Sebastian Student

Stable feature selection and classification algorithms for multiclass microarray data

BackgroundRecent studies suggest that gene expression profiles are a promising alternative for clinical cancer classification. One major problem in applying DNA microarrays for classification is the dimension of obtained data sets. In this paper we propose a multiclass gene selection method based on Partial Least Squares (PLS) for selecting genes for classification. The new idea is to solve multiclass selection problem with the PLS method and decomposition to a set of two-class sub-problems: one versus rest (OvR) and one versus one (OvO). We use OvR and OvO two-class decomposition for other recently published gene selection method. Ranked gene lists are highly unstable in the sense that a small change of the data set often leads to big changes in the obtained ordered lists. In this paper, we take a look at the assessment of stability of the proposed methods. We use the linear support vector machines (SVM) technique in different variants: one versus one, one versus rest, multiclass SVM (MSVM) and the linear discriminant analysis (LDA) as a classifier. We use balanced bootstrap to estimate the prediction error and to test the variability of the obtained ordered lists.ResultsThis paper focuses on effective identification of informative genes. As a result, a new strategy to find a small subset of significant genes is designed. Our results on real multiclass cancer data show that our method has a very high accuracy rate for different combinations of classification methods, giving concurrently very stable feature rankings.ConclusionsThis paper shows that the proposed strategies can improve the performance of selected gene sets substantially. OvR and OvO techniques applied to existing gene selection methods improve results as well. The presented method allows to obtain a more reliable classifier with less classifier error. In the same time the method generates more stable ordered feature lists in comparison with existing methods.ReviewersThis article was reviewed by Prof Marek Kimmel, Dr Hans Binder (nominated by Dr Tomasz Lipniacki) and Dr Yuriy Gusev

Gene Expression Analysis in Ovarian Cancer – Faults and Hints from DNA Microarray Study

The introduction of microarray techniques to cancer research brought great expectations for finding biomarkers that would improve patients’ treatment; however, the results of such studies are poorly reproducible and critical analyses of these methods are rare. In this study, we examined global gene expression in 97 ovarian cancer samples. Also, validation of results by quantitative RT-PCR was performed on 30 additional ovarian cancer samples. We carried out a number of systematic analyses in relation to several defined clinicopathological features. The main goal of our study was to delineate the molecular background of ovarian cancer chemoresistance and find biomarkers suitable for prediction of patients’ prognosis. We found that histological tumor type was the major source of variability in genes expression, except for serous and undifferentiated tumors that showed nearly identical profiles. Analysis of clinical endpoints [tumor response to chemotherapy, overall survival, disease-free survival (DFS)] brought results that were not confirmed by validation either on the same group or on the independent group of patients. CLASP1 was the only gene that was found to be important for DFS in the independent group, whereas in the preceding experiments it showed associations with other clinical endpoints and with BRCA1 gene mutation; thus, it may be worthy of further testing. Our results confirm that histological tumor type may be a strong confounding factor and we conclude that gene expression studies of ovarian carcinomas should be performed on histologically homogeneous groups. Among the reasons of poor reproducibility of statistical results may be the fact that despite relatively large patients’ group, in some analyses one has to compare small and unequal classes of samples. In addition, arbitrarily performed division of samples into classes compared may not always reflect their true biological diversity. And finally, we think that clinical endpoints of the tumor probably depend on subtle changes in many and, possibly, alternative molecular pathways, and such changes may be difficult to demonstrate.

Global Gene Expression Profiling in Three Tumor Cell Lines Subjected to Experimental Cycling and Chronic Hypoxia

