Thèse l'IA Agentique pour la Construction Automatique de Graphe de Connaissances en Agronomie H/F - Doctorat.Gouv.Fr

Établissement : Université de Montpellier École doctorale : I2S - Information, Structures, Systèmes Laboratoire de recherche : DIADE - Diversité, Adaptation et DEveloppement des plantes Direction de la thèse : Pierre LARMANDE ORCID 0000000229239790 Début de la thèse : 2026-10-01 Date limite de candidature : 2026-09-07T23:59:59 La question de recherche centrale est de déterminer dans quelle mesure un système multi-agents basé sur des LLM peut extraire, valider et intégrer des relations gène-phénotype issues de la littérature scientifique dans un graphe de connaissances, avec une fidélité comparable à celle d'une curation manuelle par des experts, tout en quantifiant formellement la confiance associée à ces extractions. Le travail couvre trois sous-problèmes : la fiabilité de l'extraction par des LLM d'entités biologiques, la conception d'architectures multi-agents réduisant les hallucinations par validation croisée, et la construction de benchmarks d'évaluation rigoureux.

AgroLD est un graphe de connaissances à grande échelle intégrant plus de 151 jeux de données biologiques et 1,08 milliard de triplets RDF décrivant gènes, protéines, phénotypes et voies métaboliques en sciences des plantes. Sa construction reste actuellement largement manuelle, et une grande partie des connaissances biologiques pertinentes demeure enfouie dans des sources non structurées comme la littérature scientifique.

