Nobel prize winner Eric Kandel wrote: “The task of neural science is to explain behavior in terms of the activities of the brain.”¹ Unfortunately, the currently prevalent data analysis strategies do not aim at exploring behavior in terms of neural activity per se. Instead, the majority of methods primarily explore the data by performing mass-univariate hypothesis tests, searching for statistically significant excursions of the signal from a ‘no-effect’ baseline. Such approaches often rely on restrictive modeling assumptions: e.g. the forward model of a hemodynamic (blood circulation) response function in functional magnetic resonance imaging (fMRI). Because of this, they require pre-processing steps (spatial and temporal smearing, averaging, etc.) that necessarily ignore or obliterate some information embedded in the data. Furthermore, univariate modeling of the acquired signal in terms of behavioral factors neither considers present covariance and causal structure among distinct brain areas, nor does it account for the variance of the response patterns across trials.

In recent fMRI-based research, these limitations have led to a reconsideration² of multivariate pattern analysis (MVPA) methods that had been introduced more than a decade ago in studies employing positron emission tomography.^3,4 Enabled by recent advances in the field of statistical learning theory, some striking developments have attracted considerable interest throughout the neuroscience community.^5–8 For instance, the application of regularized statistical classifiers (e.g., a support vector machine⁹ or SVM) allowed the reliable prediction of behavioral conditions based on full-brain fMRI data¹⁰ for each single trial. This reversal of the analysis strategy, where now aspects of behavior are modeled in terms of neural activity, represents a critical difference from previously established approaches (see Figure 1).

Reversing the analysis flow: classical statistical parametric mapping (SPM) performs mass-univariate testing to localize hypothetical brain responses. In contrast, multivariate pattern analysis (MVPA) offers a direct quantifiable mapping⁸ from brain activity patterns onto behavioral states.

Despite the advantages and promise of these methods, various factors have delayed their adoption. Although a growing number of studies now employ statistical learning methods, the compressed verbal descriptions of the novel and rather complex analysis pipelines—coupled with the lack of a unified and flexible software framework—have hindered straightforward replication attempts. Nevertheless, replication—and hence validation of reported results by independent research groups—is essential for scientific progress.

To provide the neuroscience community with an adequate tool for the analysis of neural data using statistical learning methods, we have developed PyMVPA (Python MVPA¹¹). This is a free, open-source, and platform-agnostic project that utilizes the Python programming language. Python is a perfect choice because of its portability, its concise and descriptive syntax, and its ability to easily interface to low-level libraries and high-level scientific scripting environments, such as R.¹² PyMVPA makes it easy to access data stored in standard data formats (e.g., NIfTI), to perform typical statistical learning procedures (such as training, testing, feature selection, and cross-validation without ‘peeking’ or ‘double-dipping’),¹³ while exploring the multitude of available learning methods and facilitating rapid development. It also makes it easy to allow contributions from any interested researcher.

We designed PyMVPA to offer a high-level programming interface that allows for a flexible combination of the provided building blocks to express complex analysis pipelines in just a few lines of code.¹² This feature enables researchers to easily replicate existing studies, and to carry out novel non-standard analyses. Moreover, the descriptive power of human-readable, yet compact, source code opens the possibility of including the complete source code of a study as a supplemental material of a publication. (Mandatory code-inclusion research papers could tremendously expedite verification and adoption of novel analysis strategies.)

To demonstrate the power and applicability of the suggested analysis methodologies we¹⁴ analyzed data from four different neural modalities and accompanied the publication with the complete source code of all of them. Essentially the same workflow was used for all neural data modalities: basic preprocessing, training and testing (by cross-validation) of statistical classifiers, and the analysis of the trained classifiers sensitivities with respect to any given input dimension. Applied to extracellular recordings data (post-stimulus time histograms of spike counts) it was possible to reliably identify eight original auditory stimuli conditions for single trials, and to obtain an assessment of the relevance of any given neuron to the processing of stimulus conditions. Applied to electroencephalography (EEG) data from a visual processing experiment, it was possible not only to confirm results of conventional event-related potential (ERP) analysis, but also to discover a late response component not revealed by ERPs. Applied to fMRI data from an event-related visual object processing experiment,¹⁵ PyMVPA allowed to identify the original stimulus condition of each trial, and to provide spatio-temporal category specificity profiles without imposing any specific hemodynamic response model.

MVPA methods are in no way limited to processing data from one modality at a time. For example, a reliable description of fMRI data in terms of a simultaneously recorded EEG signal (see Figure 2) allows for identification of areas that are active during any given task, and localization of generators and covariates of dominant EEG frequency bands.¹⁶ Furthermore, the constructed EEG-to-fMRI mapping can be used for filtering of fMRI and EEG signals, and for EEG-driven interpolation of fMRI timeseries.

