A neural network approach to multi-biomarker panel discovery by high-throughput plasma proteomics profiling of breast cancer
© Zhang et al.; licensee BioMed Central Ltd. 2013
Published: 20 December 2013
In the past several years, there has been increasing interest and enthusiasm in molecular biomarkers as tools for early detection of cancer. Liquid chromatography tandem mass spectrometry (LC/MS/MS) based plasma proteomics profiling technique is a promising technology platform to study candidate protein biomarkers for early detection of cancer. Factors such as inherent variability, protein detectability limitation, and peptide discovery biases among LC/MS/MS platforms have made the classification and prediction of proteomics profiles challenging. Developing proteomics data analysis methods to identify multi-protein biomarker panels for breast cancer diagnosis based on neural networks provides hope for improving both the sensitivity and the specificity of candidate cancer biomarkers for early detection.
In our previous method, we developed a Feed Forward Neural Network-based method to build the classifier for plasma samples of breast cancer and then applied the classifier to predict blind dataset of breast cancer. However, the optimal combination C* in our previous method was actually determined by applying the trained FFNN on the testing set with the combination. Therefore, in this paper, we applied a three way data split to the Feed Forward Neural Network for training, validation and testing based. We found that the prediction performance of the FFNN model based on the three way data split outperforms our previous method and the prediction performance is improved from (AUC = 0.8706, precision = 82.5%, accuracy = 82.5%, sensitivity = 82.5%, specificity = 82.5% for the testing set) to (AUC = 0.895, precision = 86.84%, accuracy = 85%, sensitivity = 82.5%, specificity = 87.5% for the testing set).
Further pathway analysis showed that the top three five-marker panels are associated with complement and coagulation cascades, signaling, activation, and hemostasis, which are consistent with previous findings. We believe the new approach is a better solution for multi-biomarker panel discovery and it can be applied to other clinical proteomics.
Breast cancer is the most common cancer among American women, except for skin cancers. About 1 in 8 (12%) women in the US will develop invasive breast cancer during their lifetime. In 2012, an estimated 226, 870 new cases of invasive breast cancer were expected to be diagnosed in women in the U.S., along with 63,300 new cases of non-invasive (in situ) breast cancer .
In recent years, functional genomics studies using DNA Microarrays have been shown effective in differentiating between breast cancer tissues and normal tissues by measuring thousands of differentially expressed genes simultaneously [2–4]. However, early detection and treatment of breast cancer is still challenging. One reason is that obtaining tissue samples for microarray analysis can still be difficult. Another reason is that genes are not directly involved in any physical functions. On the contrary, the proteome are the real functional molecules and the keys to understanding the development of cancer. Moreover, the fact that breast cancer is a complex disease where disease genes exhibit an increased tendency for their protein products to interact with one another [5, 6], makes the disease difficult to detect in early stages by single-marker approach. A chance of success with a multi-biomarker panel is higher than the simpler conventional single-marker approach .
Recent advances in clinical proteomics technology, particularly liquid chromatography coupled tandem mass spectrometry (LC-MS/MS) have enabled biomedical researchers to characterize thousands of proteins in parallel in biological samples. Using LC-MS/MS, it has become possible to detect complex mixtures of proteins, peptides, carbohydrates, DNA, drugs, and many other biologically relevant molecules unique to disease processes . A modern mass spectrometry (MS) instrument consists of three essential modules: an ion source module that can transform molecules to be detected in a sample into ionized fragments, a mass analyzer module that can sort ions by their masses, charges, or shapes by applying electric and magnetic fields, and a detector module that can measure the intensity or abundance of each ion fragment separated earlier. Tandem mass spectrometry (MS/MS) has additional analytical modules for bombarding peptide ions into fragment peptide ions by pipelining two MS modules together, therefore providing peptide sequencing potentials for selected peptide ions in real time. LC-MS/MS proteomics has been used to identify candidate molecular biomarkers in a diverse range of samples, including cells, tissues, serum/plasma, and other types of body fluids. Due to the inherent high variability of both clinical samples and MS/MS instruments, it is still challenging to classify and predict proteomics profiles without an advanced computational method.
Developing a proteomics data analysis method to identify multi-protein biomarker panels for breast cancer diagnosis based on neural networks, therefore, provides hope for improving both the sensitivity and the specificity of candidate disease biomarkers. Neural Networks have several unique advantages and characteristics as research tools for cancer prediction problems [8–12]. A very important feature of these networks is their adaptive nature, where "learning by example" replaces conventional "programming by different cases" in solving problems .
