Current and future directions in network biology.

Summary: Network biology is an interdisciplinary field bridging computational and biological sciences that has proved pivotal in advancing the understanding of cellular functions and diseases across biological systems and scales. Although the field has been around for two decades, it remains nascent. It has witnessed rapid evolution, accompanied by emerging challenges. These stem from various factors, notably the growing complexity and volume of data together with the increased diversity of data types describing different tiers of biological organization. We discuss prevailing research directions in network biology, focusing on molecular/cellular networks but also on other biological network types such as biomedical knowledge graphs, patient similarity networks, brain networks, and social/contact networks relevant to disease spread. In more detail, we highlight areas of inference and comparison of biological networks, multimodal data integration and heterogeneous networks, higher-order network analysis, machine learning on networks, and network-based personalized medicine. Following the overview of recent breakthroughs across these five areas, we offer a perspective on future directions of network biology. Additionally, we discuss scientific communities, educational initiatives, and the importance of fostering diversity within the field. This article establishes a roadmap for an immediate and long-term vision for network biology. Availability and implementation: Not applicable.

Chronic Lymphocytic Leukemia (CLL)-Derived Extracellular Vesicles Educate Endothelial Cells to Become IL-6-Producing, CLL-Supportive Cells.

We hypothesized that via extracellular vesicles (EVs), chronic lymphocytic leukemia (CLL) cells turn endothelial cells into CLL-supportive cells. To test this, we treated vein-derived (HUVECs) and artery-derived (HAOECs) endothelial cells with EVs isolated from the peripheral blood of 45 treatment-naïve patients. Endothelial cells took up CLL-EVs in a dose- and time-dependent manner. To test whether CLL-EVs turn endothelial cells into IL-6-producing cells, we exposed them to CLL-EVs and found a 50% increase in IL-6 levels. Subsequently, we filtered out the endothelial cells and added CLL cells to this IL-6-enriched medium. After 15 min, STAT3 became phosphorylated, and there was a 40% decrease in apoptosis rate, indicating that IL-6 activated the STAT3-dependent anti-apoptotic pathway. Phospho-proteomics analysis of CLL-EV-exposed endothelial cells revealed 23 phospho-proteins that were upregulated, and network analysis unraveled the central role of phospho-ß-catenin. We transfected HUVECs with a ß-catenin-containing plasmid and found by ELISA a 30% increase in the levels of IL-6 in the culture medium. By chromatin immunoprecipitation assay, we observed an increased binding of three transcription factors to the IL-6 promoter. Importantly, patients with CLL possess significantly higher levels of peripheral blood IL-6 compared to normal individuals, suggesting that the inducers of endothelial IL-6 are the neoplastic EVs derived from the CLL cells versus those of healthy people. Taken together, we found that CLL cells communicate with endothelial cells through EVs that they release. Once they are taken up by endothelial cells, they turn them into IL-6-producing cells.

D'or: deep orienter of protein-protein interaction networks.

MOTIVATION: Protein-protein interactions (PPIs) provide the skeleton for signal transduction in the cell. Current PPI measurement techniques do not provide information on their directionality which is critical for elucidating signaling pathways. To date, there are hundreds of thousands of known PPIs in public databases, yet only a small fraction of them have an assigned direction. This information gap calls for computational approaches for inferring the directionality of PPIs, aka network orientation. RESULTS: In this work, we propose a novel deep learning approach for PPI network orientation. Our method first generates a set of proximity scores between a protein interaction and sets of cause and effect proteins using a network propagation procedure. Each of these score sets is fed, one at a time, to a deep set encoder whose outputs are used as features for predicting the interaction's orientation. On a comprehensive dataset of oriented PPIs taken from five different sources, we achieve an area under the precision-recall curve of 0.89-0.92, outperforming previous methods. We further demonstrate the utility of the oriented network in prioritizing cancer driver genes and disease genes. AVAILABILITY AND IMPLEMENTATION: D'or is implemented in Python and is publicly available at https://github.com/pirakd/DeepOrienter.

interFLOW: maximum flow framework for the identification of factors mediating the signaling convergence of multiple receptors.

