Table 2.

Machine‐learning models of biomass traits using carbohydrate data

A. Canopy height		Information included in models
		Genotype, carbohydrates	Carbohydrates only
		All genotype replicates	All genotype replicates	Averaged by genotype
Plants	Carbohydrate fractions	R values
Mixed population	All	0.92***	0.44***	0.76***
	Nonstructural	0.81***	0.70***	0.81***
	Soluble	0.84***	0.52***	0.70**
Mapping family 2013	All carbohydrates	0.94***	0.72***	0.77***
	Nonstructural	0.92***	0.63***	0.68***
	Soluble	0.93***	0.68***	0.70***
Mapping family 2014	Nonstructural	0.88***	0.61***	0.76***
	Soluble	0.86***	0.44***	0.66***
Carbohydrate fractions	Constituents common to predictors of all modelsa
All	Glucan; Fructose (as Fru, Glc/Fru, Hex or Suc/Fru); Glucose (as Glc/Fru, Hex or Sta/Glc); Starch (as Sta, Sta/Glc or Suc/Sta)
Nonstructural	Fructose (as Fru, Glc/Fru, Hex, NSC or Suc/Fru); Glucose (as Glc/Fru, Hex, NSC, Sta/Glc or Suc/Glc); Starch (as NSC, Sta, Sta/Fru, Sta/Glc or Suc/Sta)
Soluble	Fructose (as Fru, Glc/Fru, Hex or Suc/Fru)

B. Harvest yield		Information included in models
		Genotype, carbohydrates	Carbohydrates only
		All genotype replicates	All genotype replicates	Averaged by genotype
Plants	Carbohydrate fractions	R values
Mixed populationb	All	0.79***	0.61***	0.81***
	Nonstructural	0.79***	0.62***	0.75***
	Soluble	0.79***	0.62***	0.75***
Mapping family 2013	All carbohydrates	0.85***	0.77***	0.70***
	Nonstructural	0.86***	0.65***	0.72***
	Soluble	0.86***	0.68***	0.72***
Mapping family 2014	Nonstructural	0.75***	0.56***	0.68***
	Soluble	0.74***	0.40***	0.58***
Carbohydrate fractions	Constituents common to predictors of all modelsa
All	Fructose (as Fru, Glc/Fru or NSC)
Nonstructural	Fructose (as Fru, Glc/Fru, NSC or Suc/Fru); Glucose (as Glc/Fru, NSC or Suc/Glc); Sucrose (as NSC, Suc/Fru, Suc/Glc or Suc/Sta); Starch (as NSC, Sta or Suc/Sta)
Soluble	Fructose (as Fru, Glc/Fru or Suc/Fru)

Support vector regression (SMOreg) models were trained using subsets of predictors selected (CfsSubsetEval) for individual correlation with trait but low correlation with each other. Models were evaluated in ninefold cross‐validations. R values indicate Pearson correlation between actual and predicted biomass data.

^aFull lists in Supplementary Information.

^b M. sacchariflorus genotype Mb306 was excluded from Table B.