Multivariate pattern analysis has been applied successfully to task-based and resting-based fMRI studies to investigate which neural markers distinguish individuals with autistic spectrum disorders (ASD) from controls. Given the wealth of existing fMRI datasets, there is a need for techniques that can merge information from multiple datasets. We propose a data-driven approach that we apply to 2 fMRI experiments acquired in the same group of ASD (n=15) and control subjects (n=14). Methods. The 1st experiment (Conty, Dezecache, Hugueville, & Grezes, 2012) addressed how ASD participants process anger expressions. We presented pictures of individuals expressing social cues (i.e. gaze) directed or not at the observer (factors gaze direction*emotion, total 4 conditions). The 2nd experiment (Pichon, De Gelder, & Grèzes, 2012) aimed at distinguishing between automaticity and attention in the processing of emotional stimuli varying attention while observing movie-clips depicting angry, fear and neutral body actions (factors attention*emotion, total 4 conditions). For each subject and experiment, we collected 370 EPI volumes (TR=2000ms). fMRI images were processed and analyzed using SPM8. Motion artefacts were taken into account by adding 24 motion-related regressors in each models (Friston, Williams, Howard, Frackowiak, & Turner, 1996). We first classified ASD from controls (figure 1). We trained a linear Support Vector Machine (SVM) independently for each condition using the raw beta maps as inputs. The continuous outputs of the classifiers were then averaged to obtain a final decision for each subject. The accuracy was computed for each experiment (i.e. combination of 4 conditions) and for the fusion of the two experiments (i.e. combination of 8 conditions). The significance of the accuracy was computed using a binomial test. Although SVMs achieve good performance in high dimensional spaces, they can still benefit from feature selection methods. In addition, feature selection allows investigating the location of voxel activities driving the classification performance. Therefore we employed SVM Recursive Feature Elimination (SVM RFE) (Guyon, Weston, Barnhill, & Vapnik, 2002) to order features by relevance and restrain the classification to a subset of discriminant voxels. To identify the most discriminative voxels across experiments (i.e. voxels which are well ranked by SVM RFE on several conditions), we employed an approximation of the rank product test (Koziol, 2010). Results. The accuracy, sensitivity and specificity of the different classification methods are reported in Table1. Regarding accuracy, the fusion of experiments reflected the accuracy of the most discriminative experiment. The fusion dramatically increased the significance of discriminative features and increased the number of significant discriminative voxels by roughly 50%. Discriminative voxels (figure2) were found in regions related to social cognition (FFA, OFA, EBA, STS, TPJ and premotor). These regions showed reduced contribution in ASD compared to controls. Feature selection did not improve accuracy but helped reducing the correlation of classifiers' output with motion (figure 3). Conclusions. These results show that the method is efficient for selecting features that are common across experiments, for identifying discriminative brain activity and for reducing motion artefacts. Identified areas were astonishingly consistent with brain regions of the “social-brain” known to show aberrant functioning in ASD. One of the originalities of our method is that the computed statistic is based only on the features ranks. Consequently, this method could be used to combine heterogeneous data sources such as different fMRI experiments, different BOLD-related signals (i.e. activation magnitude, connectivity) or even different brain imaging modalities (i.e. structural MRI, voxel-based morphometry, functional MRI, PET).