Validation of multivariate model of leaf ionome is fundamentally confounded.

Authors: Evens, Terence; Niedz, Randall

Submitted to: Proceedings of the National Academy of Sciences (PNAS)
Publication Type: Other
Publication Acceptance Date: 12/9/2008
Publication Date: 1/13/2009
Citation: Evens, T.J., Niedz, R.P. 2009. Validation of multivariate model of leaf ionome is fundamentally confounded. Proceedings of the National Academy of Sciences. 106(2):E6

Interpretive Summary: The multivariable signature model reported by Baxter et al. (2008, Proceedings of the National Academy of Sciences, 105:12081-12086) to predict Fe and P levels in Arabidopsis is fundamentally flawed. The objective of the study was to identify plants growing in low levels of Fe or P by measuring the levels of the other mineral nutrients in the plant. This is important since plants can adjust their biochemistry, to a certain extent, to adjust for soils with low levels of nutrients. This enhances the plant’s ability to take up nutrients from poor soils. This paper reports a statistical model to predict if a plant that appears normal has altered its nutrient uptake biochemistry to survive on low nutrient levels. The two primary flaws are 1) ion confounding and 2) not validating the statistical model with the underlying biochemical changes known to occur when a plant adapts to low nutrient conditions.

Technical Abstract: The multivariable signature model reported by Baxter et al. (1) to predict Fe and P homeostasis in Arabidopsis is fundamentally flawed for two reasons: 1) The initial experiments identified a correlation between trace metal (Mn, Co, Zn, Mo, Cd) signature and “Fe-deficiency,” which was used to train a logistic regression model with known Fe-sufficient/-deficient plants. Subsequent experiments to validate the model were insufficient and merely served as unverified negative controls. Given that Fe- and P-deficiency are defined by biochemical indicators, e.g. IRT1 accumulation and ferric chelate reductase activity, it must be these indicators, not the trace metal signatures that are the “validators” of any correlative model. Using the trace metal signature as the “validator” assumes a priori, that only plants with a similar trace metal signature are Fe-deficient. Without any data to determine if the plants grown for the validation experiments actually were Fe-deficient, as opposed to simply not possessing a trace metal signature similar to that used to train the logistic regression model, we have no way of determining if the model is valid. 2) The authors attempted to determine single ion effects in the validation experiments, but these experiments exhibited substantial confounding (2; Table 1) and cannot be used to quantify the main and/or interactive effects of these ions on Fe-deficiency/sufficiency (3). A suitable approach to validation would involve sampling the implied 5-dimensional Mn-Co-Zn-Mo-Cd experiment design space and then comparing the model’s prediction to IRT1 accumulation, ferric chelate reductase activity and the trace metal signatures.