Background: Machine learning methods are widely used in the field of bioinformatics, for example, to discover important genes for specific disease or phenotype, to classify proteins based on their structure, and to predict the risk of a disease for an individual based on its genome.However, the existence of confounders such as age, gender, data source and population structure increases the difficulty level for the tasks, especially when the data are not sampled randomly.It is still unclear how to effectively correct confounders.Methods: In this work, we propose an orthogonal projection correction (OPC) method to correct confounders.This is achieved by looking for a specific decomposition of each feature to a confounding component and non-confounding component, such that the original data can be best reconstructed by only the non-confounding components of features.We show that this can be done by an orthogonal projection of each feature to the complement of a confounder subspace, and the OPC procedure is kernelizable.We then further propose a proSVM method by integrating the OPC method with support vector machine for classification.Results: We applied the OPC method on cross-platform microarray data for tumor diagnosis and SNP data of Arabidopsis thaliana in presence of population structure for genome-wide analysis.In both experiments, the OPC method for confounder correction outperforms the other competitors.Conclusions: In this work, we proposed the OPC method for confounder correction and showed the OPC method has excellent performance for confounder correction in biological applications .