Human limits in machine learning: prediction of potato yield and disease using soil microbiome data

Area of Science:

• Agricultural Science
• Environmental Science
• Computational Biology

Background:

• Soil health is crucial for agriculture, human health, and biodiversity.
• Machine learning (ML) offers potential for predicting biological phenotypes from soil properties.
• This study investigates ML models for soil-plant performance prediction.

Purpose of the Study:

• To assess the predictive power of ML models for plant performance using soil data.
• To explore the impact of environmental features and data preprocessing on prediction accuracy.
• To provide a decision tree for optimizing ML model selection in soil science.

Main Methods:

• Utilized random forest and Bayesian neural network models.
• Integrated soil biological, chemical, and physical properties with microbiome data.
• Evaluated various data preprocessing strategies, including normalization and taxonomic levels.

Main Results:

• Prediction accuracy improved with the inclusion of environmental features and microbiome data.
• Total sum scaling normalization was identified as an optimal preprocessing strategy.
• Accurate sample labeling was found to be more critical than normalization or model choice.
• ML performance was limited by human accuracy in sample classification.

Conclusions:

• Integrating diverse environmental data and careful preprocessing enhances ML predictive power for soil-plant connections.
• This approach aids in advancing agricultural practices and soil health management.
• Domain scientists can use the provided decision tree to optimize ML model selection.