MINIE: A MACHINE LEARNING NETWORK INFERENCE TOOL FOR TIME-SERIES MULTI-OMIC DATA

Biological processes and phenotypes result from complex interactions across multiple molecular levels. Advances in high-throughput technologies allow the measurement of these levels, gener- ating multiple types of omic data such as genomics, epigenomics, transcriptomics, proteomics, and metabolomics. Although integrating multi-omic data offers the potential to unravel the full complexity of biological systems, traditional network inference methods have largely focused on single-omic studies, resulting in a partial understanding. Multi-omic network inference therefore emerges as a promising approach, providing a comprehensive view of biological regulation across molecular layers.
This thesis presents the development of MINIE, a novel computational algorithm for multi- omic network inference from time-series data. MINIE integrates multi-omic data from bulk metabolomics and single-cell transcriptomics to reconstruct regulatory networks spanning several molecular layers. By leveraging time-series data, the algorithm infers causal interactions through a Bayesian regression framework while considering biological properties, such as the timescale separation between omic layers and the pseudotime concept. This dissertation describes the development and validation of this method, using both simulated data from biological models and experimental data from a Parkinson’s disease study. Overall, MINIE exhibits accurate and robust predictive performance across and within omic layers, outperforming previously published methods. By integrating system dynamics across molecular layers and temporal scales, MINIE provides a powerful tool for comprehensive multi-omic network inference, paving the way for new insights into complex diseases and other multi-omic applications.