Consensus models to predict oral rat acute toxicity and validation on a dataset coming from the industrial context

Abstract

We report predictive models of acute oral systemic toxicity representing a follow-up of our previous work in the framework of the NICEATM project. It includes the update of original models through the addition of new data and an external validation of the models using a dataset relevant for the chemical industry context. A regression model for LD 50 and multi-class classification model for toxicity classes according to the Global Harmonized System categories were prepared. ISIDA descriptors were used to encode molecular structures. Machine learning algorithms included support vector machine (SVM), random forest (RF) and naïve Bayesian. Selected individual models were combined in consensus. The different datasets were compared using the generative topographic mapping approach. It appeared that the NICEATM datasets were lacking some relevant chemotypes for chemical industry. The new models trained on enlarged data sets have applicability domains (AD) sufficiently large to accommodate industrial compounds. The fraction of compounds inside the models' AD increased from 58% (NICEATM model) to 94% (new model). The increase of training sets improved models' prediction performance: RMSE values decreased from 0.56 to 0.47 and balanced accuracies increased from 0.69 to 0.71 for NICEATM and new models, respectively.

Visualization and Analysis of the REACH‐chemical Space with Generative Topographic Mapping

“…The REACH regulation requires its assessment even for small tonnages. Consequently, this experimental test is one of the most commonly performed animal tests which partly explains its much higher data availability compared to the other endpoints [3] …”

“…Experimental data was collected from multiple publicly available databases and scientific literature [2–4] . Among them, the main source was the database of the European Chemical Agency (ECHA), [14] which comprises the REACH‐registered substances.…”

“…The Global dataset results from the merging of all the available data on the abovementioned 11 endpoints: it counts 29433 experimentally measured data points for 17762 unique compounds (as one compound can have more than one associated experimental value). This dataset has been assembled and curated in our previous studies [2–4] . Table 1 reports the sizes of the endpoint subsets.…”

“…In the past years, several models predicting REACH‐relevant endpoints were obtained with the help of various machine‐learning methods like multiple linear regression, support vector machine or neural networks [2–5] . However, some of them suffered from absence of technical documentation complying with the REACH requirements, [6] for instance concerning the model's Applicability Domain (AD) or its validation procedure [3,7] . In our recent studies, [2,4] we found that for some endpoints currently‐existing tools have disappointing performances when applied to compounds coming from an industrial context, due to their restricted AD.…”

“…Here, a set of curated data for 11 endpoints prepared in our previous studies [2–4] (“Global dataset”) has been used to train the GTM model and to delineate the REACH chemical space. The following 11 endpoints were considered: bioconcentration factor, ready biodegradability, environmental persistence in sediment, soil and water media, acute aquatic toxicity to algae, daphnia and fish, rat acute toxicity, androgen and estrogen receptor binding potential.…”