Posters - Schedules

Posters Home

Motivation: Complex phenotypes, such as many common diseases and morphological traits, are controlled by multiple genetic factors, namely genetic mutations and genes, and are influenced by environmental conditions. Deciphering the genetics underlying such traits requires a systemic approach, where many different genetic factors and their interactions are considered simultaneously. Many association mapping techniques available nowadays follow this reasoning, but have some severe limitations. In particular, they require binary encodings for the genetic markers, forcing the user to decide beforehand whether to use, for example, a recessive or a dominant encoding. Moreover, most methods cannot include any biological prior or are limited to testing only lower-order interaction among genes for association with the phenotype, potentially missing a large number of marker combinations.
Results: We propose HOGImine, a novel algorithm that expands the class of discoverable genetic meta-markers by considering higher-order interactions of genes and by allowing multiple encodings for the genetic variants. Our experimental evaluation shows that the algorithm has a substantially higher statistical power compared to previous methods, allowing it to discover genetic mutations statistically associated with the phenotype at hand that could not be found before. Our method can exploit prior biological knowledge on gene interactions, such as protein-protein interaction networks, genetic pathways and protein complexes, to restrict its search space. Since computing higher-order gene interactions poses a high computational burden, we also develop novel algorithmic techniques to make our approach applicable in practice, leading to substantial runtime improvements compared to state-of-the-art methods.

The Molecular mechanisms underlying Alzheimer's disease (AD) are highly complex and still poorly understood. Furthermore, AD is a human-specific disorder raising the possibility that some genomic changes related to the evolution of human brain also contribute to pathology of AD. Here to identify key transcriptional changes in AD we performed gene co-expression network analysis and subsequently investigated the evolutionary history of genes.
Four genome-wide transcriptomics datasets from different cortex regions were used to construct gene co-expression network using a weighted topological overlap (wTO) method. To enrich the set of genes potentially involved in AD pathology we extracted the graph neighbourhood for thirteen well-known AD-associated genes. A wTO networks comparison revealed 2538 genes co-expressed specifically in the AD network, 66 in the control network, and 159 in both networks. PAML Branch-site model test showed 25 AD specifically co-expressed genes with positively selected codons in human lineage compared with 23 non-human primate species. Interestingly, two AD specifically co-expressed genes are evolutionary young; ZNF439 a positively selected gene that arose within the primate clade, and SRGAP2C is human-specific. Both are the first neighbours of ADAM10 in the AD wTO network. ADAM10 codes for the main α-secretase that cleaves amyloid precursor protein.

Non-small-cell lung carcinoma (NSCLC) is a common type of lung cancer that affects both men and women. Various factors can influence its development and progression. However, the differences at the molecular level between males and females in NSCLC are not well understood. In this study, we aimed to identify genes that are differentially expressed (DE) between genders in NSCLC subtypes: lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC). We performed a network analysis to identify hub genes that have a significant impact on dysregulated PPI networks. We then used a support vector machine (SVM) model to validate these hub genes. Our results showed that some of the hub genes identified by the SVM model were associated with oncogenic Gene Ontology (GO) terms. We also constructed a competing endogenous RNA (ceRNA) network to explore the regulatory roles of key genes and their target RNA-interferences (miRNA and lncRNA) in NSCLC tumorigenesis. We found novel male-specific and female-specific genes regulated by miRNAs that had an oncogenic role in the ceRNA network and were associated with poor survival in LUSC and LUAD patients. Our study provides new insights into the gender-specific molecular mechanisms of NSCLC tumorigenesis and suggests potential targets for personalized therapy.

Bacterial pneumonia is a leading cause of morbidity worldwide. While antibiotics represent the standard of care against pneumonia, their effectiveness is declining because of antimicrobial resistance. This project uses an original immunotherapy targeting the airways’ innate immune defenses in combination with antibiotics to improve treatment against pneumonia caused by Streptococcus pneumoniae. Using murine models of pneumococcal infection, we previously showed that the agonist of innate immunity flagellin, in conjunction with amoxicillin, is a promising way of boosting effectiveness of antibiotics and overcoming antimicrobial resistance. To further understand the underlying mechanisms of this combination therapy, we performed high-throughput RNA sequencing of blood and lung tissues from infected mice treated overtime with flagellin alone, amoxicillin alone or the combination of both, and applied Weighted Gene Co-expression Network Analysis (WGCNA).
Applying this pipeline with time and treatment-resolved network analysis, we expect to identify blood and respiratory markers of innate immunity specific for the combined administration of flagellin and amoxicillin. The temporal resolution provides additional insights into pharmacodynamics in local and systemic compartments. We next plan to analyze cross-species conservation of the identified markers in pig and monkey models, as well as human airway primary epithelium models.

Differentially analysing individual immune cell types would capture the hallmark characteristics of a cell. Here, we retrieved the RNAseq of 25 stimulated and unstimulated immune cells processed by Calderon et al. (2019). A complete protein interaction network (PPIN) does not contain characteristic information specific to the cell and its condition. Hence, the in-house computational tool PPIXpress was used to construct condition-specific PPIN for all 25 immune cell types by pruning the complete PPIN to those genes/transcripts that cover at least one read. We used another in-house tool PPICompare to compare the condition-specific PPINs. Results showed the rewired interactions between conditions and the causes of rewiring either differential expression or alternative splicing.
Furthermore, we developed a tool that can be used downstream of PPICompare. It executes automatic biological interpretation of PPICompare results based on Gene-Ontology and KEGG pathway enrichment analysis. It reported the enriched annotations of rewired proteins. The tool has a feature to analyze how the rewirings affect certain groups of proteins like transcription factors, chromatin readers and splicing factors and their interactors. When comparing the unstimulated and stimulated versions of an immune cell, the top enriched GO terms showed the characteristics involved in the stimulation of the immune cell.

A long-term objective of network medicine is to replace our current, mainly phenotype-based disease definitions by subtypes of health conditions corresponding to distinct pathomechanisms. For this, molecular and health data are modeled as networks and are mined for pathomechanisms. However, many such studies rely on large-scale disease association data where diseases are annotated using the very phenotype-based disease definitions the network medicine field aims to overcome. This raises the question to which extent the biases mechanistically inadequate disease annotations introduce in disease association data distort the results of studies which use such data for pathomechanism mining. We address this question using global- and local-scale analyses of networks constructed from disease association data of various types. Our results indicate that large-scale disease association data should be used with care for pathomechanism mining and that analyses of such data should be accompanied by close-up analyses of molecular data for well-characterized patient cohorts.

Diseases and phenotypic manifestations result from the combination of genetics and environmental factors. For instance, an increasing number of studies attribute a large variety of different diseases to water, soil, and air pollution. Here, we used a network-based approach to construct a comprehensive map in which 9,887 exposures are linked through their shared impact on the human genome. In this map, we can identify groups of exposures that have similar biological effects, even if they are chemically different. By using the human interactome of protein-protein interactions, we found that exposures affect well-defined neighborhoods and high interactome connectivity is the prime indicator of the harmfulness of an exposure. A systematic comparison between the interactome modules affected by exposures and disease-associated modules suggested that interactome overlap can be used to predict exposure-disease relationships. To evaluate the validity of these predictions, we cross-referenced country-wide disease prevalence data with reports of environmental exposures. We found that elevated levels of a particular exposure in air or water correlated with an increased prevalence of diseases whose interactome modules overlapped with those of the respective exposures. Taken together, we provide a framework for relating the genetic component of chemical exposures with their epidemiological observation.

Machine learning and network analysis are computational approaches that have proven to be effective in supporting different stages of drug discovery programs. When combined with high-throughput data, they can foster data-driven drug target identification pipelines. Here, we present a novel machine learning and network-based workflow to identify drug targets for cystinosis, which is the most common cause of inherited tubular dysfunction and kidney disease in children and currently lacks effective therapies. To identify genes associated with kidney tubular function, we first built an ensemble neural network. The model takes five different matrices of gene-gene similarities and known phenotype-gene associations as input and returns the probability of each gene being phenotype-related. It predicted 1331 genes as phenotype genes. Next, we built a biological network comprising transcription factor-gene, protein-protein, and protein-metabolite interactions. We mapped genes deregulated in cystinosis animal models (disease genes) and the phenotype genes onto the network to identify modules enriched in disease genes. Finally, we designated key genes connecting phenotype and disease genes as candidate drug targets. This approach allows us to recapitulate the disease mechanisms in the context of renal tubular physiology and identify candidate drug targets for further validation using a cross-species workflow and disease-relevant screening technologies.

With the astonishingly rapid advances in single-cell sequencing, there is an undeniable demand for computational methods to employ such rich data, particularly in deriving gene regulation networks (GRNs). The single-cell data allows the researchers to unravel the underlying complexities in the regulation mechanism, which can lead to understanding the origin of biological disorders.
Here, we address the caveats of previous methods by incorporating more diverse single-cell modalities. Through employing a multi-task learning algorithm, we establish a versatile statistical framework, called GAZE, that guarantees a comprehensive analysis of single-cell data in an integrative fashion. We first build a catalog of enhancer-gene interactions using sc-ATAC data. The enhancer regions are then exploited to assess TF binding affinities associated to a gene. Next, we train predictive regression models associating TFs with genes in a GRN.

By interpreting these models (regression coefficients or using explainable AI approaches), we are able to reveal novel regulatory aspects. Additionally, we designed adept tests for investigating the inferred regulatory activities to identify key genes or TFs driving cell type heterogeneity and differentiation.

We believe that GAZE allows us to broaden the current understanding of transcriptional regulatory mechanisms through identifying the key players involved in differential regulation in GRNs.

