Abstract:
Application of nanoparticles (NPs) in many different fields brought several benefits,
especially in biomedicine, physics, chemistry, and agriculture. However, nanoparticles
can exhibit toxic effects due to their very high surface-to-volume ratio causing harm
to biological systems and to their respective ecosystems. Nanotoxicity testing is an
important phase to determine the potential risks that NPs may bring. On the other
hand, the process to perform experimental assays often requires quite a lot of time and
resources. An alternative way to perform nanotoxicity testing is through in-silico testing.
In-silico methods are usually centered around the quantitative structure-activity
relationship (QSAR) modeling, however in this work, we refined the predictive capacity
of QSAR modeling by integrating the Perturbation Theory, which was a recent
novel method of testing nanotoxicity. Using different machine learning algorithms,
this work developed a QSAR-perturbation model to predict toxicity profiles of NPs
under diverse experimental conditions. The models were developed from a dataset of
5,437 NP-NP pairs derived from applying perturbation theory to 260 unique NPs. In
the results, XGBoost was the top performing model with 98.43% MCC value. This is
comparable to previous results in existing literature. The QSAR-PT model was then
employed in a web application as a final output.