Cell-cell crosstalk involves simultaneous interactions of multiple receptors and ligands, followed by downstream signaling cascades working through receptors converging at dominant transcription factors, which then integrate and propagate multiple signals into a cellular response. Single-cell RNAseq of multiple cell subsets isolated from a defined microenvironment provides us with a unique opportunity to learn about such interactions reflected in their gene expression levels. We developed the interFLOW framework to map the potential ligand-receptor interactions between different cell subsets based on a maximum flow computation in a network of protein-protein interactions (PPIs). The maximum flow approach further allows characterization of the intracellular downstream signal transduction from differentially expressed receptors towards dominant transcription factors, therefore, enabling the association between a set of receptors and their downstream activated pathways. Importantly, we were able to identify key transcription factors toward which the convergence of multiple receptor signaling pathways occurs. These identified factors have a unique role in the integration and propagation of signaling following specific cell-cell interactions.

A network-based method for associating genes with autism spectrum disorder.

Autism spectrum disorder (ASD) is a highly heritable complex disease that affects 1% of the population, yet its underlying molecular mechanisms are largely unknown. Here we study the problem of predicting causal genes for ASD by combining genome-scale data with a network propagation approach. We construct a predictor that integrates multiple omic data sets that assess genomic, transcriptomic, proteomic, and phosphoproteomic associations with ASD. In cross validation our predictor yields mean area under the ROC curve of 0.87 and area under the precision-recall curve of 0.89. We further show that it outperforms previous gene-level predictors of autism association. Finally, we show that we can use the model to predict genes associated with Schizophrenia which is known to share genetic components with ASD.

Batch correction of single-cell sequencing data via an autoencoder architecture.

Motivation: Technical differences between gene expression sequencing experiments can cause variations in the data in the form of batch effect biases. These do not represent true biological variations between samples and can lead to false conclusions or hinder the ability to integrate multiple datasets. Since there is a growing need for the joint analysis of single-cell sequencing datasets from different sources, there is also a need to correct the resulting batch effects while maintaining the true biological variations in the data. Results: We developed a semi-supervised deep learning architecture called Autoencoder-based Batch Correction (ABC) for integrating single-cell sequencing datasets. Our method removes batch effects through a guided process of data compression using supervised cell type classifier branches for biological signal retention. It aligns the different batches using an adversarial training approach. We comprehensively evaluate the performance of our method using four single-cell sequencing datasets and multiple measures for batch effect removal and biological variation conservation. ABC outperforms 10 state-of-the-art methods for this task including Seurat, scGen, ComBat, scanorama, scVI, scANVI, AutoClass, Harmony, scDREAMER, and CLEAR, correcting various types of batch effects while preserving intricate biological variations.

BERTwalk for integrating gene networks to predict gene- to pathway-level properties.

Motivation: Graph representation learning is a fundamental problem in the field of data science with applications to integrative analysis of biological networks. Previous work in this domain was mostly limited to shallow representation techniques. A recent deep representation technique, BIONIC, has achieved state-of-the-art results in a variety of tasks but used arbitrarily defined components. Results: Here, we present BERTwalk, an unsupervised learning scheme that combines the BERT masked language model with a network propagation regularization for graph representation learning. The transformation from networks to texts allows our method to naturally integrate different networks and provide features that inform not only nodes or edges but also pathway-level properties. We show that our BERTwalk model outperforms BIONIC, as well as four other recent methods, on two comprehensive benchmarks in yeast and human. We further show that our model can be utilized to infer functional pathways and their effects. Availability and implementation: Code and data are available at https://github.com/raminass/BERTwalk. Contact: roded@tauex.tau.ac.il.

From Leiden to Tel-Aviv University (TAU): exploring clustering solutions via a genetic algorithm.

