyrosine kinase 2(TYK2), each of which combines variously to control the downstream of cytokines, which play a pi...
The target, Janus Kinase (JAK), is known for four types of isozyme: Janus kinase 1(JAK1), Janus kinase 2(JAK2), Janus kinase 3(JAK3), and Tyrosine kinase 2(TYK2), each of which combines variously to control the downstream of cytokines, which play a pivotal role in essential cell functions such as cell proliferation, erosion and survival.
In this study, we build a model that finds inhibitors of Janus kinase through quantitative structural activity relationship (QSAR) of compound-based drug design methods with machine learning and deep learning using data on JAK. We used bayesian models in machine learning and artificial neural network (ANN) models in deep learning.
3D descriptors were used to exploit the chiral structure in JAK, indicating that the accuracy was significantly reduced. While thinking about how to supplement this, we referenced to look at a paper that complemented the model using a convolutional neural network (CNN) to examine whether this SMILES code can be used as a kind of expression by converting it to a number.
Before converting SMILES code to numbers, we calculated the 2D descriptor, 3D descriptor, and SMILES code for the compounds with different active values at the same target to ensure that the chiral structure can be distinguished. As a result, we find that there was a difference between the 3D representative and SMILES code, and that these two can be used as descriptors.
As SMILES codes have been added as descriptor, the accuracy has been increased through training, test sets, and external datasets.
Descriptor identified in this paper could be applied as complementary indicators for new drug development to other targets.