Developing high-dimensional statistical inference method to identify the individual features associated with the response is very important in analyzing large-scale datasets from economics, finance, medicine, cancer studies, bioinformatics. However, themodern data sets collected often not only own the high-dimension property but alsoexhibit unknown relationships between the response and its explanatory features.Reliable statistical inference depends on an accurate modeling for the observed data.Such an involved modelling task can be done by the state-of-the-art semi-parametricmodel with few model assumptions. In this paper, based on high-dimensionalsemi-parametric model, we utilize the estimators of unknown parameters’ directions andsymmetrized data aggregation approach to develop a novel and model-free featureselection method for achieving fine-mapping of risk features while controlling the falsediscovery rate (FDR) of selection. The proposed method can be applied for the analysisof both continuous and discrete response data sets. The results of simulation studiesdemonstrate the proposed method has robust feature selection performance whilecontrolling FDR very well. The analysis results of a real ocean microbiome dataindicate our method indeed is effective to detect risk features.