Effects of protein interaction data integration, representation and reliability on the use of network properties for drug target prediction

Signature reversion of three disease-associated gene signatures prioritizes cancer drug repurposing candidates

Drug repurposing is promising because approving a drug for a new indication requires fewer resources than approving a new drug. Signature reversion detects drug perturbations most inversely related to the disease-associated gene signature to identify drugs that may reverse that signature. We assessed the performance and biological relevance of three approaches for constructing disease-associated gene signatures (i.e., limma, DESeq2, and MultiPLIER) and prioritized the resulting drug repurposing candidates for four low-survival human cancers. Our results were enriched for candidates that had been used in clinical trials or performed well in the PRISM drug screen. Additionally, we found that pamidronate and nimodipine, drugs predicted to be efficacious against the brain tumor glioblastoma (GBM), inhibited the growth of a GBM cell line and cells isolated from a patient-derived xenograft (PDX). Our results demonstrate that by applying multiple disease-associated gene signature methods, we prioritized several drug repurposing candidates for low-survival cancers.

Exploration of Target Spaces in the Human Genome for Protein and Peptide Drugs

After decades of development, protein and peptide drugs have now grown into a major drug class in the marketplace. Target identification and validation are crucial for the discovery of protein and peptide drugs, and bioinformatics prediction of targets based on the characteristics of known target proteins will help improve the efficiency and success rate of target selection. However, owing to the developmental history in the pharmaceutical industry, previous systematic exploration of the target spaces has mainly focused on traditional small-molecule drugs, while studies related to protein and peptide drugs are lacking. Here, we systematically explore the target spaces in the human genome specifically for protein and peptide drugs. Compared with other proteins, both successful protein and peptide drug targets have many special characteristics, and are also significantly different from those of small-molecule drugs in many aspects. Based on these features, we develop separate effective genome-wide target prediction models for protein and peptide drugs. Finally, a user-friendly web server, Predictor Of Protein and PeptIde drugs' therapeutic Targets (POPPIT) (http://poppit.ncpsb.org.cn/), is established, which provides not only target prediction specifically for protein and peptide drugs but also abundant annotations for predicted targets.

Centrality of drug targets in protein networks

Background

In the pharmaceutical industry, competing for few validated drug targets there is a drive to identify new ways of therapeutic intervention. Here, we attempted to define guidelines to evaluate a target’s ‘fitness’ based on its node characteristics within annotated protein functional networks to complement contingent therapeutic hypotheses.

Results

We observed that targets of approved, selective small molecule drugs exhibit high node centrality within protein networks relative to a broader set of investigational targets spanning various development stages. Targets of approved drugs also exhibit higher centrality than other proteins within their respective functional class. These findings expand on previous reports of drug targets’ network centrality by suggesting some centrality metrics such as low topological coefficient as inherent characteristics of a ‘good’ target, relative to other exploratory targets and regardless of its functional class. These centrality metrics could thus be indicators of an individual protein’s ‘fitness’ as potential drug target. Correlations between protein nodes’ network centrality and number of associated publications underscored the possibility of knowledge bias as an inherent limitation to such predictions.

Conclusions

Despite some entanglement with knowledge bias, like structure-oriented ‘druggability’ assessments of new protein targets, centrality metrics could assist early pharmaceutical discovery teams in evaluating potential targets with limited experimental proof of concept and help allocate resources for an effective drug discovery pipeline.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12859-021-04342-x.

DecoPath: a web application for decoding pathway enrichment analysis

The past decades have brought a steady growth of pathway databases and enrichment methods. However, the advent of pathway data has not been accompanied by an improvement in interoperability across databases, hampering the use of pathway knowledge from multiple databases for enrichment analysis. While integrative databases have attempted to address this issue, they often do not account for redundant information across resources. Furthermore, the majority of studies that employ pathway enrichment analysis still rely upon a single database or enrichment method, though the use of another could yield differing results. These shortcomings call for approaches that investigate the differences and agreements across databases and methods as their selection in the design of a pathway analysis can be a crucial step in ensuring the results of such an analysis are meaningful. Here we present DecoPath, a web application to assist in the interpretation of the results of pathway enrichment analysis. DecoPath provides an ecosystem to run enrichment analysis or directly upload results and facilitate the interpretation of results with custom visualizations that highlight the consensus and/or discrepancies at the pathway- and gene-levels. DecoPath is available at https://decopath.scai.fraunhofer.de, and its source code and documentation can be found on GitHub at https://github.com/DecoPath/DecoPath.

