|
|
| Identifier |
000419274 |
| Title |
Causal discovery limitations in learning molecular interaction networks : an empirical study using linear mechanistic models |
| Alternative Title |
Περιορισμοί αιτιακής ανακάλυψης στην εκμάθηση δικτύων μοριακής αλληλεπίδρασης: μια εμπειρική μελέτη βασισμένη σε γραμμικά μηχανιστικά μοντέλα |
|
Author
|
Κρανά, Μυρτώ Λ.
|
|
Thesis advisor
|
Τσαμαρδινός, Ιωάννης
|
|
Reviewer
|
Τσακαλίδης, Παναγιώτης
Πανταζής, Ιωάννης
|
| Abstract |
Mechanistic models have been traditionally successful in describing biological systems. Their accuracy, however, depends more on expert knowledge about the structure of causal interactions between system components than the abundance of experimental data. At the same time, several algorithms that can learn causal structures de novo from observational and experimental data have been developed over the past decades. Despite the attracted attention, their applicability in learning biological systems has been relatively poor. In this work we systematically study the effect of violations of causal assumptions to basic structure learning algorithms using mechanistic models of protein signaling networks as testbed biological systems. Because the same network of causal interactions can be described using several mechanistic models we study several combinations of system topologies and model specifications. We calculate the analytical solution of each model at steady-state and juxtapose the solution with the fundamental causal discovery principles. We prove mathematically the conditions under which a causal learning algorithm is guaranteed to discover the structure of the system that a mechanistic model describes. Whenever there is no tractable analytic solution, a simulated one is employed. We show that the structure of interactions estimated using data from mechanistic models under steady-state conditions is, in general, inconsistent with the expected causal structure. Accordingly, we reveal that only under very specific conditions is the discovery of the structure guaranteed using a constrained-based causal discovery algorithm.
|
| Language |
English |
| Subject |
Molecular networks |
|
Μοριακά δίκτυα |
| Issue date |
2018-11-23 |
| Collection
|
School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses
|
|
|
Type of Work--Post-graduate theses
|
| Views |
1862 |
Causal discovery limitations in learning molecular interaction networks : an empirical study using linear mechanistic models