Graph clustering is a fundamental problem in machine learning with numerous applications in data science. State-of-the-art approaches to the problem, Louvain and Leiden, aim at optimizing the modularity function. However, their greedy nature leads to fast convergence to sub-optimal solutions. Here, we design a new approach to graph clustering, Tel-Aviv University (TAU), that efficiently explores the solution space using a genetic algorithm. We benchmark TAU on synthetic and real data sets and show its superiority over previous methods both in terms of the modularity of the computed solution and its similarity to a ground-truth partition when such exists. TAU is available at https://github.com/GalGilad/TAU.

A mutation-level covariate model for mutational signatures.

Mutational processes and their exposures in particular genomes are key to our understanding of how these genomes are shaped. However, current analyses assume that these processes are uniformly active across the genome without accounting for potential covariates such as strand or genomic region that could impact such activities. Here we suggest the first mutation-covariate models that explicitly model the effect of different covariates on the exposures of mutational processes. We apply these models to test the impact of replication strand on these processes and compare them to strand-oblivious models across a range of data sets. Our models capture replication strand specificity, point to signatures affected by it, and score better on held-out data compared to standard models that do not account for mutation-level covariate information.

A Biterm Topic Model for Sparse Mutation Data.

Mutational signature analysis promises to reveal the processes that shape cancer genomes for applications in diagnosis and therapy. However, most current methods are geared toward rich mutation data that has been extracted from whole-genome or whole-exome sequencing. Methods that process sparse mutation data typically found in practice are only in the earliest stages of development. In particular, we previously developed the Mix model that clusters samples to handle data sparsity. However, the Mix model had two hyper-parameters, including the number of signatures and the number of clusters, that were very costly to learn. Therefore, we devised a new method that was several orders-of-magnitude more efficient for handling sparse data, was based on mutation co-occurrences, and imitated word co-occurrence analyses of Twitter texts. We showed that the model produced significantly improved hyper-parameter estimates that led to higher likelihoods of discovering overlooked data and had better correspondence with known signatures.

Predicting genetic interactions, cell line dependencies and drug sensitivities with variational graph auto-encoder.

Large scale cancer genomics data provide crucial information about the disease and reveal points of intervention. However, systematic data have been collected in specific cell lines and their collection is laborious and costly. Hence, there is a need to develop computational models that can predict such data for any genomic context of interest. Here we develop novel models that build on variational graph auto-encoders and can integrate diverse types of data to provide high quality predictions of genetic interactions, cell line dependencies and drug sensitivities, outperforming previous methods. Our models, data and implementation are available at: https://github.com/aijag/drugGraphNet.

A comparative analysis of telomere length maintenance circuits in fission and budding yeast.

The natural ends of the linear eukaryotic chromosomes are protected by telomeres, which also play an important role in aging and cancer development. Telomere length varies between species, but it is strictly controlled in all organisms. The process of Telomere Length Maintenance (TLM) involves many pathways, protein complexes and interactions that were first discovered in budding and fission yeast model organisms (Saccharomyces cerevisiae, Schizosaccharomyces pombe). In particular, large-scale systematic genetic screens in budding yeast uncovered a network of ≈ 500 genes that, when mutated, cause telomeres to lengthen or to shorten. In contrast, the TLM network in fission yeast remains largely unknown and systematic data is still lacking. In this work we try to close this gap and develop a unified interpretable machine learning framework for TLM gene discovery and phenotype prediction in both species. We demonstrate the utility of our framework in pinpointing the pathways by which TLM homeostasis is maintained and predicting novel TLM genes in fission yeast. The results of this study could be used for better understanding of telomere biology and serve as a step towards the adaptation of computational methods based on telomeric data for human prognosis.

Exosomal telomerase transcripts reprogram the microRNA transcriptome profile of fibroblasts and partially contribute to CAF formation.

