prikaz prve stranice dokumenta Binary classification of peptides using deep neural networks and transfer learning
  Download
PDF 408.63 KB
doctoral thesis
Binary classification of peptides using deep neural networks and transfer learning
2024. urn:nbn:hr:190:986491
Otović, Erik
University of Rijeka
Faculty of Engineering

Cite this document

Otović, E. (2024). Binary classification of peptides using deep neural networks and transfer learning (Doctoral thesis). Rijeka: University of Rijeka, Faculty of Engineering. Retrieved from https://urn.nsk.hr/urn:nbn:hr:190:986491

Otović, Erik. "Binary classification of peptides using deep neural networks and transfer learning." Doctoral thesis, University of Rijeka, Faculty of Engineering, 2024. https://urn.nsk.hr/urn:nbn:hr:190:986491

Otović, Erik. "Binary classification of peptides using deep neural networks and transfer learning." Doctoral thesis, University of Rijeka, Faculty of Engineering, 2024. https://urn.nsk.hr/urn:nbn:hr:190:986491

Otović, E. (2024). 'Binary classification of peptides using deep neural networks and transfer learning', Doctoral thesis, University of Rijeka, Faculty of Engineering, accessed 17 December 2024, https://urn.nsk.hr/urn:nbn:hr:190:986491

Otović E. Binary classification of peptides using deep neural networks and transfer learning [Doctoral thesis]. Rijeka: University of Rijeka, Faculty of Engineering; 2024 [cited 2024 December 17] Available at: https://urn.nsk.hr/urn:nbn:hr:190:986491

E. Otović, "Binary classification of peptides using deep neural networks and transfer learning", Doctoral thesis, University of Rijeka, Faculty of Engineering, Rijeka, 2024. Available at: https://urn.nsk.hr/urn:nbn:hr:190:986491

Please login to the repository to save this object to your list.
Metadata
TitleBinary classification of peptides using deep neural networks and transfer learning
Title (english)Binarna klasifikacija peptida korištenjem dubokih neuronskih mreža i učenja prijenosom znanja
AuthorErik Otović
MentorGoran Mauša (mentor)
MentorDaniela Kalafatović (komentor)
Committee memberKristijan Lenac (predsjednik povjerenstva)
Committee memberMarina Ivašić-Kos (član povjerenstva)
Committee memberTell Tuttle https://orcid.org/0000-0003-2300-8921 (član povjerenstva)
GranterUniversity of Rijeka
Faculty of Engineering
Rijeka
Defense date and country2024, Croatia
Scientific / art field,
discipline and subdiscipline		TECHNICAL SCIENCES
Computing
Universal decimal classification
(UDC)		004 - Computer science and technology. Computing. Data processing
Abstract
Machine learning is increasingly used for high-throughput peptide screening, providing a rapid and efficient method to identify peptides with desired functions in contrast to traditional trial-and-error approaches that are time-consuming and resource-intensive. It streamlines the exploration of the vast peptide space in a data-driven way and accelerates the discovery of novel peptides. This thesis investigates three dominantly used peptide representation schemes and analyzes them based on the type of information they capture. Considering that machine learning models require input to be in a numerical form, the choice of peptide representation scheme is crucial as it can directly influence model performance. Therefore, anovel sequential properties representation scheme is proposed to address gaps identified in existing schemes. Additionally, a manually curated dataset comprising 126 peptides evaluated for the catalysis of ester and phosphoester hydrolysis is presented. The experimental evaluation of four peptide representation schemes in combination with deep neural networks was conducted using antimicrobial, antiviral and catalytic datasets. Results on the antimicrobial and antiviral datasets were used for statistical tests and to draw reliable conclusions due to their diversity and size. Statistical tests applied across seven evaluation metrics demonstrated that the introduced sequential properties scheme significantly outperformed other representations in 90% of cases. The antimicrobial and antiviral datasets were downsampled to create smaller target datasets to assess the effectiveness of transfer learning. Results showed that knowledge transfer was beneficial only when transferring from the more diverse antimicrobial dataset encompassing multiple subfunctions to less diverse antiviral dataset, improving the ROC-AUC score by 6.9% with statistical significance. Moreover, the results show that the transfer learning model outperforms the baseline model by more than 1% when the target dataset contains fewer than 275 peptides.
Abstract (english)
Strojno učenje se sve više koristi za visoko-propusno pregled peptida, pružajući brzu i učinkovitu metodu za identifikaciju peptida sa željenim funkcijama za razliku od tradicionalnih pristupa baziranih na postupku pokušaja i pogreške koji zahtijevaju puno vremena i resursa. Takoder pojednostavljuje istraživanje velikog prostora peptida metodama strojnog učenja koje su vođene dostupnim podacima te ubrzava otkrivanje novih peptida. Ova doktorska disertacija razmatra tri dominantno korištene sheme predstavljanja peptida i analizira ih na temelju vrsta informacija koje obuhvaćaju. Uzimajući u obzir da modeli strojnog učenja zahtijevaju ulazne podatke u numeričkom obliku, odabir sheme predstavljanja peptida je ključan odabir jer može imati izravni utjecaj na učinak modela. Stoga, nova shema predstavljanja nazvana slijedne značajke je predložena u ovoj disertaciji s ciljem premošćivanja identificiranih nedostatak u postojećim shemama. Dodatno, u disertaciji je predstavljen skup podataka koji se sastoji od 126 ručno prikupljenih peptida ispitanih za katalizu hidrolize estera i fosfoestera. Provedeno je eksperimentalno ispitivanje četiri sheme predstavljanja peptida u kombinaciji s dubokim neuronskim mrežama korištenjem antimikrobnih, antivirusnih i katalitičkih skupova podataka. Rezultati na antimikrobnom i antivirusnom skupu podataka korišteni su za statističke testove i donošenje pouzdanih zaključaka zbog raznolikosti i veličine tih skupova podataka. Statistički testovi primijenjeni na sedam metrika vrednovanja pokazali su da je predstavljena shema slijednih značajki statistički značajno nadmašila ostale sheme u 90% slučajeva. Smanjene inačice antimikrobnog i antivirusnog skupa podataka korištene su kao odredišni skupovi podataka za ispitivanje učinka učenjem prijenosom znanja. Rezultati pokazuju da je poboljšanje ostvareno samo pri prijenosu znanja iz raznovrsnijeg antimikrobnog skupa podataka, koji obuhvaća nekoliko podfunkcija na antivirusni skup podataka, na manje raznoliki antivirusni skup rezultirajući u statistički Classification of Peptides using Deep Neural Networks and Transfer Learning VIII signifikantnom povećanju ROC-AUC metrike za 6.9%. Povrh toga, rezultati su pokazali da model baziran na prijenosu znanja ostvaruje ičinak veći od 1% u usporedbi sa modelom koji nije koristio prijenos znanja kada ciljni skup podataka sadrži manje od 275 peptida.
Keywords
Keywords (english)
Languagecroatian
URN:NBNurn:nbn:hr:190:986491
Promotion2024-12-09
Study programmeTitle: Postgraduate University Doctoral Study in the area of Engineering Sciences, in the field of Computer Science
Study programme type: university
Study level: postgraduate
Academic / professional title: doktor/doktorica znanosti, područje tehničkih znanosti, polje računarstvo (doktor/doktorica znanosti, područje tehničkih znanosti, polje računarstvo)
Type of resourceText
File originBorn digital
Access conditionsOpen access
Terms of useLicence In copyright icon
Created on2024-12-16 14:06:52