Develepment and application of models for ecotoxicological risk assesment of bioactive chemical compounds

U disertaciji je razvijen novi računalni postupak procjene ekotoksikološkog rizika vodenih staništa uzrokovan bioaktivnim kemijskim onečišćivalima sastavljen od tri cjeline: utvrđivanja stanja, određivanja rizika od postojećim metodama, te razvoja i primjene novih QSAR modela za procjenu rizika. Utvrđeno je da su vodena staništa rijeke Save opterećena kemijskim onečišćivalima od rastućeg značaja za okoliš. S obzirom na nedostatke postojećih QSAR metoda za procjenu rizika razvijenih na malom broju spojeva, novi robusniji QSAR modeli razvijeni su na nekoliko tisuća kemikalija iz baza ToxCast i Tox21 s brojnim i osjetljivijim toksičnim učincima izmjerenim na staničnim linijama i embrijima zebrice. Novi modeli temeljeni su na molekularnim deskriptorima i strukturnim otiscima, a dobiveni su logističkom regresijom, neuronskim mrežama i slučajnim šumama koje su dale ponajbolje modele. Dobiveni modeli pokazuju dobru generalizaciju na vanjskim skupovima onečišćivala, a njihova kvaliteta provjerena je novouvedenim metrikama. Procjene ekotoksikološkog rizika od spojeva utvrđenih u rijeci Savi temeljena na novim modelima toksičnosti pokazuju dobro slaganje s postojećim metodama, uz značajno povećanje kemijskog strukturnog prostora kojeg pokrivaju. A novel computational approach for assessing the ecotoxicological risk for water habitats caused by bioactive chemical pollutants is developed in this work. It consists of three parts: the determination of current ectoxicological status, the risk assessment based on existing methods, and the development and application of new QSAR models for risk assessment. Water habitats of the Sava River were found to be burdened with chemical pollutants of growing importance to the environment. Given the shortcomings of existing QSAR risk assessment methods developed on a small number of compounds, new, more robust QSAR models were developed herein on several thousand chemical compounds from the ToxCast and Tox21 databases which were assessed on more sensitive toxic endpoints measured on zebrafish embryos and cell lines. The newly developed models are based on molecular descriptors and structural fingerprints and are obtained by logistic regression, neural networks and random forests which gave the best models. The trained models show good generalization on external sets of chemical compounds. Their quality is validated by newly introduced metrics. Ecotoxicological risk assessment of compounds identified in the Sava River, is based on the new toxicity models which agree well with the existing methods, supported by a significant increase of structural space covered.