In my daily practice performing clinical tests using a genomics platform, I was often surprised by the discrepancy between laboratory test predictions and clinical reality.
A major question that remains unsatisfactorily answered for the treating physician is: what is the best drug for my patient? The current companion diagnostic approach used for most targeted therapies provides limited answers, with a binary "yes"/"no" expected response to a drug. But it does not provide any element of comparison with other drugs. In addition, with the number of validated drug targets increasing, testing each patient’s tumor for all markers related to all possible targeted therapies becomes infeasible due to the limited amount of tissue usually available from needle biopsies.
We therefore embarked on our research journey seeking a universal biomarker that could enable a rational choice of therapy among a large number of drugs and that would also address the complexity of cancer biology requiring the investigation of networks of pathways to understand the variability of clinical outcomes observed. Our objective was to create a biomarker tool able to predict in a single assay the time until tumor progression or death (progression-free survival (PFS)) to multiple targeted treatments: the Digital Display Precision Predictor (DDPP).
Our approach was based on the use of transcriptomics that our group, the Worldwide Innovative Network (WIN) Consortium for personalized cancer therapy, had explored successfully in the WINTHER clinical trial where the number of advanced cancer patients treated with a personalized targeted treatment was bolstered by 30% compared to using genomics alone to identify treatment options. WINTHER was the first trial that introduced transcriptomic investigations from biopsies of fresh frozen (FF) tumor and analogous normal tissues in order to significantly reduce the transcriptomic noise. The hypothesis was that apparent overexpression of genes in tumor cells could be driven by the normal tissue from which the tumor originated (reflecting variability between individuals and not carcinogenesis). The WINTHER dataset provided an incredible amount of information: genomics, whole transcriptomics (20,000 genes), treatment and clinical outcomes (progression-free survival (PFS) and overall survival). Key success factors for the transcriptomic data analysis in WINTHER were the standard operating procedures, operator training, and stringent histologic quality control of fresh frozen tissues (with high tumor content) obtained by biopsy.
The main hurdle that needed to be overcome in developing the DDPP was the small size of the WINTHER dataset and the variety of treatments employed that resulted in a small number of patients treated with the same drug (< 10) across several tumor types thus limiting the internal validationof our findings.
Unfortunately, our search for other comparable datasets was unsuccessful. No dataset had available for each patient: paired fresh frozen (FF), that is, tumor tissue (for ex NSCLC) and matching analogous normal tissue (bronchial mucosa in this case), transcriptomic data and clinical outcome.
The unavailability of another comparable independent dataset to further test our algorithm’s prediction of PFS under a variety of treatment options obliged us to be limited to one data source: WINTHER data. As a result, the conventional correlative methods such as Cox univariate and multivariate regression models, LASSO regression, multiple linear regression, failed. We went back to the drawing board.
We explored all the transcriptome results from WINTHER data for the patients who had received the same treatment looking for the group of genes which intersection's between differential expression (between tumor and normal tissue) and PFS would be spatially aligned on a Pearson Correlation. Inspiring ourselves from Euclid postulates, we concluded that if this was possible for a minimum of 3 patients, then the PFS was a linear function of the gene expression, and PFS could therefore be predicted for other patients.
We explored the DDPP ability to predict PFS by assessing correlations with PFS associated with everolimus, axitinib, trametinib, afatinib, experimental FGFR inhibitors and anti-PD-1/PDL-1 therapies received by WINTHER trial patients.
Based on our analysis, we confirmed the following: (1) when using only tumor tissue, significance of the correlations dropped, supporting the importance of the tumor/normal analysis; (2) the predictive value versus prognostic value of DDPP, by cross-correlating combined differential expression of the selected genes with PFS of patients under another treatment; (3) ability of analyses to be used as predictors and address potential over-fitting, by performing leave-one-outcombinatorial analyses using the DDPP on 5 patients to determine the PFS of the 6th patient with a significant accuracy despite the small number of patients.
All the rest is in the paper.
Several conclusions from our journey
- WINTHER is a gold mine for transcriptomic investigations due to the availability of patient tumor and analogous normal tissue genomics, transcriptomics and clinical outcomes for each patient. It wasthe first study pioneered by the WIN Consortium that assembles 35 world-class academic medical centers, industries (pharmaceutical and diagnostic companies), research organizations and patient advocates spanning 19 countries and 4 continents, aligned to launch trials using its genomics and transcriptomics biomarker platform to bolster Precision Oncology across the world.The WINTHER study was one of the last projects of our late Chairman Emeritus Dr. John Mendelsohn (previously also the President of MD Anderson Cancer Center) and Vice Chairman Prof. Thomas Tursz (previously also General Director of Gustave Roussy) who founded the organization.
- Fresh frozen biopsies are so far the only way to perform reliable transcriptomics.
- DDPP is potentially a new global biomarker model that can apply to any type of drug alone or in combination, agnostic of tumor type, and can lead, pending further prospective validation, to a new approach to optimal treatment selection for patients with cancer.
Most importantly, DDPP and the predictive ability of transcriptomics will need to be validated prospectively in larger datasets.