Institutional Repository of Vilnius University
Gilieji neuroniniai tinklai ir jų taikymas /
Abstract
dc:descriptionThis bachelor's thesis examines deep neural networks and their application in a bioinformatics classification task. The theoretical part of the thesis discusses the concept of artificial and deep neural networks, their main operating principles, neural network layers, activation and loss functions, the problem of overfitting and methods for reducing it. The neural network architectures relevant to this work are also presented: multilayer perceptron, convolutional neural network, embedding layer and multimodal neural networks. In the practical part of the thesis, the application of deep neural networks is examined in a protein sequence classification task. The aim of the work is to develop and evaluate models for classifying proteins into proteins with an EC number and proteins without an EC number after additional annotation-based filtering. For this purpose, a balanced UniProtKB/Swiss-Prot protein sequence dataset was constructed, and sequence cleaning, standardisation, duplicate removal, class balancing, and MMseqs2 cluster-based splitting into training, validation, and test sets were performed. Several models were developed and compared in this work. A convolutional neural network was used to analyse amino acid sequences, a multilayer perceptron was used for biochemical features calculated from the sequences, and a multimodal CNN and MLP model was used to combine both types of information. In addition, a Random Forest model was used as a classical machine learning baseline method. Model quality was evaluated using accuracy, F1 score, ROC-AUC, Matthews correlation coefficient and confusion matrices. In an additional experiment, the use of AlphaFold-predicted protein structures was evaluated. Structural information was converted into C-alpha atom distance matrices, which were used as input to a convolutional neural network. This was done to examine whether the spatial structure of a protein can provide additional information for the classification task. The results of the main experiment showed that deep neural networks can be applied to protein sequence classification, but their performance depends on the chosen way of representing the data. Among the deep neural network models, the MLP model based on biochemical features performed best, while in the overall comparison of the main experiment the Random Forest model achieved the best result. In the additional AlphaFold experiment, the CNN model based on distance matrices achieved the best result, indicating that structural information can be useful for additional classification evaluation.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Černyškovaitė, Laura,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308336599&prefLang=en_US