Machine Learning Approaches for Gene Regulatory Network Inference Using Single-Cell RNA Sequencing Data - Master Thesis

Aug 5, 2026

Upload by Ezra

Gene Regulatory Networks (GRNs) explain molecular interactions driving biological processes, and can be modeled as directed gene graphs. GRN inference traditionally used bulk RNA-seq, but single-cell RNA-seq enables high-resolution study of cellular heterogeneity while adding technical noise and variability. This thesis evaluates both unsupervised methods (GENIE3, scGeneRAI) and supervised methods (GNNLink, STGRNS) for GRN inference from single-cell data, using consistent datasets for fair comparisons. Enhancements target GNNLink via TF frequency lookup tables and an unsupervised expression-only training strategy with gene filtering to improve relevance and comparability across ground-truth networks.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages97

Reading Time4 h 4 min

AccessFree to download

1 Introduction

1.1 Problem Overview

1.2 Objectives

2 Background

2.1 Biological Background

2.2 Graph Structure and GRN Inference Goal

2.3 GRN Inference with Single-Cell Data

2.4 Performance Metrics

3 Datasets

3.1 Datasets from Synthetic Networks

3.2 Datasets from Curated Models

3.3 Experimental Single-Cell RNA-Seq Datasets

4 Methods

4.1 GENIE3

4.2 ScGeneRAI

4.3 GNNLink

4.4 STGRNS

4.5 Contributions

5 Experimental Results

5.1 Chosen Evaluation Metrics

5.2 GRN Inference Methods Assessment

5.3 Variability in Ground-Truth GRNs

5.4 Training Datasets Derived from Expression Data

5.5 Impact of TF Frequency on GRN Inference

6 Conclusion

University of Padova Department of Mathematics “Tullio Levi-Civita” Master Thesis in Data Science Machine Learning Approaches for Gene Regulatory Network Inference Using Single-Cell RNA Sequencing Data Supervisor Master Candidate Prof. Nicolò Navarin Bogdana Živković University of Padova Co-supervisor Student ID Prof. Gabriele Sales 2105083 University of Padova Academic Year 2023-2024 ii iv Abstract Gene Regulatory Networks (GRNs) are essential for understanding the molecular interactions that drive biological processes, from development and metabolism to disease progression. GRNs can be represented as directed networks (or graphs), where nodes correspond to genes and directed edges indicate regulatory interactions between genes. GRN inference, the process of reconstructing these networks, has traditionally been performed using bulk RNA-sequencing (RNA-seq) data. However, the rise of single-cell RNA-sequencing (scRNA-seq) has introduced new opportunities and challenges, enabling the exploration of cellular heterogeneity at unprecedented resolution but also introducing significant technical noise and variability. Methods for GRN inference from scRNA-seq data can be classified into unsupervised and supervised approaches. Unsupervised methods identify regulatory interactions without prior knowledge of gene pairs, while supervised approaches rely on known networks to train models that predict gene interactions. This thesis investigates different machine learning approaches, GENIE3 and scGeneRAI, which are unsupervised, and GNNLink and STGRNS, which are supervised, for GRN inference using scRNA-seq data. Meaningful comparisons of methods were previously impossible because they were not originally tested on the same datasets. To address this, a significant contribution of this thesis is the evaluation and comparison of different methods using consistent datasets, ensuring direct comparability. While the performance of all methods is evaluated, specific enhancements were applied to GNNLink. These enhancements include using transcription factor frequency lookup tables to improve performance and creating an unsupervised version of GNNLink by leveraging only expression data to generate the training set based on Pearson correlation between genes. Additionally, irrelevant genes are filtered out from both the unsupervised approaches and the training set for the unsupervised version of GNNLink, ensuring that the predictions are not only more relevant but also comparable to those of supervised methods when evaluated against various ground-truth networks. By refining these computational methods, this research aims to improve the reliability and applicability of GRN inference across diverse biological contexts. vi Contents Abstract v List of figures ix List of tables xi Listing of acronyms xiii 1 Introduction 1 1. 1 Problem Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Objectives ................................... 3 2 Background 5 2.1 Biological Background ............................ 5 2.1.1 The Core Concepts of Gene Expression ............... 5 2.1.2 Gene Regulatory Networks ...................... 8 2.1.3 Gene Regulatory Network Definitions ................ 10 2.1.4 Single-cell RNA Sequencing ..................... 12 2.2 Graph Structure and GRN Inference Goal .................. 14 2.3 GRN Inference with Single-Cell Data ..................... 14 2.3.1 Correlation-Based Approaches .................... 16 2.3.2 Regression-Based Approaches .................... 16 2.3.3 Probabilistic Approaches ....................... 18 2.3.4 Dynamical Systems-Based Approaches ................ 18 2.3.5 Neural Networks and Deep Learning-Based Approaches ....... 19 2.4 Performance Metrics ............................. 21 2.4.1 Metrics in Binary Classification .................... 21 2.4.2 Imbalanced Datasets ......................... 23 3 Datasets 29

FAQs

What is the goal of gene regulatory network (GRN) inference in this thesis?

To reconstruct directed regulatory interactions among genes from single-cell RNA-sequencing data, representing GRNs as graph structures.

How are the evaluated methods grouped in the thesis?

Methods are classified into unsupervised approaches (GENIE3, scGeneRAI) and supervised approaches (GNNLink, STGRNS), with performance assessed across tasks and datasets.

Why does the thesis emphasize using consistent datasets for comparison?

Previous comparisons were limited because methods were not tested on the same datasets. Using consistent datasets enables direct, meaningful performance comparisons.

What enhancements were applied to GNNLink?

The thesis improves GNNLink using transcription factor frequency lookup tables and also builds an unsupervised variant that derives the training set from expression data using Pearson correlation, with irrelevant genes filtered out for better relevance and comparability.

Aug 5, 2026

Upload by Ezra