The prediction of subcellular locations of proteins can provide useful hints for revealing their functions as well as for understanding the mechanisms of some diseases and, finally, for developing novel drugs. As the number of newly discovered proteins has been growing exponentially, laboratory-based experiments to determine the location of an uncharacterized protein in a living cell have become both expensive and time-consuming. Consequently, to tackle these challenges, computational methods are being developed as an alternative to help biologists in selecting target proteins and designing related experiments. However, the success of protein subcellular localization prediction is still a complicated and challenging problem, particularly when query proteins may have multi-label characteristics, i.e. their simultaneous existence in more than one subcellular location, or if they move between two or more different subcellular locations as well. At this point, to get rid of this problem, several types of subcellular localization prediction methods with different levels of accuracy have been proposed. The support vector machine (SVM) has been employed to provide potential solutions for problems connected with the prediction of protein subcellular localization. However, the practicability of SVM is affected by difficulties in selecting its appropriate kernel as well as in selecting the parameters of that selected kernel. The literature survey has shown that most researchers apply the radial basis function (RBF) kernel to build a SVM based subcellular localization prediction system. Surprisingly, there are still many other kernel functions which have not yet been applied in the prediction of protein subcellular localization. However, the nature of this classification problem requires the application of different kernels for SVM to ensure an optimal result. From this viewpoint, this paper presents the work to apply different kernels for SVM in protein subcellular localization prediction to find out which kernel is the best for SVM. We have evaluated our system on a combined dataset containing 5447 single-localized proteins (originally published as part of the Höglund dataset) and 3056 multi-localized proteins (originally published as part of the DBMLoc set). This dataset was used by Briesemeister et al. in their extensive comparison of multilocalization prediction system. The experimental results indicate that the system based on SVM with the Laplace kernel, termed LKLoc, not only achieves a higher accuracy than the system using other kernels but also shows significantly better results than those obtained from other top systems (MDLoc, BNCs, YLoc+). The source code of this prediction system is available upon request.