Knowledge graph-primed deep learning to identify condition-specific gene importance

Kert, Aleš (author), Curk, Tomaž (mentor), Robyn Bleker, Carissa (comentor), Zrimec, Jan (comentor)

In this thesis, we sought to incorporate prior knowledge to improve the interpretability of deep learning models trained on gene expression data. We created a pipeline consisting of batch effect removal using a deep learning approach, training the prediction model, and the interpretation of the model using guided backpropagation. The prediction model architecture was con- structed using prior knowledge on molecular interactions. In genes relevant to tissue types and perturbation groups, the baseline achieved an AUC score of 0.629 and 0.597, respectively. The proposed CKN-based models achieved 18.6% and 23.1% relative improvement, with AUC scores of 0.746 and 0.735, respectively.

• Language: English
• Keywords: neural networks, gene expression, interaction networks, knowledge network, prior knowledge
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-164587
• Publication date in RUL: 04.11.2024

Secondary language

• Language: Slovenian
• Title: Na grafu znanja osnovano globoko učenje za identifikacijo pogojno-specifične pomembnosti genov

Abstract

Cilj naloge je bila integracija biološkega predznanja z namenom izboljšave razložljivosti modelov globokega učenja, naučenih na podatkih o genskem izražanju. Razvili smo računski cevovod, sestavljen iz odstranjevanja odstopanja serije, učenja napovedovalnega modela in interpretacije tega modela. Arhitektura napovedovalnega modela je bila zasnovana na podlagi predznanja o molekulskih interakcijah. Razložljivost modela je bila primerjana z razložljivostjo modela, naučenega brez predznanja. Neinformirani modeli so pri genih, pomembnih za tkiva in stresorje, dosegli oceni AUC 0,629 in 0,597, v tem vrstnem redu. Predlagani modeli, ki uporabljajo predznanje, pa so dosegli 18,6 % ter 23,1 % relativno izboljšavo rezultatov, z ocenama AUC 0,746 in 0,735.

• Keywords: nevronske mreže, ekspresija genov, omrežja interakcij, mreža znanja, predznanje