Ensemble classification for gene expression data based on parallel clustering

Analysis of temporal gene expression data poses a significant challenge due to the combination of high dimensionality and low sample size. The purpose of this paper is to present a methodology for classification, modelling, and analysis of short time-series gene expression data using spiking neural networks (SNN) and to uncover temporal expression patterns for knowledge discovery. The classification is based on the NeuCube SNN model. Time-series gene expression data of mouse primary cortical neurons is examined as a case study. The results of the analysis are promising, indicating that SNN methodologies can be effectively used to model and analyse temporal gene expression data with surpassing performance over traditional machine learning algorithms. Additionally, a gene interaction network is constructed from the temporal gene activity modelled using the NeuCube architecture offering a new way of knowledge discovery. Future work will be directed towards using gene interactions networks to help guide pharmacological research for dementia.

Machine Learning algorithms have been widely used for gene expression data classification, despite the fact that these data have often intrinsic limitations, such as high dimensionality and a small number of examples. Few studies try to characterize to which extent these aspects can influence the performance of the classification models induced. In this paper we compute different measures characterizing the complexity of gene expression data sets for cancer diagnosis. We then investigate how these measures relate to the classification performances achieved by support vector machines, a popular Machine Learning technique usually employed in the analysis of gene expression data. The results obtained indicate that some of the complexity indices utilized are indeed successful in explaining the difficulty involved in the classification of cancer gene expression data.

The purpose of this paper is to introduce transductive inference with support vector machines (TSVM) as a powerful methodology for classification of gene expression data, using training and prediction data sets. The following classification problems will be considered: determination of cancer diagnosis categories and classification of genes from the budding yeast Saccharomyces cerevisiae in functional groups. In the case of training samples, experts have already classified the samples in their respective classes. So, given each prediction sample, the purpose is to determine its corresponding class. The main aspect of TSVM is that the classification task will be implemented in just one step, improving the generalization capability of the classifier. The TSVM will be compared with the traditional inductive method (SVM) in a series of experiments concerning the two classification problems, with promising results.

The application of mathematics in the field of bioinformatics has been widely developed. For example Support Vector Machines (SVM) and Random Forest (RF) are state of the art for classification of cancer in many applications. One of them is Chronic Kidney Disease (CKD). CKD is one of the kidney diseases that sufferers are increasing and have symptoms that are difficult to detect at first. Later, microarrays in gene expression are important tools for this approach. Microarrays gene expression provides an overview of all transcription activities in biological samples. The purpose of this research is a hybrid model combining Random Forest (RF) and Support Vector Machine (SVM) can be used to classify gene expression data. RF can highly accurate, generelize better and are interpretable and SVM (called RF-SVM) to effectively predict gene expression data with very high dimensions. In addition, from the simulation results on data from the Gene Expression Omnibus (GEO) database, it is shown that the proposed RF-SVM is a more accurate algorithm on CKD data than RFE-SVM.

The paper presents a methodology to assess the problems behind static gene expression data modelling and analysis with machine learning techniques. As a case study, transcriptomic data collected during a longitudinal study on the effects of diet on the expression of oxidative phosphorylation genes was used. Data were collected from 60 abdominally overweight men and women after an observation period of eight weeks, whilst they were following three different diets. Real-valued static gene expression data were encoded into spike trains using Gaussian receptive fields for multinomial classification using an evolving spiking neural network (eSNN) model. Results demonstrated that the proposed method can be used for predictive modelling of static gene expression data and future works are proposed regarding the application of eSNNs for personalised modelling.