Current treatments for Inflammatory Bowel Disease (IBD) are not effective in all patients: primary non-response and loss of response are reported in a high proportion of patients. Therefore there is an urgent need to identify novel targets and to provide a comprehensive disease characterization.
Here we propose to use Bayesian approaches to build a causal gene network using omics data from Ulcerative Colitis (UC) patients. Multiple UC-relevant data sources (i.e. transcriptomics, protein-protein interactions, disease genetic association and eQTLs) were integrated to derive causal relationships among genes. The resulting network, comprised of 2577 nodes and 4640 edges was used to (i) predict key driver genes whose expression affects disease-associated pathways, (ii) explore gene neighborhoods of relevant but undruggable targets and (iii) evaluate the overlap of sub-networks around specific genes that could be targeted through bispecific approaches.
Among the key driver genes found in the causal gene network were genes previously associated with IBD, as well as new genes which can be further explored as potential targets for drug discovery. We also demonstrate that the network can be used to evaluate the overlap of genes-of-interest sub-networks and functional or cellular disease modules, thus guiding selection of gene partners for bispecific antibodies.

A better understanding of protein-protein interaction (PPI) networks representing physical interactions between proteins could be beneficial for evolutionary insights as well as for practical applications such as drug development. As a statistical model for PPI networks, duplication-divergence models have been proposed, but they suffer from resulting in either very sparse networks in which most of the proteins are isolated, or in networks which are much denser than what is usually observed. Moreover in real networks, gene loss may occur, which is not captured in duplication-divergence models to date. Here we introduce a new duplication-divergence model which includes gene loss. This mechanism results in networks in which the proportion of isolated proteins can take on values which are strictly between 0 and 1. To understand this new model we apply strong and weak attacks to networks from duplication-divergence models with and without gene loss, and compare the results to those obtained when carrying out similar attacks on real PPI networks. We find that the new model more closely reflects the damage caused by strong and weak attacks found in the empirical PPI network.

Motivation of study: Biological networks are used in cardiovascular diseases to identify causal genes, biomarkers and drug targets. However, networks on atherosclerosis-specific tissues and phenotypes have not been fully elucidated. Existing network reconstruction techniques typically use the same association threshold for all molecules when inferring positive interactions, many are not able to incorporate the study of negative associations, and others lack edge directionality. To overcome such limitations, we designed a new method for reconstructing networks from proteomics datasets.

Methods & Results: We introduce DiRec-AP, which combines two mutual information(MI)-based algorithms (ARACNe-AP, SIREN), integrating a data-driven process for the identification of a different MI threshold per molecule. DiRec-AP was benchmarked against three representative network reconstruction methods, using golden standard datasets, and significantly outperformed them. We applied DiRec-AP to reconstruct atherosclerotic plaque networks and phenotype-, sex- and vasculature-specific networks. Among significant findings, the regulation of Insulin-like Growth Factor transport pathway was found to be upregulated in symptomatic plaques while the fibrillar collagens formation was upregulated in asymptomatic.

Conclusion/Significance: We developed a network reconstruction pipeline which overcomes the limitations of existing methods and substantially improves the quality of the reconstructed networks from quantitative omics data, and allowed resolving of atherosclerotic networks.

The search for effective therapeutic targets in fields like regenerative medicine and cancer research has generated interest in cell fate reprogramming. We present NETISCE, a novel computational tool for identifying cell fate reprogramming targets in static networks. In combination with control theory and machine learning algorithms, NETISCE estimates the attractor landscape and predicts reprogramming targets using Signal Flow Analysis and Feedback Vertex Set Control, respectively. Through validations in studies of cell fate reprogramming from developmental, stem cell, and cancer biology, we show that NETISCE can predict previously identified cell fate reprogramming targets and identify potentially novel combinations of targets. NETISCE extends cell fate reprogramming studies to larger-scale biological networks without the need for full model parameterization and can be implemented by experimental and computational biologists to identify parts of a biological system relevant to the desired reprogramming task.

Deciphering the mechanisms of gene expression and regulation is a fundamental challenge of molecular biology. A particular area of regulation that remains largely opaque in its functions is the spatial organization of chromatin. While technologies such as chromatin conformation capture provide insight into this organization, there is undoubtedly much that is yet unclear about the mechanisms at play. Chromatin interaction graphs, a graph-based representation of chromatin contact maps generated through experiments such as Hi-C, offer a potential avenue of interpretation and integration of data on this topic. In our work we specifically study the usefulness of cliques of size 3 in locating and characterizing spatially dense regions of chromatin, finding that in three out of four of our studied datasets the cliques appeared to have some measure in overlap in their distribution. We then assessed overrepresentation and the relative abundance of ChromHMM annotations of gene ontologies within cliques compared to edges, and found significant differences in both, most notably in cliques being significantly enriched in chromatin annotations for active transcription and enhancers. Finally, we develop a measure of edge centrality based on involvement in cliques, seeking to identify key areas of dense, hairball-like aggregations of chromatin.

Deciphering the mechanisms of gene expression and regulation is a fundamental challenge of molecular biology. A particular area of regulation that remains largely opaque in its functions is the spatial organization of chromatin. While technologies such as chromatin conformation capture provide insight into this organization, there is undoubtedly much that is yet unclear about the mechanisms at play. Chromatin interaction graphs, a graph-based representation of chromatin contact maps generated through experiments such as Hi-C, offer a potential avenue of interpretation and integration of data on this topic. In our work we specifically study the usefulness of cliques of size 3 in locating and characterizing spatially dense regions of chromatin, finding that in three out of four of our studied datasets the cliques appeared to have some measure in overlap in their distribution. We then assessed overrepresentation and the relative abundance of ChromHMM annotations of gene ontologies within cliques compared to edges, and found significant differences in both, most notably in cliques being significantly enriched in chromatin annotations for active transcription and enhancers. Finally, we develop a measure of edge centrality based on involvement in cliques, seeking to identify key areas of dense, hairball-like aggregations of chromatin.

Cancer hallmark genes (CHG) related to inflammation and immunity are promising as prognostic biomarkers and therapeutical targets. However, no CHG-based study has been conducted on the prognosis and diagnosis of breast cancer (BC). Here, we clustered inflammation and immunity CHGs based on their expression levels in BC tissues and examined the prognostic accuracy of each module, which is the gene cluster with high correlation.
Totally, 366 microarray data collected from patients with BC, and of those, 382 inflammation and immunity CHGs from within top 25% variance were analyzed using weighted gene correlation network analysis (WGCNA). Then, survival analysis was conducted using the Module Eigengene (ME) which is the first principal component of the expression levels in each module.
WGCNA identified five modules and the ME of one module comprised 45 genes showed high prognostic accuracy (RFS: P=8.76×10-4, HR=0.43; OS: P=1.99×10-3, HR=0.53). Co-expression of inflammation and immunity CHGs contributed to cancer suppression in the tumor microenvironment (TME). Multiple indicators, such as Module Membership and STRING, identified seven key genes. These genes had functions related to lymphocyte infiltration, indicating that maintenance of immune homeostasis in the TME would contribute to a favorable prognosis.

Several experimentally identified biomarker panels are practically used for prognosis prediction of breast cancer. Characterization of those biomarkers on the protein-protein-interaction (PPI) networks may lead to the efficient identification of new biomarker. In this study, we applied centrality analysis to the PPI network of breast cancer and investigated centrality indices that can effectively extract characteristics of existing biomarkers (PAM50, MammaPrint, and Oncotype-DX). Specifically, we created the PPI network consisting of the 2,000 breast cancer related genes by Cytoscape and performed centrality-analyses using Cytohubba. For the genes with the top centrality score, we investigated the number of existing biomarkers, survival time by Kaplan-Meier-Plotter, and expression levels in each subtype of breast cancer by Bc-geneExMiner. In the case of degree-centrality, only 2 of the existing 63 biomarkers were included in the top30 score. In contrast, in the case of Maximal-Clique- Centrality (MCC), 13 biomarker genes were included. That is, MCC is suggested to be an effective indicator for the selection of biomarker candidates for breast cancer. Among the 17 genes other than existing biomarkers, 16 genes showed significant differences both in survival analysis and expression levels on each subtype. These results suggest that these 16 genes are new biomarker candidates.

An abundance of proteogenomic data enables cancer researchers to identify an increasing number of genes related to cancer progression and treatment outcomes. However, widely used gene enrichment analysis databases, such as Gene Ontology (GO) and HALLMARK, do not provide gene interaction information and assign equal importance to each gene in the gene sets. Network propagation algorithms are essential for analyzing biological networks and offer valuable insights into gene sets. We developed an R package NetProp that combines five prominent network propagation algorithms – Steiner Algorithm (SA), Diffusion Score (DS), Random Walk with Restart (RWR), Shortest Path (SP), and Clustering (CL) – with protein-protein interactions (PPI). This package aims to expand gene sets and pinpoint additional genes of potential biological and clinical relevance. To showcase its usefulness, we conducted data analyses using genes with the highest mutation frequency in Lung Adenocarcinoma (LUAD) and Prostate Adenocarcinoma (PRAD) from TCGA. Our findings revealed an increase in the number of significant pathways and differentially expressed genes, underscoring our package's ability to reveal new associations in biological networks and select features for subsequent machine learning and classification studies.

