Gökmen Altay

Revealing differences in gene network inference algorithms on the network level by ensemble methods

MOTIVATION The inference of regulatory networks from large-scale expression data holds great promise because of the potentially causal interpretation of these networks. However, due to the difficulty to establish reliable methods based on observational data there is so far only incomplete knowledge about possibilities and limitations of such inference methods in this context. RESULTS In this article, we conduct a statistical analysis investigating differences and similarities of four network inference algorithms, ARACNE, CLR, MRNET and RN, with respect to local network-based measures. We employ ensemble methods allowing to assess the inferability down to the level of individual edges. Our analysis reveals the bias of these inference methods with respect to the inference of various network components and, hence, provides guidance in the interpretation of inferred regulatory networks from expression data. Further, as application we predict the total number of regulatory interactions in human B cells and hypothesize about the role of Myc and its targets regarding molecular information processing.

A comprehensive comparison of association estimators for gene network inference algorithms

MOTIVATION Gene network inference (GNI) algorithms enable the researchers to explore the interactions among the genes and gene products by revealing these interactions. The principal process of the GNI algorithms is to obtain the association scores among genes. Although there are several association estimators used in different applications, there is no commonly accepted estimator as the best one for the GNI applications. In this study, 27 different interaction estimators were reviewed and 14 most promising ones among them were evaluated by using three popular GNI algorithms with two synthetic and two real biological datasets belonging to Escherichia coli bacteria and Saccharomyces cerevisiae yeast. Influences of the Copula Transform (CT) pre-processing operation on the performance of the interaction estimators are also observed. This study is expected to assist many researchers while studying with GNI applications. RESULTS B-spline, Pearson-based Gaussian and Spearman-based Gaussian association score estimators outperform the others for all datasets in terms of the performance and runtime. In addition to this, it is observed that, when the CT operation is used, inference performances of the estimators mostly increase, especially for two synthetic datasets. Detailed evaluations and discussions are given in the experimental results. CONTACT gokmen.altay@bahcesehir.edu.tr SUPPLEMENTARY INFORMATION Supplementary data are available at Bioinformatics online.

Feature Selection For The Prediction Of Tropospheric Ozone Concentration Using A Wrapper Method

High concentrations of ozone (03) in the lower troposphere increase global warming, and thus affect climatic conditions and human health. Especially in metropolitan cities like Istanbul, ozone leve...

DepEst: an R package of important dependency estimators for gene network inference algorithms

Gene network inference algorithms (GNI) are popular in bioinformatics area. In almost all GNI algorithms, the main process is to estimate the dependency (association) scores among the genes of the dataset. We present a bioinformatics tool, DepEst (Dependency Estimators), which is a powerful and flexible R package that includes 11 important dependency score estimators that can be used in almost all GNI Algorithms. DepEst is the first bioinformatics package that includes such a large number of estimators that runs both in parallel and serial. DepEst is currently available at https://github.com/altayg/Depest. Package access link, instructions, various workflows and example data sets are provided in the supplementary file.

Trellis decoding of some decomposable codes

This paper presents a maximum likelihood (ML) soft decision decoding scheme to implement Viterbi Algorithm for some constructed decomposable codes of Hamming distance four and show that the decoding complexity is simple, hence, it may be employed in trellis - based decoders. Bit error rate performances of some decomposable codes that were obtained employing the technique in AWGN channel are also presented.

Directed Conservative Causal Core Gene Networks

Motivation: Inferring large scale directional networks with higher accuracy has important applications such as gene regulatory network or finance. Results: We modified a well-established conservative causal core network inference algorithm, C3NET, to be able to infer very large scale networks with direction information. This advanced version is called Ac3net. We demonstrate that Ac3net outperforms C3NET and many other popular algorithms when considering the directional interaction information of gene/protein networks. We provide and R package and present performance results that are reproducible via the Supplementary file. Availability: Ac3net is available on CRAN and at github.com/altayg/Ac3net Contact: altay@lji.org Supplementary information: Supplementary file is available online.

Tensorflow Based Deep Learning Model and Snakemake Workflow for Peptide-Protein Binding Predictions

In this study, we first present a Tensorflow based Deep Learning (DL) model that provides high performances in predicting the binding of peptides to major histocompatibility complex (MHC) class I protein. Second, we provide the necessary Python codes to run the model and also easily input large train and test peptide binding benchmark dataset. Third, we provide Snakemake based workflow that allows to run all the model and performance analysis over all the different test alleles at once in parallel over computer and clusters. We also provide comparison analysis of the performances of various models. Finally, in order to help attaining to the best possible DL model by a community effort, this work is intended to be a ready to modify base model and workflow for the global Deep Learning community with no domain knowledge in MHC-peptide binding problem and thus provides all the necessary reference code templates and benchmarking data sets for further developments on the presented model architecture. All the reproducible Python codes, Snakemake workflow and benchmark data sets and a tutorial are available online at https://github.com/altayg/Deep-Learning-MHCI.