The classification problem of breast cancer can be restricted to consideration of the two-class problem without loss of generality (breast cancer and normal). In the early case study , we developed a Feed Forward Neural Network-based method to build the classifier for plasma samples of breast cancer and then applied the classifier to predict blind dataset of breast cancer. However, the optimal combination C* was actually determined by applying the trained Feed Forward Neural Network (FFNN) on the testing set with the combination, which maximizes the AUC. Therefore, in this paper, we applied a three way data split to the FFNN for training, validation and testing based. Our results show the prediction performance of the FFNN model based on the three way data split outperforms our previous method in the earlier study .
We present the multi-marker panel development solution for early detection of breast cancer using the FFNN model based on three way data split, and show how to use it to model the classification and prediction problem of early detection of breast cancer in plasma proteomics.
Materials and methods
Human plasma samples
Comparison of clinical distribution for study A and B and C
ER- PR- HER2+
LC/MS/MS plasma proteomics analysis
We performed the MS database search against the International Protein Index (IPI, version 3.6, Figure 2d) . Protein quantification was carried out using the same algorithm mentioned before . Briefly, first all extracted ion chromatograms (XIC) were aligned by retention time. Each aligned peak was matched by parent ion, charge state, daughter ions (MS/MS data) and retention time within a one-minute window. Then, the area-under-the-curve (AUC) for each individually aligned peak was measured, normalized, and compared for their relative abundance using methods described in [14, 15].
Linear mixed model
where Here, μ is the mean intensity value, T j is the fixed group effect (caused by the experimental conditions or treatments being evaluated), S k is the random sample effect (random effects from either individual biological samples or sample preparations), I i is the random replicate effect (random effects from replicate injections of the same sample), and ε ijk is the within-groups errors. All of the injections were in random order and the instrument was operated by the same operator. All random effects are assumed independent of each other and independent of the within-group errors ε ijk .
Feed forward neural network
The classification problem of breast cancer can be restricted to consideration of the two-class problem without loss of generality (breast cancer and normal). An FFNN-based method  was used to develop the classifier for plasma samples of breast cancer and then applied the classifier to predict blind dataset of breast cancer. Based on the FFNN-based method, however, we apply in the paper a three way data split for training, validation and testing, instead of directly using Study A as a training set and Study B as a testing set in the . Briefly, Study A is used as the training set for learning to fit the parameters of the classifier, Study C as the validation set to tune the parameters of the classifier, and Study B as the testing set only to assess the performance of the fully-trained classifier.
The enumeration method based on FFNN was built to identify optimal biomarkers panel by us . Similarly, we designed an enumeration method based on the three way data split and the FFNN to find the optimal classifier, which measures the area under the curve (AUC) for Receiver Operating Characteristics (ROC).
Each combination of N (N = 5 for five-marker panel) out of all the 32 genes differentially expressed in the training set is chosen as inputs to the FFNN. We select N = 5 because 1) Li estimated that five or six genes rather than 37 or 738 would be sufficient for the early detection of breast cancer, based on colon cancer, leukemia, and breast cancer , 2) we expect to achieve high prediction accuracy for breast cancer with as few genes as possible, and 3) we applied N = 5 to the prediction of breast cancer in our previous study and achieved satisfied prediction performance .
where AUC is the area under the ROC curve of the FFNN, C is combination of picking five out of the 32 genes, and V is the validation set.
Results and discussions
The plasma proteome sets from Study A, B, and C contains 1423, 1389, and 1249 proteins, respectively. 246 proteins are in common between the three datasets. After ANOVA analysis of the 246 proteins in the Study A, we obtained 32 candidate markers in the training set with pvalue < 0.01. No data from the testing set were utilized in 1) identification of breast cancer markers or 2) development of the FFNN model. The validation set was used to tune the parameters of the FFNN model.
Based on an FFNN model that was built on all 60 markers differentially expressed in Study A and Study B, a high performance (AUC = 0.8713, precision = 86.8%, accuracy = 85%, sensitivity = 82.5%, specificity = 87.5% for the training set; AUC = 0.8706, precision = 82.5%, accuracy = 82.5%, sensitivity = 82.5%, specificity = 82.5% for the testing set) was obtained . However, the optimal combination C* was actually determined by applying the trained FFNN on the testing set with the combination, which maximizes the AUC. This step obtained an objective optimization with training set and testing set. Therefore, in this paper, we applied the three way data split for training, validation and testing and constructed an FFNN for each combination of five out of the 32 markers and trained with plasma samples derived from 40 women diagnosed with breast cancer and 40 control women in the training set. The optimal combinations were obtained by our optimization model based on the training set and validation set instead of the training set and testing set we used in .