Sarcopenia, a progressive muscle wasting disorder resulting from aging, presents significant challenges in identifying and validating potential therapeutic targets. To address this issue, we created a novel network-based model of microRNA:target interactions for the efficient selection of potential targets for sarcopenia treatment. Using RNAseq and microarray data from five studies on healthy skeletal muscle participants, aged 19 to 85 years old, we conducted a comprehensive analysis of microRNA involvement in aging. We identified gene interactions and predicted microRNAs that are likely to regulate shortlisted genes using target prediction tools. We then calculated the relevance of each node using multiple metrics and centrality measures to identify the key regulatory microRNAs and their targets during aging. Our model is tailored to the context of muscle aging and integrates multiple layers of biological information, providing a more comprehensive and accurate representation of the complex regulatory mechanisms involved. By identifying a few tens of microRNAs and genes with potential therapeutic power, our model offers an efficient approach to target identification and validation for the treatment of sarcopenia.

X-linked centronuclear myopathy (XLCNM) is a severe muscle disorder affecting children. It is characterized by general muscle weakness, muscle fiber atrophy, and abnormal nuclear positioning in myofibers. The transcriptome, proteome, and metabolome have been separately investigated in the Mtm1-/y murine model of XLCNM, but the pathomechanism of the disease remains poorly understood and no curative treatment is available.
Here, we used network-based approaches to perform multi-omics integrated analysis of XLCNM. The nodes of the network include proteins, metabolites, and pathways. Relationships between the nodes were retrieved using custom SPARQL querying on molecular interaction and pathway databases. Differential gene, protein and metabolite expression analyses were performed to identify entities associated with the disease. In total, 1981 genes, 609 proteins and 417 metabolites are differentially expressed in Mtm1-/y mice compared to WT. Quantitative dysregulation data (log2 fold-changes) were integrated in the biological network as node scores. Active module identification uncovered several heterogeneous subsets of nodes closely linked to specific phenotypes of the disease.
Overall, our study shows that the use of network-based integration improves the knowledge we gain from multi-omics datasets. By including additional layers of information such as phenotypic and imaging data, this approach may uncover novel pathomechanisms underlying centronuclear myopathies.

Metabolic networks are complex systems of chemical reactions proceeding via physical interactions between metabolites and proteins. We aimed to predict previously unknown compound–protein interactions (CPI) in metabolic networks by applying biclique extension, a network-structure-based prediction method.
We developed a workflow, named BiPredict, to predict CPIs based on biclique extension and applied it to Escherichia coli and human using their respective known CPI networks as input. Depending on the chosen biclique size and using a STITCH-derived E.coli CPI network as input, a sensitivity of 39% and an associated precision of 59% was reached. For the larger human STITCH network, a sensitivity of 78% with a false-positive rate of <5% and precision of 75% was obtained. High performance was also achieved when using KEGG metabolic-reaction networks as input. Prediction performance significantly exceeded that of randomized controls and compared favorably to state-of-the-art deep-learning methods. Regarding metabolic process involvement, TCA-cycle and ribosomal processes were found enriched among predicted interactions. BiPredict can be used for network curation, may help increase the efficiency of experimental testing of CPIs, and can readily be applied to other species. BiPredict and related datasets are available at https://github.com/SandraThieme/BiPredict.

Enrichment analysis (EA) is the standard approach to functionally characterize omics results. As a consequence, a large number of EA methods have been developed, yet it is unclear which method is the best for a given dataset. This is due to the lack of unbiased gold standards and appropriate evaluation metrics, for which the rank of a single target pathway is commonly used. We here present a generalized EA benchmark of all four categories of currently existing approaches. We compiled a new set of 82 curated gene expression datasets available in Gemma (https://gemma.msl.ubc.ca/), from RNA-Seq and DNA microarray for 26 diseases, of which 13 are cancers. In order to address the shortcomings of the single target pathway approach and to enhance the sensitivity evaluation, we introduced a “disease pathway network” based on functional associations between KEGG pathways. By making the assumption that no pathway should be enriched in a randomized gene expression dataset, we measured the specificity of the methods in an independent manner. By focusing on the results under the null hypothesis, we could also investigate whether any method is biased towards certain pathways. Network-based approaches like ANUBIX, BinoX, and NEAT showed overall best performances.

Multi-omic Graph Diagnosis (MOGDx) is a tool for the integration of omic data and classification of heterogeneous diseases. MOGDx exploits a patient similarity network framework to integrate omic data using Similarity Network Fusion (SNF). One autoencoder per omic modality is trained, and the latent embeddings from each autoencoder are concatenated. These reduced vectors are used as node features in the integrated network. Classification is performed on the fused network using the Graph Convolutional Network (GCN) deep learning algorithm. GCN is a novel paradigm for learning from both network structure and node features. Heterogeneity in diseases confounds clinical trials, treatments, genetic association and more. Accurate stratification of these diseases is therefore critical to optimize treatment strategies for patients with heterogeneous diseases. Previous research has shown that accurate classification of heterogenous diseases has been achieved by integrating and classifying multi-omic data. MOGDx improves upon this research. MOGDx's advantages are that it can handle both a variable number of modalities and missing patient data in one or more modalities. Performance of MOGDx was benchmarked on the BRCA TCGA dataset, with competitive performance compared to its counterparts. In summary, MOGDx combines patient similarity network integration with graph neural network learning for accurate disease classification.

Representation learning has revolutionized sequence-based protein function prediction. Protein interaction networks are another crucial source of information on protein function; however, applying representation learning to networks in a manner that works across species is much harder.

The STRING database provides protein networks as well as orthology relations for over 1,500 eukaryotes. The aim of the SPACE project is to network embeddings for all proteins in these species, which are directly comparable between species and are complementary to existing sequence embeddings.

We propose a method in which we first use node2vec to generate embeddings for each species, and subsequently use the FedCoder model to align them across species based on orthologous proteins.

To test SPACE, we generated embeddings for three very distantly related eukaryotes, namely human, yeast, and Arabidopsis. We assessed both how well they are aligned and how well they capture network topology by using a simple logistic model to perform link prediction from embeddings.

The results showed that the three species-specific networks were successfully aligned in the same space with minimal loss of topological information before alignment. This means that the embeddings can be used to train a single model to solve network-based cross-species problems such as function prediction.

Heterogeneous disorders such as cancers and neurological diseases are often characterized by dysregulations in several biomolecules. Previous studies suggest that these alterations are synchronized and locally grouped within molecular networks. To identify these coordinated network activity changes in omics data, we propose a feature selection approach that uses input expression data and a graph encoding functional associations between biomolecules, such as protein-protein interactions. We frame this as a two-class classification problem, using a new Pairwise Elastic Net penalty that favors selecting discriminative biomolecules based on their connectedness in the network graph. Empirical results on Parkinson’s disease gene expression data show that the proposed approach achieves similar classification performance to other state-of-the-art methods, while providing significant improvements in detecting locally grouped network activity changes.

Lack of trust in artificial intelligence (AI) models is still the key blockage for the use of AI in clinical decision support systems (CDSS). Although AI models are performing excellently in medicine, their black-box nature entails that it is often impossible to understand why a particular decision was made. In the field of explainable AI (XAI), many algorithms have been developed to ""explain"" which input features influenced a specific prediction. However, for clinical use it is essential that these explanations lead to some degree of causal understanding.

Therefore, we developed the CLARUS platform, aiming to promote human understanding of graph neural network (GNN) predictions. CLARUS enables the visualization and interaction with biological networks used to train and test a GNN. Relevance values of genes and interactions are computed by XAI models and highlighted in the visualized graph. More importantly, the expert can interactively manipulate the patient PPI based on their understanding and initiate a retraining or re-prediction.

We will present the first interactive XAI platform prototype, CLARUS, that enables domain experts to gain deeper insights into what parts of the biological network were most influential in a GNN decision-making process.
CLARUS is hosted by the GWDG: https://rshiny.gwdg.de/apps/clarus/.

Proteins interact with other proteins and biomolecules to carry out their cellular functions. Interactions among proteins could conveniently be represented in the form of protein-protein interaction networks (PPINs) where proteins involved in the interactions are represented by nodes and their interaction(s) with other proteins as edges. Degree centrality is the number of edges on a node. Betweenness centrality is the proportion of all the shortest paths between all the pairs of nodes in the network that pass through a given node. Nodes with high degree values are called "hubs", and those with high betweenness values are called "bottlenecks". In interaction networks, there is a significant overlap between hubs and bottlenecks. A look at the literature reveals that the studies that focus on deciphering the properties of hubs have ignored that some may be bottlenecks. This lacuna motivated us to undertake the following study. We constructed a human PPIN and found three distinct groups, namely a) Hub-bottlenecks, b) Pure bottlenecks and c) Pure hubs. Our study, clearly establishes trichotomy among the proteins with high degree and high betweenness values associated with distinct molecular-level properties. Such information is helpful for target prioritisation while designing new drugs or repositioning known or withdrawn drugs.

Therapy-induced senescence leads to sustained growth arrest (i.e., cytostasis) of cancer cells. This cytostasis is reversible, and cells that evade senescence can enhance the aggressiveness of cancer. Senolytics, chemicals selectively targeting senescent cells, offer a promising path for improved cancer treatment. Understanding how cancer cells evade senescence is needed to develop efficient senolytics and optimize their clinical benefits.

To investigate the mechanisms by which cancer cells evade senescence, we characterized the response of three different NRAS mutant melanoma cell lines to a combination of CDK4/6 and MEK inhibitors over 33 days using bulk RNA-seq, small RNA-seq, qCLASH and kinome profiling. We also performed data integration of bulk and single-cell RNA-seq data of the cell lines by adapting a matrix factorization methodology, iCell.

Our study provides new insights into the intricate relationship between senescence, interferon and insulin signaling, and the EMT process. Our findings suggest that insulin signaling contributes to senescence persistence, while interferon-gamma activates ERK5 signaling to promote senescence escape. Using iCell-based integration, we also predicted 90 novel genes associated with senescence escape. Future studies could validate our findings in vitro or in vivo. Our methodology could be expanded to include additional multi-omics time-series data or investigate other disease mechanisms.