Cinq objectifs structurent la recherche: l'extraction de connaissances biologiques par LLM (modeles comme BioMedLM et PubMedBERT); une architecture multi-agents (roles d'extracteur, de lieur, de validateur et de critique) avec validation croisee par requetes SPARQL sur AgroLD; l'ancrage dans des ontologies specifiques au domaine (Gene Ontology, Plant Trait Ontology) pour limiter les hallucinations; l'integration garantissant la coherence semantique avec les donnees existantes; et la construction d'un benchmark curate pour l'extraction d'associations gene-trait, valide sur des cas d'usage reels comme la decouverte de genes candidats dans des loci GWAS. Agricultural and plant science research generates large volumes of heterogeneous data, including genomic sequences, phenotypic measurements, biological networks, and scientific literature. Integrating these diverse resources is essential for understanding complex genotype-phenotype relationships and improving crop resilience in the context of climate change.
Knowledge Graphs (KGs) have become a powerful approach for integrating such heterogeneous data using Semantic Web technologies. AgroLD (Agronomic Linked Data) is a large-scale knowledge graph designed for plant science research, integrating more than 151 biological datasets and over 1.08 billion RDF triples describing genes, proteins, phenotypes, pathways, and biological interactions [1,1b]. The platform is actively maintained by the South Green Bioinformatics Platform and constitutes the primary experimental ground for this PhD. In the plant science domain, KnetMiner [8] provides a comparable graph-based approach for evidence-based gene discovery, illustrating both the potential and the current limitations of such systems before the integration of LLM-based methods.
A central challenge remains that the current KG construction process is largely manual or semi-automated. A substantial amount of biologically relevant knowledge is embedded in unstructured sources such as scientific publications, technical reports, and database annotations. Extracting, structuring, and validating this knowledge at scale is an open scientific problem.
Recent advances in Large Language Models (LLMs) and agent-based AI systems open new opportunities to automate knowledge extraction, integration, and reasoning. However, LLMs applied to the biological domain introduce specific risks: hallucinated gene names, incorrect genomic identifiers, fabricated citations, and semantically inconsistent relations. Developing rigorous grounding and confidence mechanisms tailored to the specificities of plant genomics is a critical and largely unsolved research problem, which constitutes the scientific core of this PhD.
Recent surveys confirm that agentic AI is rapidly transforming scientific discovery across chemistry, biology, and materials science [2]. High-profile systems such as the Google AI co-scientist [3] demonstrate the potential of multi-agent architectures to generate novel, verifiable hypotheses by combining literature synthesis with autonomous reasoning. In oncology, AI agents are already being applied to complex multi-step workflows ranging from drug design to clinical decision support [4]. However, these initiatives share a common limitation: they operate either at a general domain level or in well-resourced biomedical contexts, and none address the specific challenges of plant sciences - a domain characterized by high species diversity, specialized ontologies (Plant Trait Ontology, Plant Ontology, Gene Ontology), and heterogeneous agronomic data. Furthermore, none tackle the problem of automatically constructing and updating a large-scale, semantically structured knowledge graph from unstructured literature while enforcing consistency with formal biological ontologies. This gap defines the unique scientific contribution of this PhD.
The PhD pursues five interconnected objectives:
1. Objective 1 - LLM-based biological knowledge extraction. Develop and benchmark methods based on fine-tuned and prompted LLMs (including domain-specific models such as BioMedLM [7] and PubMedBERT) to extract biological entities and relationships from scientific literature, with a specific focus on gene-trait associations in crops.
2. Objective 2 - Multi-agent architecture for KG construction. Design a multi-agent AI architecture separating distinct roles (extractor, linker, validator, critic) with explicit disagreement protocols. The critic agent will cross-check extractions against AgroLD using SPARQL queries to detect semantic inconsistencies.
3. Objective 3 - Grounding and hallucination mitigation. Develop domain-specific grounding mechanisms to anchor LLM outputs to controlled vocabularies and ontologies (Gene Ontology, Plant Trait Ontology, Gene Ontology), and design confidence scoring methods for extracted triples.
4. Objective 4 - Integration and semantic consistency. Integrate the validated extracted knowledge into the AgroLD knowledge graph while ensuring semantic consistency with existing data, including entity linking and ontology alignment.
5. Objective 5 - Evaluation benchmark and biological validation. Construct a manually curated benchmark dataset for gene-trait association extraction in plant sciences, and evaluate the system on real biological use cases such as candidate gene discovery in GWAS loci.
4. Work Plan (36 months)
Year 1 - Knowledge extraction and benchmark construction
- Study the AgroLD architecture, ontologies, and datasets in depth.
- Collect and preprocess plant science literature (PubMed, bioRxiv, full-text corpora).
- Develop and compare LLM pipelines for entity and relation extraction (fine-tuned vs. prompted models).
- Develop grounding mechanisms linked to biological ontologies and controlled vocabularies.
- Construct a manually curated benchmark dataset for gene-trait associations (publishable contribution).
- Evaluate extraction performance on benchmark and existing NLP datasets (BioRED [6], CRAFT).
Year 2 - Multi-agent architecture and KG integration
- Design the multi-agent architecture (extractor / linker / validator / critic agents) with formal disagreement protocols.
- Implement SPARQL-based cross-validation of extracted triples against AgroLD.
- Implement automated pipelines for integrating extracted knowledge into AgroLD.
- Improve entity linking and ontology alignment methods, including similarity-based entity profiling.
Year 3 - Validation, reasoning, and biological applications
- Develop graph-based validation and reasoning methods for KG quality assessment.
- Apply the system to candidate gene prioritization in GWAS loci and transcriptomic studies.
- Evaluate the scientific utility of the system with plant genomics researchers at DIADE and partner labs.
- Release open-source software integrated into the AgroLD platform.
- Write and submit scientific publications.

Given the technical nature of this PhD, the ideal candidate has a strong background in natural language processing or knowledge representation, with plant genomics knowledge being acquirable during the thesis.
Required:
- Master's degree in Bioinformatics, Artificial Intelligence, Computer Science, or Data Science.
- Strong programming skills in Python (NLP libraries, graph databases, REST APIs).
- Solid knowledge of machine learning and/or natural language processing.
- Ability to work in a multidisciplinary environment (biology, computer science, data management).
Advantageous:
- Experience with knowledge graphs or semantic web technologies (RDF, SPARQL, OWL).
- Familiarity with biological ontologies (Gene Ontology, Plant Ontology, Plant Trait Ontology).
- Prior exposure to plant genomics data (GWAS, transcriptomics, gene annotation).

• Python
• SPARQL
• RDF
• API
• Intelligence artificielle
• OWL
• Machine learning
• REST