Strojno učenje kemijskih reakcij proteinov v interakciji z RNA

HENIGMAN, JERNEJ (author), Curk, Tomaž (mentor)

V diplomski nalogi smo uporabili metode strojnega učenja za klasifikacijo kemijskih reakcij. Hkrati smo identificirali najpomembnejše strukturne spremembe molekul, ki nastajajo v kemijskih reakcijah proteinov, ki sicer vstopajo tudi v interakcijo z RNA. V prvem delu smo na podlagi šestih osnovnih encimatskih skupin kemijskih reakcij, določili optimalen nabor parametrov modeliranja. Testirali smo tri skupine parametrov: metode za uravnoteževanje učne množice (sedem metod), metode za opisovanje kemijskih profilov (sedem vrst opisov) in metode za izgradnjo napovednih modelov (pet metod). Za najboljši nabor smo izbrali kombinacijo tistih parametrov, pri kateri je mera AUC najvišja. Empirično smo pokazali, da najboljši nabor sestavlja kombinacija parametrov: uravnoteževalna metoda naključnega podvzorčenja, združena kemijska profila Morgan in MorganBitVector, napovedni model naključnega gozda. Dosežen povprečen AUC na šestih osnovnih skupinah kemijskih reakcij je znašal 0,97. V drugem delu smo uporabili predhodno določen nabor parametrov za modeliranje skupine kemijskih reakcij proteinov v interakciji z RNA, pri čemer smo dosegli AUC 0,77.

• Language: Slovenian
• Keywords: strojno učenje, kemijske reakcije, uravnoteževalne metode, kemijski opisi molekul, napovedni modeli, AUC, RNA, struktura molekule
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2015
• PID: 20.500.12556/RUL-72049
• Publication date in RUL: 20.08.2015

Secondary language

• Language: English
• Title: Modeling chemical reactions of RNA-binding proteins with machine learning

Abstract

In this thesis machine learning methods are used to classify chemical reactions. At the same time the most important changes in molecular structure are identified that are typical for chemical reactions of RNA-binding proteins. In the first part, six basic groups of chemical reactions were used to determine the optimal set of parameters for modeling and prediction. Three groups of parameter sets were tested: methods for balancing the learning set (seven methods), methods for molecular fingerprinting (seven methods) and predictive models (five methods). Empirically is shown that the best combination consists of the following parameters: random undersampling as balancing method, Morgan+MorganBitVector for molecular fingerprinting and random forest as predictive model, with which average AUC 0.97 was achieved. For the second part, the optimal set of parameters is used to discriminate between chemical reactions associated with RNA-binding proteins and those chemical reactions associated with non RNA-binding proteins. AUC score 0.77 was achieved.

• Keywords: machine learning, chemical reactions, balancing methods, molecular fingerprints, predictive models, AUC, RNA, molecular structure