Plantgraph: ai-driven multiomics knowledge graph for plant phenotype prediction and herbicide analysis

Overview

PlantGraph is an AI-powered platform that constructs multiomics knowledge graphs to model plant biology and predict phenotypes. By combining foundational Graph Neural Networks with Large Language Models trained on genomic, transcriptomic, proteomic, and metabolomic data, the platform enables cross-species reasoning about metabolic pathways and their interactions with agricultural chemicals. The system supports phenotype predictions including weight, height, biomass, and disease states, while also modeling how pesticides and herbicides influence plant metabolism. This approach addresses a critical need in precision agriculture and crop science by providing a computational framework for understanding plant responses at the molecular level.

Technical specifications

Core architecture:

• Graph Neural Networks (GNNs) for generalizing across multiomics knowledge graphs of different plant species
• Large Language Models for embedding node information from genomes, transcriptomes, proteomes, and metabolomes
• Literature-based enrichment using LLMs to extract knowledge from academic publications
• Edge prediction and updating based on combined LLM and GNN approaches
• Uncertainty quantification for quality control of embedded connections
• ADMET property prediction for herbicides and pesticides on graph elements
• Phenotype prediction incorporating environmental conditions

Target plant species for initial development:

• Lemna (duckweed)
• Arabidopsis thaliana
• Alopecurus myosuroides (slender foxtail)
• Amaranthus palmeri (Palmer amaranth)
• Glycine max (soybean)
• Brassica napus (rapeseed)
• Triticum aestivum (wheat)

The platform leverages existing foundational models including DNABert2 and Ntformer for genomics, tGPT for transcriptomics, AlphaFold2, RosettaFold, and ESM for proteins, and ChemBerta and ElemBert for small molecules. ElemBert has been validated for ADMET property prediction with AUC scores exceeding 0.95.

Technology readiness level

Individual components of the architecture have been validated in their respective domains. LLMs across different omics types have been independently evaluated, with ElemBert achieving state-of-the-art performance in structure-based prediction of biological properties. The integrated PlantGraph system is at an early stage, with a planned one-year development roadmap covering six sequential phases: omics LLM setup, knowledge graph creation, literature enrichment, reasoning with uncertainty quantification, ADMET prediction, and phenotype prediction. Future extensions could include generative models for plant morphogenesis prediction, though this is currently out of scope.