Studying cortical anatomy by examining the deepest part of cortical sulci, the sulcal pits, has recently raised a growing interest. In particular, constructing structural representations from patterns of pits has proved a promising approach. This study follows up in this direction and brings two main contri-butions. First, we introduce a graph kernel adapted to sulcal pit graphs, in order to perform classification of patterns of sul-cal pits using support vector machines directly in graph space. Second, we design a multivariate searchlight technique that enables the localization of informative patterns of sulcal pits. We demonstrate the relevance of our approach by studying cortical differences between male and female subjects using a large dataset of 134 subjects.