Viruses are a leading cause of infections worldwide, often resulting in severe and sometimes deadly diseases. The infectivity of viruses is attributed to their individual proteins and how they interact with one another and with their host. To mitigate the risks posed by viruses through targeted therapeutics, understanding the network of protein-protein interactions (PPIs) they form is crucial. In particular, targeting intra-viral PPIs presents an attractive treatment strategy, as these interactions are unique to viruses and not present in the host. An equally effective strategy would be to target host-viral complexes that do not have structural equivalents within the host system.
To supplement experimental approaches, we have developed a pipeline for predicting the PPI network of any virus. This pipeline not only allows for intra-viral, but also host-viral PPI detection. Our method integrates data from molecular interaction repositories and uses interlog mapping to predict these PPIs. Additionally, we have modeled the predicted interactions of representative viruses using AlphaFold2 to evaluate the physical interaction probability. Overall, our approach offers a means to explore the PPIs formed by a given virus of interest and to uncover new interactions.

Anticipating the toxicity of compounds in humans is crucial for drug discovery and chemical risk assessment. However, current in silico methods often focus on specific toxicity endpoints, providing limited insights into the mechanisms of toxicity, and ignoring human genomic variability. To overcome these challenges, we developed the Chemical Effect Predictor (CEP), which employs AI and systems biology to predict the toxicity of compounds and generate mechanistic hypotheses for their mode of action. CEP leverages a multiscale network that models the effect of compounds through a biological network, spanning different levels of organisation. Using this network as a scaffold, CEP computes diffusion profiles for compounds and disease phenotypes, which are compared to identify the most likely adverse effects elicited by compounds and pinpoint potential mechanisms of toxicity. CEP features a machine learning classifier trained on features derived from the network as well as chemical and genetic variability descriptors. CEP achieved a ROC of 0.75 in predicting adverse effects elicited by a compound, covering adverse effects related to all organ systems. With CEP we aim to support the early assessment of the toxicity of compounds using innovative in silico tools and contribute to reducing the number of animals used in research.

Estimating microbial association networks from high-throughput sequencing data is a common exploratory data analysis approach aimed at understanding the complex interplay of microbial communities in their natural habitat. Statistical network estimation workflows comprise several analysis steps, including methods for zero handling, data normalization, and computing microbial associations. Since microbial interactions are likely to change between conditions, e.g., between healthy individuals and patients, identifying network differences between groups is often an integral secondary analysis step.

With my poster, I will present the R package NetCoMi (Network Construction and comparison for Microbiome data), which integrates our proposed analysis workflow of constructing, analyzing, and comparing microbial association networks from high-throughput sequencing data. It also provides the ability to quantify network differences and thus enables insights into whether single taxa, groups of taxa, or the overall network structure change between groups. Using data from the American Gut Project, I will give an overview of NetCoMi’s extensive functionality as well as recent enhancements of the package.

CRISPR/Cas systems have advanced genome editing, but designing specific and efficient guide RNAs remains challenging. Our proposed approach, CRISPRnet, enhances two current challenges in CRISPR screens: target accuracy and genetic redundancy.

CRISPRnet incorporates a deep learning model that predicts on-target effects while also considering genome-wide off-targets. To improve model accuracy, we perform efficiency experiments using data from methods like Surro-seq to complement the predictions. The model is trained on gRNA primary sequences and their corresponding binding energy at different sites, both on-target and off-targets, and uses a classifier to discriminate between them, which operates on RNA embeddings. The generation of these embeddings for gRNA has the potential to be extended to RNA types beyond gRNAs, such as siRNA and miRNA.

To address genetic redundancy, we are developing a tool based on graph neural networks. We do this by training a graph neural network model to predict genetic interactions based on other types of networks. These include sequence similarity, physical protein interactions, and functional associations from the STRING database.

CRISPRnet has the potential to significantly improve CRISPR-based gene editing and functional genomics research by enhancing guide RNA design and identifying functionally related genes.

Achilles' heel relationships arise when the status of one gene exposes a cell's vulnerability to perturbation of a second gene, such as chemical inhibition, providing opportunities for precision oncology. The MultiSEp algorithm and R package were developed to interrogate mutually exclusive loss signatures in multiomics data. MultiSEp performs unsupervised assignment of cell lines into gene expression clusters, providing for partitioning of CRISPR scores and mutational status in order to propose candidate Gene Dependency Relationships (GDRs). Almost all Multiple Myeloma (MM) patients relapse and succumb to therapy-resistant disease; accordingly, more effective treatments are urgently needed. We predicted MM GDRs by MultiSEp analysis of gene expression (RNA-seq) from the CoMMpass study (n=859 patients) to generate an all vs all gene dependency network of 372,303,828 edges (n=27,288 genes). Statistical filtering and GDR pattern analysis produced a high confidence predicted synthetic lethality network. Analysis of the predicted MM GDR network revealed individual ‘nexus’ genes where the neighbourhood of genes were collectively mutated in a relatively high proportion of MM patients, representing attractive drug targets. Further filtering adds confidence to predicted therapeutic targets including assessment of the impact of mutations on protein function and predicted population coverage. Results identified context-specific biochemical interactions, illuminating fundamental biology.

Advancements in technology have led to “all-by-all” proteome-scale protein-protein interaction (PPI) experiments, which typically involve large investments of time, money, and resources. Such experiments are usually performed on a select few well-known model organisms, leading to a significant disparity in PPI information. In silico methods represent a potential solution to narrow this information gap by requiring fewer resources and time. However, machine learning methods have shown an inability to generalize predictions of PPIs leading to the development of the RAPPPID model, a deep regularized PPI prediction model.

We show here for the first time that RAPPPID makes accurate interaction predictions between proteins, independent of their species of origin or their presence in the training dataset. Additionally, RAPPPID maintains comparable performance when tested on various species independent of their evolutionary distance. The model is calibrated against strict datasets which are carefully controlled for data leakage. The RAPPPID online interface offers an accessible service for interactive predictions. All together, we show that RAPPPID is able to make predictions that are more accurate than existing methods, effective across species, and computationally efficient.

Coexpression networks are a powerful tool for the characterization of gene function and are increasingly available across a range of plant species. However, cross-species comparisons have been hindered by the scarcity of 1-1 gene orthologs in plants, as repeated genome duplications have expanded gene families. We utilize coexpression networks generated from publicly available bulk RNA-sequencing data to directly compare the conservation of function between two genes across different species without relying on clear tissue homology. Leveraging these networks, we improve cross-species analysis on two fronts. First, we demonstrate the identification of gene pairs with similar expression profiles using coexpression, improving the performance of traditional single-cell RNAseq integration methods and reducing barriers to integration between 78 pairs of plant species. We highlight the utility of this approach by integrating a dataset that has been split and disassociated to have no shared genes, such that it would be uncorrectable with traditional integration methods. Second, we incorporate Gene Ontology terms into our coexpression analysis to rapidly identify conserved and rewired gene modules across 11 plant species without computationally intensive unsupervised clustering. This work lays the foundation for easier cross-species comparisons in plants, which have historically been impeded by the complexity of plant genomes.

There are several approaches for analyzing Hi-C data to identify biologically significant features. One of these approaches is the study of Hi-C interaction networks using graph theory, wherein Hi-C data is rendered as a chromatin interaction graph that can be studied topologically.
Validating the results of topological studies, however, is a non-trivial task. While observable properties of the chromatin interaction graph may reflect functional biological properties, they may also reflect technical artifacts arising from either the Hi-C methodology itself or experimental specifics. For topological study result validation, graph randomization techniques that preserve general graph topology are needed.
Typical graph randomization techniques can not be trivially applied to chromatin interaction graphs, because each node in a chromatin interaction graph corresponds to a locus. To our knowledge, none of the existing solutions associate graph nodes with a position on the chromosome, therefore interaction length distribution changes after randomization.
We propose an algorithm to randomize a chromatin interaction graph while preserving the graph’s most basic topological features - node degrees and interaction length distribution. We achieve this by modifying an existing graph to obtain a random graph that can be adequately matched to a given Hi-C dataset to control for biologically significant topologies.

Complex biological systems are often represented as biological networks containing several types of molecular entities. Analyzing and visualizing such networks can advance our knowledge of the underlying cellular mechanisms. This can be accomplished with online databases and software tools like STRING, Cytoscape, and its many apps. Specifically, Cytoscape stringApp has focused on providing intra-species protein–protein interactions from STRING to facilitate the interpretation of data from omics studies, such as proteomics and transcriptomics.
Here, we highlight the latest stringApp and STRING functionality that opens new avenues for exploring results from high-throughput experiments. Version 2.0 of stringApp greatly improves the support for heterogeneous networks, thus making it possible to create networks that contain both proteins and interactions from STRING as well as other molecular entities or associations from external sources. We exemplify this on a published SARS-CoV-2 interactome in Cytoscape using stringApp. Thereby, we also showcase the new group-wise enrichment analysis, which can be performed automatically on several subnetworks, as well as the retrieval of both functional associations and physical interactions from STRING. Finally, the latest stringApp version offers an improved user interface as well as support for cross-species queries to retrieve protein–protein interactions between eukaryotic parasites and their hosts.

Cancer is a heterogeneous disease caused by disruptions in normal cellular functions. The role of gene regulation in the development and progression of cancer has been well established. Therefore, understanding gene regulation and how it differs in disease is a key challenge. Gene regulation can be represented as biological networks and various tools for gene regulatory network (GRN) inference have been developed in recent years.

