Napovedovanje topnosti spojin v vodi s protitočno nevronsko mrežo

Velkov, Katarina (author), Venko, Katja (mentor), Lukšič, Miha (comentor)

V magistrskem delu sem z uporabo protitočne nevronske mreže razvila napovedne modele za napovedovanje topnosti spojin v vodi. Napovedovanje topnosti spojin v vodi je pomemben korak v postopku razvoja novih zdravilnih učinkovin, na področju toksikologije in okoljske varnosti. Kljub pomembnosti pa ostaja topnost spojin v vodi fizikalno-kemijska količina, ki jo je težko natančno napovedovati. Z razvojem metod umetne inteligence in predvsem globokih nevronskih mrež se razvijajo izboljšani modeli v primerjavi s preteklimi. Kljub temu pa so vrednosti korena srednje kvadratne napake (RMSE) objavljenih modelov dokaj visoke. Gibljejo se v razponu 0,4 - 1,2 logaritmiranih enot topnosti v vodi (logS). Topnost v vodi se ponavadi podaja kot logS, kjer je $S$ numerična vrednost topnosti izražena v mol/L. V sklopu magistrske naloge sem razvila tri napovedne modele. Za namene modeliranja sem uporabila program za protitočno nevronsko mrežo \textit{CPANNatNIC software}. V dveh optimizacijskih korakih je bil uporabljen genetski algoritem. Pripravila sem kvaliteten podatkovni set kemijskih spojin, na podlagi podatkov o spojinah iz članka Sluga et al. 2020 ter pripadajočih eksperimentalnih vrednosti za topnost v vodi, ki so pridobljene iz podatkovne zbirke AqSolDB. Za opis strukturnih značilnosti molekul sem uporabila molekulske deskriptorje iz programov alvaDesc in RDKit. Za vrednotenje uspešnosti razvitih modelov sem izračunala vrednosti RMSE na dveh zunanjih validacijskih setih, na validacijskem setu AqSol (na podlagi AqSolDB podatkovne zbirke) in na Zunanjem validacijskem setu (zbirka spojin, obljavljena v Cui et al. 2020). Za te validacijske sete sem pripravila tudi napovedi z že objavljenimi modeli - model VEGA, program ALOGPS in modeloma NN-A in NN-D, obljavljenima v Sluga et al. 2020. Opravljena je bila analiza domene uporabnosti na obeh zunanjih validacijskih setih s pomočjo podatkov o evklidskih razdaljah. Pripravila sem tudi napovedi z modelom soglasja, ki združuje tri modele (imenovan K3) in tudi model soglasja K5, ki združuje tri razvite modele in modela NN-A in NN-D. Ugotovljeno je bilo, da se trije modeli, razviti v magistrski nalogi, izkažejo dobro, v primerjavi z že objavljenimi modeli. Najbolje pa se izkaže model soglasja K3, ki je, glede na vrednosti RMSE na Zunanjem validacijskem setu najuspešnejši, glede na Validacijski set AqSol pa drugi najuspešnejši model.

• Language: Slovenian
• Keywords: Protitočna nevronska mreža, topnost v vodi, strojno učenje, modeli QSPR, molekulski deskriptorji
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FKKT - Faculty of Chemistry and Chemical Technology
• Year: 2025
• PID: 20.500.12556/RUL-171937
• COBISS.SI-ID: 253674499
• Publication date in RUL: 04.09.2025

Secondary language

• Language: English
• Title: Predicting aqueous solubility of compounds using counter-propagation neural network

Abstract

In this master thesis predictive models were developed using counter propagation neural networks for prediction of aqueous solubility of compounds. Predicting aqueous solubility is an important step in the process of drug discovery and development of new active compounds, in the fields of toxicology and environmental safety. Despite its importance, aqueous solubility remains a physicochemical quantity that is difficult to predict accurately. With the development of artificial intelligence methods and especially deep neural networks, improved models are being developed. However, the values of the root mean square error (RMSE) of currently published models remain high, ranging from 0.4 to 1.2 logarithmic units of aqueous solubility (logS). Aqueous solubility is commonly denoted as logS, where $S$ is the numerical value of solubility expressed in mol/L. As part of my master's thesis, I developed three predictive models. For modeling purposes, I used the \textit{CPANNatNIC software} for counter propagation networks. A genetic algorithm was used in two optimization steps. I prepared a high-quality dataset of chemical compounds, based on the data on compounds from the article Sluga et al. 2020 and the corresponding experimental values for aqueous solubility (from the AqSolDB database). To describe the structural features of the molecules, I used molecular descriptors from the programs alvaDesc and RDKit. To evaluate the performance of the developed models, RMSE values were calculated on two external validation sets, the AqSol validation set, which is based on the AqSolDB database, and the External Validation Set (compounds published in Cui et al. 2020). For these validation sets, I prepared predictions with previously published models - the VEGA model, the ALOGPS program and the NN-A and NN-D models published in Sluga et al. 2020. Analysis of the applicability domain of models was performed on both external validation sets using Euclidean distance data. I generated predictions using a consensus model that combines the three models developed in this thesis (called K3) and also a consensus model K5 that combines the three developed models, the NN-A and NN-D models. It was found that the three developed models perform well compared to previously published models. However, the consensus model K3 performs the best. K3 is the most successful model according to the RMSE values on the External Validation Set and the second most successful model according to the AqSol validation Set.

• Keywords: Counter-propagation neural network, aqueous solubility, machine learning, QSPR models, molecular descriptors