Limits...
Optimally discriminative subnetwork markers predict response to chemotherapy.

Dao P, Wang K, Collins C, Ester M, Lapuk A, Sahinalp SC - Bioinformatics (2011)

Bottom Line: However, prediction of response to cancer treatment has proved to be more challenging and novel approaches with improved generalizability are still highly needed.We show that our OptDis method improves over previously published subnetwork methods and provides better and more stable performance compared with other subnetwork and single gene methods.We also show that our subnetwork method produces predictive markers that are more reproducible across independent cohorts and offer valuable insight into biological processes underlying response to therapy.

View Article: PubMed Central - PubMed

Affiliation: School of Computing Science, Simon Fraser University.

ABSTRACT

Motivation: Molecular profiles of tumour samples have been widely and successfully used for classification problems. A number of algorithms have been proposed to predict classes of tumor samples based on expression profiles with relatively high performance. However, prediction of response to cancer treatment has proved to be more challenging and novel approaches with improved generalizability are still highly needed. Recent studies have clearly demonstrated the advantages of integrating protein-protein interaction (PPI) data with gene expression profiles for the development of subnetwork markers in classification problems.

Results: We describe a novel network-based classification algorithm (OptDis) using color coding technique to identify optimally discriminative subnetwork markers. Focusing on PPI networks, we apply our algorithm to drug response studies: we evaluate our algorithm using published cohorts of breast cancer patients treated with combination chemotherapy. We show that our OptDis method improves over previously published subnetwork methods and provides better and more stable performance compared with other subnetwork and single gene methods. We also show that our subnetwork method produces predictive markers that are more reproducible across independent cohorts and offer valuable insight into biological processes underlying response to therapy.

Availability: The implementation is available at: http://www.cs.sfu.ca/~pdao/personal/OptDis.html

Contact: cenk@cs.sfu.ca; alapuk@prostatecentre.com; ccollins@prostatecentre.com.

Show MeSH

Related in: MedlinePlus

Bar charts show the average MCCs of different predictive models. Single gene marker models include one based on t-test (SGM) and models from MAQC project (MAQC). Subnetwork marker models include Chuang et al. (2007) (GreedyMI), Dao et al. (2010) (Dense), and our method (OptDis). The yellow bars and blue bars show the classification performance in FXD and BXD analyses respectively. The green bars show the overall average performance, calculated as the average of the yellow and blue bars.
© Copyright Policy - creative-commons
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC3117373&req=5

Figure 3: Bar charts show the average MCCs of different predictive models. Single gene marker models include one based on t-test (SGM) and models from MAQC project (MAQC). Subnetwork marker models include Chuang et al. (2007) (GreedyMI), Dao et al. (2010) (Dense), and our method (OptDis). The yellow bars and blue bars show the classification performance in FXD and BXD analyses respectively. The green bars show the overall average performance, calculated as the average of the yellow and blue bars.

Mentions: To assess the predictive performance, we performed two analyses. In the forward cross-dataset (FXD) analysis, we treated the 130 patient cohort as the training set used for deriving markers, and validated their performance on the 100 patient cohort. We also performed the complementary backward cross-dataset (BXD) analysis and swapped the cohorts used in training and validation. In Figure 2, we compare the performance of OptDis against single gene marker models. The single gene marker classifier constructed using t-test is denoted by SGM and includes only genes that map to the PPI network. For each mappable gene, the corresponding probe with the lowest P value was used in the model. We also compared the performance of our method OptDis against implementations of existing subnetwork-based methods, one based on mutual information (GreedyMI) (Chuang et al., 2007), and another based on dense subnetworks (we denote as Dense) using the STRING functional network (Dao et al., 2010). The density threshold to extract all dense subnetworks is set at 0.7 as implemented in Dao et al. (2010). Note that, top 50 subnetworks for GreedyMI and Dense are ranked based on their mutual information scores. Starting from around 20 features, the performance of OptDis is better than competing methods. While the maximum MCC value is not that high, it is still significant compared with the random classifier which has an MCC value of 0. Moreover, predicting response to chemotherapy has been shown as a difficult endpoint to predict in the recent MAQC publications (Shi, 2010). The difficulties might be due to the known heterogeneity within tumors of the same cancer type, subtype-specific response, differences in drug metabolism between individuals and variations in chemotherapy schedules between patients (Popovici, 2010). Figure 3 shows the average performance of models in cross-dataset validation of FXD and BXD analyses. Here, the average performance for a model is the average MCC of 50 models generated using the top 1–50 features. The MAQC performance was derived from the average of top model from each participating group. As shown in Figure 3, OptDis outperforms all the other competitors on the average classification performance in FXD and BXD analyses.Fig. 2.