In this study, we integrated sample specific GRNs with multi-omics data in 10 cancers from The Cancer Genome Atlas (TCGA) using joint dimensionality reduction (JDR) approaches to find latent factors across omic modalities. We added network summary metrics for our GRNs to the JDR model and compared to the model without networks, to assess the usefulness of our GRNs in integrative approaches.

We found that the contribution of the GRNs to the JDRs varied across cancer types, with the strongest contribution being in colon cancer, where it accounted for over 50% of the variance explained. Furthermore, the addition of GRNs yielded factors more strongly associated with patient survival in 5/10 cancers. We were able to identify two survival associated factors in colon cancer that were enriched in homeobox and inflammatory response genes.

Chromosome conformation capture (3C) techniques exploit digestion and subsequent
relegation of cross-linked chromatin in cell nuclei, allowing the identification of spatial proximity between DNA sequences. The output of 3C techniques is a matrix,
representing the chromatin interactome of a population of cells. In order to better
understand intrinsic relationships between chromatin interactions, we decided to take
advantage of network analysis: contact maps are interpreted as distance matrices and
easily transformed into adjacency matrices of chromatin networks.
We developed HiCONA, a python package which is able to perform several network
analysis tasks, starting from standard 3C inputs. In particular, HiCONA generates
chromatin network, it performs network sparsification, in order to improve network
handling.
Starting from the sparse network, we are able to weight the contribution of
some standard chromatin annotation (ChIP-peaks, ChromHMM states) to global network
measures. On the other hand, it is possible to associate local network measures to
chromatin annotations. Furthermore, HiCONA enables differential analysis between two HiC experiments, by extracting differentially dense subgraphs.
With HiCONA we dissected the major rules of chromatin dynamics and regulation, by
comparing different 3C-derived datasets.

Genome-wide loss-of-function screens are powerful tools for deciphering gene knockout effects and identifying therapeutic targets. To what extent such data can be used to refine regulatory pathways, however, is unknown. We demonstrate that constructing a correlation network from the DepMap CRISPR knockout screens reveals genes with common biological functions. Extracting subnetworks that represent well-studied pathways uncovers a hierarchical substructure that coincides with the compartmentalisation of glycolysis. The incompleteness of pathway annotation data raises the question whether we can build on the loss-of-function correlation network to refine these annotations. Using network propagation, specifically the well-established personalised PageRank algorithm, we identify genes that are in proximity to selected seed genes. We verify this approach with a cross-validation and outperform baseline ranking across a wide range of parameters. We then demonstrate the method by using members of the Ragulator--Rag complex and identify genes functionally associated with this complex. The presented method is a general tool for identifying a ranking of genes from a list of seed genes, based on a similarity in loss-of-function screens. As such, we anticipate that it can be used as hypothesis creator for biologists that aim to extend an identified list of genes.

Boolean networks are graphical representations of gene connections at regulatory level. As such, Boolean networks are a key tool for investigating the impact of drug treatment on gene activities, and potential drug candidates. The open-source Python package NORDic aims at providing reproducible and modular pipelines for the inference and analysis of Boolean networks. Moreover, NORDic also implements a ready-to-use scoring of drug effect on Boolean networks, with a computationally efficient procedure for recommending top candidates. That procedure relies on bandit algorithms, which are adapted to the optimization of resource allocation. As an example, we have performed a quick study of breast cancer. As a result, we were able to effortlessly build a large model associated with breast cancer on 251 genes. That model was able to predict around 68% of previously unseen experiments from the literature. Furthermore, we have tested the drug repurposing approach on a list of anticancer drugs. Scoring relying on predictions from Boolean networks was consistently able to improve upon the benchmark. In addition, the bandit-based recommendation system offers theoretical guarantees on the quality of candidates, while being more sample-efficient than a grid-search approach. These results represent a step towards faster development of network-based drug repurposing.

We are interested in deciphering the complexity of intra-cellular signaling mediated by (tyrosine)-phosphorylation. We propose to model the propagation of the perturbation of a kinase activity as a Random Walk with Restart (RWR) in a Protein-Protein Interaction network. This choice is motivated since RWR arises naturally when describing diffusive processes. We show that RWR parameters can be optimized as a Linear Program, meaning that we can learn the weights of a RWR to best match observations on the network’s nodes. RWR edge weights are often set as a prior, we show that we can choose them to best fit data. Using our theoretical result, we developed an approach which uses both a phosphoproteomics dataset and an interaction network built from Kinase-Substrate interaction databases to propose the optimal weights to propagate the signal from the perturbed protein to the differentially phosphorylated proteins. We have first applied our approach to a dataset [10.1242/jcs.254599] to propose pathways explaining the differential phenotypic response of cells subject to the activation of an optogenetic mutant of the Src protein. We are currently comparing our approach to other methods (PHONEMeS [10.1021/acs.jproteome.0c00958], CausalPath [10.1016/j.patter.2021.100257]), on an experimental dataset built to assess the quality of causal networks (HPN-DREAM [10.1038/nmeth.3773]).

We present a new hybrid approach network modeling approach, called pyPARAGON. pyPARAGON has three steps: i. Graphlet-guided network (GGN) construction, ii. Propagation via the personalized PageRank (PPR) algorithm, iii. Edge scoring and selection via flux calculation. pyPARAGON takes a list of seed genes/proteins (initial nodes) as input, which are specific to the biological context of interest. Seeds can be obtained from but not limited to omics experiments, drug perturbation analysis, or disease-associated proteins. Graphlets help in reducing the load of possible false positives, while network propagation via PPR scores all other proteins in the reference interactome by orienting them from the given seed list. This hybrid approach, on one hand, trims the reference interactome to the most relevant interactions, on the other hand, quantifies the importance of other proteins based on a given seed list. Resulting node scores are used for calculating fluxes on each edge, which consider degrees of the node pair and edge confidence scores in a function. pyPARAGON outperformed other state-of-the-art approaches on a benchmark set consisting of cancer signaling pathways. We applied it to integrate phosphoproteomic data of 105 breast cancer tumors and stratified the tumors based on reconstructed networks. pyPARAGON is available at https://github.com/metunetlab/pyPARAGON.

Viral infections continue to cause pandemics, highlighting the importance of uncovering their disease mechanisms to enable drug re-purposing to treat them. In our previous work, we discovered that "common neighbours" (CNs) genes, which directly connect viral interactors (VIs) and differentially expressed genes (DEGs) in the human interactome, are key to COVID-19 mechanisms and promising targets for drug re-purposing.
Here, we use our CN concept across five well-studied viruses and extend it for viral infections without DEG data, uncovering disease genes for 13 viruses (8 without DEG data). We find a large overlap across the identified genes for each virus with significant enrichment in viral and immune system-related processes, uncovering genes related to "pan-viral" disease mechanisms.
Finally, we re-purpose existing drugs to target our identified "pan-viral" genes using a two-step model, an NMTF embedding method and an XGBoost classifier. We confirmed our predictions with molecular docking. For our top 10 "pan-viral" genes, we re-purpose drugs with antiviral evidence, such as Artenimol, an antimalarial agent investigated for treating viral infections. Our CN concept allowed us to find "pan-viral" genes and re-purpose drugs for treating viral infections by disrupting "pan-viral" disease mechanisms giving a unique opportunity for drug re-purposing in a broad-spectrum way.

Mechanisms behind complex diseases, as multiple sclerosis (MS), cannot be attributed to individual genes/proteins. OMICs data provide insights at gene/protein level, but the phenotypic effect is elucidated at the system level (“genes do not work alone”). There are multiple ways to affect biological systems behind a disease and single/few variants can rarely explain a phenotype. Furthermore, a disease signal spread over multiple genes/proteins can be difficult to identify without large cohorts.

Here we used the inBio Map PPI resource based on a strong experimental foundation and extracted a high-confidence subset of interactions to find biological systems enriched in MS disease signal. Next, gene-level p-values for association to MS (UK_biobank) were mapped to the interactome and our own SystemSignificance algorithm was applied to identify sub-networks enriched in signal. Finally, we evaluated the performance based on rediscovering previously reported MS drug targets.

We identified 10 networks (219 unique proteins) enriched in disease signal and observed six times more MS drug targets compared to random expectance and gene-level data alone. The retrieved biological functions aligned with known MS biology (e.g. neuronal degeneration and muscular dystrophy) and can be used for selecting novel drug target candidates.

Modeling microbial systems as sparse and reproducible networks is a major challenge in network inference. Direct interactions between the microbial species of a given biome can help to understand how the microbial communities influence the system and through which mechanisms. Most state of the art methods reconstruct network from abundance data using Gaussian Graphical Models for which statistically grounded and efficient inference approaches are available: MB neighborhood selection, glasso, tree averaging approach,...

In this article, we consider 8 such inference techniques and introduce a two-step consensus method to combine them. The individual methods all rely on stability selection: a resampling-based procedures to select a regularization parameter from the selection frequency of edges in the networks constructed. Rather than selecting the optimal regularization parameter and discarding the selection frequencies, we combine edge frequencies directly to reconstruct the network.

The effectiveness of the consensus compared to individual methods is proved on synthetic data derived from a set of healthy microbiota and then applied to a real study-case of microbiota from patients with liver cirrhosis. The performances of all methods in terms of precision/recall are quite similar, but MB-based ones tend to outperform glasso-based ones and be outperformed by the consensus.

