kinCSM-RTK: Machine Learning-Based Screening of Receptor Tyrosine Kinase Inhibitors in Drug Discovery

Receptor tyrosine kinases (RTKs) are key regulators of
cellular
functions, such as differentiation, migration and proliferation. Dysregulated
RTK activity contributes to various diseases, including neurological
disorders and cancer, for which small molecule inhibitors are often
used as therapeutic drugs to manage conditions involving constitutively
active RTKs. Preclinical development of these inhibitors, unfortunately,
often faces uncertainty, high costs, and time constraints, primarily
due to the extensive in vitro and in vivo testing of numerous chemical
compounds. To tackle the challenge of shortlisting potential drug
candidates, we developed kinCSM-RTK, which incorporates two machine
learning models to predict small molecule pKi and pIC50 values, estimating the inhibitory potency
of small molecules against RTKs. Our proposed machine learning models
have shown a robust and generalizable predictive performance. Specifically,
the models achieved Pearson’s correlation coefficients of 0.773
and 0.762 for pKi prediction, under 10-fold
cross-validation (CV) and an independent blind test, respectively.
Similarly, for pIC50 prediction, the models yielded coefficients
of 0.773 in 10-fold CV and 0.768 in the independent blind test. In
addition, we aimed to understand these results through post hoc explanation
analyses. In our explanation analyses, we observed that the proximity
and quantity of aromatic interactions have correlated with stronger
RTK inhibition, providing important insights for drug design targeting
RTKs. Accordingly, we made our models publicly accessible on a user-friendly
web server at https://biosig.lab.uq.edu.au/kincsm_rtk/.