Animal models have facilitated the investigation of the mechanisms of several diseases. For drug addiction in particular, inbred mouse strains have proved to be invaluable [1, 2], and have facilitated the mapping of aspects of addiction-related behaviour to specific genetic loci. Inbred Long Sleep (ILS) and Inbred Short Sleep (ISS) mice, for instance, present many contrasts with respect to a number of alcoholism related phenotypes [3–6]. They have been widely used to model ethanol sensitivity [7, 8]. Ethanol sensitivity has a genetic basis [9], the comprehensive workings of which remain elusive. Consequently, a comparison of relevant brain region transcriptomes of ILS and ISS mice has the potential of revealing unique patterns of gene expression [10] that could be relevant to the mechanisms of alcoholism.

The cerebellum has long been almost exclusively associated with balance and motor co-ordination. It has relatively recently been found to be more involved with cognition than previously thought [11]. During neurodevelopment, the cerebellum is especially susceptible to ethanol toxicity [12]. Studies indicate a role for activation of the cerebellum in alcoholism. A Functional Magnetic Resonance Imaging study has indicated that ethanol odour-induced craving in untreated recently abstinent male alcoholics involves activation of the cerebellum along with the subcortical-limbic region of the right amygdala/hippocampal area [13]. Positron Emission Tomography studies in drug addiction similarly indicate a role for cerebellar activation [14, 15]. The identification of specific pathways contributing to alcoholism-related events in the cerebellum would, therefore, be important.

The phosphoinositide (PI) and cyclic adenosine 3',5'-monophosphate (cAMP) signalling pathways have long been thought to be important in the development of ethanol dependence and tolerance [16]. There are several pieces of evidence suggesting a role for the cAMP/protein kinase A (PKA)/cAMP-response-element-binding protein (CREB) signalling pathway in addiction, even though they do not necessarily involve the cerebellum: Alcohol preferring (P) rats have lower levels of CREB and the transcriptionally-active phospho-CREB in the medial amygdala and central amygdala (CeA) than non-preferring (NP) rats [17]. Ethanol administration (or PKA activator [Sp-cAMP] administration into the CeA) increases CREB function in the CeA of P (but not NP) rats. Also, 24 hours following a single intra-peritoneal 2 mg/kg ethanol dose to C57BL/6J mice, there is long-term potentiation of GABA synaptic transmission at Ventral Tegmental Area dopaminergic neurons, via a cAMP-PKA-dependent mechanism [18]. One mechanism by which ethanol increases CREB levels involves inhibition of adenosine reuptake which results in increases in extracellular adenosine and activation of the adenosine A2 receptor, leading to increases in cAMP levels [19]. The ethanol-induced increase in CRE-mediated gene transcription requires PKA and involves an adenosine receptor-dependent phase and a later adenosine receptor-independent phase [20].

The emergence of high throughput data has facilitated the study of patterns of transcription. Machine Learning (ML) is one such avenue for mining such data [21]. It concentrates on methods for computer programs to improve their performance (i.e. modifying behaviour) by learning from previous data examples. ML is useful for the purpose of class prediction. During the learning process, structural patterns in the "training set" are established; these then constitute the basis upon which predictions are made when presented with data of unknown classification ("test set").

Four ML schemes were evaluated: a Decision Tree (J48, an implementation of the C4.5 algorithm), a Support Vector Machine (SVM), a Naïve Bayes classifier (NN) and a Multi-layer Perceptron (MLP). Two alternate models for ML were tested in this study, using a dataset of 46 instances and two classes. These were:

• -a two-class model with classifications: "CREB-regulated" and "NOT CREB-regulated", and

• -a three-class model with a third classification "Nrf2-regulated" [24]

Nrf2 (NF-E2-related factor 2), the primary transcription factor that binds the Antioxidant Response Element (ARE), was selected because, like CREB, Nrf2 is a ubiquitous transcription factor. Secondly, it has a requirement for CREB Binding Protein for enhanced transcription activity [25]. Using the leave-one-out cross-validation technique, the two-class model had lower Mean Absolute Error rates for all learning schemes explored than the three-class model (Figure 1A). Also, of the four schemes and two models evaluated, the area under the Receiver Operating Characteristic (ROC) curve, a measure of test accuracy, was highest for the C4.5 scheme under the two-class model (Figure 1A).

Of the four ML schemes, using the leave-one-out cross-validation technique and the two-class model, the C4.5 Decision Tree algorithm had the lowest overall predicted error rate (Figure 1B; Table 2). Its ROC curve was closest to the left-hand border and the top border of the ROC space (Figure 2 and Additional File 1), indicating that it had the most optimal trade-off between sensitivity and specificity among the four schemes evaluated. It also had the highest area under the ROC curve (Table 3). The C4.5 Decision Tree algorithm [26] works top-down, seeking at each stage an attribute that best separates the classes. The attribute with the greatest information gain is chosen. It then recursively processes the sub-problems resulting from the split until the information either reaches a maximum or is zero. The information measure (entropy) is calculated thus:

	C4.5	SVM	NN	MLP
PERCENT CORRECT	69.57 (46.06)	50.00 (50.05) *	58.70 (49.29)	55.00 (49.80)
MEAN ABSOLUTE ERROR	0.18 (0.28)	0.33 (0.11) *	0.28 (0.33)	0.30 (0.30) *
RELATIVE ABSOLUTE ERROR	52.51 (80.99)	96.00 (32.03) *	79.30 (94.63)	85.53 (86.77) *
ROOT MEAN SQUARED ERROR	0.22 (0.34)	0.41 (0.14) *	0.34 (0.40)	0.36 (0.37) *
ROOT RELATIVE SQUARED ERROR	53.54 (82.59)	97.90 (32.67) *	80.86 (96.50)	87.11 (88.44) *