It is now well accepted that cancer cells change their microenvironment from normal to tumor-supportive state to provide sustained tumor growth, metastasis and drug resistance. These processes are partially carried out by exosomes, nano-sized vesicles secreted from cells, shuttled from donor to recipient cells containing a cargo of nucleic acids, proteins and lipids. By transferring biologically active molecules, cancer-derived exosomes may transform microenvironmental cells to become tumor supportive. Telomerase activity is regarded as a hallmark of cancer. We have recently shown that the transcript of human telomerase reverse transcriptase (hTERT), is packaged in cancer cells derived- exosomes. Following the engulfment of the hTERT transcript into fibroblasts, it is translated into a fully active enzyme [after assembly with its RNA component (hTERC) subunit]. Telomerase activity in the recipient, otherwise telomerase negative cells, provides them with a survival advantage. Here we show that exosomal telomerase might play a role in modifying normal fibroblasts into cancer associated fibroblasts (CAFs) by upregulating [Formula: see text]SMA and Vimentin, two CAF markers. We also show that telomerase activity changes the transcriptome of microRNA in these fibroblasts. By ectopically expressing microRNA 342, one of the top identified microRNAs, we show that it may mediate the proliferative phenotype that these cells acquire upon taking-up exosomal hTERT, providing them with a survival advantage.

Multi-task learning for predicting SARS-CoV-2 antibody escape.

The coronavirus pandemic has revolutionized our world, with vaccination proving to be a key tool in fighting the disease. However, a major threat to this line of attack are variants that can evade the vaccine. Thus, a fundamental problem of growing importance is the identification of mutations of concern with high escape probability. In this paper we develop a computational framework that harnesses systematic mutation screens in the receptor binding domain of the viral Spike protein for escape prediction. The framework analyzes data on escape from multiple antibodies simultaneously, creating a latent representation of mutations that is shown to be effective in predicting escape and binding properties of the virus. We use this representation to validate the escape potential of current SARS-CoV-2 variants.

Gene architecture directs splicing outcome in separate nuclear spatial regions.

How the splicing machinery defines exons or introns as the spliced unit has remained a puzzle for 30 years. Here, we demonstrate that peripheral and central regions of the nucleus harbor genes with two distinct exon-intron GC content architectures that differ in the splicing outcome. Genes with low GC content exons, flanked by long introns with lower GC content, are localized in the periphery, and the exons are defined as the spliced unit. Alternative splicing of these genes results in exon skipping. In contrast, the nuclear center contains genes with a high GC content in the exons and short flanking introns. Most splicing of these genes occurs via intron definition, and aberrant splicing leads to intron retention. We demonstrate that the nuclear periphery and center generate different environments for the regulation of alternative splicing and that two sets of splicing factors form discrete regulatory subnetworks for the two gene architectures. Our study connects 3D genome organization and splicing, thus demonstrating that exon and intron definition modes of splicing occur in different nuclear regions.

ScalpelSig Designs Targeted Genomic Panels from Data to Detect Activity of Mutational Signatures.

Over the past decade, a promising line of cancer research has utilized machine learning to mine statistical patterns of mutations in cancer genomes for information. Recent work shows that these statistical patterns, commonly referred to as "mutational signatures," have diverse therapeutic potential as biomarkers for cancer therapies. However, translating this potential into reality is hindered by limited access to sequencing in the clinic. Almost all methods for mutational signature analysis (MSA) rely on whole genome or whole exome sequencing data, while sequencing in the clinic is typically limited to small gene panels. To improve clinical access to MSA, we considered the question of whether targeted panels could be designed for the purpose of mutational signature detection. Here we present ScalpelSig, to our knowledge the first algorithm that automatically designs genomic panels optimized for detection of a given mutational signature. The algorithm learns from data to identify genome regions that are particularly indicative of signature activity. Using a cohort of breast cancer genomes as training data, we show that ScalpelSig panels substantially improve accuracy of signature detection compared to baselines. We find that some ScalpelSig panels even approach the performance of whole exome sequencing, which observes over 10 × as much genomic material. We test our algorithm under a variety of conditions, showing that its performance generalizes to another dataset of breast cancers, to smaller panel sizes, and to lesser amounts of training data.

