Deliverable

6.3Scenario B-Retrospective use of data

Scenario B
Objective	A researcher wants to define if a gene expression signature can be used to predict the toxic effects (grade 3 (G3) or more) of an investigational class of drugs (e.g. mAb, TKI) used in the neoadjuvant treatment of a specific breast cancer subtype.
Steps	The researcher logs into the system. The researcher filters by type of cancer (i.e. breast), the treatment setting (i.e. neoadjuvant), the selected class of treatment and the toxicity (i.e. G3 or more). The researcher selects for the following outputs; gene expression data, type of drug, toxicity type and grade, clinical trial and patients baseline characteristics. The researcher either downloads the results on his computer (i.e. an excel file in csv format) and the gene expression data in the relevant format or works directly on the INTEGRATE platform using the provided tools.
Results	The researcher analyses gene expression data and tries to confirm his hypothesis: “A gene signature can predict the toxicity of a class of drugs”.

Table Scenario B-Retrospective use of data

The objective of this study is slightly changed compared to the previous one, but the overall prediction framework remains the same. As in previous scenarios, the researcher logs into the INTEGRATE platform and exports the examined dataset which consists of patients with breast cancer, treated by an investigational class of drugs with a specific toxicity type and grade per drug, under neoadjuvant therapy. All available patients that received the investigational drug are dichotomised into two classes based on the toxicity grade of the drugs; a class with high grade (grade 3 or more) and a class with low grade toxic effect. The overall dataset enters the prediction model as in Figure and the researcher can get access to the results by an excel file with all the available information as mentioned in the description of the previous scenario (chapter 6.2).

6.4Scenario C-Retrospective use of data

In this scenario, data generated from samples collected in a neo adjuvant clinical trial run by Institute Jules Bordet (IJB), are used to create a model that predicts response/resistance to a specific preoperative drug (i.e. epirubicin) in estrogen receptor-negative (ER-) breast cancer patients. Again the dataset is dichotomized based on the independent clinical variable of pCR. The researcher logs into the INTEGRATE platform and exports the examined dataset which consists of patients with breast cancer, treated by a specific preoperative drug (i.e. epirubicin), under neoadjuvant therapy. This scenario encapsulates the adaptation of both kernel-based and ensemble trees classification techniques for prediction analysis.
When using kernel-based approaches, both implementations as depicted in Figure , can be applied to the examined dataset (a basis kernel for each data source or a basis kernel for each feature). Specifically, a first model of the MKL using support vector machines will compute separately a basis kernel for each data source (i.e. a kernel for each clinical, microarray and proteomic dataset respectively) and a unique kernel from the linear combination of the individual kernels projects the overall data to the feature space for further analysis. A prediction model is then constructed by using all the available data (no feature analysis is performed), and its accuracy is assessed using the statistics in chapter 6.1.6 under the iterative validation techniques described in chapter 6.1.7.
A slightly changed kernel-based framework will be constructed using an individual kernel for each feature of all the available data sources. As in both scenarios described in 6.2 and 6.3, feature selection for reducing the high dimensionality of the microarray and molecular data will be first implemented. Then, a weighted basis kernel is computed for each feature and an iterative analysis is performed to estimate the most relevant features that give the highest classification accuracy.
A third prediction model will be provided using the ensemble of the decision trees (see chapter 6.1.5 for further details). Using several methods from the field of decision trees like the RotBoost, Random Forest and Rotation Forest, we aim in assessing the performance of them using an iterative evaluation process (i.e. bootstrapping or cross validation) and choose the “tree model” that shows the maximum performance.
Finally, the most accurate prediction model from both fields will be selected, becomes a part of the INTEGRATE platform and could be used as a predictive model for response to a specific preoperative drug (i.e. epirubicin) in estrogen receptor-negative (ER-) breast cancer patients.

7Conclusion

This deliverable summarised the main objectives of the WP proposing an approach of building the framework for INTEGRATE VPH predictive modelling development. The clinically relevant questions that have been defined so far with concern the development of scenario-driven prediction models that given a set of characteristics will be able to predict in an accurate way the response to a drug and/or the response/resistance to a specific preoperative drug. This deliverable highlighted the main techniques that will be exploited giving emphasis to multi-kernel techniques that will allow the integration of multi-level heterogeneous data and subsequently the development of predictive models beyond the sate-of-the-art.

8Appendix

Additional scenarios that do not represent highest priorities for users but they pose interesting challenges from the computational or methodological view, and will therefore by utilized for the definition of the generic framework, are presented in the following chapters. A statistical analysis of the heterogeneous data along with the implementation of the prediction modelling framework offers a thorough study to the researchers and constitutes a relevant and tool for assessing the clinical behaviour of patients’ response to disease.