Gene Regulatory Cross Networks: Inferring Gene Level Cell-to-Cell Communications of Immune Cells

Background Gene level cell-to-cell communications are crucial part of biology as they may be potential targets of drugs and vaccines against a disease condition of interest. Yet, there are only few studies that propose algorithms on this particularly important research field. Results In this study, we first overview the current literature and define two general terms for the types of approaches in general for gene level cell-to-cell communications: Gene Regulatory Cross Networks (GRCN) and Gene Co-Expression Cross Networks (GCCN). We then propose two algorithms for each type, named as GRCNone and GCCNone. We applied them to reveal communications among 8 different immune cell types and evaluate their performances mainly via membrane protein database. Also, we show the biological relevance of the predicted cross-networks with pathway enrichment analysis. We then provide an approach that prioritize the targets by ranking them before experimental validations. Conclusions We establish two main approaches and propose algorithms for genome-wide scale gene level cell-to-cell communications between any two different cell-types. This study aims accelerating this relatively new avenue of research in cross-networks and points out the gap of it with the well-established single cell type gene networks. The proposed algorithms have the potential to reveal gene level interactions between normal and disease cell types. For instance, they might reveal the interaction of genes between tumor and normal cells, which are the potential drug-targets and thus can help finding new cures that might prevent the prevailing of tumor cells.

Genome-Wide Differential Gene Network Analysis R Software And Its Application In LnCap Prostate Cancer

We introduce an R software package for condition-specific gene regulatory network analysis based on DC3NET algorithm. We also present an application of it on a real prostate dataset and demonstrate the benefit of the software. We performed genome-wide differential gene network analysis with the software on the LnCap androgen stimulated and deprived prostate cancer gene expression datasets (GSE18684) and inferred the androgen stimulated prostate cancer specific differential network. As an outstanding result, CXCR7 along with CXCR4 appeared to have the most important role in the androgen stimulated prostate specific genome-wide differential network. This blind estimation is strongly supported from the literature. The critical roles for CXCR4, a receptor over-expressed in many cancers, and CXCR7 on mediating tumor metastasis, along with their contributions as biomarkers of tumor behavior as well as potential therapeutic target were studied in several other types of cancers. In fact, a pharmaceutical company had already developed a therapy by inhibiting CXCR4 to block non-cancerous immuno-suppressive and pro-angiogenic cells from populating the tumor for disrupting the cancer environment and restoring normal immune surveillance functions. Considering this strong confirmation, our inferred regulatory network might reveal the driving mechanism of LnCap androgen stimulated prostate cancer. Because, CXCR4 appeared to be in the center of the largest subnetwork of our inferred differential network. Moreover, enrichment analyses for the largest subnetwork of it appeared to be significantly enriched in terms of axon guidance, fc gamma R-mediated phagocytosis and endocytosis. This also conforms with the recent literature in the field of prostate cancer. We demonstrate how to derive condition-specific gene targets from expression datasets on genome-wide level using differential gene network analysis. Our results showed that differential gene network analysis worked well in a prostate cancer dataset, which suggest the use of this approach as essential part of current expression data processing. Availability: The introduced R software package available in CRAN at https://cran.r-project.org/web/packages/dc3net and also at https://github.com/altayg/dc3netWe present an R software package that performs at genome-wide level differential network analysis and infers only disease-specific molecular interactions between two different cell conditions. This helps revealing the disease mechanism and predicting most influential genes as potential drug targets or biomarkers of the disease condition of interest. As an exemplary analysis, we performed an application of the software over LNCaP datasets and, out of approximately 25000 genes, predicted CXCR7 and CXCR4 together as drug targets of LNCaP prostate cancer dataset. We further successfully validated them with our initial wet-lab experiments. The introduced software can be applied to all the diseases, especially cancer, with gene expression data of two different conditions (e.g. tumor vs normal) and thus has the potential of a global benefit. As a distinct remark, our software provide the causal disease mechanism with multiple potential drug-targets rather than a single independent target prediction. Availability The introduced R software package for the analysis is available in CRAN at https://cran.r-project.org/web/packages/dc3net and also at https://github.com/altayg/dc3net

Inferring Genome-Wide Interaction Networks

The inference of gene regulatory networks is an important process that contributes to a better understanding of biological and biomedical problems. These networks aim to capture the causal molecular interactions of biological processes and provide valuable information about normal cell physiology. In this book chapter, we introduce GNI methods, namely C3NET, RN, ARACNE, CLR, and MRNET and describe their components and working mechanisms. We present a comparison of the performance of these algorithms using the results of our previously published studies. According to the study results, which were obtained from simulated as well as expression data sets, the inference algorithm C3NET provides consistently better results than the other widely used methods.