Results from both simulated and real datasets showed that the structure of a dataset influences to a large extent the efficiency of the methods that use projection on discriminant axes.

In testing a new method, simulated and real datasets play complementary roles. Simulation of data with known properties is useful to study the influence of the dataset characteristics and the performance of a given method, and could be considered as a practical guide to understand results from real situations. For choosing an analysis method to discriminate two groups of patients, we think it is necessary to have a prior examination of the structure of the data to analyze. This will enable an informed choice between the available methods.

We propose here a new simulation approach that allows exploring known structures with control through several parameters. Nguyen [26] proposed to simulate datasets to compare the performance of PCA and PLS as prior procedure before logistic discrimination. However, his method of simulation did not allow a discussion on the influence of the data structure. Our simulations allow generating different structures of different degrees of complexity and assessing the impact of three parameters: the distance between the clusters, the eccentricity of these clusters, and their relative positions in a two-dimensional component space. The major source of complexity in real microarray datasets is the existence of regulation networks. In our simulations, this may be described by a component with a very large variance; that is, a large eccentricity. This corresponds usually to a common effect on all the genes. A high variance on one component corresponds also to a cluster of highly correlated genes. Whether a network of genes exists or not would determine the relative importance of the other components with respect to the first one. Nevertheless, we are aware that our simulations have limits. Therefore, a compromise has to be found between the uncontrolled nature of real datasets and the controlled nature of simulated datasets as research tools. This will be the object of future works.

The use of real datasets to prove the superiority of any method should be considered with caution. For example, the leukaemia dataset from Golub, very often used to demonstrate the efficiency of a new method, may not be used for that purpose because of its very strong between-group structure. This structure is such that we expect the groups to be distinguished whatever the method used (e.g., BGA that simply joins the barycenters of the groups). We believe that, in such situations, the good performance of a particular method does not only inform on its ability to discriminate between groups. If the structure of the dataset had been previously examined before its analysis, for example with the graphical tool we propose, this dataset would not have been chosen to validate new prediction methods. Thus, bioinformaticians should be cautious in choosing the datasets to use for method comparisons. The proposed visualization tool helps in choosing the dataset, by having an idea of its structure. The prostate or ALL datasets for example may be appropriate for that purpose.

Besides, the structure of a given dataset may depend on the type of disease. In diagnosis, some pathophysiological entities may be already clearly identified; if their origin is a metabolic activation, they will induce different processes that will be easy to distinguish (e.g., ALL vs. AML). However, differentiating patients with or without multidrug resistance may be even more difficult because no pathophysiological entities are involved. In prognosis, distinguishing good from bad prognosis patients would be more difficult because they often share the same pathophysiological characteristics.

Three main configurations of the data structure may be identified. When the clusters of points are quite distinct the between-group difference is so obvious that the within-group structure will have no impact; BGA and DA will give good prediction results. The simple method that consists in drawing an axis between the barycenters is sufficient. In fact, the way of projecting patients on the discriminant axis does not come into consideration. On the opposite, there are situations in which both methods are inappropriate. This corresponds to superposed clusters of points obtained in plotting the within-group versus the between-group coordinates. In other situations, we believe that DA is more advantageous than BGA because it allows taking into account the partition of the total variance into between and within variances. However, in case the variances of the two groups are not the same, the total variance will not reflect the variance in each group, so there will be no advantage of favoring DA over BGA. Moreover, keeping more than one component in the first dimension reduction step using PLS or PCA is a way to capture more information than the single projection in BGA, particularly with PLS. This is illustrated with the ALL dataset; by keeping ten PLS components, DA outperforms BGA to a large extent (respectively 0.97% and 0.70% of well-classified patients). These observations illustrate the fact that the first PLS component and the BGA discriminant axis are identical. This was demonstrated by Barker and Rayens [27], and by Boulesteix [12]. Thus, using PLS with one component followed by DA gives a final component that is collinear to that of PLS alone, and also to the BGA axis. This is illustrated with the leukaemia dataset, where PLS+DA and BGA give equivalent results (respectively 0.97% and 0.98% of well-classified patients). However, in simulations, PLS+DA seemed to yield, on average, slightly better results than BGA. In fact, due to random sampling, some simulated datasets needed more than one component to optimize prediction because dimensions other than those simulated may be informative by chance alone. Note that in case of a spherical cluster of points, a second PLS component will not capture more information than the first one and both methods will be equally efficient.

Overall, DA becomes advantageous when the structure of the variance is such that the way of projecting patients on the discriminant axis needs to come into consideration. This leads to conclude that DA is the most suitable method; it provides better or at least equivalent results in a diversity of datasets because it ensures that the within-group variance will be taken into account, when relevant. The diversity of real datasets encountered confirms the fact that, unlike DA, BGA is unable to deal with too complex data structures. The only advantage of BGA is its ease of use and interpretation: a single projection enables to go from the original variable space to a one-dimension axis on which inter-group variance is maximum.

This axis is also a direct linear combination of genes where a high coefficient means that the gene is important to classify the patients into one of the groups. With DA, the samples are first expressed in a component space, which makes interpretation more difficult.

BGA and DA used with more than two groups provide *k* - 1 discriminant axes, which enables each of the *k* groups to be separated from the *k* - 1 others. By plotting these groups in successive two-dimensional graphs, the structure assessment described here may be applied to each of the two-dimension spaces so obtained.