Training of the FFNN was performed using back propagation algorithm for two-variable encoding scheme, because we had verified that the two-variable encoding scheme performed better than single-variable encoding scheme . Five performance measurements: (1) Sensitivity; (2) Specificity; (3) Precision; (4) Accuracy; and (5) Area Under the Curve were computed in order to evaluate the prediction performance of the FFNN.
Best three five-marker panels identified
C4BPA; HP; ORM1; SAMD9; SRCRB4D
C4BPA; STBD1; DDX24; GRASP; CFI
C4BPA; CNO; FGG; SERPING1; SRCRB4D
Pathway analysis for the best three five-marker panels.
Complement and coagulation cascades
Intrinsic Prothrombin Activation Pathway
Formation of Fibrin Clot (Clotting Cascade)
Response to elevated platelet cytosolic Ca2+
Staphylococcus aureus infection
GRB2:SOS provides linkage to MAPK signaling for Intergrins
Platelet activation, signaling and aggregation
p130Cas linkage to MAPK signaling for integrins
Extrinsic Prothrombin Activation Pathway
Acute Myocardial Infarction
Beta2 integrin cell surface interactions
Ephrin B reverse signalling
Integrin alphaIIb beta3 signaling
Platelet Aggregation (Plug Formation)
Beta3 integrin cell surface interactions
Regulation of RhoA activity
Urokinase-type plasminogen activator (uPA) and uPAR-mediated signalling
IL6-mediated signaling events
Golgi Associated Vesicle Biogenesis
trans-Golgi Network Vesicle Budding
Clathrin derived vesicle budding
Beta1 integrin cell surface interactions
Prediction result for the best 5-marker panel
We also compared the three way data split with other combination such as mixing three datasets. When mixing three datasets together, the best marker panel has the same performance in the training mode (precision = 87.5%, accuracy = 87.5%, sensitivity = 87.5%, specificity = 87.5%), and a little higher performance in the testing mode (Study B as testing set, precision = 87.18%, accuracy = 86.25%, sensitivity = 85%, specificity = 87.5%). The reason why the performance is higher in the testing mode is because the mixture of three datasets already contains the testing set and the testing set is not independent of the training set. The three way data split method is more close to real applications where testing data are blind or unknown. The prediction performance of the testing set in a three-way data split can better reflect the outcome in a real application than other combination such as the mixture of three datasets. We believe the new approach is a better solution for multi-biomarker panel discovery and it can be applied to other clinical proteomics.
We developed a Feed Forward Neural Network approach that addressed a challenging multi-panel biomarker development problem in the early detection of breast cancer. The approach that we used combined the three way data split with an optimization model of FFNN. We found that the prediction performance of the FFNN model combined with the three way data split outperforms our previous method. Further pathway analysis showed that the top three five-marker panels are associated with complement and coagulation cascades, signaling, activation, and hemostasis, which are consistent with previous findings. We believe the new method is a better solution for multi-biomarker panel discovery and can provide general guidance for future molecular medicine multi-marker panel discovery applications in other diseases. In the future, we will follow up with biological experiments to validate these biomarkers with our collaborators.
We thank Hoosier Oncology Group for collecting breast cancer plasma samples. The proteomics study for biomarker discovery was supported by the National Cancer Institute Clinical Proteomics Technology Assessment for Cancer program (U24 CA126480). We also thank the support of Bioinformatics Program at University of North Texas Health Science Center.
Publication of this work was supported by the Bioinformatics Program at University of North Texas Health Science Center.
This article has been published as part of BMC Proceedings Volume 7 Supplement 7, 2013: Proceedings of the Great Lakes Bioinformatics Conference 2013. The full contents of the supplement are available online at http://www.biomedcentral.com/bmcproc/supplements/7/S7.
