In order to reduce the amount of trace metals such as cadmium in human food, it is useful to predict the trace metal composition of cereal grains from well-chosen topsoil variables. Statistical relationships between soil properties and metal concentration in plant organs have already been studied. However, such studies involved only a few common soil properties such as pH, organic carbon content and cation exchange capacity. Here, we compared metal contents in grain and in topsoil samples. 198 samples of grains of winter wheat were collected from paired topsoil and crop surveys in the northern half of France. The soils belong to 18 contrasting pedogeological families. Grain and topsoil samples were analysed for Cd, Cr, Cu, Ni, Pb, Zn, Fe, Mg and Mn. For soil samples, three pools of metal were determined: total concentration by HF digestion, and two available pools by partial extraction using either diethylenetriamine pentaacetic acid (DTPA) or NH4NO3. Canonical correlation analysis including multiple linear regression was used to study relationships between soil and grain data. Our findings show the occurrence of six significant and independent relations between the topsoil variables and the grain variables. The adjustment quality of those relations is measured by six coefficients equivalent to individual determination coefficients. The maximum coefficient of 0.79 is higher than any individual ones. Our findings show in particular an excellent regression model for grain Cd with a small number of topsoil variables, thus allowing an accurate prediction for winter wheat grains. Practically, farmers can use the prediction model to increase soil pH and by growing a less metal-accumulating cultivar upon detection of a threat. The strengths of our study are the following: (1) we studied a very large number of sites, including soils with very diverse parent materials, soil-forming processes and geochemical properties; (2) we measured and tested a large number of variables, including in particular available metal fractions; and (3) we applied unusual statistical methods such as canonical correlation analysis.