Information
-
Patent Application
-
20040083084
-
Publication Number
20040083084
-
Date Filed
November 12, 200221 years ago
-
Date Published
April 29, 200420 years ago
-
CPC
-
US Classifications
-
International Classifications
Abstract
The statistical analysis of the invention is a predictive statistical tree model that overcomes several problems observed in prior statistical models and regression analyses, while ensuring greater accuracy and predictive capabilities. The claimed model can be used for a variety of applications including the prediction of disease states, susceptibility of disease states or any other biological state of interest, as well as other applicable non-biological states of interest. The model as applied to genetic applications generates a statistically significant number of cluster-derived singular factors called metagenes, that characterize multiple patterns of expression of the genes across samples. Formal predictive analysis then uses the metagenes in a Bayesian classification tree analysis which generates multiple recursive partitions of the sample into subgroups (the “leaves” of the classification tree), and associates Bayesian predictive probabilities of outcomes with each subgroup. Overall predictions for an individual sample are then generated by averaging predictions, with appropriate weights, across many such tree models. The model includes the use of iterative out-of-sample, cross-validation predictions leaving each sample out of the data set one at a time, refitting the model from the remaining samples and using it to predict the hold-out case. This rigorously tests the predictive value of a model and mirrors the real-world prognostic context where prediction of new cases as they arise is the major goal.
Description
FIELD OF THE INVENTION
[0001] The field of this invention is the application of classification tree models incorporating Bayesian analysis to the statistical prediction of binary outcomes.
BACKGROUND OF THE INVENTION
[0002] Bayesian analysis is an approach to statistical analysis that is based on the Bayes's law, which states that the posterior probability of a parameter p is proportional to the prior probability of parameter p multiplied by the likelihood of p derived from the data collected. This increasingly popular methodology represents an alternative to the traditional (or frequentist probability) approach: whereas the latter attempts to establish confidence intervals around parameters, and/or falsify a-priori null-hypotheses, the Bayesian approach attempts to keep track of how a-priori expectations about some phenomenon of interest can be refined, and how observed data can be integrated with such a-priori beliefs, to arrive at updated posterior expectations about the phenomenon.
SUMMARY OF THE INVENTION
[0003] This invention discusses the generation and exploration of classification tree models, with particular interest in problems involving many predictors. Problems involving multiple predictors arise in situations where the prediction of an outcome is dependent on the interaction of numerous factors (predictors), such as the prediction of clinical or physiological states using various forms of molecular data. One motivating application is molecular phenotyping using gene expression and other forms of molecular data as predictors of a clinical or physiological state.
[0004] The invention addresses the specific context of a binary response Z and many predictors xi; in which the data arises via case-control design, i.e., the numbers of 0/1 values in the response data are fixed by design. This allows for the successful relation of large-scale gene expression data (the predictors) to binary outcomes, such as a risk group or disease state. The invention elaborates on a Bayesian analysis of this particular binary context, with several key innovations. The analysis of this invention addresses and incorporates case-control design issues in the assessment of association between predictors and outcome with nodes of a tree. With categorical or continuous covariates, this is based on an underlying non-parametric model for the conditional distribution of predictor values given outcomes, consistent with the case-control design. This uses sequences of Bayes' factor based tests of association to rank and select predictors that define significant “splits” of nodes, and that provides an approach to forward generation of trees that is generally conservative in generating trees that are effectively self-pruning. An innovative element of the invention is the implementation of a tree-spawning method to generate multiple trees with the aim of finding classes of trees with high marginal likelihoods, and where the prediction is based on model averaging, i.e., weighting predictions of trees by their implied posterior probabilities. The advantage of the Bayesian approach is that rather than identifying a single “best” tree, a score is attached to all possible trees and those trees which are very unlikely are excluded. Posterior and predictive distributions are evaluated at each node and at the leaves of each tree, and feed into both the evaluation and interpretation tree by tree, and the averaging of predictions across trees for future cases to be predicted. To demonstrate the utility and advantages of this tree classification model, several embodiments are provided. The first concerns the prediction of levels of fat content (higher than average versus lower than average) of biscuits based on reflectance spectral measures of the raw dough. The second and third examples concern gene expression profiling using DNA microarray data as predictors of a clinical states in breast cancer. The clinical states include estrogen receptor (“ER”) prediction, tumor recurrence, and lymph node metastases. The example of ER status prediction demonstrates not only predictive value but also the utility of the tree modeling framework in aiding exploratory analysis that identify multiple, related aspects of gene expression patterns related to a binary outcome, with some interesting interpretation and insights. Embodiments 2 through 4 also illustrate the use of metagene factors—multiple, aggregate measures of complex gene expression patterns—in a predictive modeling context. The fourth embodiment relates to the prediction of atherosclerotic phenotype determinative genes.
[0005] In the case of large numbers of candidate predictors, in particular, model sensitivity to changes in selected subsets of predictors are ameliorated though the generation of multiple trees, and relevant, data-weighted averaging over multiple trees in prediction. The development of formal, simulation-based analyses of such models provides ways of dealing with the issues of high collinearity among multiple subsets of predictors, and challenging computational issues.
BRIEF DESCRIPTION OF THE FIGURES
[0006]
FIG. 1: An example prediction tree for cookie fat outcomes. The root node splits on predictor/factor 92, followed by two subsequent splits on additional predictors 330 and 305. The Π values are point estimates of the predictive probabilities of high fat versus low fat at each of the nodes, with suffixes simply indexing nodes. The labels Z(0=1) indicate the numbers of low fat (0) and high fat (1) samples within each node, and the F# symbols indicate the thresholds that define the predictor based splits within each node.
[0007]
FIG. 2: Two predictive factors in cookie dough analysis. All samples are represented by index numbers 1 through 78. Training data are denoted by blue (low fat) and red (high fat), and validation data by cyan (low fat) and magenta (high fat). The two full lines (black)demark the thresholds on the two predictors in this example tree.
[0008]
FIG. 3: Scatter plot of cookie data on three factors in example tree. Samples are denoted by blue (low fat) and red (high fat), with training data represented by filled circles and validation data by open circles.
[0009]
FIG. 4: Three ER related metagenes in 49 primary breast tumors. Samples are denoted by blue (ER negative) and red (ER positive), with training data represented by filled circles and validation data by open circles.
