Accurate prediction of kinase-substrate networks using knowledge graphs

Network-based elucidation of colon cancer drug resistance mechanisms by phosphoproteomic time-series analysis

Aberrant signaling pathway activity is a hallmark of tumorigenesis and progression, which has guided targeted inhibitor design for over 30 years. Yet, adaptive resistance mechanisms, induced by rapid, context-specific signaling network rewiring, continue to challenge therapeutic efficacy. Leveraging progress in proteomic technologies and network-based methodologies, we introduce Virtual Enrichment-based Signaling Protein-activity Analysis (VESPA)-an algorithm designed to elucidate mechanisms of cell response and adaptation to drug perturbations-and use it to analyze 7-point phosphoproteomic time series from colorectal cancer cells treated with clinically-relevant inhibitors and control media. Interrogating tumor-specific enzyme/substrate interactions accurately infers kinase and phosphatase activity, based on their substrate phosphorylation state, effectively accounting for signal crosstalk and sparse phosphoproteome coverage. The analysis elucidates time-dependent signaling pathway response to each drug perturbation and, more importantly, cell adaptive response and rewiring, experimentally confirmed by CRISPR knock-out assays, suggesting broad applicability to cancer and other diseases.

Predicting protein and pathway associations for understudied dark kinases using pattern-constrained knowledge graph embedding

The 534 protein kinases encoded in the human genome constitute a large druggable class of proteins that include both well-studied and understudied "dark" members. Accurate prediction of dark kinase functions is a major bioinformatics challenge. Here, we employ a graph mining approach that uses the evolutionary and functional context encoded in knowledge graphs (KGs) to predict protein and pathway associations for understudied kinases. We propose a new scalable graph embedding approach, RegPattern2Vec, which employs regular pattern constrained random walks to sample diverse aspects of node context within a KG flexibly. RegPattern2Vec learns functional representations of kinases, interacting partners, post-translational modifications, pathways, cellular localization, and chemical interactions from a kinase-centric KG that integrates and conceptualizes data from curated heterogeneous data resources. By contextualizing information relevant to prediction, RegPattern2Vec improves accuracy and efficiency in comparison to other random walk-based graph embedding approaches. We show that the predictions produced by our model overlap with pathway enrichment data produced using experimentally validated Protein-Protein Interaction (PPI) data from both publicly available databases and experimental datasets not used in training. Our model also has the advantage of using the collected random walks as biological context to interpret the predicted protein-pathway associations. We provide high-confidence pathway predictions for 34 dark kinases and present three case studies in which analysis of meta-paths associated with the prediction enables biological interpretation. Overall, RegPattern2Vec efficiently samples multiple node types for link prediction on biological knowledge graphs and the predicted associations between understudied kinases, pseudokinases, and known pathways serve as a conceptual starting point for hypothesis generation and testing.

KSFinder-a knowledge graph model for link prediction of novel phosphorylated substrates of kinases

Background

Aberrant protein kinase regulation leading to abnormal substrate phosphorylation is associated with several human diseases. Despite the promise of therapies targeting kinases, many human kinases remain understudied. Most existing computational tools predicting phosphorylation cover less than 50% of known human kinases. They utilize local feature selection based on protein sequences, motifs, domains, structures, and/or functions, and do not consider the heterogeneous relationships of the proteins. In this work, we present KSFinder, a tool that predicts kinase-substrate links by capturing the inherent association of proteins in a network comprising 85% of the known human kinases. We also postulate the potential role of two understudied kinases based on their substrate predictions from KSFinder.

Methods

KSFinder learns the semantic relationships in a phosphoproteome knowledge graph using a knowledge graph embedding algorithm and represents the nodes in low-dimensional vectors. A multilayer perceptron (MLP) classifier is trained to discern kinase-substrate links using the embedded vectors. KSFinder uses a strategic negative generation approach that eliminates biases in entity representation and combines data from experimentally validated non-interacting protein pairs, proteins from different subcellular locations, and random sampling. We assess KSFinder's generalization capability on four different datasets and compare its performance with other state-of-the-art prediction models. We employ KSFinder to predict substrates of 68 "dark" kinases considered understudied by the Illuminating the Druggable Genome program and use our text-mining tool, RLIMS-P along with manual curation, to search for literature evidence for the predictions. In a case study, we performed functional enrichment analysis for two dark kinases - HIPK3 and CAMKK1 using their predicted substrates.