Topological network measures for drug repositioning

Drug repositioning has received increased attention since the past decade as several blockbuster drugs have come out of repositioning. Computational approaches are significantly contributing to these efforts, of which, network-based methods play a key role. Various structural (topological) network measures have thereby contributed to uncovering unintuitive functional relationships and repositioning candidates in drug-disease and other networks. This review gives a broad overview of the topic, and offers perspectives on the application of topological measures for network analysis. It also discusses unexplored measures, and draws attention to a wider scope of application efforts, especially in drug repositioning.

A deep learning model based on sparse auto-encoder for prioritizing cancer-related genes and drug target combinations

Prioritization of cancer-related genes from gene expression profiles and proteomic data is vital to improve the targeted therapies research. Although computational approaches have been complementing high-throughput biological experiments on the understanding of human diseases, it still remains a big challenge to accurately discover cancer-related proteins/genes via automatic learning from large-scale protein/gene expression data and protein-protein interaction data. Most of the existing methods are based on network construction combined with gene expression profiles, which ignore the diversity between normal samples and disease cell lines. In this study, we introduced a deep learning model based on a sparse auto-encoder to learn the specific characteristics of protein interactions in cancer cell lines integrated with protein expression data. The model showed learning ability to identify cancer-related proteins/genes from the input of different protein expression profiles by extracting the characteristics of protein interaction information, which could also predict cancer-related protein combinations. Comparing with other reported methods including differential expression and network-based methods, our model got the highest area under the curve value (>0.8) in predicting cancer-related genes. Our study prioritized ~500 high-confidence cancer-related genes; among these genes, 211 already known cancer drug targets were found, which supported the accuracy of our method. The above results indicated that the proposed auto-encoder model could computationally prioritize candidate proteins/genes involved in cancer and improve the targeted therapies research.

Predicting clinically promising therapeutic hypotheses using tensor factorization

BACKGROUND

Determining which target to pursue is a challenging and error-prone first step in developing a therapeutic treatment for a disease, where missteps are potentially very costly given the long-time frames and high expenses of drug development. With current informatics technology and machine learning algorithms, it is now possible to computationally discover therapeutic hypotheses by predicting clinically promising drug targets based on the evidence associating drug targets with disease indications. We have collected this evidence from Open Targets and additional databases that covers 17 sources of evidence for target-indication association and represented the data as a tensor of 21,437 × 2211 × 17.

RESULTS

As a proof-of-concept, we identified examples of successes and failures of target-indication pairs in clinical trials across 875 targets and 574 disease indications to build a gold-standard data set of 6140 known clinical outcomes. We designed and executed three benchmarking strategies to examine the performance of multiple machine learning models: Logistic Regression, LASSO, Random Forest, Tensor Factorization and Gradient Boosting Machine. With 10-fold cross-validation, tensor factorization achieved AUROC = 0.82 ± 0.02 and AUPRC = 0.71 ± 0.03. Across multiple validation schemes, this was comparable or better than other methods.

CONCLUSION

In this work, we benchmarked a machine learning technique called tensor factorization for the problem of predicting clinical outcomes of therapeutic hypotheses. Results have shown that this method can achieve equal or better prediction performance compared with a variety of baseline models. We demonstrate one application of the method to predict outcomes of trials on novel indications of approved drug targets. This work can be expanded to targets and indications that have never been clinically tested and proposing novel target-indication hypotheses. Our proposed biologically-motivated cross-validation schemes provide insight into the robustness of the prediction performance. This has significant implications for all future methods that try to address this seminal problem in drug discovery.

Systematic interrogation of diverse Omic data reveals interpretable, robust, and generalizable transcriptomic features of clinically successful therapeutic targets