Recent research has focused on incorporating biological function and pathway information into the analysis of gene expression data, partly as a means of compensating for insufficient experimental replications, low signal to noise, lack of reproducibility and/or multiple testing confounds. A Bayesian approach seems to be ideal for incorporating functional information into gene expression data analysis. In this study, we tested the feasibility of using literature derived gene relationships in a Bayesian model to analyze gene expression data. Prior distributions were constructed based on gene associations derived from the biomedical literature using Latent Semantic Indexing (LSI). The LSI model was built using more than 1 million Medline abstracts corresponding to 22,000 human and mouse genes. A key advantage of LSI is that both explicit and implicit gene relationships can be derived from the literature. Gene neighborhoods were determined using latent Gaussian Markov random fields and logistic transformation of the latent variables. We tested the procedure on a microarray dataset for interferon-stimulated genes in mouse embryonic fibroblasts. By integrating functional information from literature, Bayesian approach identified relevant genes that previously did not meet the 0.05 significance level. In comparison to a standard mixture model, spatial mixture model has more power for identifying direct and indirect interferon regulated genes. The spatial model enhanced the ranks of some genes which are known to be affected by interferon treatment, such as Nmi (NMI N-myc and STAT interactor) and ifi35 (interferon-induced protein 35). It also identified some genes that previously were ignored because of the marginal p-values, such as dpysl2, map2k1, msn, Psck5, and Il6st. Interestingly, these genes appear to be indirectly related to interferon treatment. In summary, we show that our procedure increases statistical power and produces more biologically meaningful gene lists. These results suggest that Bayesian methods which incorporate functional information from the literature may improve analysis of gene expression data.

Classification of gene expression data is the common denominator of various biomedical recognition tasks. However, obtaining class labels for large training samples may be difficult or even impossible in many cases. Therefore, semi-supervised classification techniques are required as semi-supervised classifiers take advantage of the unlabeled data. Furthermore, gene expression data is high dimensional which gives rise to the phenomena known under the umbrella of the curse of dimensionality, one of its recently explored aspects being the presence of hubs or hubness for short. Therefore, hubness-aware classifiers were developed recently, such as Naive Hubness Bayesian k-Nearest Neighbor (NHBNN). In this paper, we propose a semi-supervised extension of NHBNN and show in experiments on publicly available gene expression data that the proposed classifier outperforms all its examined competitors.

BackgroundSerial Analysis of Gene Expression (SAGE) is a DNA sequencing-based method for large-scale gene expression profiling that provides an alternative to microarray analysis. Most analyses of SAGE data aimed at identifying co-expressed genes have been accomplished using various versions of clustering approaches that often result in a number of false positives.Principal FindingsHere we explore the use of seriation, a statistical approach for ordering sets of objects based on their similarity, for large-scale expression pattern discovery in SAGE data. For this specific task we implement a seriation heuristic we term ‘progressive construction of contigs’ that constructs local chains of related elements by sequentially rearranging margins of the correlation matrix. We apply the heuristic to the analysis of simulated and experimental SAGE data and compare our results to those obtained with a clustering algorithm developed specifically for SAGE data. We show using simulations that the performance of seriation compares favorably to that of the clustering algorithm on noisy SAGE data.ConclusionsWe explore the use of a seriation approach for visualization-based pattern discovery in SAGE data. Using both simulations and experimental data, we demonstrate that seriation is able to identify groups of co-expressed genes more accurately than a clustering algorithm developed specifically for SAGE data. Our results suggest that seriation is a useful method for the analysis of gene expression data whose applicability should be further pursued.