[0010]
FIG. 5: Three ER related metagenes in 49 primary breast tumors. All samples are represented by index number in 1-78. Training data are denoted by blue (ER negative) and red (ER positive), and validation data by cyan (ER negative) and magenta (ER positive).
[0011]
FIG. 6: Honest predictions of ER status of breast tumors. Predictive probabilities are indicated, for each tumor, by the index number on the vertical probability scale, together with an approximate 90% uncertainty interval about the estimated probability. All probabilities are referenced to a notional initial probability (incidence rate) of 0.5 for comparison. Training data are denoted by blue (ER negative) and red (ER positive), and validation data by cyan (ER negative) and magenta (ER positive).
[0012]
FIG. 7: Table of 491 ER metagenes in initial (random) order.
[0013]
FIG. 8: Table of 491 ER metagenes ordered in terms of nonlinear association with ER status.
[0014]
FIG. 9: Cross-validation probability predictions of lymph node status. Samples (tumors) are plotted by index number, and the plotted numbers are marked on the vertical scale at the estimated predictive probabilities of high risk (red) versus low risk (blue). Approximate 90% uncertainty(?confidence) intervals about these estimated probabilities are indicated by vertical dashed lines.
[0015]
FIG. 10: Gene expression patterns from the major metagene that predicts lymph node status. Samples are plotted by sample index number and by color (color coding as in FIG. 9).
[0016]
FIG. 11: Cross-validation probability predictions of 3-year recurrence. Samples (tumors) are plotted by index number, and the plotted numbers are marked on the vertical scale at the estimated predictive probabilities of 3 year recurrence (red) versus 3 year recurrence free survival (blue). Approximate 90% uncertainty intervals about these estimated probabilities are indicated by vertical dashed lines.
[0017]
FIG. 12: Genes associated with metagene predictors of lymph node metastasis
[0018]
FIG. 13: Genes associated with metagene predictors of breast cancer recurrence.
DETAILED DESCRIPTION OF THE INVENTION
Development of the Tree Clarification Model: Model Context and Methodology
[0019] Data {Zi, xi} (i=1, . . . ,n) are available on a binary response variable Z and a p-dimensional covariate vector x: The 0/1 response totals are fixed by design. Each predictor variable xj could be binary, discrete or continuous.
1. Bayes' Factor Measures of Association
[0020] At the heart of a classification tree is the assessment of association between each predictor and the response in subsamples, and we first consider this at a general level in the full sample. For any chosen single predictor x; a specified threshold on the levels of x organizes the data into the 2×2 table.
1|
|
Z = 0Z = 1
|
|
x ≦ rn00n01N0
x > rn10n11N1
M0M1
|
[0021] With column totals fixed by design, the categorized data is properly viewed as two Bernoulli sequences within the two columns, hence sampling densities With column totals fixed by design, the categorized data is properly viewed as two Bernoulli sequences within the two columns, hence sampling densities
p
(n0z, n1z|Mz, θz,r)=θn0z(1−θz,r)n1z
[0022] for each column z=0, 1. Here, of course θ0,r=Pr(x≦|Z=0) and θz,rn=Pr(x≦r|Z=1). A test of association of the threshold predictor with the response will now be based on assessing the difference between the Bernoulli probabilities.
[0023] The natural Bayesian approach is via the Bayes' fact Br comparing the null hypothesis θo,r=θ1,r to the full alternative θo,r≈θ1,r. We adopt the standard conjugate beta prior model and require that the null hypothesis be nested within the alternative. Thus, assuming θo,r≈θ1,r, we take θo,r and θ1,r, to be independent with common prior Be(a,r,br) with mean mr,=a,r,/(a,r+br). On the null hypothesis θ0,r=θ1,r, the common value has the same beta prior. The resulting Bayes' factor in favour of the alternative over the null hypothesis is then simply
1
[0024] As a Bayes' factor, this is calibrated to a likelihood ratio scale. In contrast to more traditional significance tests and also likelihood ratio approaches, the Bayes' factor will tend to provide more conservative assessments of significance, consistent with the general conservative properties of proper Bayesian tests of null hypotheses (see Sellke, T., Bayarri, M. J. and Berger, J. O., Calibration of p_values for testing precise null hypotheses, The American Statistician, 55, 62-71, (2001) and references therein).
[0025] In the context of comparing predictors, the Bayes' factor Bτ may be evaluated for all predictors and, for each predictor, for any specified range of thresholds. As the threshold varies for a given predictor taking a range of (discrete or continuous) values, the Bayes' factor maps out a function of τ and high values identify ranges of interest for thresholding that predictor. For a binary predictor, of course, the only relevant threshold to consider is τ=0.
2. Model Consistency with Respect to Varying Thresholds
[0026] A key question arises as to the consistency of this analysis as we vary the thresholds. By construction, each probability θZτ is a non-decreasing function of τ, a constraint that must be formally represented in the model. The key point is that the beta prior specification must formally reflect this. To see how this is achieved, note first that θZτ is in fact the cumulative distribution function of the predictor values χ; conditional on Z=z; (z=0; 1); evaluated at the point χ=τ. Hence the sequence of beta priors, Be(aτ, bτ) as τ varies, represents a set of marginal prior distributions for the corresponding set of values of the cdfs. It is immediate that the natural embedding is in a non-parametric Dirichlet process model for the complete cdf. Thus the threshold-specific beta priors are consistent, and the resulting sets of Bayes' factors comparable as τ varies, under a Dirichlet process prior with the betas as margins. The required constraint is that the prior mean values mτ are themselves values of a cumulative distribution function on the range of χ, one that defines the prior mean of each θτ as a function. Thus, we simply rewrite the beta parameters (aτ, bτ) as aτ=amτ and bτ=a(1−mτ) for a specified prior mean cdf mτ, and where a is the prior precision (or “total mass”) of the underlying Dirichlet process model. Note that this specialises to a Dirichlet distribution when χ is discrete on a finite set of values, including special cases of ordered categories (such as arise if χ is truncated to a predefined set of bins), and also the extreme case of binary χ when the Dirichlet is a simple beta distribution.
