ARTIFICIAL INTELLIGENCE FOR PREDICTIVE ECOTOXICITY ASSESSMENT IN ARTEMIA SALINA

Resumo

The fresh water crustacean Artemia salina (brine shrimp) widelyused as a test organism in aquatic toxicology, being a filter crustacean of thalassic ecosystems, which feeds mainly on bacteria, unicellular algae, small protozoa and debris that are dissolved in the medium, which makes it widely used as a bioindicator of toxicity due to the high sensitivity in increasing particle concentration. The main objective of this study was to build robust and predictive machine learning models for predicting ecotoxicity of organic compounds in A. salina. All in silico steps developed in this study were performed in Python v.3.6 (https://www.python.org). The framework used is complete, such as the module for data preparation, data set balancing, construction and validation of machine learning models and machine interpretation. At the end of the balanced process, a dataset with classes containing the same weight, 232 toxic compounds, and 232 non-toxic compounds was obtained. The combination of Morgan and FeatMorgan fingerprints (radius 2: FeatMorgan_2, Morgan_2; radius 4: FeatMorgan_4, Morgan_4, radius 6: FeatMorgan_6, Morgan_6) with the Random Forest method led to highly predictive models with correct classification rate values. (RCC) ranging from 0.81–0.82, sensitivity (SE) of 0.78–0.81, specificity (SP) of 0.82–0.86, positive predictive value (PPV) of 0.82– 0.85, negative predictive value (NPV) of 0.79–0.82, Cohen's kappa coefficient (κ) of 0.63–0.65 and a coverage of 0.57–0.66. The increase in radius (ie parameter that defines the size of the molecular fragments computed during the calculation of descriptors) had a negative influence on the accuracy rate in the prediction of toxic compounds (2% reduction in SEs compared to the best model) and positive influence. on the classification rate of non-toxic compounds (~ 2% in SPs). The combination of Morgan and Fea fingerprints tMorgan (radius 2: FeatMorgan_2, Morgan_2; radius 4: FeatMorgan_4, Morgan_4, radius 6: FeatMorgan_6, Morgan_6) with the SVM method also led to predictive QSTR models, with (CCR) values ranging from 0.79–0, 81, (SE) 0.81–0.85, (SP) 0.76–0.80, (PPV) 0.77–0.80, NPV value of 0.80–0, 84, (κ) of 0.59–0.62 and a coverage of 0.57–0.66. Through this approach it was possible to identify molecules and fragments with positive contribution to toxicity in A. salina (structural alerts) such as: 5-nitrofuran (1), ethylenediamine (3), quinone (9) and 1,2-dichlorobenzene (11). Considering the potential of these molecules and fragments for toxicity, it is highly recommended that these substructures be avoided during the development process of new pesticides or have their use denied if they are being registered with environmental regulatory agencies. The models developed were robust and predictive. In the first phase of the study, the original data set was rigorously prepared to prevent experimental and compilation errors from being perpetuated and/or interfering with the predictability of the models. Increasing the radius of Morgan and FeatMorgan fingerprints is observed to decrease sensitivity and improve specificity.