Artificial intelligence and machine learning in plant tissue culture: from lab-specific models toward unified and scalable tools for intelligent protocol design

Artificial intelligence (AI) and machine learning (ML) tools are poised to solve some of the stickiest research challenges in in vitro plant science: standardizing the subjective, clarifying the complex, and making plant growth predictable. The complexity of in vitro plant systems, marked by high dimensionality and variable interactivity, limits the efficacy of traditional experimentation and slows research progress. Emerging AI/ML tools may improve research decision-making by advancing standardized data collection, interpretation and design of genomic sequences, and predictive experimental design. This work discusses key concepts in AI/ML for plant scientists, and reviews recent implementations in in vitro workflows. The application of AI/ML tools in media optimization is examined in depth, highlighting lab-specific, single-use models developed for this purpose. Finally, this work introduces a new architecture for easy-to-adopt ML media prediction across labs and experimental objectives termed MERISTEM: Multi-context Evaluation and Representation via Ingredient SeT EMbeddings. MERISTEM chains two features: (1) a recipe encoder (Set Transformer) facilitating learning across recipes of varying composition, and (2) a multi-head neural network (multi-layer perceptron) supporting prediction of multiple outcomes. A proof-of-concept MERISTEM model was trained on data aggregated from eight studies on shoot induction across multiple species. Using a Non-dominated Sorting Genetic Algorithm II for multi-objective optimization, the proof-of-concept model makes biologically meaningful recommendations for media recipes even when prior data about a predicted ingredient is unavailable. This approach represents an important step toward a unified tool for media design, enabling researchers to identify optimal media recipes for multiple use cases and any combination of ingredients without the need for prior experimental screens.