3. Generating a Tree
[0027] The above development leads to a formal Bayes' factor measure of association that may be used in the generation of trees in a forward-selection process as implemented in traditional classification tree approaches. Consider a single tree and the data in a node that is a candidate for a binary split. Given the data in this node, construct a binary split based on a chosen (predictor, threshold) pair (χ, τ) by (a) finding the (predictor, threshold) combination that maximizes the Bayes' factor for a split, and (b) splitting if the resulting Bayes' factor is sufficiently large. By reference to a posterior probability scale with respect to a notional 50:50 3 prior, Bayes' factors of 2.2,2.9,3.7 and 5.3 correspond, approximately, to probabilities of 0.9, 0.95, 0.99 and 0.995, respectively. This guides the choice of threshold, which may be specified as a single value for each level of the tree. We have utilised Bayes' factor thresholds of around 3 in a range of analyses, as exemplified below. Higher thresholds limit the growth of trees by ensuring a more stringent test for splits.
[0028] The Bayes' factor measure will always generate less extreme values than corresponding generalized likelihood ratio tests (for example), and this can be especially marked when the sample sizes M0 and M1 are low. Thus the, propensity to split nodes is always generally lower than with traditional testing methods, especially with lower samples sizes, and hence the approach tends to be more conservative in extending existing trees. Post-generation pruning is therefore generally much less of an issue, and can in fact generally be ignored. Index the root node of any tree by zero, and consider the full data set of n observations, representing Mz outcomes with Z=z in 0,1. Label successive nodes sequentially: splitting the root node, the left branch terminates at node 1, the right branch at node 2; splitting node 1, the consequent left branch terminates at node 3, the right branch at node 4; splitting node 2, the consequent left branch terminates at node 5, and the right branch at node 6, and so forth. Any node in the tree is labelled numerically according to its “parent” node; that is, a node j splits into two children, namely the (left, right) children (2j+1; 2j+2): At level m of the tree (m=0; 1; : : : ;) the candidates nodes are, from left to right, as 2m—1; 1m; : : : ;2m+1−2.
[0029] Having generated a “current” tree, we run through each of the existing terminal nodes one at a time, and assess whether or not to create a further split at that node, stopping based on the above Bayes' factor criterion. Unless samples are very large (thousands) typical trees will rarely extend to more than three or four levels.
4. Inference and Prediction with a Single Tree
[0030] Suppose we have generated a tree with m levels; the tree has some number of terminal nodes up to the maximum possible of L=2m+1−2. Inference and prediction involves computations for branch probabilities and the predictive probabilities for new cases that these underlie. We detail this for a specific path down the tree, i.e., a sequence of nodes from the root node to a specified terminal node.
[0031] First, consider a node j that is split based on a (predictor, threshold) pair labeled (χj, τj), (note that we use the node index to label the chosen predictor, for clarity). Extend the notation of Section 2.1 to include the subscript j indexing this node. Then the data at this node involves M0j cases with Z=0 and M1j cases with Z=1. Based on the chosen (predictor, threshold) pair (χj, τj) these samples split into cases n00j, n01j, n10j, n11j as in the table of Section 2.1, but now indexed by the node label j. The implied conditional probabilities θz,τj=Pr(χj≦τj|A=z), for z=0, 1 are the branch probabilities defined by such a split (note that these are also conditional on the tree and data subsample in this node, though the notation does not explicitly reflect this for clarity). These are uncertain parameters and, following the development of Section 2.1, have specified beta priors, now also indexed by parent node j, i.e., Be(aτ,j, bτ,j). Assuming the node is split, the two sample Bernoulli setup implies conditional posterior distributions for these branch probability parameters: they are independent with posterior beta distributions
θ0,τ,j˜Be(aτ,j+n00j, bτ,j+n10j) and θ1,τj˜Be(aτ,j+n01j, bτ,j+n11j).
[0032] These distributions allow inference on branch probabilities, and feed into the predictive inference computations as follows.
[0033] Consider predicting the response Z* of a new case based on the observed set of predictor values x*. The specified tree defines a unique path from the root to the terminal node for this new case. To predict requires that we compute the posterior predictive probability for Z*=1/0. We do this by following x* down the tree to the implied terminal node, and sequentially building up the relevant likelihood ratio defined by successive (predictor, threshold) pairs.
[0034] For example and specificity, suppose that the predictor profile of this new case is such that the implied path traverses nodes 0, 1, 4, 9, terminating at node 9. This path is based on a (predictor, threshold) pair (χ0, τ0) that defines the split of the root node, (χ1, τ1)that defines the split of node 1, and (χ4, τ4) that defines the split of node 4. The new case follows this path as a result of its predictor values, in sequence: () and (). The implied likelihood ratio for Z=1 relative to Z=0 is then the product of the ration of branch probabilities to this terminal node, namely
2
[0035] Hence, for any specified prior probability a Pr(Z=1), this single tree model implies that, as a function of the branch probabilities, the updated probability is, on the odds scale, given by
3
[0036] Hence, for any specified prior probability πPr(Z*=1), this single tree model implies that, as a function the branch probabilities, the updated probability π* is, on the odds scale, give by
π* λ*=Pr
(Z* =1)
({overscore (1−π* ))} {overscore (Pr(Z* =0))}
[0037] The case-control design provides no information about Pr(Z*=1) so it is up to the user to specify this or examine a range of values; one useful summary is obtained by simply taking a 50:50 prior odds as benchmark, whereupon the posterior probability is
π*=λ*/(1+λ*).
[0038] Prediction follows by estimating π* based on the sequence of conditionally independent posterior distributions for the branch probabilities that define it. For example, simply “plugging-in” the conditional posterior means of each θ. will lead to a plug-in estimate of λ* and hence π*. The full posterior for π* is defined implicitly as it is a function of the θ.. Since the branch probabilities follow beta posteriors, it is trivial to draw Monte Carlo samples of the θ. and then simply compute the corresponding values of λ* and hence π* to generate a posterior sample for summarization. This way, we can evaluate simulation-based posterior means and uncertainty intervals for π* that represent predictions of the binary outcome for the new case.
5. Generating and Weighting Multiple Trees
[0039] In considering potential (predictor, threshold) candidates at any node, there may be a number with high Bayes' factors, so that multiple possible trees with difference splits at this node are suggested. With continuous predictor variables, small variations in an “interesting” threshold will generally lead to small changes in the Bayes' factor—moving the threshold so that a single observation moves from one side of the threshold to the other, for example. This relates naturally to the need to consider thresholds as parameters to be inferred; for a given predictor χ, multiple candidate splits with various different threshold values T reflects the inherent uncertainty about τ, and indicates the need to generate multiple trees to adequately represent that uncertainty. Hence, in such a situation, the tree generation can spawn multiple copies of the “current” tree, and then each will split the current node based on a different threshold for this predictor. Similarly, multiple trees may be spawned this way with the modification that they may involve different predictors.