Background - In the NeoTRIPaPDL1 phase III trial, triple-negative breast cancer (TNBC) patients were randomized to receive nab-paclitaxel/carboplatin for 8 cycles (CT arm) with or without atezolizumab (CT/A arm). We previously reported that the majority of patients (65%) with PD-L1- at baseline converted to PD-L1+ (the majority with IC2/IC3) after the first cycle of treatment in the atezolizumab arm. Here we studied treatment associated changes in the molecular tumour features and immune microenvironment by PD-L1 expression. Methods – A total of 158 (56.4%) of patients enrolled in the NeoTRIPaPDL1 trial (76 and 82 from CT/A and CT arm, respectively) were included in this analysis by satisfying the following criteria: i) availability of a paired core biopsy at baseline and at day 1 of the second treatment cycle (D1C2); ii) evaluation of stromal TILs (sTILs) and staining for PD-L1 (Ventana SP142); iii) successful RNA-seq gene expression profiling. PD-L1 groups were defined as IC0/IC1 (<5% PD-L1low) vs IC2/IC3 (>=5% PD-L1high). Presence of different immune cell populations was quantified by gene expression profile deconvolution using ConsensusTME R package. Cancer hallmark gene set collection and custom signature activation status were estimated in each sample using singscore R package. Differences in score distribution were evaluated by 2-sided t-test. Results At baseline, both sTILs and immune cell signatures were upregulated in PD-L1high compared to PD-L1low tumours (p<0.001). No significant differences were found between the two treatment arms in the PD-L1high subpopulation. In the PD-L1low cases, both sTILs and immune related signatures were slightly downregulated in the CT/A arm (p<0.05), suggesting a modest unbalance among treatment arms with CT arm being slightly more inflamed than CT/A arm.At D1C2, PD-L1high tumours in the CT arm systematically had high sTILs (median=70%, range=30-90%), while PD-L1low tumours receiving CT/A had low sTILs in a significant proportion of cases (median=30%, range=0-90%; p<0.001). Similarly, at D1C2 several gene expression-estimated immune cell populations and immune-related signatures were upregulated in the CT arm compared to CT/A arm in PD-L1high tumours, with the weakest association observed for M2 macrophages (p=0.058). No tumour-related signatures were differentially expressed among the two treatment arms within groups with PD-L1high, suggesting that different treatment. modulate PD-L1 by engaging a different immune mileau instead of modulating tumor related features.Considering PD-L1low groups at D1C2, in the CT arm 21 tumors had >30% sTILs (n=21/69, 30.4%), while in the CT/A arm only 2 had sTILs >30% (2/27, 7.4%) (p<0.001). Analysis of gene expression data identified IFN-related signatures as the most upregulated in the CT compared to CT/A arm in PD-L1low cases at D1C2 (p<0.05). Conclusions Integrated dynamic analysis of PD-L1 expression and gene expression data highlighted significant treatment-specific changes of the immune landscape according to PD-L1 expression, when this biomarker is assessed during treatment. This indicates that the immune milieu associated with PD-L1 status is strongly dependent on the time of assessment in relationship to treatment received. Such observation may explain why in the NeoTRIP trial, baseline PD-L1 but not on-treatment PD-L1 was predictive of pCR in CT/A arm. In addition, our findings could have implications related to the use of PD-L1 as a predictive biomarker in pre-treated patients, especially when assessed early on during treatment. Citation Format: Maurizio Callari, Chiun-Sheng Huang, Daniel Egle, Begoña Bermejo, Claudio Zamagni, Matteo Dugo, Marc Thill, Antonio Anton, Marco Barreca, Stefania Russo, Eva Maria Ciruelos, Richard Greil, Stefania Zambelli, Balázs Gyorffy, Chanel Smart, Olivia Biasi, Pinuccia Valagussa, Giuseppe Viale, Luca Gianni, Giampaolo Bianchini. Immune milieu associated with PD-L1 status in TNBC is dependent on time of biomarker assessment and treatment received: A secondary analysis of the NeoTRIPaPDL1 trial [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr P1-04-02.

Rapid technological advancements in microarray analysis continue to generate enormous amounts of genomic data. However, neither hardware nor software computational capabilities have kept pace with this drastic increase. This paper presents a novel framework designed to achieve fast, robust, and accurate (biologically-significant) multi-class classification of gene expression data using distributed knowledge discovery and computational integration routines, specifically for cancer applications. The proposed paradigm consists of the following key computational steps: (a) preprocessing normalization and discretization of gene expression data, (b) partition data using two methods: overlapped windows and adaptive selection, (c) perform association rule discovery on partitioned data-spaces using FP-growth method, (d) integrate derived association rules on distributed processor nodes using a novel knowledge integration algorithm, (e) further prune rules to reduce dimensionality using parametric significance estimation, and (f) cluster remaining rules using a novel clustering algorithm for enhanced visualization and interpretation of discovered gene rule sets.

The gene ontology (GO) contains three hierarchies that represent gene function based on categories of molecular function, biological process, and cellular component. However, additional knowledge about the mechanisms underlying gene function is buried in the GO term definitions and is not computationally accessible. We describe a process for adding new links to the GO between molecular function terms and the biological process terms that those molecular functions are involved in. These new links enable inference engines to reason about relationships between genes that have disparate molecular functions but participate in the same biological process. In particular, we demonstrate how these new links enable more effective automated analysis of gene expression data.

