Sieci Bayesa w kontekście przewidywania toksyczności związków chemicznych

Abstract

Bayesian Belief Networks (BBN) are a graph structure for representing the relationships between descriptors and a target variable in a given data set. This study investigated the use of this algorithm in modeling the Quantitative Structure- Activity Relationship (QSAR) of chemical compounds using toxicity assessment as an example. Selected tests from a database containing toxicity information of 12,000 compounds for 12 different toxicity tests were used. The influence of parameters describing the representation of the compound was checked. In addition, toxicity prediction models were created for selected tests and a model predicting the outcome of several tests simultaneously (multi-target). The results indicate that the obtained models have relatively good accuracy and precision. Using multi-target models based on the BBN algorithm can lead to better predictions than models for predicting a single target variable.