Trait-based Machine-Learning Identifies Functional predictors of Plant Responsiveness to Arbuscular Mycorrhizal Fungi

Technical Summary: Trait-based Machine-Learning Identifies Functional Predictors of Plant Responsiveness to Arbuscular Mycorrhizal Fungi

Problem Statement
Arbuscular mycorrhizal fungi (AMF) are critical symbionts that enhance plant nutrient acquisition, growth, and stress tolerance. However, the magnitude and direction of plant growth response to AMF colonization—quantified as Mycorrhizal Growth Response (MGR)—vary substantially across species, genotypes, and environmental conditions. This variability is difficult to predict because MGR is frequently decoupled from colonization intensity; a plant may be heavily colonized without deriving significant growth benefits. Conventional linear statistical models are often ill-suited to capture the nonlinear relationships and complex interactions among plant functional traits and environmental variables that drive this variation. There is a need for a predictive framework that can integrate these complex factors to identify the functional trait syndromes associated with AMF responsiveness.

Methodology
The authors developed Trait2Myco, a Random Forest-based machine-learning framework designed to predict MGR using plant functional traits and environmental covariates.

• Data Source: The primary dataset comprised 80 observations derived from four grass species grown under four precipitation regimes (five biological replicates per combination). Data were sourced from the Dryad Digital Repository (Zhao et al., 2026).
• Predictors: The model incorporated ten functional traits representing leaf and root economics (e.g., specific leaf area, specific root length, root carbon concentration, leaf nitrogen) and one environmental variable (precipitation).
• Response Variable: MGR was calculated as the proportional difference in biomass between AMF-inoculated and non-inoculated control plants.
• Model Architecture:
  • A Full Model utilized all ten traits plus precipitation.
  • A Reduced Ecological Model utilized only the four most important predictors identified via permutation-based variable importance analysis: specific leaf area (SLA), specific root area (SRA), root carbon concentration, and precipitation.
  • Models were trained on 80% of the data (n=64) and evaluated on a held-out test set (20%, n=16). Ten-fold cross-validation was also performed on the training data.
• Analytical Techniques: The study employed Random Forest for prediction and variable importance ranking (%IncMSE), Partial Dependence Plots (PDPs) to visualize nonlinear relationships, and Principal Component Analysis (PCA) to map observations within multivariate trait space.
• External Validation: An exploratory analysis applied the workflow to an independent dataset (n=8) relating growth traits to AMF root colonization scores (Werner & Kiers, 2015).

Key Results

• Predictive Performance: On the held-out test set, the Full Model explained 41% of the variation in MGR ($R^2 = 0.41$). Notably, the Reduced Ecological Model, using only four variables, outperformed the full model, explaining 51.7% of the variation ($R^2 = 0.517$). Ten-fold cross-validation on the training data yielded a high $R^2$ of 0.983, though the authors note this likely overestimates generalization due to shared biological characteristics among replicates.
• Variable Importance: Specific Leaf Area (SLA) was identified as the strongest predictor, followed by precipitation, root carbon concentration, and specific root area.
• Trait Category Contribution: Leaf traits and root traits contributed nearly equally to model importance (41.9% and 41.3%, respectively), while precipitation contributed 16.8%.
• Nonlinear Relationships: Partial dependence analysis revealed:
  • A sharp decline in predicted MGR as SLA increased, plateauing at higher values.
  • A unimodal relationship with precipitation, where MGR peaked at approximately 800 mm and declined at both lower and higher levels.
  • A threshold-like relationship with root carbon concentration, where MGR increased markedly above ~4.35%.
• Multivariate Trait Space: PCA indicated that highly responsive plants occupied distinct regions of functional trait space characterized by more resource-acquisitive syndromes, showing partial separation from less responsive plants along the primary axis of variation.
• External Application: The workflow applied to the independent colonization dataset showed high apparent performance ($R^2 = 0.94$), with plant length being the strongest predictor, though the authors caution this is exploratory due to the small sample size and different response variable.

Significance and Claims
The paper claims that Trait2Myco provides a reproducible, trait-based framework for exploring variation in plant responsiveness to AMF. Its primary significance lies not in the development of a new machine-learning algorithm, but in the integration of functional trait theory with Random Forest modeling to identify predictive syndromes.

Key claims include:

1. Predictive Utility: A small, ecologically interpretable set of four variables (SLA, SRA, root carbon concentration, and precipitation) can predict MGR more effectively than a larger set of ten traits in this specific experimental context.
2. Mechanistic Insight: The identification of root carbon concentration as a major predictor highlights a potential link between belowground carbon investment and symbiotic benefit, a trait rarely evaluated explicitly in this context.
3. Whole-Plant Perspective: The results support the view that both leaf and root economic strategies are coordinated and jointly shape AMF outcomes, rather than root traits acting in isolation.
4. Context Dependence: The nonlinear relationship with precipitation reinforces the concept that AMF benefits are context-dependent, peaking at intermediate water availability.

The authors maintain a modest stance, emphasizing that the framework identifies associations rather than causal effects. They acknowledge limitations, including the small number of species (four), the use of species-level trait averages, and the exclusion of fungal identity and soil properties. They conclude that while Trait2Myco offers a useful framework for the current experimental system, broader validation across diverse taxa, environments, and fungal communities is required to determine the generalizability of these trait-MGR relationships.