(eng): In this paper, we propose a new method for detecting relevant variables from a priori given high-dimensional data under the assumption that input-output relation is described by a nonlinear function depending on a few variables. The method is based on the inspection of the behavior of discrepancies of a multi-penalty regularization with a component-wise penalization for small and large values of regularization parameters. We provide a justification of the proposed method under a certain condition on sampling operators. The effectiveness of the method is demonstrated in an example with simulated data and in the reconstruction of a gene regulatory network. In the latter example, the obtained results provide clear evidence of the competitiveness of the proposed method with respect to the state-of-the-art approaches.