The standard deviation of each attribute evaluated is located in brackets

**Leave-one-out technique, i.e. 46-fold cross-validation, performed with ten iterations each

*Use of Corrected Resampled T-test [44]; difference from corresponding C4.5 value is statistically significant (p = 0.05, two tailed)

	INDEPENDENT EVALUATION SET	TEN-FOLD CROSS VALIDATION	LEAVE-ONE-OUT
C4.5	0.8563	0.7722	0.7883
NN	0.7875	0.5936	0.6352
SVM	0.9063	0.5217	0.5
MLP	0.85	0.6711	0.5577

Entropy (p₁, p₂, .... p_n) = -p₁log ₂ p₁-p₂log ₂ p₂....-p_nlog ₂ p_n

where p₁, p₂, .... p_n are fractions representing the data distribution at a node (attribute) and sum up to 1.

F-measure = (2 * Precision * Sensitivity)/(Precision + Sensitivity)

where Precision = True Positives/(True Positives + False Positives)

Sensitivity (or Recall) is a measure of the probability that the test would reject a false null hypothesis:

Sensitivity = True Positives/(True Positives + False Negatives)

Lowered sensitivity to the effects of ethanol increases the risk of developing alcoholism. Differing sensitivities to ethanol is, at least in part, attributable to heredity [9], and inbred mouse strains have facilitated the investigation of this complex behavioral phenomenon. In studying CREB's gene regulating activity in ethanol sensitivity, a set of differentially expressed genes in the ILS/ISS mouse model of ethanol sensitivity were examined. The two-class model had lower error rates than the three-class model (Figure 1A). This is probably due to the inherent difficulty of distinguishing between the classifications "CREB-regulated" and "Nrf2-regulated". Indeed the case can be made that Nrf2 genes are dependent on CREB for enhanced transcription activity [25]. The complexity of the machinery for transcription makes the two-class model the preferred model for this study.

Properties of stretches of nucleotides can impact their affinity for specific transcription factors; this principle can be exploited for its therapeutic promise [27]. A central premise of this observation is the fact that the characteristics of individual nucleotide bases in any such oligonucleotide contribute to its structure and function [28]. As an example, hydrogen-bonded base pairs help determine the structure and function of nucleic acids. Strength of hydrogen bonding and other nucleotide base classifications used in generating the characteristics of each gene's promoter sequence for this ML study have been outlined in Table 1.

Of the four learning schemes evaluated using the two-class model, C4.5 was the most consistent performer, having the lowest overall error rates (Figure 1B), and the highest accuracy (Figure 2; Table 3), area under the ROC curves being measures of test accuracy. Because of variability between independent evaluation sets, performance evaluations based on evaluation sets are only instructive when such evaluation sets are large in size. Since the evaluation set used consisted of only 21 instances, the cross-validation techniques are better indicators of each learning scheme's performance. The Ten-fold Cross Validation technique is a standard way for predicting the error rate of a learning scheme [29, 30]. When applying this technique, an average value is obtained for ten different sets of the re-organized data such that in each case, 90% of the data is used for training and 10% used for testing. The leave-one-out technique is, in essence, an n-fold cross-validation technique (n being the number of instances in the dataset) and, for a small dataset, a good predictor of a scheme's performance on an independent dataset. In this study, 81% of genes of known classification used as an evaluation set were correctly classified by C4.5 (Table 4), with F-measures of 0.87 and 0.67 respectively for the classes "CREB-regulated" and "NOT CREB-regulated" respectively.

The stretch of nucleotides between the cAMP Response Element (CRE) and the Transcription Start Site (TSS) and the stretch between the CRE and the Transcription Factor II D (TFIID) bind site were identified as important determinants of a gene's CREB regulation status (Figure 3). Two types of CRE with different affinities for the transcription factor CREB have been reported. One class containing the symmetrical TGACGTCA site shows a high binding affinity for CREB; the other type has asymmetric and weak binding sites ("CGTCA") [31]. The TATA-binding protein (TBP) and TBP-associated factors (TAFs) constitute the TFIID complex. The TFIID complex is a major component of the general RNA polymerase II (RNAP II) transcription machinery with intrinsic sequence-specific DNA-binding activity [32]. The binding of TFIID to a gene's core promoter region is an important rate-limiting step in the assembly of the transcription initiation complex. With the notable exception of the stretch between the CRE and the TFIID bind site, CREB target promoter regions have relatively high levels of nucleotide bases with strong Hydrogen Bonding (data not shown).

Taken together, the observations made suggest that, in the cerebellum, CREB plays a key role in ethanol sensitivity and presents the field with a central hypothesis that needs to be further tested. CREB's role in mediating a number of complex behaviours has been documented [33]. Events in the extended amygdala have long been associated with the reinforcing effects of addicting drugs [36]. It is evident that the cerebellum, though less well studied in this regard, is involved in addiction [13–15]. Since CREB's transcription regulating activity differs from cell type to cell type [37], pursuit of the implications of a key role for CREB in this addiction model's cerebellar molecular milieu would be both promising and instructive.