Hypoxia is one of the most important features of the tumor microenvironment, exerting an adverse effect on tumor aggressiveness and patient prognosis. Two types of hypoxia may occur within the tumor mass, chronic (prolonged) and cycling (transient, intermittent) hypoxia. Cycling hypoxia has been shown to induce aggressive tumor cell phenotype and radioresistance more significantly than chronic hypoxia, though little is known about the molecular mechanisms underlying this phenomenon. The aim of this study was to delineate the molecular response to both types of hypoxia induced experimentally in tumor cells, with a focus on cycling hypoxia. We analyzed in vitro gene expression profile in three human cancer cell lines (melanoma, ovarian cancer, and prostate cancer) exposed to experimental chronic or transient hypoxia conditions. As expected, the cell-type specific variability in response to hypoxia was significant. However, the expression of 240 probe sets was altered in all 3 cell lines. We found that gene expression profiles induced by both types of hypoxia were qualitatively similar and strongly depend on the cell type. Cycling hypoxia altered the expression of fewer genes than chronic hypoxia (6,132 vs. 8,635 probe sets, FDR adjusted p<0.05), and with lower fold changes. However, the expression of some of these genes was significantly more affected by cycling hypoxia than by prolonged hypoxia, such as IL8, PLAU, and epidermal growth factor (EGF) pathway-related genes (AREG, HBEGF, and EPHA2). These transcripts were, in most cases, validated by quantitative reverse transcription polymerase chain reaction (qRT-PCR). Our results indicate that experimental cycling hypoxia exerts similar, although less intense effects, on the examined cancer cell lines than its chronic counterpart. Nonetheless, we identified genes and molecular pathways that seem to be preferentially regulated by cyclic hypoxia.

Cell Death in the Epithelia of the Intestine and Hepatopancreas in Neocaridina heteropoda (Crustacea, Malacostraca)

The endodermal region of the digestive system in the freshwater shrimp Neocaridina heteropoda (Crustacea, Malacostraca) consists of a tube-shaped intestine and large hepatopancreas, which is formed by numerous blind-ended tubules. The precise structure and ultrastructure of these regions were presented in our previous studies, while here we focused on the cell death processes and their effect on the functioning of the midgut. We used transmission electron microscopy, light and confocal microscopes to describe and detect cell death, while a quantitative assessment of cells with depolarized mitochondria helped us to establish whether there is the relationship between cell death and the inactivation of mitochondria. Three types of the cell death were observed in the intestine and hepatopancreas–apoptosis, necrosis and autophagy. No differences were observed in the course of these processes in males and females and or in the intestine and hepatopancreas of the shrimp that were examined. Our studies revealed that apoptosis, necrosis and autophagy only involves the fully developed cells of the midgut epithelium that have contact with the midgut lumen–D-cells in the intestine and B- and F-cells in hepatopancreas, while E-cells (midgut stem cells) did not die. A distinct correlation between the accumulation of E-cells and the activation of apoptosis was detected in the anterior region of the intestine, while necrosis was an accidental process. Degenerating organelles, mainly mitochondria were neutralized and eventually, the activation of cell death was prevented in the entire epithelium due to autophagy. Therefore, we state that autophagy plays a role of the survival factor.

Development of a population of cancer cells: Observation and modeling by a Mixed Spatial Evolutionary Games approach.

Living cells, like whole living organisms during evolution, communicate with their neighbors, interact with the environment, divide, change their phenotypes, and eventually die. The development of specific ways of communication (through signaling molecules and receptors) allows some cellular subpopulations to survive better, to coordinate their physiological status, and during embryonal development to create tissues and organs or in some conditions to become tumors. Populations of cells cultured in vitro interact similarly, also competing for space and nutrients and stimulating each other to better survive or to die. The results of these intercellular interactions of different types seem to be good examples of biological evolutionary games, and have been the subjects of simulations by the methods of evolutionary game theory where individual cells are treated as players. Here we present examples of intercellular contacts in a population of living human cancer HeLa cells cultured in vitro and propose an evolutionary game theory approach to model the development of such populations. We propose a new technique termed Mixed Spatial Evolutionary Games (MSEG) which are played on multiple lattices corresponding to the possible cellular phenotypes which gives the possibility of simulating and investigating the effects of heterogeneity at the cellular level in addition to the population level. Analyses performed with MSEG suggested different ways in which cellular populations develop in the case of cells communicating directly and through factors released to the environment.

Integrated System Supporting Research on Environment Related Cancers

There are many impediments to progress in cancer research. Insufficient or low quality data and computational tools that are dispersed among various sites are one of them. In this paper we present an integrated system that combines all stages of cancer studies, from gathering of clinical data, through elaborate patient questionnaires and bioinformatics tools, to data warehousing and preparation of analysis reports.