[0040] In problems with many predictors, this naturally leads to the generation of many trees, often with small changes from one to the next, and the consequent need for careful development of tree-managing software to represent the multiple trees. In addition, there is then a need to develop inference and prediction in the context of multiple trees generated this way. The use of “forests of trees” has recently been urged by Breiman, L., Statistical Modeling: The two cultures (with discussion), Statistical Science, 16 199-225 (2001), and our perspective endorses this. The rationale here is quite simple: node splits are based on specific choices of what we regard as parameters of the overall predictive tree model, the (predictor, threshold) pairs. Inference based on any single tree chooses specific values for these parameters, whereas statistical learning about relevant trees requires that we explore aspects of the posterior distribution for the parameters (together with the resulting branch probabilities).
[0041] Within the current framework, the forward generation process allows easily for the computation of the resulting relative likelihood values for trees, and hence to relevant weighting of trees in prediction. For a given tree, identify the subset of nodes that are split to create branches. The overall marginal likelihood function for the tree is then the product of component marginal likelihoods, one component from each of these split nodes. Continue with the notation of Section 2.1 but now, again, indexed by any chosen node j: Conditional on splitting the node at the defined (predictor, threshold) pair (χj, τj), the marginal likelihood component is
4
[0042] where p(θz,rj,j) is the Be(arj, br,j)prior to each z=0,1. This clearly reduces to
5
[0043] The overall marginal likelihood value is the product of these terms over all nodes j that define branches in the tree. This provides the relative likelihood values for all trees within the set of trees generated. As a first reference analysis, we may simply normalise these values to provide relative posterior probabilities over trees based on an assumed uniform prior. This provides a reference weighting that can be used to both assess trees and as posterior probabilities with which to weight and average predictions for future cases.
DESCRIPTION OF THE SPECIFIC EMBODIMENTS
[0044] Before the subject invention is described further, it is to be understood that the invention is not limited to the particular embodiments of the invention described below, as variations of the particular embodiments may be made and still fall within the scope of the appended claims. It is also to be understood that the terminology employed is for the purpose of describing particular embodiments, and is not intended to be limiting. Instead, the scope of the present invention will be established by the appended claims.
[0045] In this specification and the appended claims, the singular forms “a,” “an” and “the” include plural reference unless the context clearly dictates otherwise. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood to one of ordinary skill in the art to which this invention belongs.
[0046] Where a range of values is provided, it is understood that each intervening value, to the tenth of the unit of the lower limit unless the context clearly dictates otherwise, between the upper and lower limit of that range, and any other stated or intervening value in that stated range, is encompassed within the invention. The upper and lower limits of these smaller ranges may independently be included in the smaller ranges, and are also encompassed within the invention, subject to any specifically excluded limit in the stated range. Where the stated range includes one or both of the limits, ranges excluding either or both of those included limits are also included in the invention.
[0047] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood to one of ordinary skill in the art to which this invention belongs. Although any methods, devices and materials similar or equivalent to those described herein can be used in the practice or testing of the invention, the preferred methods, devices and materials are now described.
[0048] All publications mentioned herein are incorporated herein by reference for the purpose of describing and disclosing the subject components of the invention that are described in the publications, which components might be used in connection with the presently described invention.
Analysis of Biscuit Dough Data
[0049] A first example concerns the application of biscuit dough data (publicly available at Osborne, B. G., Fearn, T., Miller, A. R. and Douglas, S., Applications of near infrared reflectance spectroscopy to compositional analysis of biscuits and biscuit doughs, J. Sci. Food Agric., 35, 99-105 (1984); Brown, P. J., Fearn, T. and Vannucci, M., The choice of variables in multivariate regression: A non-conjugate Bayesian decision theory approach, Biometrika, 86, 635-648 (1999)) in which interest lies in relating aspects of near infrared (“NIR”) spectra of dough to the fat content of the resulting biscuits. The data set provides 78 samples, of which 39 are taken as training data and the remaining 39 as validation cases to be predicted, precisely as in Brown et al (1999). The binary outcome is 0/1 according to whether the measured fat content exceeds a threshold, where the threshold is the mean of the sample of fat values. As predictors, each xi comprises 300 values of the spectrum of dough sample i, augmented by the set of singular factors (principal components) of the 78 sample spectra, so that p=378; with singular factors indexed 301; : : : ; 378.
[0050] The analysis was developed repeatedly, exploring aspects of model fit and prediction of the validation sample as the number of control parameters were varied. The particular parameters of key interest varied were the Bayes' factor thresholds that define splits, and controls on the number of such splits that may be made at any one node. It was determined that across ranges of these control parameters, that there was a good degree of robustness. The Bayes' factor threshold was fixed at 3 on the log scale, after which and two-level trees were explored allowing at most 10 splits of the root node and then at most 4 splits of each of nodes 1 and 2. This allowed up to 160 trees, with this analysis generating 148 trees.
[0051] Many of the trees identified had one or two of the predictors in common, and represent variation in the threshold values for those predictors. FIGS. 1-3 display some summaries. FIG. 1 represents one of the 148 trees, split at the root node by the spectral predictor labeled factor 92 (corresponding to a wavelength of 1566 nm). Multiple wavelength values appear in the 148 trees, with values close to this appearing commonly, reflecting the underlying continuity of the spectra. The key second level predictor is factor 305, one of the principal component predictors. The data are scatter plotted on these two predictors in FIG. 2 with corresponding levels of the predictor-specific thresholds from this tree marked.
[0052] The data appears also against the three predictors in this tree in FIG. 3. Evidently there is substantial overlap in predictor space between the 0/1 outcomes, and cases close to the boundaries defined by any single tree are hard to accurately predict. Nevertheless, in terms of posterior predictive probabilities for the 39 validation samples, accuracy is good. By simply establishing the predictive probability threshold at 0.5 it is determined that 18 of 20 (90%) low fat (blue) cases are “correctly” predicted, as are 19 of 20 (95%) high fat (red) cases. Predictive accuracy is high in this example with considerable overlap between predictor patterns among the two outcome groups. This is a positive example of the use of the predictive tree approach in a context where standard methods, such as logistic regression, would be less useful. Furthermore, the We end with a note that the 50:50 split of the 78 samples into training and validation sets followed the previous authors as references. Curious about this, we reran the analysis 500 times, each time randomly splitting the data 50:50 into training and validation samples. Predictive accuracy, as measured above, was generally not so good as reported for the initial sample split, varying from a little below 50% to 100% across this set of 500 analyses. The average accuracy for low fat (blue) cases was 80%, and that for high fat (red) cases 76%.
