The curse of dimensionality: a blessing to personalized medicine.

TO THE EDITOR: We wish to provide an alternative consideration to the recommendations raised in the editorial by Sikic et al in Journal of Clinical Oncology relating to the study by Bhojwani et al. The study reports gene expression profiles that predict early response and longterm outcome in children with acute lymphoblastic leukemia. Bhojwani et al identified 41 genes out of a 38,500-gene expression data set that were predictive of long-term patient outcome treated on a single protocol and a 24-gene set predictive of early therapeutic response. Of most significance to the field however, was their conclusion that gene expression signatures provide no greater prognostic value compared with current factors such as patient age, WBC count, or karyotype. Most clinical stratification protocols stream patients into a limited number of treatment strategies with individuals being managed within risk groups. However, to realize the potential for personalized medicine, patients in need of specific clinical intervention should be distinctively identified out of the crowd. So, by personalized medicine are we aiming to improve risk stratification with patient cohorts being divided into ever smaller risk groups, or do we consider that medicine can be truly tailored for individual patients? The former is based on the assumption that we can classify complex disease into specific subgroups where patients within each group have similar genetic activity and are homogenous with respect to treatment response. This fails to recognize that individual patients within subgroups will have unique clinical responses to treatment strategies. If the latter, we then need to be able to realistically compare individual patients to each other. This will require identifying individual genetic differences, as much as similarities. Sikic et al identified the fundamental problem is the “traditional paradigm of sifting data to identify one or a few markers to use prospectively for prognosis of outcomes or prediction of therapies,” that is, reductionism. Traditionally, reductionism guides observational deductive science, data derived from complex biologic samples being broken down to single points of study. Many interesting diseaserelated genes or biomarkers have been discovered as a result. The premise of investigations such as Bhojwani et al is that gene expression microarray datasets can be reduced using statistical or informatics approaches to identify a smaller set of genes which define patients at risk of poor clinical response. We have reported that while such gene expression signatures can be identified which segregate different acute lymphoblastic leukemia patient subgroups in one study, such signatures could not be validated in independent cohorts. To date, reductionist approaches have yet to realize signatures that have universal diagnostic, prognostic, nor therapeutic applications which would be the basis for personalized medicine. The nature of microarray data and its acquisition means that it is subject to the curse of dimensionality, the situation where there are vastly more measurable features (genes) than there are samples (Fig 1A). Reductionist analysis (Fig 1B) of highly dimensional microarray data, such as used by Bhojwani et al lead to feature selection approaches that are liable to extreme type I error, and will not identify enough features to provide for individual differences within the patient cohort. For leukemia, this has been address by the Microarray Innovations in Leukemia (MILE) international multicenter study, which aims to expand and standardize the number of samples analyzed (Fig 1C) and to assess the clinical accuracy of gene expression profiling across leukemia subtypes. It is anticipated that a more robust predictor will be derived from reductionist analyses seeking gene expression similarities within disease subgroups. However, despite the large number of genes which we can derive expression data from in a single experiment, it only informs us about one cellular process, transcription. It is unclear exactly why particular genes are good indicators of patient outcome as it is not known whether these genes directly cause the metabolic effects or whether they are in the coregulating