Results

KSFinder shows improved performance over other kinase-substrate prediction models and generalized prediction ability on different datasets. We identified literature evidence for 17 novel predictions involving an understudied kinase. All of these 17 predictions had a probability score ≥0.7 (nine at >0.9, six at 0.8-0.9, and two at 0.7-0.8). The evaluation of 93,593 negative predictions (probability ≤0.3) identified four false negatives. The top enriched biological processes of HIPK3 substrates relate to the regulation of extracellular matrix and epigenetic gene expression, while CAMKK1 substrates include lipid storage regulation and glucose homeostasis.

Conclusions

KSFinder outperforms the current kinase-substrate prediction tools with higher kinase coverage. The strategically developed negatives provide a superior generalization ability for KSFinder. We predicted substrates of 432 kinases, 68 of which are understudied, and hypothesized the potential functions of two dark kinases using their predicted substrates.

Network-based elucidation of colon cancer drug resistance by phosphoproteomic time-series analysis

Aberrant signaling pathway activity is a hallmark of tumorigenesis and progression, which has guided targeted inhibitor design for over 30 years. Yet, adaptive resistance mechanisms, induced by rapid, context-specific signaling network rewiring, continue to challenge therapeutic efficacy. By leveraging progress in proteomic technologies and network-based methodologies, over the past decade, we developed VESPA-an algorithm designed to elucidate mechanisms of cell response and adaptation to drug perturbations-and used it to analyze 7-point phosphoproteomic time series from colorectal cancer cells treated with clinically-relevant inhibitors and control media. Interrogation of tumor-specific enzyme/substrate interactions accurately inferred kinase and phosphatase activity, based on their inferred substrate phosphorylation state, effectively accounting for signal cross-talk and sparse phosphoproteome coverage. The analysis elucidated time-dependent signaling pathway response to each drug perturbation and, more importantly, cell adaptive response and rewiring that was experimentally confirmed by CRISPRko assays, suggesting broad applicability to cancer and other diseases.

Prediction of Kinase-Substrate Associations Using The Functional Landscape of Kinases and Phosphorylation Sites

Protein phosphorylation is a key post-translational modification that plays a central role in many cellular processes. With recent advances in biotechnology, thousands of phosphorylated sites can be identified and quantified in a given sample, enabling proteome-wide screening of cellular signaling. However, for most (> 90%) of the phosphorylation sites that are identified in these experiments, the kinase(s) that target these sites are unknown. To broadly utilize available structural, functional, evolutionary, and contextual information in predicting kinase-substrate associations (KSAs), we develop a network-based machine learning framework. Our framework integrates a multitude of data sources to characterize the landscape of functional relationships and associations among phosphosites and kinases. To construct a phosphosite-phosphosite association network, we use sequence similarity, shared biological pathways, co-evolution, co-occurrence, and co-phosphorylation of phosphosites across different biological states. To construct a kinase-kinase association network, we integrate protein-protein interactions, shared biological pathways, and membership in common kinase families. We use node embeddings computed from these heterogeneous networks to train machine learning models for predicting kinase-substrate associations. Our systematic computational experiments using the PhosphositePLUS database shows that the resulting algorithm, NetKSA, outperforms two state-of-the-art algorithms, including KinomeXplorer and LinkPhinder, in overall KSA prediction. By stratifying the ranking of kinases, NetKSA also enables annotation of phosphosites that are targeted by relatively less-studied kinases.Availability: The code and data are available at compbio.case.edu/NetKSA/.

Phosphoproteomics data-driven signalling network inference: Does it work?

