Jointly conducted by UGB and IRD [Île-de-France], this research explores deep learning for metagenomics.
Thanh Hai Nguyen, researcher at UMMISCO; ICAN (Université Gaston Berger de Saint-Louis Sénégal).
Thesis defended in 2018 at the doctoral school École doctorale Informatique, télécommunications et électronique de Paris (1992-....).
This research is the result of an international co-supervision between several partner institutions.
Referenced in the ABES/STAR network, this thesis meets the rigorous criteria of French higher education.
Deep learning has revolutionized many fields, and metagenomics is no exception. This discipline, which studies the entirety of genomes present in an environmental sample, has seen its analysis methods transformed by technological advances in artificial intelligence. Thanh Hai Nguyen's thesis, defended at Gaston Berger University in Saint-Louis, highlights how human microbiome data can transform the diagnosis and prognosis of diseases. Indeed, this data represents a source of valuable information, but its complexity poses major challenges. Microbiomes, which vary from one individual to another, are influenced by factors such as diet, environment, and lifestyle, making their analysis all the more delicate. Research focuses on the efficient extraction of heterogeneous biomedical signatures through a feature selection framework. This framework uses self-organizing maps to visualize and merge varied data, thus demonstrating reasonable classification accuracy on real datasets.
This approach not only allows for a better understanding of microbial diversity but also helps identify potential biomarkers for specific diseases. For example, studies have shown that certain variations in the gut microbiome may be correlated with diseases such as diabetes or inflammatory bowel diseases. By integrating this data into a deep learning model, it becomes possible to predict with increased accuracy the risks of developing these conditions.
The thesis also presents an innovative deep learning approach that leverages artificial image representations for disease prediction. By visualizing metagenomic data through a simple filling method, combined with dimensionality reduction techniques, the results show that this representation allows for the application of advanced methods such as convolutional neural networks. These networks, which are particularly effective for processing visual data, can also be adapted to analyze unstructured data such as that derived from metagenomics. The results obtained often exceed the performance of existing benchmarks in the field of metagenomics, highlighting the importance of innovation in this sector.
This research raises crucial questions about the future of personalized medicine. How can these advances be integrated into clinical practices? Personalized medicine, which aims to tailor treatments to the individual characteristics of patients, could greatly benefit from deep learning tools. Indeed, the ability to quickly analyze large amounts of biological data could enable physicians to make informed decisions based on specific microbial profiles.
However, it is essential to consider the ethical implications of using complex biological data. The collection and analysis of microbiome data raise questions of privacy and informed consent. Policymakers must reflect on how these technologies can be leveraged to improve healthcare while protecting patients' rights. For example, clear protocols must be established to ensure that patient data is not used for unauthorized purposes.
The dual validation of this research, both through international co-supervision and ABES/STAR certification, attests to its level of academic rigor. This reinforces the credibility of the results and paves the way for international collaborations to deepen research in this field. The promising results open the door to concrete applications in disease diagnosis but require support for their implementation in healthcare systems.
It is crucial that governments and health institutions invest in the necessary infrastructure to integrate these technologies into medical practices. This could include training healthcare professionals in the use of these tools, as well as developing platforms for the secure sharing of data. In summary, deep learning in the field of metagenomics represents a major advancement that could transform our approach to diseases, but it must be accompanied by ethical reflection and strong institutional support.
Access the full study
Sources and Access
Thanh Hai Nguyen. Some contributions to deep learning for metagenomics. Artificial Intelligence [cs.AI]. Sorbonne Université, 2018. English. ⟨NNT : 2018SORUS102⟩. ⟨tel-02505179⟩