Computational prediction of immune cell cytotoxicity

- A computational prediction model for immune cell toxicity was developed.
- Training set: growth inhibition data on B- and T-cells for about 40 000 chemicals.
- Molecular fingerprints as descriptors in a Naïve Bayes Classifier.
- Validated using four independent data sets from literature.
- Model can be used as primary screen for immunotoxicity.

Immunotoxicity, defined as adverse effects of xenobiotics on the immune system, is gaining increasing attention in the approval process of industrial chemicals and drugs. In-vivo and ex-vivo experiments have been the gold standard in immunotoxicity assessment so far, so the development of in-vitro and in-silico alternatives is an important issue.In this paper we describe a widely applicable, easy-to use computational approach which can serve as an initial immunotoxicity screen of new chemical entities. Molecular fingerprints describing chemical structure were used as parameters in a machine-learning approach based on the Naïve-Bayes learning algorithm.The model was trained using blood-cell growth inhibition data from the NCI database and validated externally with several in-house and literature-derived data sets tested in cytotoxicity assays on different types on immune cells. Both cross-validations and external validations resulted in areas under the receiver operator curves (ROC/AUC) of 75% or higher.The classification of the validation data sets occurred with excellent specificities and fair to excellent selectivities, depending on the data set. This means that the probability of actual immunotoxicity is very high for compounds classified as immunotoxic, while the fraction of false negative predictions might vary. Thus, in a multistep immunotoxicity screening scheme, the classification as immunotoxic can be accepted without additional confirmation, while compounds classified as not immunotoxic will have to be subjected to further investigation.