Burook Misganaw

Sparse feature selection for classification and prediction of metastasis in endometrial cancer

BackgroundMetastasis via pelvic and/or para-aortic lymph nodes is a major risk factor for endometrial cancer. Lymph-node resection ameliorates risk but is associated with significant co-morbidities. Incidence in patients with stage I disease is 4–22% but no mechanism exists to accurately predict it. Therefore, national guidelines for primary staging surgery include pelvic and para-aortic lymph node dissection for all patients whose tumor exceeds 2cm in diameter. We sought to identify a robust molecular signature that can accurately classify risk of lymph node metastasis in endometrial cancer patients. 86 tumors matched for age and race, and evenly distributed between lymph node-positive and lymph node-negative cases, were selected as a training cohort. Genomic micro-RNA expression was profiled for each sample to serve as the predictive feature matrix. An independent set of 28 tumor samples was collected and similarly characterized to serve as a test cohort.ResultsA feature selection algorithm was designed for applications where the number of samples is far smaller than the number of measured features per sample. A predictive miRNA expression signature was developed using this algorithm, which was then used to predict the metastatic status of the independent test cohort. A weighted classifier, using 18 micro-RNAs, achieved 100% accuracy on the training cohort. When applied to the testing cohort, the classifier correctly predicted 90% of node-positive cases, and 80% of node-negative cases (FDR = 6.25%).ConclusionResults indicate that the evaluation of the quantitative sparse-feature classifier proposed here in clinical trials may lead to significant improvement in the prediction of lymphatic metastases in endometrial cancer patients.

Exploiting Ordinal Class Structure in Multiclass Classification: Application to Ovarian Cancer

In multiclass machine learning problems, one needs to distinguish between the nominal labels that do not have any natural ordering and the ordinal labels that are ordered. Ordinal labels are pervasive in biology, and some examples are given here. In this note, we point out the importance of making use of the order information when it is inherent to the problem. We demonstrate that algorithms that use this additional information outperform the algorithms that do not, on a case study of assigning one of four labels to the ovarian cancer patients on the basis of their time of progression-free survival. As an aside, it is also pointed out that the algorithms that make use of ordering information require fewer data normalizations. This aspect is important in biological applications, where data are plagued by variations in platforms and protocols, batch effects, and so on.

A transfer learning approach for integrating biological data across platforms

Transfer learning refers to situations where a classifier is trained on one set of data and tested on another set of data that may have an entirely different probability distribution. Biological data derived from diverse platforms, and possibly using diverse technologies, is a natural candidate for applying transfer learning methodologies. In this paper, we adapt the ℓ1-norm SVM to fit into the paradigm of Transfer Learning, by using the importance weighting approach. Our aim is to integrate biological data from diverse platforms. To validate our approach, we applied the proposed algorithm to the problem of classifying breast cancer tumors as Estrogen- Receptor-positive (ER-positive) or Estrogen-Receptor-negative (ER-negative), which is the first step in personalizing therapy to the patient. The standard approach used in Biology is to convert data to Z-scores, that is, to subtract the mean and divide by the standard deviation. The algorithm proposed here shows better performance than using Z-scores to account for platform variations.

Prediction of Time to Tumor Recurrence in Ovarian Cancer: Comparison of Three Sparse Regression Methods

Ovarian cancer is the most fatal gynecological malignancy among women. Making a reliable prediction of time to tumor recurrence would be a valuable contribution to post-surgery follow-up care. In this paper we study three well-known data sets, known as TCGA, Tothill and Yoshihara, and compare three sparse regression methods, two of which (LASSO and EN) are well-known and the third (CLOT) is from our laboratory. It is established that the three data sets are very different from each other. Therefore a two-stage predictor is built, whereby each test sample is first assigned to the most likely data set and then the corresponding predictor is used. The weighted concordance of each regression method is computed to compare the methods and select the best one. CLOT uses a biomarker panel of 103 genes and achieves a concordance index of 0.7829, which is higher than that achieved by the other two methods.

A multi-view ℓ 1 -norm SVM algorithm for data integration in biological applications

A multi-view ℓ1-norm Support Vector Machine (SVM) for integrating data from different views to improve binary classification performance in a given view is proposed. The performance of the proposed algorithm is evaluated by integrating biological data from two different gene expression measurement technologies. The experimental results show that the data integration method proposed leads to a better classification performance in comparison to the traditional ℓ1-norm SVM.

Integrating biological data across multiple platforms using importance-weighted transfer learning and applications to breast cancer data sets

Traditional machine learning approaches are based on the premise that the training and testing samples come from a common probability distribution. Transfer learning refers to situations where this assumption does not necessarily hold. Integrating biological data measured on diverse platforms is a major challenge. Transfer learning is a natural candidate for achieving such integration. In this paper, we adapt the ¿1 — norm SVM using the importance weighting approach to fit into the paradigm of Transfer Learning under Covariate Shift, with the aim of integrating biological data sources from diverse platforms. The conditional probability of the testing data with respect to the training data is estimated using a small number of testing samples. The weights of the ℓ1-norm SVM are adapted using this estimated conditional probability, also known as the importance weight. To validate our approach, we applied the proposed algorithm to the problem of classifying breast cancer tumors as ERpositive or ER-negative, which is the first step in personalizing therapy to the patient. Then we compared it against conversion to Z-scores, which is the current best practice. The ℓ1-norm SVM modified via importance weighting shows better performance than using Z-scores, on five different test data sets.

Sparse Feature Selection for Classification and Prediction of Metastasis in Endometrial Cancer: Extended Abstract

Metastasis via pelvic and/or para-aortic lymph nodes is a major risk factor for endometrial cancer. Lymph-node resection ameliorates risk but is associated with significant co-morbidities. Incidence in patients with stage I disease is 4-22% but no mechanism exists to accurately predict it. Therefore, national guidelines for primary staging surgery include pelvic and para-aortic lymph node dissection for all patients whose tumor exceeds 2cm in diameter. We sought to identify a robust molecular signature that can accurately classify risk of lymph node metastasis in endometrial cancer patients. We introduce a new feature selection algorithm, lone star, for applications where the number of samples is far smaller than the number of measured features per sample. We applied lone star to develop a predictive miRNA expression signature on a training. When applied on an independent testing cohort, the classifier correctly predicted 90% of node-positive cases, and 80% of node-negative cases (FDR= 6.25%). Our results indicate that the evaluation of the quantitative sparse-feature classifier proposed here in clinical trials may lead to significant improvement in the prediction of lymphatic metastases in endometrial cancer patients.

A robust platform-independent gene signature for single-sample breast cancer subtyping

The first step in personalizing therapy for a breast cancer patient is to assign the patient to the appropriate subtype. At present this is done through immunohistochemistry (IHC) which is both expensive as well as wasteful of tumor material. Moreover, error rates can be as high as 20%. In this paper, we propose a new approach to the subtyping of breast cancer patients by using data-driven “reference genes” so that the classification can be carried out for one patient at a time, and across multiple platforms. The validity of the proposed approach is established by training a binary classifier on the TCGA breast cancer data set measured on an Agilent platform, and then applying this classifier to five independent test data sets, on both the Agilent as well as the Affymetrix platforms. In all cases, the results are excellent.