Our understanding of biological systems can be modelled in knowledge graphs (KGs), which encode entities and their relationships as triplets, like (“gene” - “involved in” - “cellular pathway”). However, our knowledge remains incomplete and biased in certain research areas, despite efforts of many consortia; due to limited exploration of concepts, such as genes and diseases, and high experimental costs. Provided that conceptual associations can be abstracted and learned from, an automated framework that infers new knowledge from existing data would be impactful.
To this end, we adapted a graph neural network tool to reason on biomedical KG predicting links and closing gaps in our understanding of health and disease. We showcase our tool on two applications: 1) predicting links between genes and their functions for functional annotation, and 2) identifying therapeutic opportunities by linking drugs and diseases. Our path-based method outperforms existing node embedding-based methods, notably by better exploitation of additional information on the underlying biological regulation. In addition, it provides interpretability by visualizing the gradients along the paths composing an association, which can guide verification and validation of the predictions. Our tool has the potential to generate new knowledge in any biomedical KG and advance our understanding of biology.

Colorectal Cancer has the second-highest mortality rate worldwide, which requires advanced diagnostics and individualized therapies to be developed. Interactions between molecular entities provide valuable information to detect the responsible genes driving cancer progression. Graph Convolutional Neural Networks are able to utilize the prior knowledge provided by interaction networks and the Spektral library adds a performance increase in contrast to standard implementations.

Machine learning technology shows great potential to assist medical professionals, however, the deep learning models are limited in their application due to their lack to explain the factors contributing to a prediction. Adaption of the Graph Layer-Wise Relevance Propagation methodology to graph-based deep learning models allows attributing the learned outcome to single genes comprising potential targets.

We present an implementation of Graph Convolutional Neural Networks using the Spektral library with adapted functions for Graph Layer-Wise Relevance Propagation. Deep learning models were trained on a large gene expression dataset of Colorectal Cancer patients with molecular interaction networks as prior knowledge: Protein-protein interactions from the Human Protein Reference Database and STRING, and pathways from the Reactome database. Our implementation shows great performance with reduced computation time, especially for large networks, and reveals possible, and even more distant, biomarkers and drug targets.

Protein-protein interaction (PPI) networks have been found to be power-law-distributed, i.e., in observed PPI networks, the fraction of nodes with degree k often follows a power-law (PL) distribution k-α. The emergence of this property is typically explained by evolutionary or functional considerations. However, the experimental procedures used to detect PPIs are known to be heavily affected by technical and study bias. For instance, proteins known to be involved in cancer are often heavily overstudied and proteins used as baits in large-scale experiments tend to have many false-positive interaction partners. This raises the question of whether PL distributions in observed PPI networks could be explained by these biases alone. Here, we address this question using statistical analyses of the degree distributions of 1000s of observed PPI networks of controlled provenance as well as simulation studies. Our results indicate that study bias and technical bias can indeed largely explain the fact that observed PPI networks tend to be PL-distributed. This implies that it is problematic to derive hypotheses about the degree distribution and emergence of the true biological interactome from the PL distributions in observed PPI networks.

Low-dimensional embeddings are a cornerstone in the modelling and analysis of biological networks. Embedding biological networks is challenging, as it involves capturing both topological (similar wiring patterns) and homophily similarity (proximity in the network). These two similarities are complementary; for instance, a function is shared between the proteins that physically interact, but also between the proteins that are similarly wired. However, current network embedding techniques preserve either topological or homophily similarity. Preserving both similarities would lead to more functionally organized spaces and better performance in downstream analysis tasks. In this study, we introduce GPMI and DeepGraphlets, two novel network embeddings based on graphlets (small, induced and connected subgraphs) that simultaneously capture topological and neighbourhood membership information. Our newly introduced methods incorporate graphlets in the LINE and DeepWalk, two state-of-the-art random walk-based network embeddings. First, we show that in the graphlet-based representations of the networks, the genes are grouped in a more homophilic way (i.e., more genes with similar functions are adjacent) than in the standard representations that capture either neighbourhood, or topological similarity. Then we demonstrate that GPMI and DeepGraphlets that rely on these more homophilic representations lead to better functionally organized gene embedding spaces than LINE and DeepWalk methods.

Network-based analysis of omics data offers a systems-level perspective on complex biological processes, aiding in the interpretation of large datasets. For instance, such analytical methods can be used to find rare driver mutations in genomics data or to detect changes in the activity of pathways using transcriptomics. Among the available network algorithms that combine target genes with interaction networks to find relevant subnetworks, GoNetic stands out for its versatility and computational efficiency. GoNetic can leverage a large variety of feature types by converting available data to probabilities assigned to nodes and edges. Furthermore, GoNetic is able to analyze multiple samples together while retaining information that is sample-specific in the final result.
GoNetic was applied successfully on both genomics and transcriptomics data. Using RNA-Seq data from a cohort of patients with locally advanced prostate cancer, we found subnetworks connecting genes related to aggressiveness in primary lesions. Furthermore, using genomics data from the Hartwig Medical Foundation, we found low frequency mutations in genes that potentially drive metastatic prostate cancer [1].

[1] de Schaetzen van Brienen, L., Miclotte, G., Larmuseau, M., Van den Eynden, J., & Marchal, K. (2021). Network-based analysis to identify drivers of metastatic prostate cancer using GoNetic. Cancers, 13(21), 5291.

Epidemiological evidence shows that some diseases tend to co-occur; more exactly, certain groups of patients with a given disease are at a higher risk of developing a specific secondary condition. However, previous systematic studies using molecular information have captured a small proportion of the comorbidities among their analyzed diseases, being unable to provide a general interpretation for them.
Here we develop a new approach to generate a disease network that uses the accumulating RNA-seq data on human diseases to significantly match an unprecedented proportion of known comorbidities, providing plausible biological models for such co-occurrences. Of note, more than 95% of the captured comorbidities show immune system involvement.
Furthermore, 64% of the known disease pairs can be explained by analyzing groups of patients with similar expression profiles, highlighting the importance of patient stratification in the study of comorbidities. These results solidly support the existence of molecular mechanisms behind many of the known comorbidities. We provide a functional and comprehensive resource to explore comorbidities at different resolution levels at http://disease-perception.bsc.es/rgenexcom/.
(All the information is accessible at https://doi.org/10.1101/2021.07.22.21260979)

1 out of every 3 adults suffers from two or more diseases worldwide, which impacts life quality and expectancy, clinical management and health care cost. Epidemiological evidence shows that numerous diseases co-occur more than expected by chance. Very recently, a methodology based on RNA-seq proved that an unprecedented percentage of these epidemiological interactions (EIs) among the analyzed diseases (up to 64%) can be significantly recalled with molecular information. Nonetheless, it left unclear the nature of EIs transcriptomic similarities, which might reflect a range of genetic to environmental factors.
Here, we used the UK Biobank to disentangle the genetic component under the EIs transcriptomic similarities and to compare the ability of transcriptomics and genetics to recall EIs. First, we built massive disease and stratified transcriptomic networks that robustly and significantly recall an unparalleled proportion of diverse epidemiological networks (80%). We found that almost the entire universe of EIs (93.3%) have a molecular basis, where gene expression is the single most informative layer by far, able to cover and further expand the genetically-interpretable EIs. Furthermore, half of the EIs may be rooted in molecular yet non-genetic factors, suggesting the potential importance of transcriptomics, epigenetics or the environment in disease co-occurrences.

High-throughput liquid chromatography mass spectrometry enables large-scale phosphoproteomics data analysis to study cell signalling networks and their role in disease. However, interpreting the data remains a challenge as existing data-driven methods only provide associations among signalling proteins without implying causality. Therefore, network interpretation remains the researcher's responsibility. However, considering all prior knowledge is time-consuming, leading to under-analysis of available measurements and datasets.

In this study, we explore different knowledge representation and automated reasoning methods and their potential for facilitating data interpretation and scientific discovery from phosphoproteomics data and prior knowledge. First, we devise a Bayesian network structure learning approach that incorporates knowledge of basic cell signalling principles and information from biological databases to infer causal enzyme-substrate relationships from phosphorylation data. We then apply probabilistic logic modelling to prioritise possible explanations for each datapoint within the context of the learned network structure and elucidate underlying molecular mechanisms using Bayesian inference.

We demonstrate our approach on data acquired from three cell lines treated with 61 kinase inhibitors. Our findings highlight the potential of logic modelling for automatically extracting human-interpretable explanations from large-scale experimental data and the importance of including probabilistic modelling strategies for reflecting uncertainty in experimental data and prior knowledge.

Vascular Cognitive Impairment (VCI) is a growing public health concern as the population ages, and it substantially impacts a person's quality of life. Because VCI can impair cognitive abilities, identifying associated pathways and proteins is crucial. VCI is a complex disorder characterized by both cardiovascular and dementia symptoms. Computational approaches such as network medicine can assist in understanding its complexities by exploring the crosstalk between cardiovascular and cognitive disorders. We studied this crosstalk by building protein-protein interaction networks related to cardiovascular and cognitive diseases. Additionally, we examined the role of oxidative stress in this crosstalk. Phenotype-specific networks were constructed with GUILDify, which uses the guilt-by-association principle. We suggested that oxidative stress-related proteins might be crucial to the crosstalk. The crosstalk between networks suggested that the following proteins could be contributing to VCI: DOLK, TSC1, ATP1A1, MAPK14, YWHAZ, CREB3, HSPB1, PRDX6, and LMNA. Furthermore, our results suggest that glycative stress has a significant role in the generation of oxidative processes and post-translational protein changes via advanced glycation end products (AGEs). We hypothesize that these products may interact with their particular receptors, RAGE, and Notch signaling, contributing to the etiology of vascular cognitive impairment.