Metagene Expression Profiling to Predict Estrogen Receptor Status of Breast Cancer Tumors
[0053] This example illustrates not only predictive utility but also exploratory use of the tree analysis framework in exploring data structure. Here, the tree analysis is used to predict estrogen receptor (“ER”) status of breast tumors using gene expression data. Prior analyses of such data involved binary regression models which utilized Bayesian generalized shrinkage approaches to factor regression. Specifically, prior statistical models involved the use of probit linear regression linking principal components of selected subsets of genes to the binary (ER positive/negative) outcomes. See West, M., Blanchette, C., Dressman, H., Ishida, S., Spang, R., Zuzan, H., Marks, J. R. and Nevins, J. R. Utilization of gene expression profiles to predict the clinical status of human breast cancer. Proc. Natl. Acad. Sci., 98, 11462-11467 (2001). However, the tree model presents some distinct advantages over Bayesian linear regression models in the analysis of large non-linear data sets such as these.
[0054] Primary breast tumors from the Duke Breast Cancer SPORE frozen tissue bank were selected for this study on the basis of several criteria. Tumors were either positive for both the estrogen and progesterone receptors or negative for both receptors. Each tumor was diagnosed as invasive ductal carcinoma and was between 1.5 and 5 cm in maximal dimension. In each case, a diagnostic axillary lymph node dissection was performed. Each potential tumor was examined by hematoxylin/eosin staining and only those that were >60% tumor (on a per-cell basis), with few infiltrating lymphocytes or necrotic tissue, were carried on for RNA extraction. The final collection of tumors consisted of 13 estrogen receptor (ER)+lymph node (LN)+tumors, 12 ER LN+tumors, 12 ER+LN tumors, and 12 ER LN tumors
[0055] The RNA was derived from the tumors as follows: Approximately 30 mg of frozen breast tumor tissue was added to a chilled BioPulverizer H tube (Bio101) (Q-Biogene, La Jolla, Calif.). Lysis buffer from the Qiagen (Chatsworth, Calif.) RNeasy Mini kit was added, and the tissue was homogenized for 20 sec in a MiniBeadbeater (Biospec Products, Bartlesville, Okla.). Tubes were spun briefly to pellet the garnet mixture and reduce foam. The lysate was transferred to a new 1.5-ml tube by using a syringe and 21-gauge needle, followed by passage through the needle 10 times to shear genomic DNA. Total RNA was extracted by using the Qiagen RNeasy Mini kit. Two extractions were performed for each tumor, and total RNA was pooled at the end of the RNeasy protocol, followed by a precipitation step to reduce volume. Quality of the RNA was checked by visualization of the 28S:18S ribosomal RNA ratio on a 1% agarose gel. After the RNA preparation, the samples were subject to Affymetrix GENECHIP analysis. Affymetrix GENECHIP Analysis: The targets for Affymetrix DNA microarray analysis were prepared according to the manufacturer's instructions. All assays used the human HuGeneFL GENECHIP microarray. Arrays were hybridized with the targets at 45° C. for 16 h and then washed and stained by using the GENECHIP Fluidics. DNA chips were scanned with the GENECHIP scanner, and signals obtained by the scanning were processed by GENECHIP Expression Analysis algorithm (version 3.2) (Affymetrix, Santa Clara, Calif.). The same set of n=49 samples used in the binary regression analysis described in West et al (2001) is analyzed in this study, using predictors based on metagene summaries of the expression levels of many genes. Metagenes are useful aggregate, summary measures of gene expression profiles. The evaluation and summarization of large-scale gene expression data in terms of lower dimensional factors of some form is utilized for two main purposes: first, to reduce dimension from typically several thousand, or tens of thousands of genes to a more practical dimension; second, to identify multiple underlying “patterns” of variation across samples that small subsets of genes share, and that characterize the diversity of patterns evidenced in the full sample. Although, the analysis is conducive to the use of various factor model approaches known to those skilled in the art, a cluster-factor approach is used here to define empirical metagenes. This defines the predictor variables x utilized in the tree model.
[0056] Metagenes can be obtained by combining clustering with empirical factor methods. The metagene summaries used in the ER example in this disclosure, are based on the following steps.
[0057] Assume a sample of n profiles of p genes;
[0058] Screen genes to reduce the number by eliminating genes that show limited variation across samples or that are evidently expressed at low levels that are not detectable at the resolution of the gene expression technology used to measure levels. This removes noise and reduces the dimension of the predictor variable;
[0059] Cluster the genes using k_means, correlated-based clustering. Any standard statistical package may be used. This analysis uses the xcluster software created by Gavin Sherlock (http://genomewww.stanford.edu/sherlock/cluster.html). A large number of clusters are targeted so as to capture multiple, correlated patterns of variation across samples, and generally small numbers of genes within clusters;
[0060] Extract the dominant singular factor (principal component) from each of the resulting clusters. Again, any standard statistical or numerical software package may be used for this; this analysis uses the efficient, reduced singular value decomposition function (“SVD”) in the Matlab software environment (http.//www.mathworks.com/products/matlab).
[0061] In the analysis of the ER data in this disclosure, the original data was developed using Affymetrix arrays with 7129 sequences, of which 7070 were used (following removal of Affymetrix controls from the data.). The expression estimates used were log2 values of the signal intensity measures computed using the dChip software for post-processing Affymetrix output data (See Li, C. and Wong, W. H. Model-based analysis of oligonucleotide arrays: Expression index computation and outlier detection. Proc. Natl. Acad. Sci., 98, 31-36 (2001), and the software site http.//www.biostat.harvard edu/complab/dchip/). With a target of 500 clusters, the xcluster software implementing the correlation-based k_means clustering produced p=491 clusters. The corresponding p metagenes were then evaluated as the dominant singular factors of each of these cluster, as referenced above. See FIGS. 7-8 that provide tables detailing the 491 metagenes.