Unsupervised analysis reveals two molecular subgroups of serous ovarian cancer with distinct gene expression profiles and survival

PurposeOvarian cancer is typically diagnosed at late stages, and thus, patients’ prognosis is poor. Improvement in treatment outcomes depends, at least partly, on better understanding of ovarian cancer biology and finding new molecular markers and therapeutic targets.MethodsAn unsupervised method of data analysis, singular value decomposition, was applied to analyze microarray data from 101 ovarian cancer samples; then, selected genes were validated by quantitative PCR.ResultsWe found that the major factor influencing gene expression in ovarian cancer was tumor histological type. The next major source of variability was traced to a set of genes mainly associated with extracellular matrix, cell motility, adhesion, and immunological response. Hierarchical clustering based on the expression of these genes revealed two clusters of ovarian cancers with different molecular profiles and distinct overall survival (OS). Patients with higher expression of these genes had shorter OS than those with lower expression. The two clusters did not derive from high- versus low-grade serous carcinomas and were unrelated to histological (ovarian vs. fallopian) origin. Interestingly, there was considerable overlap between identified prognostic signature and a recently described invasion-associated signature related to stromal desmoplastic reaction. Several genes from this signature were validated by quantitative PCR; two of them—DSPG3 and LOX—were validated both in the initial and independent sets of samples and were significantly associated with OS and disease-free survival.ConclusionsWe distinguished two molecular subgroups of serous ovarian cancers characterized by distinct OS. Among differentially expressed genes, some may potentially be used as prognostic markers. In our opinion, unsupervised methods of microarray data analysis are more effective than supervised methods in identifying intrinsic, biologically sound sources of variability. Moreover, as histological type of the tumor is the greatest source of variability in ovarian cancer and may interfere with analyses of other features, it seems reasonable to use histologically homogeneous groups of tumors in microarray experiments.

A Holistic Approach to Testing Biomedical Hypotheses and Analysis of Biomedical Data

Testing biomedical hypotheses is performed based on advanced and usually many-step analysis of biomedical data. This requires sophisticated analytical methods and data structures that allow to store intermediate results, which are needed in the subsequent steps. However, biomedical data, especially reference data, often change in time and new analytical methods are created every year. This causes the necessity to repeat the iterative analyses with new methods and new reference data sets, which in turn causes frequent changes of the underlying data structures. Such instability of data structures can be mitigated by the use of the idea of data lake, instead of traditional database systems.

Large-Scale Data Classification System Based on Galaxy Server and Protected from Information Leak

In this work we present SPICY (SPecialized Classification sYstem) application for a supervised data analysis (feature selection, classification, model validation and model selection) with the structure preventing the data processing work-flow from so called information leak. The information leak may result in optimistically biased classification quality assessment, especially for large-scale, small-sample data sets. The application uses the Galaxy Server environment that originally allows the user to manual data processing and is not prevented from the information leak. The way how the classification model is built by the user and the specific structure of all implemented methods makes the information leak impossible. The lack of information leak in the presented supervised data analysis tool is demonstrated on numerical examples, where synthetic and real data sets are used.

The structure of the desiccated Richtersius coronifer (Richters, 1903)

Tun formation is an essential morphological adaptation for entering the anhydrobiotic state in tardigrades, but its internal structure has rarely been investigated. We present the structure and ultrastructure of organs and cells in desiccated Richtersius coronifer by transmission and scanning electron microscopy, confocal microscopy, and histochemical methods. A 3D reconstruction of the body organization of the tun stage is also presented. The tun formation during anhydrobiosis of tardigrades is a process of anterior-posterior body contraction, which relocates some organs such as the pharyngeal bulb. The cuticle is composed of epicuticle, intracuticle and procuticle; flocculent coat; and trilaminate layer. Moulting does not seem to restrict the tun formation, as evidenced from tardigrade tuns that were in the process of moulting. The storage cells of desiccated specimens filled up the free inner space and surrounded internal organs, such as the ovary and digestive system, which were contracted. All cells (epidermal cells, storage cells, ovary cells, cells of the digestive system) underwent shrinkage, and their cytoplasm was electron dense. Lipids and polysaccharides dominated among reserve material of storage cells, while the amount of protein was small. The basic morphology of specific cell types and organelles did not differ between active and anhydrobiotic R. coronifer.