Target selection is the first and pivotal step in drug discovery. An incorrect choice may not manifest itself for many years after hundreds of millions of research dollars have been spent. We collected a set of 332 targets that succeeded or failed in phase III clinical trials, and explored whether Omic features describing the target genes could predict clinical success. We obtained features from the recently published comprehensive resource: Harmonizome. Nineteen features appeared to be significantly correlated with phase III clinical trial outcomes, but only 4 passed validation schemes that used bootstrapping or modified permutation tests to assess feature robustness and generalizability while accounting for target class selection bias. We also used classifiers to perform multivariate feature selection and found that classifiers with a single feature performed as well in cross-validation as classifiers with more features (AUROC = 0.57 and AUPR = 0.81). The two predominantly selected features were mean mRNA expression across tissues and standard deviation of expression across tissues, where successful targets tended to have lower mean expression and higher expression variance than failed targets. This finding supports the conventional wisdom that it is favorable for a target to be present in the tissue(s) affected by a disease and absent from other tissues. Overall, our results suggest that it is feasible to construct a model integrating interpretable target features to inform target selection. We anticipate deeper insights and better models in the future, as researchers can reuse the data we have provided to improve methods for handling sample biases and learn more informative features. Code, documentation, and data for this study have been deposited on GitHub at https://github.com/arouillard/omic-features-successful-targets.

A novel framework for the identification of drug target proteins: Combining stacked auto-encoders with a biased support vector machine

The identification of drug target proteins (IDTP) plays a critical role in biometrics. The aim of this study was to retrieve potential drug target proteins (DTPs) from a collected protein dataset, which represents an overwhelming task of great significance. Previously reported methodologies for this task generally employ protein-protein interactive networks but neglect informative biochemical attributes. We formulated a novel framework utilizing biochemical attributes to address this problem. In the framework, a biased support vector machine (BSVM) was combined with the deep embedded representation extracted using a deep learning model, stacked auto-encoders (SAEs). In cases of non-drug target proteins (NDTPs) contaminated by DTPs, the framework is beneficial due to the efficient representation of the SAE and relief of the imbalance effect by the BSVM. The experimental results demonstrated the effectiveness of our framework, and the generalization capability was confirmed via comparisons to other models. This study is the first to exploit a deep learning model for IDTP. In summary, nearly 23% of the NDTPs were predicted as likely DTPs, which are awaiting further verification based on biomedical experiments.

In silico identification of essential proteins in Corynebacterium pseudotuberculosis based on protein-protein interaction networks

Background

Corynebacterium pseudotuberculosis (Cp) is a gram-positive bacterium that is classified into equi and ovis serovars. The serovar ovis is the etiological agent of caseous lymphadenitis, a chronic infection affecting sheep and goats, causing economic losses due to carcass condemnation and decreased production of meat, wool, and milk. Current diagnosis or treatment protocols are not fully effective and, thus, require further research of Cp pathogenesis.

Results

Here, we mapped known protein-protein interactions (PPI) from various species to nine Cp strains to reconstruct parts of the potential Cp interactome and to identify potentially essential proteins serving as putative drug targets. On average, we predict 16,669 interactions for each of the nine strains (with 15,495 interactions shared among all strains). An in silico sanity check suggests that the potential networks were not formed by spurious interactions but have a strong biological bias. With the inferred Cp networks we identify 181 essential proteins, among which 41 are non-host homologous.

Conclusions

The list of candidate interactions of the Cp strains lay the basis for developing novel hypotheses and designing according wet-lab studies. The non-host homologous essential proteins are attractive targets for therapeutic and diagnostic proposes. They allow for searching of small molecule inhibitors of binding interactions enabling modern drug discovery. Overall, the predicted Cp PPI networks form a valuable and versatile tool for researchers interested in Corynebacterium pseudotuberculosis.

Electronic supplementary material

The online version of this article (doi:10.1186/s12918-016-0346-4) contains supplementary material, which is available to authorized users.

HVint: A Strategy for Identifying Novel Protein-Protein Interactions in Herpes Simplex Virus Type 1

Human herpesviruses are widespread human pathogens with a remarkable impact on worldwide public health. Despite intense decades of research, the molecular details in many aspects of their function remain to be fully characterized. To unravel the details of how these viruses operate, a thorough understanding of the relationships between the involved components is key. Here, we present HVint, a novel protein-protein intraviral interaction resource for herpes simplex virus type 1 (HSV-1) integrating data from five external sources. To assess each interaction, we used a scoring scheme that takes into consideration aspects such as the type of detection method and the number of lines of evidence. The coverage of the initial interactome was further increased using evolutionary information, by importing interactions reported for other human herpesviruses. These latter interactions constitute, therefore, computational predictions for potential novel interactions in HSV-1. An independent experimental analysis was performed to confirm a subset of our predicted interactions. This subset covers proteins that contribute to nuclear egress and primary envelopment events, including VP26, pUL31, pUL40, and the recently characterized pUL32 and pUL21. Our findings support a coordinated crosstalk between VP26 and proteins such as pUL31, pUS9, and the CSVC complex, contributing to the development of a model describing the nuclear egress and primary envelopment pathways of newly synthesized HSV-1 capsids. The results are also consistent with recent findings on the involvement of pUL32 in capsid maturation and early tegumentation events. Further, they open the door to new hypotheses on virus-specific regulators of pUS9-dependent transport. To make this repository of interactions readily accessible for the scientific community, we also developed a user-friendly and interactive web interface. Our approach demonstrates the power of computational predictions to assist in the design of targeted experiments for the discovery of novel protein-protein interactions.