The advent of global phosphoproteome profiling has led to wide phosphosite coverage and therefore the opportunity to predict kinase-substrate associations from these datasets. However, the regulatory kinase is unknown for most substrates, due to biased and incomplete database annotations. In this study we compare the performance of six pairwise measures to predict kinase-substrate associations using a data driven approach on publicly available time resolved and perturbation mass spectrometry-based phosphoproteome data. First, we validated the performance of these measures using as a reference both a literature-based phosphosite-specific protein interaction network and a predicted kinase-substrate (KS) interactions set. The overall performance in predicting kinase-substrate associations using pairwise measures across both these reference sets was poor. To expand into the wider interactome space, we applied the approach on a network comprising pairs of substrates regulated by the same kinase (substrate-substrate associations) but found the performance to be equally poor. However, the addition of a sequence similarity filter for substrate-substrate associations led to a significant boost in performance. Our findings imply that the use of a filter to reduce the search space, such as a sequence similarity filter, can be used prior to the application of network inference methods to reduce noise and boost the signal. We also find that the current gold standard for reference sets is not adequate for evaluation as it is limited and context-agnostic. Therefore, there is a need for additional evaluation methods that have increased coverage and take into consideration the context-specific nature of kinase-substrate associations.

A knowledge graph representation learning approach to predict novel kinase-substrate interactions

The human proteome contains a vast network of interacting kinases and substrates. Even though some kinases have proven to be immensely useful as therapeutic targets, a majority are still understudied. In this work, we present a novel knowledge graph representation learning approach to predict novel interaction partners for understudied kinases. Our approach uses a phosphoproteomic knowledge graph constructed by integrating data from iPTMnet, protein ontology, gene ontology and BioKG. The representations of kinases and substrates in this knowledge graph are learned by performing directed random walks on triples coupled with a modified SkipGram or CBOW model. These representations are then used as an input to a supervised classification model to predict novel interactions for understudied kinases. We also present a post-predictive analysis of the predicted interactions and an ablation study of the phosphoproteomic knowledge graph to gain an insight into the biology of the understudied kinases.

Proteasomal down-regulation of the proapoptotic MST2 pathway contributes to BRAF inhibitor resistance in melanoma

The loss of MST2 pathway protein expression in BRAF inhibitor resistant melanoma cells is due to ubiquitination and subsequent proteasomal degradation and prevents MST2-mediated apoptosis.

The RAS-RAF-MEK-ERK pathway is hyperactivated in most malignant melanomas, and mutations in BRAF or NRAS account for most of these cases. BRAF inhibitors (BRAFi) are highly efficient for treating patients with BRAF^V600E mutations, but tumours frequently acquire resistance within a few months. Multiple resistance mechanisms have been identified, due to mutations or network adaptations that revive ERK signalling. We have previously shown that RAF proteins inhibit the MST2 proapoptotic pathway in a kinase-independent fashion. Here, we have investigated the role of the MST2 pathway in mediating resistance to BRAFi. We show that the BRAF^V600E mutant protein, but not the wild-type BRAF protein, binds to MST2 inhibiting its proapoptotic signalling. Down-regulation of MST2 reduces BRAFi-induced apoptosis. In BRAFi-resistant cell lines, MST2 pathway proteins are down-regulated by ubiquitination and subsequent proteasomal degradation rendering cells refractory to MST2 pathway–induced apoptosis. Restoration of apoptosis can be achieved by increasing MST2 pathway protein expression using proteasome inhibitors. In summary, we show that the MST2 pathway plays a role in the acquisition of BRAFi resistance in melanoma.

