Dynemo was validated in two separate breast cancer studies using publicly available protein-protein interaction data.
Non-familial breast cancer
Dynemo’s effectiveness in its capacity as a prognostic test was demonstrated in two studies of tumors from surgically treated cancer patients. Data for the first cohort of 295 samples was collected from a single site, the Netherlands Cancer Institute tissue bank, and included primary invasive carcinoma tumors less than 5 cm in diameter in women 52 years of age and younger at diagnosis. Data for the second cohort of 291 samples was collected from 39 sites in Europe and Latin America and included tumor samples from patients under the age of 60 who had received surgical and/or radiotherapy treatment only.
Patients were divided into groups, and the algorithm was trained with gene-expression data from tumor samples in one group to predict prognosis in the other groups. The process was repeated several times with new training and prediction groups each time (this cross-validation procedure avoids statistical overfitting). The analysis resulted in a prognostic accuracy of 76.1% with a positive predictive value of 81.2% and sensitivity of 86.1% in the first study. The experiment was repeated in the second cohort resulting in an accuracy of 78.5%. The concordance of both the performance and the signature networks determined by the independent repeats of the Dynemo algorithm confirms the robustness and suitability of the algorithm for direct use in the clinic.
Interactome data
PPI data for the study was compiled from three publicly available databases: the Online Predicted Human Interaction Database (OPHID), a database of literature-curated and experimentally derived predicted protein interactions; the Molecular INTeraction database (MINT), a literature-curated database of experimentally verified protein interactions; and other protein interaction data from the European Molecular Biology Laboratory (EMBL).
The scientific community’s knowledge of the human interactome is not complete, but it is an active field of research and Dynemo’s accuracy is expected to improve as more PPI data becomes available. To test this concept, we incrementally removed random pieces of the interactome and found that Dynemo’s prediction accuracy dropped exponentially as the size of the interactome available to the algorithms was reduced.
Expected network changes
The breast cancer study revealed expected changes in certain sub-networks between patient outcomes. One such network is BRCA1 and its interacting partners. Dynemo found that the expression of BRCA1 and its interactors is highly correlated in tumors from good-outcome patients, while dissociated in poor-outcome patients. Another example is the network of SRC and its interactors: Dynemo found significant differences in the co-expression of SRC and its interactors between tumour groups; however, the difference in expression level of SRC between tumour groups is insignificant. In this case, SRC would not be a suitable as a traditional standalone biomarker, but Dynemo is able infer prognosis using SRC.
Incorporating histopathology
Resected tumor samples are routinely examined by pathologists to understand and predict disease severity and prognosis. While Dynemo is a molecular diagnostic tool with a high degree of accuracy on its own, clinical data can be incorporated to improve the test. This was demonstrated in the breast cancer study, wherein the accuracy of the result was improved by including information about the tumour grade and stage as well as patient age. Thus, Dynemo is a significant and independent predictor of clinical outcome that can improve the predictive value of prognostic indicators already in use in the clinic.
Non-familial breast cancer
Dynemo’s effectiveness in its capacity as a prognostic test was demonstrated in two studies of tumors from surgically treated cancer patients. Data for the first cohort of 295 samples was collected from a single site, the Netherlands Cancer Institute tissue bank, and included primary invasive carcinoma tumors less than 5 cm in diameter in women 52 years of age and younger at diagnosis. Data for the second cohort of 291 samples was collected from 39 sites in Europe and Latin America and included tumor samples from patients under the age of 60 who had received surgical and/or radiotherapy treatment only.
Patients were divided into groups, and the algorithm was trained with gene-expression data from tumor samples in one group to predict prognosis in the other groups. The process was repeated several times with new training and prediction groups each time (this cross-validation procedure avoids statistical overfitting). The analysis resulted in a prognostic accuracy of 76.1% with a positive predictive value of 81.2% and sensitivity of 86.1% in the first study. The experiment was repeated in the second cohort resulting in an accuracy of 78.5%. The concordance of both the performance and the signature networks determined by the independent repeats of the Dynemo algorithm confirms the robustness and suitability of the algorithm for direct use in the clinic.
Interactome data
PPI data for the study was compiled from three publicly available databases: the Online Predicted Human Interaction Database (OPHID), a database of literature-curated and experimentally derived predicted protein interactions; the Molecular INTeraction database (MINT), a literature-curated database of experimentally verified protein interactions; and other protein interaction data from the European Molecular Biology Laboratory (EMBL).
The scientific community’s knowledge of the human interactome is not complete, but it is an active field of research and Dynemo’s accuracy is expected to improve as more PPI data becomes available. To test this concept, we incrementally removed random pieces of the interactome and found that Dynemo’s prediction accuracy dropped exponentially as the size of the interactome available to the algorithms was reduced.
Expected network changes
The breast cancer study revealed expected changes in certain sub-networks between patient outcomes. One such network is BRCA1 and its interacting partners. Dynemo found that the expression of BRCA1 and its interactors is highly correlated in tumors from good-outcome patients, while dissociated in poor-outcome patients. Another example is the network of SRC and its interactors: Dynemo found significant differences in the co-expression of SRC and its interactors between tumour groups; however, the difference in expression level of SRC between tumour groups is insignificant. In this case, SRC would not be a suitable as a traditional standalone biomarker, but Dynemo is able infer prognosis using SRC.
Incorporating histopathology
Resected tumor samples are routinely examined by pathologists to understand and predict disease severity and prognosis. While Dynemo is a molecular diagnostic tool with a high degree of accuracy on its own, clinical data can be incorporated to improve the test. This was demonstrated in the breast cancer study, wherein the accuracy of the result was improved by including information about the tumour grade and stage as well as patient age. Thus, Dynemo is a significant and independent predictor of clinical outcome that can improve the predictive value of prognostic indicators already in use in the clinic.