“Development of a personalized medicine approach to improve the diagnosis of mitochondrial diseases”

President of the jury :                       

• Michel RIVEILL, I3S, Sophia Antipolis

Rapporteurs :                     

• Anaïs BAUDOTL, MMG, Marseille
• Marie-Laure MARTIN NÉGRIER, IMN, Bordeaux 

Examiners

• Christine BRUN, TAGC, Marseille
• Pauline GAIGNARD, AP-HP, Paris

Thesis Director :

• Sylvie BANNWARTH, IRCAN, Nice
• Silvia BOTTINI, ISA, Sophia Antipolis

Abstract :

Mitochondrial diseases (MDs) are a group of rare and highly heterogeneous disorders caused by dysfunction of the mitochondria, the cellular organelles responsible for energy production. These diseases can result from variants in either the nuclear or mitochondrial genome and affect virtually any organ system, leading to a broad and often overlapping spectrum of symptoms. The diagnosis of MDs remains a significant challenge for clinicians due to this genetic and phenotypic complexity, the frequent involvement of multiple organ systems, and the limitations of conventional diagnostic approaches. In this context, precision medicine offers promising avenues, for integrating multi-omics data and advanced computational methods to improve diagnostic accuracy and guide therapeutic strategies.
This thesis aimed to contribute to the development of a personalized medicine approach for MDs through the design and application of bioinformatics and artificial intelligence (AI) methods tailored to rare disease contexts. The work is structured around several complementary axes.
First, we focused on proteomics and metabolomics data to address the challenge of patient classification and biomarker discovery. We performed a comprehensive benchmark of feature selection and feature extraction techniques to identify those most effective at improving machine learning model performance in the classification of patients versus controls. This benchmark was first validated on cancer datasets before being adapted to proteomic profiles associated with MDs. Our results highlighted the critical importance of dimensionality reduction and the choice of algorithms in enhancing classification accuracy and interpretability.
In a second axis, we developed a novel transcriptomics analysis tool, ABEILLE (ABerrant Expression Identification empLoying machine LEarning from RNA-sequencing data), designed to identify aberrant gene expression (AGEs) in individual patients. Existing tools often struggle with small cohort sizes typical of rare diseases or lack sensitivity for single-patient analysis like traditional differential expression analysis. ABEILLE leverages variational autoencoders to detect AGEs while accounting for noise and missing data, thus enabling the discovery of new candidate genes potentially implicated in disease mechanisms.
Finally, we addressed the complex task of variant prioritization by creating VIOLA (Variant PrIOritization using Latent spAce), a tool that integrates genomic, transcriptomic, and phenotypic data to identify potentially pathogenic variants from the millions identified by sequencing technologies. VIOLA incorporates machine learning and data integration approaches to guide the interpretation of unsolved cases and to support the diagnostic decision-making process.
Overall, this thesis demonstrates how the integration of multi-omics data, bioinformatics, and AI techniques can significantly advance the field of precision medicine for mitochondrial diseases. The tools and methods developed contribute to more accurate, personalized diagnoses and provide the basis for future translational applications aimed at improving patient care.

Keywords :

Mitochondrial diseases ; Precision medicine ; Multi-omics integration ; Machine learning ; Variant prioritization

The thesis can also be followed remotely via Zoom:

https://inrae-fr.zoom.us/j/98294413970