The human genome is composed of genes that appeared at different evolutionary ages for 3.5 billion years. Those genes have progressively been integrated across time, they acquired new functions and made species genomes more sophisticated. Evolutionary scientists have been able to precisely estimate current human gene ages by studying duplication events through time. Whether genes with the same evolutionary age are associated within the genome, have similar expression or share a same chromatin structure is still not well characterized. Inspired by the atavistic theory of cancer, which relates malignancy to the expression of evolutionary ancient phenotypes, we investigate whether cell differentiation and cancer can alter the associations between gene age and other (epi)genomic characteristics. We therefore investigate whether genes with different evolutionary ages could show specific epigenome properties, expression regulation, variability and location in 3D chromatin structures that can be potentially altered in cancer and during differentiation. We identify consistent changes during oncogenesis in the spatial organization of genes from different evolutionary classes, correlated with changes in their expression variability across individuals, reinforcing the important role that old genes and their organization in the nucleus play in cancer phenotypes.

Tissues are complex systems made of multiple cells spatially and temporally organized that interact with each other. Recently, single-cell RNA-sequencing has emerged as a powerful tool to study such cellular communications and several bioinformatics methods have been proposed to infer intercellular signaling between groups of cells, combining ligand and receptor expression levels. Only few methods investigate also the intracellular signalling activated by the ligand–receptor binding, i.e. the signalling cascade triggering cell response and transcriptional activation/inhibition of specific genes.
We proposed scSeqComm (R-package https://gitlab.com/sysbiobig/scseqcomm), a computational method to identify, quantify and characterize cellular communication at both intercellular and intracellular signalling level from scRNA-seq data. Compared to previous approaches, scSeqComm quantifies intracellular signalling not only to characterize the functional effects of communication, but also to support the evidence that the communication has occurred.
With respect to the original publication, the proposed framework was extended to enable analysis of differential cellular communication in multi-condition and multi-patient scenarios. Moreover, parallelism and in-memory computation features were introduced, along with a user-friendly and interactive dashboard to support analysis and interpretations of results.
We show applications of the approach to scRNA-seq datasets, its validation using spatial transcriptomics data and the comparison with state-of-the-art intercellular scoring schemes.

Signal transduction networks manage various biological processes and are often aberrated in disease such as cancer. Obtaining a mechanistic, quantitative understanding of those networks, and their differences, is a key challenge in cellular biology.
Here we introduce single cell Comparative Network Reconstruction (scCNR), which uses single cell protein count data to reconstruct their signal transduction network and infer differences in signalling between cell states.
The method is based on modular response analysis, which derives networks from perturbation data. Similarly, scCNR exploits stochastic variability of total protein and treats those differences as ‘natural perturbation experiments’. Those differences in total protein result in differences of active protein and lead to unique steady states of the network for every single cell. Taking the total and active protein counts of single cells as input scCNR reconstructs a shared network topology for several cell states and recovers separate interaction strengths between active proteins for every cell state.
We evaluated scCNR on simulated single cell data of the EGFR/ ERK signalling pathway. We additionally simulated mutant networks (e.g., RAS/ BRAF mutants) and evaluated the ability of scCNR to recover the mayor differences in signalling between the wildtype and mutant cell states. The method achieves high accuracy in the reconstruction of the network topology even in the presence of considerable noise and can further recover the mayor differences between wildtype and mutant networks.
We expect this method to assist in the understanding of cellular heterogeneity and identify cell state dependent differences in signal transduction within seemingly homogeneous populations of cells.

Distinction of drivers from passengers, their cooperativity and exclusivity, and the temporal order of accumulation of molecular alterations is a crucial yet daunting, unsolved task. Early alterations in temporal order can inform about network rewiring and direct the identification of drug targets. The challenge in temporal modelling of cancer and the evolution of networks beyond the alterations remain elusive. The focus of this study is directly devoted to address this challenge. We developed an integrative approach to reveal the network-based history of tumor progression and to design personalized therapeutic strategies based on the validated models. The core method is the graph-based cellular automata (GCA), which is a discrete dynamic model. GCA model can simulate complex systems by considering local interaction rules, and it gives insights into the underlying mechanism of complex behavior. The reference graph is a tissue-specific interactome consisting of protein-protein interactions and regulatory interactions. Proteins have a state of being on/off based on the rules of their neighbors derived from multi-omic data sets such as co-occurring or mutually exclusive alterations and interactions. We applied this model to the cell line-dependent data in DepMap which consists of both molecular alterations (mutation transcriptomic and proteomic data) and drug response.

We present a new network-based method for predicting context-specific transcription factor (TF) activity. Our method integrates H3K27ac ChIP-seq and public ChIP-Atlas data to construct a dense and diverse gene regulatory network (GRN) that reflects biologically relevant TF-target gene relationships. The context-specific GRN is used in statistical inference methods to predict upstream TF activity from downstream gene expression data. To demonstrate the utility of our method, we applied our approach to a uterine leiomyoma dataset (n = 20 patients). The constructed context-specific network contains over 5 million unique connections across 1,643 TFs and 9,724 target genes. We predicted TF activity at the full transcriptome and differential gene level, using consensus scores from multiple methods (AUCell and GSEA for full transcriptome; ULM and WSUM for differential gene level). Differential activity analysis results are supported by literature, such as enhanced PGR and ESR1 activity. The results also identified novel TFs potentially linked to leiomyoma biology, such as E2F family members. To evaluate the impact of our context-specific GRN, we compared our approach with using DoRothEA, a widely-used curated GRN. The results highlight the importance of incorporating large-scale databases and context-specific information. Our integrative multi-omics approach serves as an informative tool for hypothesis generation.

O-GlcNAc is a post-translational modification that is attached to serine and threonine residues of intracellular proteins by O-GlcNAc transferase (OGT) and removed by O-GlcNAcase (OGA). It is particularly hard to study, since manipulation of O-GlcNAc causes pleiotropic effects. Thus, we combined multi-omics and biological networks with novel statistical algorithms to elucidate the important features implicated in O-GlcNAc manipulation and their relationships with each other.
Transcriptomic, proteomic, phospho-proteomic, and metabolomic data were collected from mouse liver samples treated with one of three treatments at 1, 2, or 4 weeks: OGT knockout (KO), OGA-KO, and TMG (an OGA inhibitor). Results from differential expression analyses were combined with biological networks of protein-protein, metabolite-metabolite, and protein-metabolite interactions to find active subnetworks, which were functionally characterized through over-representation analysis. Further wet lab experiments, such as flow cytometry-based ploidy analysis, were performed to validate findings.
Subnetworks revealed explicit connections between proteins and metabolites that are associated with mitotic regulation. These findings were corroborated by ploidy analysis, showing aneuploidy in OGT-KO + PHX hepatocytes compared to WT. The integration of multi-omics with biological networks provides a powerful framework for understanding the complex responses produced from O-GlcNAc perturbations.

Single-cell RNA sequencing has exposed extensive and complex heterogeneity at the cell-type-level. Given such complexity, detecting gene coexpression is essential for understanding biological mechanisms because it focuses on gene-gene pairings that contribute functional regulation to a condition or phenotype. However, measuring coexpression rarely takes into account the variability or stability of this value. Here, we introduce scRoGG, which leverages the essential genes to measure coexpression based on the proportion-based correlation under reference genes.
scRoGG can capture cell-type-specific coexpressed gene pairs, which are robust across experiments and sequencing methods. We showcase that cell-type-specific coexpression patterns in innate and adaptive immunities. Additionally, shared genes identified from pancreatic β cells across experiments consolidate the robustness of scRoGG under batch effects. scRoGG can also extend to identify differentially correlated (DC) gene pairs among conditions, uniquely based on variability changes. Using T2D data, scRoGG identifies DC pairs that are independent of mean changes with significantly enriched peptide and hormone secretion function. Furthermore, scRoGG reveals DC changes pattern in the mouse brain endothelial arteriovenous zonation with novel transcription factors co-regulation changes under the spatial trajectory.
Overall, scRoGG provides a novel framework for interrogating robust cell-type-specific coexpression and significant DC gene pairs under different biological systems.

We propose a novel strategy for provenance tracing in network diffusion algorithms, a problem that has been surprisingly overlooked in spite of the wide use of diffusion algorithms in biological applications. We develop a path-based approach where for each node we rank a path by the magnitude of its contribution to that node's score. We demonstrate two applications of this analysis: (i) a quantitative measure called “path-based effective diffusion” for evaluating a diffusion algorithm’s ability to exploit the full topology of a network and (ii) a method to analyze a node’s “diffusion betweenness”, i.e., its importance in diffusing scores across a network. We applied these ideas to SARS-CoV-2 protein interactors and human PPI networks. Provenance tracing of the Regularized Laplacian and Random walks with restarts revealed that a substantial amount of a node’s score is contributed via paths with more than one edge. Additionally, our analysis showed that nodes with high diffusion betweenness were enriched in essential human genes and interactors of other viruses, demonstrating the usefulness of this new concept.

The growing availability of multiple molecular modalities for individual patients in large cohorts requires effective methods for combining and analyzing these datasets, particularly in complex diseases. While network-based approaches like Similarity Network Fusion show promise, we propose an alternative multilayer network approach providing potential benefits. Our method involves three steps: constructing single-modality similarity networks using a resampling approach (COGENT) for optimality; coupling networks based on shared information; and applying the Leiden algorithm to identify patient groupings. We applied this method to COMBAT, a multimodal dataset of both COVID-19 and sepsis patients, and healthy volunteers, spanning single cell and bulk gene expression, plasma and serum proteomics, cell composition measured by mass cytometry, and detailed clinical information. Our analysis identified five clusters that recovered major clinical diagnoses and revealed connections between biological pathways and clinical features, including regulation of complement cascade, interleukin signaling, Toll-like Receptor cascade, length of hospital stay, and oxygenation. Furthermore, we applied our method to a separate sepsis dataset, integrating transcriptomic and proteomic modalities, and identified clusters with differential protein levels, gene activities, and distinct clinical outcomes. Our multilayer network approach offers potential for use with other multimodal molecular health datasets, providing a versatile tool for understanding complex diseases.