[0062] The data comprised 40 training samples and 9 validation cases. Among the latter, 3 were initial training samples that presented conflicting laboratory tests of the ER protein levels, so casting into question their actual ER status; these were therefore placed in the validation sample to be predicted, along with an initial 6 validation cases selected at random. These three cases are numbers 14, 31 and 33. The color coding in the graphs is based on the first laboratory test (immunohistochemistry). Additional samples of interest are cases 7, 8 and 11, cases for which the DNA microarray hybridizations were of poor quality, with the resulting data exhibiting major patterns of differences relative to the rest. The metagene predictor has dimension p=491: the analysis generated trees based on a Bayes' factor threshold of 3 on the log scale, allowing up to 10 splits of the root node and then up to 4 at each of nodes 1 and 2. Some pertinent summaries appear in the following figures. FIGS. 4 and 5 display 3-D and pairwise 2-D scatterplots of three of the key metagenes, all clearly strongly related to the ER status and also correlated. However, there are in fact five or six metagenes that quite strongly associate with ER status and it is evident that they reflect multiple aspects of this major biological pathway in breast tumors. In the study reported in West et al (2001), Bayesian probit regression models were utilized with singular factor predictors which identified a single major factor predictive of ER. That analysis identified ER negative tumors 16, 40 and 43 as difficult to predict based on the gene expression factor model; the predictive probabilities of ER positive versus negative for these cases were near or above 0.5, with very high uncertainties reflecting real ambiguity.
[0063] In contrast to the more more traditional regression models, the current tree model identifies several metagene patterns that together combine to define an ER profile of tumors, and that when displayed as in FIGS. 4 and 5 isolate these three cases as quite clearly consistent with their designated ER negative status in some aspects, yet conflicting and much more in agreement with the ER positive patterns on others. Metagene 347 is the dominant ER signature; the genes involved in defining this metagene include two representations of the ER gene, and several other genes that are coregulated with, or regulated by, the ER gene. Many of these genes appeared in the dominant factor in the regression prediction. This metagene strongly discriminates the ER 11 negatives from positives, with several samples in the mid-range. Thus, it is no surprise that this metagene shows up as defining root node splits in many high-likelihood trees. This metagene also clearly defines these three cases—16, 40 and 43 —as appropriately ER negative. However, a second ER associated metagene, number 352, also defines a significant discrimination. In this dimension, however, it is clear that the three cases in question are very evidently much more consistent with ER positives; a number of genes, including the ER regulated PS2 protein and androgen receptors, play roles in this metagene, as they did in the factor regression; it is this second genomic pattern that, when combined together with the first as is implicit in the factor regression model, breeds the conflicting information that fed through to ambivalent predictions with high uncertainty. The tree model analysis here identifies multiple interacting patterns and allows easy access to displays such as those shown in FIGS. 4 to 6 that provide insights into the interactions, and hence to interpretation of individual cases. In the full tree analysis, predictions based on averaging multiple trees are in fact dominated by the root level splits on metagene 347, with all trees generated extending to two levels where additional metagenes define subsidiary branches. Due to the dominance of metagene 347, the three interesting cases noted above are perfectly in accord with ER negative status, and so are well predicted, even though they exhibit additional, subsidiary patterns of ER associated behaviour identified in the figures. FIG. 6 displays summary predictions. The 9 validation cases are predicted based on the analysis of the full set of 40 training cases. Predictions are represented in terms of point predictions of ER positive status with accompanying, approximate 90% intervals from the average of multiple tree models. The training cases are each predicted in an honest, cross-validation sense: each tumor is removed from the data set, the tree model is then refitted completely to the remaining 39 training cases only, and the hold-out case is predicted, i.e., treated as a validation sample. Excellent predictive performance is observed for both these one-at-a-time honest predictions of training samples and for the out of sample predictions of the 9 validation cases. One ER negative, sample 31, is firmly predicted as having metagene expression patterns completely consistent with ER positive status. This is in fact one of the three cases for which the two laboratory tests conflicted. The other two such cases, however agree with the initial ER negative test result—number 33, for which the predictions firmly agree with the initial ER negative test result, and number 14, for which the predictions agree with the initial ER positive result though not quite so forcefully. The lack of conformity of expression patterns in some cases (Case 8, 11 and 7) are due to major distortions in the data on the DNA microarray due to hybridization problems.
Prediction of Lymph Node Metastases and Cancer Recurrence
[0064] This study assesses complex, multivariate patterns in gene expression data from primary breast tumor samples that can accurately predict nodal metastatic states and relapse for the individual patient using the statistical tree model of the invention.
[0065] DNA microarray data on samples of primary breast tumors was generated to which non-linear statistical analyses embodied by the tree model of the invention was applied to evaluate multiple patterns of interactions of groups of genes that have true predictive value, at the individual patient level, with respect to lymph node metastasis and cancer recurrence. For both lymph node metastasis and cancer recurrence, patterns of gene expression (metagenes) were identified that associate with outcome. Much more importantly, these patterns were capable of honestly predicting outcomes in individual patients with about 90% accuracy, based on a simple threshold of 0.5 probability in each case. The metagenes that predict lymph node metastasis and recurrence identify distinct groups of genes, suggesting different biological processes underlying these two characteristics of breast cancer.
[0066] Patients and biopsy specimens: The analyses of gene expression phenotypes drew samples from 171 primary tumor biopsies at the Koo Foundation Sun Yat-Sen Cancer Center (KF-SYSCC) in Taipei, Taiwan, collected and banked from 1991 to 2001. Samples from eleven patients who received preoperative chemotherapy and one with in-situ carcinoma were excluded from analysis. These 159 samples represent a heterogeneous population, though patient selection was enriched with cases of longer-term follow-up and observed recurrences. By September 2002, 62 patients developed recurrence whereas 97 remain disease free. The median follow-up was 49 months. Full details of clinical characteristics are shown in Table 1.
[0067] Microarray analysis: Tumor total RNA was extracted with Qiagen RNEasy kits, and assessed for quality with an Agilent Lab-on-a-Chip 2100 Bioanalyzer. Hybridization targets were prepared from total RNA according to Affymetrix protocols and hybridized to Affymetrix Human U95 GeneChip arrays See West M, Blanchette C, Dressman H, Huang E, Ishida S, Spang R et al. Predicting the clinical status of human breast cancer by using gene expression profiles, Proc Natl Acad Sci, 98:11462-11467 (2001).