- What are the key statistics about breast cancer?. [http://www.cancer.org/Cancer/BreastCancer/DetailedGuide/breast-cancer-key-statistics]
- Hu X, Zhang Y, Zhang A, Li Y, Zhu Z, Shao Z, Zeng R, Xu LX: Comparative serum proteome analysis of human lymph node negative/positive invasive ductal carcinoma of the breast and benign breast disease controls via label-free semiquantitative shotgun technology. OMICS. 2009, 13 (4): 291-300. 10.1089/omi.2009.0016.View ArticlePubMedGoogle Scholar
- Zeidan BA, Cutress RI, Murray N, Coulton GR, Hastie C, Packham G, Townsend PA: Proteomic analysis of archival breast cancer serum. Cancer Genomics Proteomics. 2009, 6 (3): 141-147.PubMedGoogle Scholar
- Lebrecht A, Boehm D, Schmidt M, Koelbl H, Schwirz RL, Grus FH: Diagnosis of breast cancer by tear proteomic pattern. Cancer Genomics Proteomics. 2009, 6 (3): 177-182.PubMedGoogle Scholar
- Polyak K: Breast cancer: origins and evolution. J Clin Invest. 2007, 117 (11): 3155-3163. 10.1172/JCI33295.PubMed CentralView ArticlePubMedGoogle Scholar
- Zhang F, Chen JY: Discovery of pathway biomarkers from coupled proteomics and systems biology methods. BMC genomics. 2010, 11 (Suppl 2): S12-10.1186/1471-2164-11-S2-S12.View ArticleGoogle Scholar
- Klampfl CW: Review coupling of capillary electrochromatography to mass spectrometry. J Chromatogr A. 2004, 1044 (1-2): 131-144. 10.1016/j.chroma.2004.04.072.View ArticlePubMedGoogle Scholar
- Lai KC, Chiang HC, Chen WC, Tsai FJ, Jeng LB: Artificial neural network- based study can predict gastric cancer staging. Hepatogastroenterology. 2008, 55 (86-87): 1859-1863.PubMedGoogle Scholar
- Amiri Z, Mohammad K, Mahmoudi M, Zeraati H, Fotouhi A: Assessment of gastric cancer survival: using an artificial hierarchical neural network. Pak J Biol Sci. 2008, 11 (8): 1076-1084. 10.3923/pjbs.2008.1076.1084.View ArticlePubMedGoogle Scholar
- Chi CL, Street WN, Wolberg WH: Application of artificial neural network- based survival analysis on two breast cancer datasets. AMIA Annu Symp Proc. 2007, 130-134.Google Scholar
- Anagnostopoulos I, Maglogiannis I: Neural network-based diagnostic and prognostic estimations in breast cancer microscopic instances. Med Biol Eng Comput. 2006, 44 (9): 773-784. 10.1007/s11517-006-0079-4.View ArticlePubMedGoogle Scholar
- Wang HQ, Wong HS, Zhu H, Yip TT: A neural network-based biomarker association information extraction approach for cancer classification. J Biomed Inform. 2009Google Scholar
- Fan Z, Chen JY: A neural network approach to multi-biomarker panel development based on LC/MS/MS proteomics profiles: A case study in breast cancer. Computer-Based Medical Systems, 2009 CBMS 2009 22nd IEEE International Symposium on: 2-5 Aug. 2009. 2009, 1-6.Google Scholar
- Wang M, You J, Bemis KG, Tegeler TJ, Brown DP: Label-free mass spectrometry-based protein quantification technologies in proteomic analysis. Brief Funct Genomic Proteomic. 2008, 7 (5): 329-339. 10.1093/bfgp/eln031.View ArticlePubMedGoogle Scholar
- Higgs RE, Knierman MD, Gelfanova V, Butler JP, Hale JE: Comprehensive label-free method for the relative quantification of proteins from biological samples. Journal of proteome research. 2005, 4 (4): 1442-1450. 10.1021/pr050109b.View ArticlePubMedGoogle Scholar
- Kersey PJ, Duarte J, Williams A, Karavidopoulou Y, Birney E, Apweiler R: The International Protein Index: an integrated database for proteomics experiments. Proteomics. 2004, 4 (7): 1985-1988. 10.1002/pmic.200300721.View ArticlePubMedGoogle Scholar
- Bolstad BM, Irizarry RA, Astrand M, Speed TP: A comparison of normalization methods for high density oligonucleotide array data based on variance and bias. Bioinformatics. 2003, 19 (2): 185-193. 10.1093/bioinformatics/19.2.185.View ArticlePubMedGoogle Scholar
- Heaton J: Introduction to Neural Networks for Java. 2008, Heaton Research, Inc.;, 2Google Scholar
- Li W: How many genes are needed for early detection of breast cancer, based on gene expression patterns in peripheral blood cells?. Breast cancer research : BCR. 2005, 7 (5): E5-10.1186/bcr1295.PubMed CentralView ArticlePubMedGoogle Scholar
- Zhang F, Drabier R: IPAD: the Integrated Pathway Analysis Database for Systematic Enrichment Analysis. BMC Bioinformatics. 2012, 13 (14):Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.