The immune system is controlled through complex networks of interactions involving many individual components across multiple cell types. The action of these networks serve to connect the underlying genotype to the organisms immune phenotype. This project employs integrative network biology to explore molecular determinants of immune response and disease susceptibility. The use of phenotypic, genomic, and high throughput sequencing in cattle breeding has improved dairy herd performance. However, cattle health and welfare issues continue to harm the industry. A single-cell sequencing cattle immune cell dataset was selected, alongside gene pair GO semantic similarity measurements to train statistical models for functional gene network inference with gold standard data taken from Reactome. We predict context-specific biochemical interactions between genes in these immune cell types. Post-GWAS analysis will combine with these networks and inform construction of risk stratification models predicting bovine immune status, disease susceptibility and risk. These will help development of improved treatment interventions for common cattle diseases and conditions, improving cattle health and reducing zoonotic disease and antibiotic-resistant risk, while application of our approach in other immune contexts is expected to inform immuno-oncology.

At AstraZeneca, we follow a strategy where decision-making is driven by five technical determinants or the 5Rs - the right target, right tissue, right safety, right patient, and right commercial potential. At the start of the drug discovery pipeline is our first R – the right target. In our team, we use the world’s data in the form of biomedical knowledge graphs to identify new targets in respiratory and fibrotic disease.

At the heart of our target discovery knowledge graph is the link between genes (protein- protein interactions), as well as genes and diseases that are drawn from both internal and external resources, of which there are 10s-100s. Of these, we believe it is important to choose the information that is relevant to the machine learning task at hand but also of high quality, thus enabling us to predict high-quality targets with lower attrition rates as they progress through the drug discovery pipeline.

To generate evidence for this claim, we have begun performing information ablation studies using a set of well-known targets as the expected outcome in respiratory disease. I will be presenting the initial results of these in my poster and plans for future work in the area.

Knowledge graphs (KGs) provide a structured way to store, query, and organize
complex biomedical information, capturing entities (e.g. genes, diseases,
molecules, or patient-level data), their attributes, and the relationships between
them. This knowledge representation is crucial for machine learning tasks in the
biomedical domain, including information retrieval, community detection, and
attribute or link prediction.
Despite the remarkable performance of deep graph neural networks in the
representation of KGs in recent years, they suffer from over-smoothing and
struggle to handle heterogeneity as commonly found in biomedical KGs. Here, we
propose a novel approach, based on geometric scattering with learnable scales, to
effectively capture the geometry and topology of biomedical KGs at both local and
global scales. Our approach leverages the scattering transform, a signal processing
technique inspired by wavelet theory, on a KG integrating literature, genomics,
expression, disease, and ontology data to connect genes, diseases, tissues, and
biological pathways.
Our method shows state-of-the-art performance in the prediction of novel, high-confidence gene-disease links for drug repurposing, or indication expansion. Our
method also provides insightful low-dimensional representations of the KG and
prediction explainability, which are key features in biomedical applications.
This work was funded by a Yale–Boehringer Ingelheim Biomedical Data Science
Fellowship.

In multi-cellular organism, cells constantly send and receive signals to coordinate and regulate all biological processes. The breakthrough of single-cell technologies offers the unprecedented opportunity of investigating cell-cell communications in biological systems. Simultaneously, an increasing number of computational tools and resources have been proposed.
In this review, we provide a comprehensive and in-depth overview of cell-cell communication analyses. Since all the computational tools strongly rely on the prior knowledge of intercellular interactions, the choice of such a resource has a considerable impact on the results. Thus, firstly, we review more than 20 resources with particular focus on the level of curation (manual/computational), cardinality of annotated interactions, and type of biological information provided (protein complexes, cofactors, etc.). Secondly, we systematically review more than 40 computational tools for cellular communication inference, highlighting their peculiarities, the adopted computational strategies, and the modeled biological characteristics (e.g. ligand-receptor interactions, downstream intracellular signaling) and the possibility to consider multi-conditions experiment design and 2D/3D spatial information.
The goal is to provide an exhaustive panorama of cell-cell communication analyses, highlighting the current challenges and the open issues that need to be faced to make cellular communication inference a standard and fully integrated analysis step in single-cell transcriptomics studies.

Exploring the molecular basis of disease severity in rare disease scenarios is a challenging task provided the limitations on data availability. Causative genes have been described for Congenital Myasthenic Syndromes (CMS), a group of diverse minority neuromuscular junction (NMJ) disorders; yet a molecular explanation for the phenotypic severity differences remains unclear. Here, we present a workflow to explore the functional relationships between CMS causal genes and altered genes from each patient, based on the analysis of a general knowledge multilayer network presenting relevant biomedical aspects for the disease, namely protein-protein interactions, pathways and metabolomics.

Our results show that CMS severity can be ascribed to the personalized impairment of specific classes of NMJ proteins, namely extracellular matrix components (proteoglycans, tenascins, chromogranins) and postsynaptic modulators of AChR clustering (LRP4, PLEC). Moreover, we provide experimental evidence for the potential modifying effect of a gene previously unknown to be a NMJ interactor, USH2A, through expression knock-down of the zebrafish orthologue (ush2a), examining effects on movement and NMJ morphology.

This work showcase how coupling multilayer network analysis with personalized -omics information provides molecular explanations to the varying severity of rare diseases; paving the way for sorting out similar cases in other rare diseases.

The complete elucidation of a biological phenomena goes through the possibility to computationally explore the mechanisms and dynamics insight the molecular networks and interaction maps. Traditionally, high-throughput experimental data is a resource for understanding high-level functions and utilities. To mine knowledge from these maps, we must consider them as a deep and dynamic hierarchy of subgraphs strongly interconnected over several orders of scale. The challenge is to translate this hierarchy into a label node integrating multiple layers of biological, clinical, and annotated information in order to describe a particular biological condition. In our work, we aim to integrate omic data into databases of cured structure modeling drug response. As a proof of concept, we will exploit data on response to a targeted therapy in a cohort of 231 metastatic patients whose patient-driven xenograft had been characterized at the genomic, transcriptomic, and epigenetic level, as well as in response to cetuximab for a total of 5 different omics.

Abnormal alterations in a kinome-mediated signaling networks are signatures in many cancer types. Discovery of significant network motifs and elaborating them context-specifically can unveil clinically important target proteins or pathways as well as help for tumor stratification. In this study, we constructed ovarian tumor-specific networks by integrating reference kinase interactome and kinase-substrate interaction in NetworKIN with phosphoproteomic profiles from TCGA. We screened all possible network motifs including feed-forward and feed-back loops and tested their presence compared to random interactomes. As a result, we found 5 significant motifs in which positive cascade, positive feedback loops, and coherent type-1, coherent type-2, and incoherent type-1 feed-forward loops (FFLs) were present. Tumor-specific network of each patient was constructed by mapping differentially phosphorylated proteins to significant motifs and filtering out edges which are possibly not context-specific. All pair comparison and clustering of networks show that 40% node and 2.5% network motif similarity among patients. Finally, we aligned the patient groups with available ovarian cancer cell lines to assess the efficacy of kinase inhibitors in a clinically relevant context, providing valuable insights for developing personalized treatment strategies with personalized network motif profiles.

While immune checkpoint inhibitors show success in treating a subset of patients with certain late-stage cancers, these treatments fail in many other patients, due to mechanisms that have yet to be fully characterized. The process of CD8 T cell exhaustion, by which T cells become dysfunctional in response to extended antigen exposure, has been implicated in immunotherapy resistance. Single-cell RNA sequencing (scRNA-seq) produces an abundance of data to analyze this process; however, due to the complexity of the process, contributions of other cell types to a process within a single cell type cannot be simply inferred. We constructed an analysis framework to first rank human skin tumor samples by degree of exhaustion in tumor-infiltrating CD8 T cells and then identify immune cell type-specific gene-regulatory network patterns significantly associated with T cell exhaustion. Using this framework, we further analyzed scRNA-seq data from human tumor and chronic viral infection samples to compare the T cell exhaustion process between these two contexts. In doing so, we identified transcription factor activity in the macrophages of both tissue types associated with this process. Our framework can be applied to other biological contexts, facilitating insights into key processes that underpin the effective treatment of disease.

Introduction: Spermatogenesis caused by endocrine, genetic, immunological and environmental influences is one of the major causes of male infertility. The egg case of the praying mantis, is a product of insect farming that is used in a steamed form in korean traditional medicine, which used to treat male infertility. Here, we taken forward the pseudo-medicinal practice to genome wide molecular differential expressions.


Materials and methods: In this study, we observed genome wide methylation profiles of irradiated and Mantidis Ootheca (MO) treated rats to understand the differential methylation patterns which could associated with male infertility. In total, 30 samples (10 groups) with replication were sequenced with Illumina-Novaseq-6000.


Result: resulted with 1,753,742 CpG , 5,080,116 CHG, and 14,302,927 CHH genomic locations. Among those, 22 genes were identified with differential methylated regions and those involved in sperm motility, cell cycle, and DNA repair, could associated with spermatogenesis.

Conclusion: Among those, we observed there is no similar molecular patterns of MO treatments while at prevention and treatment of irradiations. Which shows, the MO treatment and prevention molecular mechanism are different. So, this data set will help researcher to track the detail molecular mechanisms, which involved in MO treatments.