Distinctive Behaviors of Druggable Proteins in Cellular Networks

The interaction environment of a protein in a cellular network is important in defining the role that the protein plays in the system as a whole, and thus its potential suitability as a drug target. Despite the importance of the network environment, it is neglected during target selection for drug discovery. Here, we present the first systematic, comprehensive computational analysis of topological, community and graphical network parameters of the human interactome and identify discriminatory network patterns that strongly distinguish drug targets from the interactome as a whole. Importantly, we identify striking differences in the network behavior of targets of cancer drugs versus targets from other therapeutic areas and explore how they may relate to successful drug combinations to overcome acquired resistance to cancer drugs. We develop, computationally validate and provide the first public domain predictive algorithm for identifying druggable neighborhoods based on network parameters. We also make available full predictions for 13,345 proteins to aid target selection for drug discovery. All target predictions are available through canSAR.icr.ac.uk. Underlying data and tools are available at https://cansar.icr.ac.uk/cansar/publications/druggable_network_neighbourhoods/.

The need for well-validated targets for drug discovery is more pressing than ever, especially in cancer in view of resistance to current therapeutics coupled with late stage drug failures. Target prioritization and selection methodologies have typically not taken the protein interaction environment into account. Here we analyze a large representation of the human interactome comprising almost 90,000 interactions between 13,345 proteins. We assess these interactions using an extensive set of topological, graphical and community parameters, and we identify behaviors that distinguish the protein interaction environments of drug targets from the general interactome. Moreover, we identify clear distinctions between the network environment of cancer-drug targets and targets from other therapeutics areas. We use these distinguishing properties to build a predictive methodology to prioritize potential drug targets based on network parameters alone and we validate our predictive models using current FDA-approved drug targets. Our models provide an objective, interactome-based target prioritization methodology to complement existing structure-based and ligand-based prioritization methods. We provide our interactome-based predictions alongside other druggability predictors within the public canSAR resource (cansar.icr.ac.uk).

Protein-protein interaction networks identify targets which rescue the MPP+ cellular model of Parkinson's disease

Neurodegenerative diseases are complex multifactorial disorders characterised by the interplay of many dysregulated physiological processes. As an exemplar, Parkinson’s disease (PD) involves multiple perturbed cellular functions, including mitochondrial dysfunction and autophagic dysregulation in preferentially-sensitive dopamine neurons, a selective pathophysiology recapitulated in vitro using the neurotoxin MPP⁺. Here we explore a network science approach for the selection of therapeutic protein targets in the cellular MPP⁺ model. We hypothesised that analysis of protein-protein interaction networks modelling MPP⁺ toxicity could identify proteins critical for mediating MPP⁺ toxicity. Analysis of protein-protein interaction networks constructed to model the interplay of mitochondrial dysfunction and autophagic dysregulation (key aspects of MPP⁺ toxicity) enabled us to identify four proteins predicted to be key for MPP⁺ toxicity (P62, GABARAP, GBRL1 and GBRL2). Combined, but not individual, knockdown of these proteins increased cellular susceptibility to MPP⁺ toxicity. Conversely, combined, but not individual, over-expression of the network targets provided rescue of MPP⁺ toxicity associated with the formation of autophagosome-like structures. We also found that modulation of two distinct proteins in the protein-protein interaction network was necessary and sufficient to mitigate neurotoxicity. Together, these findings validate our network science approach to multi-target identification in complex neurological diseases.