[0068] Statistical analysis: This analysis used the predictive statistical tree model of this invention. The method of the invention first screens genes to reduce noise, applies k-means correlation-based clustering targeting a large number of clusters, and then uses singular value decompositions (“SVD”) to extract the single dominant factor (principal component) from each cluster. This generated 496 cluster-derived singular factors (metagenes) that characterize multiple patterns of expression of the genes across samples. The strategy aimed to extract multiple such patterns while reducing dimension and smoothing out gene-specific noise through the aggregation within clusters. Formal predictive analysis then uses these metagenes in a Bayesian classification tree analysis. This generates multiple recursive partitions of the sample into subgroups (the “leaves” of the classification tree), and associates Bayesian predictive probabilities of outcomes with each subgroup. Overall predictions for an individual sample are then generated by averaging predictions, with appropriate weights, across many such tree models. Iterative out-of-sample, cross-validation predictions are then performed leaving each tumor out of the data set one at a time, refitting the model from the remaining tumors and using it to predict the hold-out case. This rigorously tests the predictive value of a model and mirrors the real-world prognostic context where prediction of new cases as they arise is the major goal.
[0069] Although, clinico-pathologic parameters such as the presence or absence of positive axillary nodes represent the best means available to classify patients into broad subgroups by recurrence and survival, such methods remain an imperfect tool. Among patients with no detectable lymph node involvement, a population thought to be in a low risk category, between 22 and 33% develop recurrent disease after a 10-year follow-up. See Polychemotherapy for early breast cancer: an overview of the randomized trials, Early Breast Cancer Trialists' Collaborative Group, Lancet; 352:930-942 (2001). Thus, properly identifying individuals out of this group who are at risk for recurrence is beyond the current capabilities of most predictive diagnostics.
[0070] The question of lymph node diagnosis is part of the broader issue of more accurately predicting breast cancer disease course and recurrence. Recently, genomic-scale measures of gene expression, using microarrays and other technologies have opened a new avenue for cancer diagnosis. They identify patterns of gene activity that sub-classify tumors, and such patterns may correlate with the biological and clinical properties of the tumors. The utility of such data in improving prognosis will rely on analytical methods that accurately predict the behavior of the tumors based on expression patterns. Credible predictive evaluation is critical in establishing valid and reproducible results and implicating expression patterns that do indeed reflect underlying biology. This predictive perspective is a key step towards integrating complex data into the process of prognosis for the individual patient, a step that can be accomplished through the practice of the present invention.
[0071] Furthermore, an ultimate goal is to integrate molecular and genomic information with traditional clinical risk factors, including lymph node status, patient age, hormone receptor status, and tumor size, in comprehensive models for predicting disease outcomes. Rather than supplant traditional clinical appraisal, genomic data adds data to traditional risk factors, and assessing individuals based on combinations of relevant traditional risk factors with identified genomic factors could potentially improve predictions. The present invention allows this goal to be realized by demonstrating the ability of genomic data to accurately predict lymph node involvement and disease recurrence in defined patient subgroups. Most importantly, these predictions are relevant for the individual patient and can provide a quantitative measure of the probability for the clinical phenotype and outcome of disease. Such predictions may ultimately facilitate treating patients as individuals rather than as unidentifiable members of a risk profile.
[0072] The present invention was applied to the analysis of gene expression patterns in primary breast tumors that predict lymph node metastasis, as well as tumor recurrence. The first study compares traditional “low-risk” versus “high-risk” patients, primarily based on age, primary tumor size, lymph node status, and Estrogen receptor (“ER”) status. Among ER positive individuals, the “high-risk” clinical profile is represented by advanced lymph node metastases (10 or more positive nodes); the “low-risk profile” identifies node-negative women of age greater than 40 years with tumor size below 2 cm. The number of samples in the tumor collection that met these criteria reduced down to 18 high-risk and 19 low-risk cases. Expression data were generated and metagenes identified and used in the Bayesian statistical tree analysis. FIG. 9 displays summary predictions from the resulting total of 37 cross-validation analyses. For each individual tumor, this graph illustrates the predicted probability for “high-risk” versus “low-risk” (red versus blue) together with an approximate 90% confidence interval, based on analysis of the 36 remaining tumors performed successively 37 times as each tumor prediction is made. It is important to recognize that each sample in the data set, when assayed in this manner, constitutes a validation set that accurately assesses the robustness of the predictive model. The metagene model accurately predicts metastatic potential; about 90% of cases are accurately predicted based on a simple threshold at 0.5 on the estimated probability in each case. Case number 7 is in the intermediate zone, exhibiting patterns of expression of the selected metagenes that relate equally well to those of “high-“ and “low-risk” cases, while case 22 is a clinical “high-risk” case with genomic expression patterns that relate more closely to “low-risk” cases. In contrast, node negative patients 5 and 11 have gene expression patterns more strongly indicative of “high-risk”, and are key cases for follow-up investigations. The details of clinical information in these apparently discordant cases are shown in Table 2. Clinical features of these “discordant” cases are illuminating, and suggestive of how a broader investigation of clinical data combined with molecular model-based predictions may aid in the eventual decision-making process. Although case 22 did in fact recur, 6 years post-surgery; this patient's clinical classification as high risk for recurrence based on purely clinical parameters was moderated by a lower risk based on metagenes, as demonstrated by this patient having survived recurrence-free for a longer time. Thus the lower probability prediction assigned to patient 22 based on the gene expression profiles is reflected in the clinical behavior of her disease. The “low-risk” patient 7 recurred at 31 months, and patient 11 at 38 months, whereas case 5 is currently disease-free after only 12 months of follow-up. Again, case 7, and to some degree case 11, thus partly corroborate the predictions based on genomic criteria. data. With such predictions as part of a prognostic model, more intensive or innovative post-surgical therapy should perhaps have been recommended for these two cases. A critical aspect of the analyses described here is allowing the complexity of distinct gene expression patterns to enter the predictive model. Tumors are graphed against metagene levels for three of the highest scoring metagene factors as shown in FIG. 10. This analysis highlights the need to analyze multiple aspects of gene expression patterns. For example, if the low-risk cases 1, 3 and 11 are assessed against metagene 146 alone, their levels are more consistent with high-risk cases. However, when additional dimensions are considered, the picture changes. The second frame (upper right) shows that low-risk is consistent with low levels of metagene 130 or high levels of metagene 146; hence, cases 1 and 3 are not inconsistent in the overall pattern, though case 11 is consistent. An analysis that selects one set of genes, summarized here as one metagene, as a “predictor” would be potentially misleading, as it ignores the broader picture of multiple interlocked genomic patterns that together characterize a state. In the predictions, these two metagenes play key roles: low levels of metagene 146 coupled with higher levels of metagene 130 are strongly predictive of high-risk cases. Combined use of multiple metagenes, in the context of the tree selection model building process, ultimately yields a pattern that has the capacity to accurately predict the clinical outcome.
