Universität zu Köln

Hajseyed Nasrollah, Zahra Sadat (2025). Learning the Topology of Latent Signaling Networks from High Dimensional Transcriptional Intervention Effects. PhD thesis, Universität zu Köln.

PDF main.pdf Download (7MB)

Abstract

Biological systems rely on complex networks of interacting molecules, genes, proteins, and other regulators to govern cellular processes. However, inferring the causal structure of these networks is challenging. Particularly when important regulatory factors are only indirectly measured or entirely hidden. This difficulty is combined with the need to distinguish direct from indirect interactions and to capture dynamic responses under different perturbations such as gene knockouts or drug treatments. Consequently, developing a method that can incorporate both time resolved data and partial observability is critical for inference of the intricate architectures underlying signal transduction and gene regulation. To address these challenges, we propose an Ordinary Differential Equation based Nested Effects Model (odeNEM). Our approach combines the established Nested Effects Model framework,which maps hidden regulatory nodes to observable downstream effects, with a continuous time ODE formulation. This combination can capture saturable, non-linear kinetics. By explicitly modeling the propagation of perturbation signals over time, odeNEM infers which hidden regulators are causally upstream of others. Additionally, we developed mixture models for the downstream expression data which is considering the noisy high throughput nature of biological experiments. We validate our method on two key applications. First, we reconstruct signaling pathways in breast cancer cells using phosphoprotein time course data from the HPN-DREAM challenge. This shows how odeNEM captures context specific interactions among kinases such as AKT, mTOR, and MEK. Second, we infer a gene regulatory network in pluripotent stem cells from CRISPR single cell transcriptomics (RENGE), highlighting both canonical regulators like POU5F1 and SOX2 and novel interactions that warrant further biological investigation. Across both use cases, our results align with known biology and external data sources (e.g., ChIP-seq). This also confirms that a continuous time hidden node model can robustly uncover causal relationships. Overall, odeNEM expands the applicability of NEM approache by explicitly modeling. The model’s synergy with perturbation data, prior knowledge, and advanced inference strategies (e.g., MCMC) enables a more robust and biologically realistic reconstruction of latent signaling networks. This contributes a promising picture for unraveling the complex interplay of molecular interactions in systems biology, with broad implications for identifying therapeutic targets and understanding disease mechanisms.

Item Type:	Thesis (PhD thesis)
Translated title:	Title Language Learning the Topology of Latent Signaling Networks from High Dimensional Transcriptional Intervention Effects English
Creators:	Creators Email ORCID ORCID Put Code Hajseyed Nasrollah, Zahra Sadat zhajseye@smail.uni-koeln.de UNSPECIFIED UNSPECIFIED
URN:	urn:nbn:de:hbz:38-784973
Date:	2025
Language:	English
Faculty:	Faculty of Mathematics and Natural Sciences
Divisions:	Faculty of Medicine > Medizinische Statistik und Bioinformatik > Institut für Medizinische Statistik und Bioinformatik – IMSB
Subjects:	Natural sciences and mathematics
Uncontrolled Keywords:	Keywords Language Latent Network Reconstruction UNSPECIFIED Nested Effect Models (NEMs) UNSPECIFIED Signaling Pathway Networks UNSPECIFIED
Date of oral exam:	24 April 2025
Referee:	Name Academic Title Tresch, Achim Prof. Dr Hofmann, Kay Prof. Dr
Refereed:	Yes
URI:	http://kups.ub.uni-koeln.de/id/eprint/78497