Though DNA microarray technology simultaneously measures the expression levels of thousands of genes, only a few underlying gene features may account for significant data variation in gene classification problems. Selection of features from huge data set is difficult and so dimension reduction of gene expression data set is essential in order to determining important features, which play key role in predicting an outcome. Rough set theory (RST) has been used recently for dimension reduction of data, however, the existing methods are inadequate to finding minimal reduct. The paper proposes a RST based technique, applied on gene expression data for dimension reduction by obtaining single reduct in one pass. The gene expression data are discretized using linguistic terms with proper semantics and represented by fuzzy sets. The discretized values are calculated using Gaussian membership function with varied mean and standard deviation in order to eliminate the ambiguity between different linguistic terms. The genes are classified using linguistic decision attribute values based on the frequency of gene expression data. Discritization and classification of gene expression data are performed simultaneously, which significantly reduces time complexity of the procedure. Thus, the proposed framework selects the most significant samples for gene classification, resulting dimension reduction. The Proposed method produces output, which exhibits no variation with experimental microarray gene information unlike other existing methods.

One of the most significant research topics in bioinformatics is the classification of gene expression. Gene expression data commonly have a large number of features and a small number of samples. The gene expression data are very different from one to another, this differentiation among data and the feature's large number make the classification for gene expression data challenging. In this study, for classification we assessed the accuracy for most powerful deep learning's algorithms such as Deep Neural Network, Recurrent Neural Network, Convolutional Neural Network and improved Deep Neural Network with the preprocessing technique. The DNN was improved by adding Dropout to it by which the overfitting problem was overcame. Our results showed that the proposed improved-DNN outperforms the other algorithms among all used datasets.

In order to understand even the simplest cellular processes, we need to integrate proteomic, gene expression and other biomolecular data. To date, most computational approaches aimed at integrating proteomics and gene expression data use direct gene/protein correlation measures. However, due to post-transcriptional and translational regulations, the correspondence between the expression of a gene and its protein is complicated. We apply a multivariate statistical method, co-inertia analysis (CIA), to visualise gene and proteomic expression data stemming from the same biological samples. Principal components analysis or correspondence analysis can be used for data exploration on single datasets. CIA is then used to explore the relationships between two or more datasets. We further explore the data by projecting gene ontology (GO) information onto these plots to describe the cellular processes in action. We apply these techniques to gene expression and protein abundance data from studies of the human malarial parasite life cycle and the NCI-60 cancer cell lines. In each case, we visualise gene expression, protein abundance and GO classes in the same low dimensional projections and identify GO classes that are likely to be of biological importance.

BackgroundWith the recent advancements in high-throughput experimental procedures, biologists are gathering huge quantities of data. A main priority in bioinformatics and computational biology is to provide system level analytical tools capable of meeting an ever-growing production of high-throughput biological data while taking into account its biological context. In gene expression data analysis, genes have widely been considered as independent components. However, a systemic view shows that they act synergistically in living cells, forming functional complexes and more generally a biological system.ResultsIn this paper, we propose LatNet, a signal transformation framework that, starting from an initial large-scale gene expression data, allows to generate new representations based on latent network-based relationships between the genes. LatNet aims to leverage system level relations between the genes as an underlying hidden structure to derive the new transformed latent signals. We present a concrete implementation of our framework, based on a gene regulatory network structure and two signal transformation approaches, to quantify latent network-based activity of regulators, as well as gene perturbation signals. The new gene/regulator signals are at the level of each sample of the input data and, thus, could directly be used instead of the initial expression signals for major bioinformatics analysis, including diagnosis and personalized medicine.ConclusionMultiple patterns could be hidden or weakly observed in expression data. LatNet helps in uncovering latent signals that could emphasize hidden patterns based on the relations between the genes and, thus, enhancing the performance of gene expression-based analysis algorithms. We use LatNet for the analysis of real-world gene expression data of bladder cancer and we show the efficiency of our transformation framework as compared to using the initial expression data.