Reliable mapping from EEG onto fMRI data.¹⁶ The upper plot outlines the analysis workflow, where for each fMRI voxel a mapper (multiple regression) is trained on the joint EEG signal. The lower part shows thresholded maps of correlation coefficients between predicted and actual fMRI data from an auditory experiment.

To improve the understanding of brain function, neuroscience research requires versatile computing environments and advanced methods that make efficient use of acquired data. Methods developed in the domain of machine and statistical learning are generic, powerful, and their application to neural research has already provided new insights about the brain. Our PyMVPA analysis framework aims to provide a convenient, extensive, and expandable environment to apply existing and to develop new methods for the analysis of neural data. PyMVPA's user base has been constantly growing and new data analysis methods and methodologies are continuously added to the framework. Future development will further enrich the available techniques and offer promising analysis strategies. One of the immediate next steps will allow for an improved transparent and unbiased model selection. This new functionality will especially help to apply complex non-linear methods while ensuring valid results.¹⁷

2. E. R. Kandel, J. H. Schwartz and T. M. Jessell, Principles of Neural Science 4^th ed., McGraw-Hill, New York, 2000.
   
   
3. J. Haxby, M. Gobbini, M. Furey, A. Ishai, J. Schouten and P. Pietrini, Distributed and overlapping representations of faces and objects in ventral temporal cortex, Science 293, pp. 2, 2001.
   
   
4. J. R. Moeller and S. Strother, A regional covariance approach to the analysis of functional patterns in positron emission tomographic data, J. Cerebral Blood Flow and Metabolism 11, pp. 121-135, 1991.
   
   
5. J. S. Kippenhahn, W. W. Barker, S. Pascal, J. Nagel and R. Duara, Evaluation of a neural-network classifier for PET scans of normal and Alzheimer's disease subjects, J. Nuclear Medicine 33, pp. 1, 1992.
   
   
6. S. Hanson, T. Matsuka and J. Haxby, Combinatorial codes in ventral temporal lobe for object recognition: Haxby (2001) revisited: is there a “face” area?, NeuroImage 23, pp. 156-166, 2004.
   
   
7. K. A. Norman, S. M. Polyn, G. J. Detre and J. V. Haxby, Beyond mind-reading: multi-voxel pattern analysis of fMRI data, Trends in Cog. Sci. 10, pp. 424-430, 2006.
   
   
8. J.-D. Haynes and G. Rees, Decoding mental states from brain activity in humans, Nature Reviews Neuroscience 7, pp. 523-534, 2006.
   
   
9. A. J. O'Toole, F. Jiang, H. Abdi, N. Penard, J. P. Dunlop and M. A. Parent, Theoretical, statistical, and practical perspectives on pattern-based classification approaches to the analysis of functional neuroimaging data, J. Cog. Neurosci. 19, pp. 1, 2007.
   
   
10. V. Vapnik, The Nature of Statistical Learning Theory, Springer, New York, 1995.
    
    
11. S. J. Hanson and Y. O. Halchenko, Brain reading using full brain support vector machines for object recognition: there is no “face” identification area, Neural Comp. 20, pp. 486-503, 2008.
    
    
12. http://www.pymvpa.org
    
    
13. M. Hanke, Y. O. Halchenko, P. B. Sederberg, S. J. Hanson, J. V. Haxby and S. Pollmann, PyMVPA: A Python toolbox for multivariate pattern analysis of fMRI data, Neuroinformatics 7 (1), pp. 37-53 Mar, 2009.
    
    
14. N. Kriegeskorte, W. K. Simmons, P. S. F. Bellgowan and C. I. Baker, Circular analysis in systems neuroscience: the dangers of double dipping, Nature Neuroscience 12 (5), pp. 535-540, 2009.
    
    
15. M. Hanke, Y. O. Halchenko, P. B. Sederberg, E. Olivetti, I. Fründ, J. W. Rieger, C. S. Herrmann, J. V. Haxby, S. J. Hanson and S. Pollmann, PyMVPA: A Unifying Approach to the Analysis of Neuroscientific Data, Front. Neuroinformatics 3 (3), 2009.
    
    
16. M. Hanke, Advancing the Understanding of Brain Function with Multivariate Pattern Analysis, PhD thesis Otto-von-Guericke-University Magdeburg, Germany May, 2009.
    
    
17. Y. O. Halchenko, Predictive Decoding of Neural Data, PhD thesis NJIT Newark, NJ, USA May, 2009. http://www.onerussian.com/Sci/disser/yoh-phd09.pdf
    
    
18. F. Pereira, T. Mitchell and M. Botvinick, Machine learning classifiers and fMRI: A tutorial overview, NeuroImage 45, pp. 199-209, 2009.