This thesis investigates the performance of data-driven and Physics-Informed Neural Networks (PINNs) for inferring Gene Regulatory Networks (GRNs). By focusing on two distinct biological models—multi-cell communication and isolated protein interaction dynamic—the study evaluates the effectiveness of these approaches in accurately capturing GRN parameters, including activation and inhibition strengths. Both methods produced good results in terms of accuracy, demonstrating their capability to model GRNs effectively. However, PINNs demonstrated superior predictive performance in the presence of noisy data by embedding physical constraints within the learning process. This approach not only enhances robustness to noise but also enables the simultaneous estimation of system states over time and inference of parameters, providing biologically plausible insights into the dynamics of the studied systems. These findings underscore the potential of PINNs as powerful tools for accurate and interpretable GRN modeling, particularly in scenarios with noisy or limited data, which could lead to significant advancements in developmental biology.
Questa tesi analizza le prestazioni di modelli di reti neurali puramente data-driven e delle Physics-Informed Neural Networks (PINNs) nell’inferenza delle Gene Regulatory Networks (GRNs). Lo studio si focalizza su due distinti modelli biologici: comunicazione multicellulare e regolazione genica isolata, valutando l’efficacia di questi approcci nel catturare con precisione i parametri delle GRNs, inclusi i coefficienti di attivazione e inibizione. Entrambi i metodi hanno mostrato buone prestazioni in termini di accuratezza, dimostrando la loro capacità di modellare efficacemente le GRNs. Tuttavia, le PINNs si sono distinte per la loro superiore capacità predittiva in presenza di dati rumorosi, grazie all'integrazione di vincoli fisici nel processo di apprendimento. Questo approccio non solo migliora la robustezza del modello al rumore, ma consente anche la stima simultanea degli stati del sistema nel tempo e dei parametri, fornendo una visione biologicamente realistica della dinamica dei sistemi studiati. Questi risultati evidenziano il grande potenziale delle PINNs come strumenti potenti per un’inferenza accurata e interpretabile delle GRNs, specialmente in scenari con dati limitati o rumorosi, aprendo la strada a nuove applicazioni nel campo della biologia dello sviluppo.
Data-driven and Physics-Informed machine learning models for gene regulatory network inference
NEGRONI, SABRINA
2023/2024
Abstract
This thesis investigates the performance of data-driven and Physics-Informed Neural Networks (PINNs) for inferring Gene Regulatory Networks (GRNs). By focusing on two distinct biological models—multi-cell communication and isolated protein interaction dynamic—the study evaluates the effectiveness of these approaches in accurately capturing GRN parameters, including activation and inhibition strengths. Both methods produced good results in terms of accuracy, demonstrating their capability to model GRNs effectively. However, PINNs demonstrated superior predictive performance in the presence of noisy data by embedding physical constraints within the learning process. This approach not only enhances robustness to noise but also enables the simultaneous estimation of system states over time and inference of parameters, providing biologically plausible insights into the dynamics of the studied systems. These findings underscore the potential of PINNs as powerful tools for accurate and interpretable GRN modeling, particularly in scenarios with noisy or limited data, which could lead to significant advancements in developmental biology.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Negroni_Tesi.pdf
non accessibile
Descrizione: Testo tesi
Dimensione
4.41 MB
Formato
Adobe PDF
|
4.41 MB | Adobe PDF | Visualizza/Apri |
2024_12_Negroni_Executive Summary.pdf
non accessibile
Descrizione: Testo executive summary
Dimensione
1.52 MB
Formato
Adobe PDF
|
1.52 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/231111