Deep Unfolding for Non-Negative Matrix Factorization with Application to Mutational Signature Analysis.

Non-negative matrix factorization (NMF) is a fundamental matrix decomposition technique that is used primarily for dimensionality reduction and is increasing in popularity in the biological domain. Although finding a unique NMF is generally not possible, there are various iterative algorithms for NMF optimization that converge to locally optimal solutions. Such techniques can also serve as a starting point for deep learning methods that unroll the algorithmic iterations into layers of a deep network. In this study, we develop unfolded deep networks for NMF and several regularized variants in both a supervised and an unsupervised setting. We apply our method to various mutation data sets to reconstruct their underlying mutational signatures and their exposures. We demonstrate the increased accuracy of our approach over standard formulations in analyzing simulated and real mutation data.

A mixture model for signature discovery from sparse mutation data.

Mutational signatures are key to understanding the processes that shape cancer genomes, yet their analysis requires relatively rich whole-genome or whole-exome mutation data. Recently, orders-of-magnitude sparser gene-panel-sequencing data have become increasingly available in the clinic. To deal with such sparse data, we suggest a novel mixture model, Mix. In application to simulated and real gene-panel sequences, Mix is shown to outperform current approaches and yield mutational signatures and patient stratifications that are in higher agreement with the literature. We further demonstrate its utility in several clinical settings, successfully predicting therapy benefit and patient groupings from MSK-IMPACT pan-cancer data. Availability: https://github.com/itaysason/Mix-MMM .

Effect of SARS-CoV-2 proteins on vascular permeability.

Severe acute respiratory syndrome (SARS)-CoV-2 infection leads to severe disease associated with cytokine storm, vascular dysfunction, coagulation, and progressive lung damage. It affects several vital organs, seemingly through a pathological effect on endothelial cells. The SARS-CoV-2 genome encodes 29 proteins, whose contribution to the disease manifestations, and especially endothelial complications, is unknown. We cloned and expressed 26 of these proteins in human cells and characterized the endothelial response to overexpression of each, individually. Whereas most proteins induced significant changes in endothelial permeability, nsp2, nsp5_c145a (catalytic dead mutant of nsp5), and nsp7 also reduced CD31, and increased von Willebrand factor expression and IL-6, suggesting endothelial dysfunction. Using propagation-based analysis of a protein-protein interaction (PPI) network, we predicted the endothelial proteins affected by the viral proteins that potentially mediate these effects. We further applied our PPI model to identify the role of each SARS-CoV-2 protein in other tissues affected by coronavirus disease (COVID-19). While validating the PPI network model, we found that the tight junction (TJ) proteins cadherin-5, ZO-1, and ß-catenin are affected by nsp2, nsp5_c145a, and nsp7 consistent with the model prediction. Overall, this work identifies the SARS-CoV-2 proteins that might be most detrimental in terms of endothelial dysfunction, thereby shedding light on vascular aspects of COVID-19.

A data-driven approach for constructing mutation categories for mutational signature analysis.

Mutational processes shape the genomes of cancer patients and their understanding has important applications in diagnosis and treatment. Current modeling of mutational processes by identifying their characteristic signatures views each base substitution in a limited context of a single flanking base on each side. This context definition gives rise to 96 categories of mutations that have become the standard in the field, even though wider contexts have been shown to be informative in specific cases. Here we propose a data-driven approach for constructing a mutation categorization for mutational signature analysis. Our approach is based on the assumption that tumor cells that are exposed to similar mutational processes, show similar expression levels of DNA damage repair genes that are involved in these processes. We attempt to find a categorization that maximizes the agreement between mutation and gene expression data, and show that it outperforms the standard categorization over multiple quality measures. Moreover, we show that the categorization we identify generalizes to unseen data from different cancer types, suggesting that mutation context patterns extend beyond the immediate flanking bases.