The human plasma membrane peripherome: visualization and analysis of interactions

A major part of membrane function is conducted by proteins, both integral and peripheral. Peripheral membrane proteins temporarily adhere to biological membranes, either to the lipid bilayer or to integral membrane proteins with noncovalent interactions. The aim of this study was to construct and analyze the interactions of the human plasma membrane peripheral proteins (peripherome hereinafter). For this purpose, we collected a dataset of peripheral proteins of the human plasma membrane. We also collected a dataset of experimentally verified interactions for these proteins. The interaction network created from this dataset has been visualized using Cytoscape. We grouped the proteins based on their subcellular location and clustered them using the MCL algorithm in order to detect functional modules. Moreover, functional and graph theory based analyses have been performed to assess biological features of the network. Interaction data with drug molecules show that ~10% of peripheral membrane proteins are targets for approved drugs, suggesting their potential implications in disease. In conclusion, we reveal novel features and properties regarding the protein-protein interaction network created by peripheral proteins of the human plasma membrane.

An empirical study of different approaches for protein classification

Many domains would benefit from reliable and efficient systems for automatic protein classification. An area of particular interest in recent studies on automatic protein classification is the exploration of new methods for extracting features from a protein that work well for specific problems. These methods, however, are not generalizable and have proven useful in only a few domains. Our goal is to evaluate several feature extraction approaches for representing proteins by testing them across multiple datasets. Different types of protein representations are evaluated: those starting from the position specific scoring matrix of the proteins (PSSM), those derived from the amino-acid sequence, two matrix representations, and features taken from the 3D tertiary structure of the protein. We also test new variants of proteins descriptors. We develop our system experimentally by comparing and combining different descriptors taken from the protein representations. Each descriptor is used to train a separate support vector machine (SVM), and the results are combined by sum rule. Some stand-alone descriptors work well on some datasets but not on others. Through fusion, the different descriptors provide a performance that works well across all tested datasets, in some cases performing better than the state-of-the-art.

Protein-protein interaction detection: methods and analysis

Protein-protein interaction plays key role in predicting the protein function of target protein and drug ability of molecules. The majority of genes and proteins realize resulting phenotype functions as a set of interactions. The in vitro and in vivo methods like affinity purification, Y2H (yeast 2 hybrid), TAP (tandem affinity purification), and so forth have their own limitations like cost, time, and so forth, and the resultant data sets are noisy and have more false positives to annotate the function of drug molecules. Thus, in silico methods which include sequence-based approaches, structure-based approaches, chromosome proximity, gene fusion, in silico 2 hybrid, phylogenetic tree, phylogenetic profile, and gene expression-based approaches were developed. Elucidation of protein interaction networks also contributes greatly to the analysis of signal transduction pathways. Recent developments have also led to the construction of networks having all the protein-protein interactions using computational methods for signaling pathways and protein complex identification in specific diseases.

Tamoxifen integromics and personalized medicine: dynamic modular transformations underpinning response to tamoxifen in breast cancer treatment

Recent advances in pharmacogenomics technologies allow bold steps to be taken towards personalized medicine, more accurate health planning, and personalized drug development. In this framework, systems pharmacology network-based approaches offer an appealing way for integrating multi-omics data and set the basis for defining systems-level drug response biomarkers. On the road to individualized tamoxifen treatment in estrogen receptor-positive breast cancer patients, we examine the dynamics of the attendant pharmacological response mechanisms. By means of an "integromics" network approach, we assessed the tamoxifen effect through the way the high-order organization of interactome (i.e., the modules) is perturbed. To accomplish that, first we integrated the time series transcriptome data with the human protein interaction data, and second, an efficient module-detecting algorithm was applied onto the composite graphs. Our findings show that tamoxifen induces severe modular transformations on specific areas of the interactome. Our modular biomarkers in response to tamoxifen attest to the immunomodulatory role of tamoxifen, and further reveal that it deregulates cell cycle and apoptosis pathways, while coordinating the proteasome and basal transcription factors. To the best of our knowledge, this is the first report that informs the fields of personalized medicine and clinical pharmacology about the actual dynamic interactome response to tamoxifen administration.

Protein interaction networks as starting points to identify novel antimicrobial drug targets

Novel classes of antimicrobials are needed to address the challenge of multidrug-resistant bacteria. Current bacterial drug targets mainly consist of specific proteins or subsets of proteins without regard for either how these targets are integrated in cellular networks or how they may interact with host proteins. However, proteins rarely act in isolation, and the majority of biological processes are dependent on interactions with other proteins. Consequently, protein-protein interaction (PPI) networks offer a realm of unexplored potential for next-generation drug targets. In this review, we argue that the architecture of bacterial or host-pathogen protein interactomes can provide invaluable insights for the identification of novel antibacterial drug targets.