[0073] The second analysis concerns 3 year recurrence following primary surgery among the challenging and varied subset of patients with 1-3 positive lymph nodes. Such patients typically receive adjuvant chemotherapy alone, but more than 20% suffer relapse within five years. Hence, improved prognosis for this heterogeneous group is of critical importance; patients identified with a high probability of relapse could be targeted for more intensive treatment. The dataset provided 52 ER-positive cases in this lymph node category (34 non-recurrent, 18 recurrent). The aggregate predictions from the sets of generated statistical tree models defines a rather accurate picture; once again, there is an approximate 90% overall predictive accuracy in the 52 separate one-at-a-time, cross-validation prediction assessments as shown in FIG. 11. Based on the gene expression analysis, the 3 year non-recurrent cases 6 and 23, having profiles more akin to recurrent cases, would be candidates for intensive treatment. hese patients did receive adjuvant chemotherapy based on additional clinical risk factors (especially tumor size). Thus traditional clinical risk factors other than lymph node status also indicate higher risk of recurrence for these two cases, consistent with the molecular predictions. Each actually survived recurrence-free for over three years; case 6 recurred at 42 months and case 23 remains disease-free after over 6 years. Cases with low genomic criteria for recurrence would be 36, 38 and 42. They, however, experienced recurrence within three years. These are cases that, under prognosis informed by only the genomic model, would have been indicated as more benign and not candidates for intensive treatment, whereas such a treatment might have proven to be more beneficial.
[0074] The tree model of the invention identified subsets of genes related to the metagene predictors of lymph node involvement. These are replete with those involved in cellular immunity, including a high proportion of genes that function in the interferon pathway. They include genes that are induced by interferon such as various chemokines and chemokine receptors (Rantes, CXCL10, CCR2), other interferon-induced genes (IFI30, IFI35, IFI27, IFIT1, IFIT4, IFITM3), as well as interferon effectors (2′-5′ oligoA synthetase), and genes encoding proteins mediating the induction of these genes in response to interferon (STAT1 and IRF1). This connection is intriguing given the role of interferon as a mediator of the anti-tumor response and, together with the fact that many genes involved in T cell function (TCRA, CD3D, IL2R, MHC) are also included within the group that predict lymph node metastasis. This may reflect the distinct nature of these tumors that have acquired a metastatic potential that elicits an anti-tumor response that is ultimately unsuccessful or an aberration of the normal anti-tumor response.
[0075] Genes implicated in recurrence prediction as identified by the tree model of the invention do not exhibit such a striking functional clustering but do include many examples previously associated with breast cancer. Moreover, this group of genes is clearly distinct set from those that predict lymph node involvement. They include genes associated with cell proliferation control, both cell cycle specific activities (CDKN2D, Cyclin F, E2F4, DNA primase, DNA ligase), more general cell growth and signaling activities (MK2, JAK3, MAPK8IP, and EF1□), and a number of growth factor receptors and G-protein coupled receptors, some of which have been shown to facilitate breast tumor growth (EpoR). Possibly, the poor prognosis with respect to survival reflects a more vigorous proliferative capacity of the tumor.
[0076] Thus, the genes implicated in the prediction of lymph node metastasis and overall recurrence of disease, although clearly representing interrelated phenomena, nevertheless reflect the participation of distinct biological processes. The tree model is thus flexible in that regard as it only selects those metagenes that are most relevant to the prediction in hand. By contrast, traditional statistical testing perspectives that focus on significant differences at a population parameter level may say little of practical significance in terms of an individual patient's prognosis. Furthermore, the present invention takes into account the relevant multiple features of the complex patterns of gene expression, especially in a context such as breast cancer where multiple, interacting biological and environmental processes define physiological states, and individual dimensions provide only partial information. The tree model of the present invention assesses the complex, multivariate patterns in gene expression data from primary tumor biopsies, exploring the value of such patterns in predicting lymph node metastasis and relapse, two critically important aspects of breast cancer, at the individual patient level. The tree model identifies multivariate patterns of gene expression that, in this realistic context of substantial patient heterogeneity, deliver predictive accuracy of about 90%. The probabilistic models highlight cases where uncertainty is high, and generate subsets of implicated genes that relate to the biology of metastasis and tumor evolution.
[0077] To ascertain the success of the tree model, an out-of-sample predictive assessment via cross-validation is always conducted. Any selection of gene, metagene or clinical variables must be part of each cross-validation analysis. The results of such “feature selection” will vary each time a tumor is analyzed, and can dramatically impact on predictive accuracy. Analyses that select a set of predictors based on the entire dataset, including the individual to be predicted, in advance of predictive evaluation are inappropriate, and lead to misleadingly over-optimistic conclusions about predictive value. For breast cancer recurrence, the results provide evidence for gene expression profiles associated with recurrence in a homogeneous cohort of low risk patients. There are, however, several distinctions. First is the evaluation of models on the basis of accuracy in prediction at the individual level, with predictions made in formal probabilistic terms. Second, multiple, related and interacting biological patterns, here represented as separate and distinct metagenes, together represent a clinical state. Reducing high-dimensional genomic data to a single index may sacrifice opportunity for understanding complex interactions (see FIG. 2) that are truly predictive. Thirdly, we believe that the integration of molecular profiles with clinical risk factors—rather than the replacement of clinical data with molecular data—will define the major step towards personalized prognosis utilizing genomic data, hence the need for stratification using clinical variables.
[0078] <<<INSERT TABLES 1 & TABLE 2 from 7163557>>>
Identifying Atherosclerotic Phenotype Determinative Genes Related to Atherosclerosis Disease Progression and Susceptibility to Atherosclerosis.
Claims
- 1. The application of classification tree models incorporating Bayesian analysis to the statistical prediction of binary outcomes
Provisional Applications (3)
|
Number |
Date |
Country |
|
60420729 |
Oct 2002 |
US |
|
60421062 |
Oct 2002 |
US |
|
60424718 |
Nov 2002 |
US |