6533b851fe1ef96bd12a94a2

RESEARCH PRODUCT

Apprentissage automatique de réseaux d'interaction à partir de données de séquences de nouvelle génération

subject

description

Climate change and other human-induced processes are modifying ecosystems, globally, at an ever increasing rate. Microbial communities play an important role in the functioning ecosystems, maintaining their diversity and services. These communities are shaped by the different abiotic environmental effects to which they are subjected and the biotic interactions between all community members. The ANR Next-Generation Biomonitoring (NGB) project proposed to reconstruct interaction networks from abundance measures obtained sequencing environmental DNA (eDNA) and to use these networks to monitor ecosystem change. In this thesis, conducted as part of the NGB project, I evaluate the potential of two existing statistical network reconstruction tools, SparCC and SPIEC-EASI, to reconstruct microbial networks in order to evaluate ecosystem change. Microbial communities from grapevine leaves were used to differentiate between two different agricultural practices, identifying the appropriate network metrics to capture ecosystem change. The experiments showed that although it is difficult to obtain replicate networks, even from the same environmental conditions, it is still possible to differentiate networks from different agricultural practices using some network metrics. Although statistically-based network reconstruction tools can obtain networks of associations between microorganisms, with accuracy, these statistical associations are not direct indicators of the underlying ecological processes of interaction. To address this issue, I developed a new network reconstruction tool called Interaction Inference using Explainable Machine Learning (InfIntE), based upon Explainable Machine Learning (EML). EML is a branch of machine learning which uses the prior knowledge from a scientific domain, such as Ecology, to declare logical statements of concept (hypotheses) to carry out human-understandable inference. InfIntE uses ecological rules of interaction together with the abundance information obtained from sequencing eDNA to reconstruct networks by logical inference. In contrast to statistically-based network reconstruction, the use of interaction rules allows direct classification the inferred interactions to their type (e.g. mutualism, competition), obtaining more informative and objective interaction networks. The performance of InfIntE was evaluated using computer-generated data as well as datasets obtained by eDNA sampling of grapevine leaf microbiome. My results show that InfIntE can detect interaction networks with similar accuracy to the tested statistically-based tools, SparCC and SPIEC-EASI, with the significant benefit of direct classification of the type of each interaction.