Optimally discriminative subnetwork markers predict response to chemotherapy.

Dao P, Wang K, Collins C, Ester M, Lapuk A, Sahinalp SC - Bioinformatics (2011)

Bar charts show the average MCCs of different predictive models. Single gene marker models include one based on t-test (SGM) and models from MAQC project (MAQC). Subnetwork marker models include Chuang et al. (2007) (GreedyMI), Dao et al. (2010) (Dense), and our method (OptDis). The yellow bars and blue bars show the classification performance in FXD and BXD analyses respectively. The green bars show the overall average performance, calculated as the average of the yellow and blue bars.
© Copyright Policy - creative-commons
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC3117373&req=5

Figure 3: Bar charts show the average MCCs of different predictive models. Single gene marker models include one based on t-test (SGM) and models from MAQC project (MAQC). Subnetwork marker models include Chuang et al. (2007) (GreedyMI), Dao et al. (2010) (Dense), and our method (OptDis). The yellow bars and blue bars show the classification performance in FXD and BXD analyses respectively. The green bars show the overall average performance, calculated as the average of the yellow and blue bars.
Mentions: To assess the predictive performance, we performed two analyses. In the forward cross-dataset (FXD) analysis, we treated the 130 patient cohort as the training set used for deriving markers, and validated their performance on the 100 patient cohort. We also performed the complementary backward cross-dataset (BXD) analysis and swapped the cohorts used in training and validation. In Figure 2, we compare the performance of OptDis against single gene marker models. The single gene marker classifier constructed using t-test is denoted by SGM and includes only genes that map to the PPI network. For each mappable gene, the corresponding probe with the lowest P value was used in the model. We also compared the performance of our method OptDis against implementations of existing subnetwork-based methods, one based on mutual information (GreedyMI) (Chuang et al., 2007), and another based on dense subnetworks (we denote as Dense) using the STRING functional network (Dao et al., 2010). The density threshold to extract all dense subnetworks is set at 0.7 as implemented in Dao et al. (2010). Note that, top 50 subnetworks for GreedyMI and Dense are ranked based on their mutual information scores. Starting from around 20 features, the performance of OptDis is better than competing methods. While the maximum MCC value is not that high, it is still significant compared with the random classifier which has an MCC value of 0. Moreover, predicting response to chemotherapy has been shown as a difficult endpoint to predict in the recent MAQC publications (Shi, 2010). The difficulties might be due to the known heterogeneity within tumors of the same cancer type, subtype-specific response, differences in drug metabolism between individuals and variations in chemotherapy schedules between patients (Popovici, 2010). Figure 3 shows the average performance of models in cross-dataset validation of FXD and BXD analyses. Here, the average performance for a model is the average MCC of 50 models generated using the top 1–50 features. The MAQC performance was derived from the average of top model from each participating group. As shown in Figure 3, OptDis outperforms all the other competitors on the average classification performance in FXD and BXD analyses.Fig. 2.

Bottom Line: However, prediction of response to cancer treatment has proved to be more challenging and novel approaches with improved generalizability are still highly needed.We show that our OptDis method improves over previously published subnetwork methods and provides better and more stable performance compared with other subnetwork and single gene methods.We also show that our subnetwork method produces predictive markers that are more reproducible across independent cohorts and offer valuable insight into biological processes underlying response to therapy.

View Article: PubMed Central - PubMed

Affiliation: School of Computing Science, Simon Fraser University.

ABSTRACT

Motivation: Molecular profiles of tumour samples have been widely and successfully used for classification problems. A number of algorithms have been proposed to predict classes of tumor samples based on expression profiles with relatively high performance. However, prediction of response to cancer treatment has proved to be more challenging and novel approaches with improved generalizability are still highly needed. Recent studies have clearly demonstrated the advantages of integrating protein-protein interaction (PPI) data with gene expression profiles for the development of subnetwork markers in classification problems.

Results: We describe a novel network-based classification algorithm (OptDis) using color coding technique to identify optimally discriminative subnetwork markers. Focusing on PPI networks, we apply our algorithm to drug response studies: we evaluate our algorithm using published cohorts of breast cancer patients treated with combination chemotherapy. We show that our OptDis method improves over previously published subnetwork methods and provides better and more stable performance compared with other subnetwork and single gene methods. We also show that our subnetwork method produces predictive markers that are more reproducible across independent cohorts and offer valuable insight into biological processes underlying response to therapy.

Availability: The implementation is available at: http://www.cs.sfu.ca/~pdao/personal/OptDis.html

Contact: cenk@cs.sfu.ca; alapuk@prostatecentre.com; ccollins@prostatecentre.com.

Show MeSH
Related in: MedlinePlus