Interaction of LATS1 with SMAC links the MST2/Hippo pathway with apoptosis in an IAP-dependent manner

Metastatic malignant melanoma is the deadliest skin cancer, and it is characterised by its high resistance to apoptosis. The main melanoma driving mutations are part of ERK pathway, with BRAF mutations being the most frequent ones, followed by NRAS, NF1 and MEK mutations. Increasing evidence shows that the MST2/Hippo pathway is also deregulated in melanoma. While mutations are rare, MST2/Hippo pathway core proteins expression levels are often dysregulated in melanoma. The expression of the tumour suppressor RASSF1A, a bona fide activator of the MST2 pathway, is silenced by promoter methylation in over half of melanomas and correlates with poor prognosis. Here, using mass spectrometry-based interaction proteomics we identified the Second Mitochondria-derived Activator of Caspases (SMAC) as a novel LATS1 interactor. We show that RASSF1A-dependent activation of the MST2 pathway promotes LATS1-SMAC interaction and negatively regulates the antiapoptotic signal mediated by the members of the IAP family. Moreover, proteomic experiments identified a common cluster of apoptotic regulators that bind to SMAC and LATS1. Mechanistic analysis shows that the LATS1-SMAC complex promotes XIAP ubiquitination and its subsequent degradation which ultimately results in apoptosis. Importantly, we show that the oncogenic BRAFV600E mutant prevents the proapoptotic signal mediated by the LATS1-SMAC complex while treatment of melanoma cell lines with BRAF inhibitors promotes the formation of this complex, indicating that inhibition of the LATS1-SMAC might be necessary for BRAFV600E-driven melanoma. Finally, we show that LATS1-SMAC interaction is regulated by the SMAC mimetic Birinapant, which requires C-IAP1 inhibition and the degradation of XIAP, suggesting that the MST2 pathway is part of the mechanism of action of Birinapant. Overall, the current work shows that SMAC-dependent apoptosis is regulated by the LATS1 tumour suppressor and supports the idea that LATS1 is a signalling hub that regulates the crosstalk between the MST2 pathway, the apoptotic network and the ERK pathway.

In Vitro Kinase-to-Phosphosite Database (iKiP-DB) Predicts Kinase Activity in Phosphoproteomic Datasets

Phosphoproteomics routinely quantifies changes in the levels of thousands of phosphorylation sites, but functional analysis of such data remains a major challenge. While databases like PhosphoSitePlus contain information about many phosphorylation sites, the vast majority of known sites is not assigned to any protein kinase. Assigning changes in the phosphoproteome to the activity of individual kinases therefore remains a key challenge. A recent large-scale study systematically identified in vitro substrates for most human protein kinases. Here, we reprocessed and filtered these data to generate an in vitro Kinase-to-Phosphosite database (iKiP-DB). We show that iKiP-DB can accurately predict changes in kinase activity in published phosphoproteomic data sets for both well-studied and poorly characterized kinases. We apply iKiP-DB to a newly generated phosphoproteomic analysis of SARS-CoV-2 infected human lung epithelial cells and provide evidence for coronavirus-induced changes in host cell kinase activity. In summary, we show that iKiP-DB is widely applicable to facilitate the functional analysis of phosphoproteomic data sets.

Accurate, high-coverage assignment of in vivo protein kinases to phosphosites from in vitro phosphoproteomic specificity data

Phosphoproteomic experiments routinely observe thousands of phosphorylation sites. To understand the intracellular signaling processes that generated this data, one or more causal protein kinases must be assigned to each phosphosite. However, limited knowledge of kinase specificity typically restricts assignments to a small subset of a kinome. Starting from a statistical model of a high-throughput, in vitro kinase-substrate assay, I have developed an approach to high-coverage, multi-label kinase-substrate assignment called IV-KAPhE (“In vivo-Kinase Assignment for Phosphorylation Evidence”). Tested on human data, IV-KAPhE outperforms other methods of similar scope. Such computational methods generally predict a densely connected kinase-substrate network, with most sites targeted by multiple kinases, pointing either to unaccounted-for biochemical constraints or significant cross-talk and signaling redundancy. I show that such predictions can potentially identify biased kinase-site misannotations within families of closely related kinase isozymes and they provide a robust basis for kinase activity analysis.

Proteins can pass around information inside cells about changes in the environment. This process, called intracellular signaling, helps to trigger appropriate cellular responses to environmental changes. One of the main ways information is passed to proteins is through chemical “tagging,” called phosphorylation, by enzymes called protein kinases. We can measure the phosphorylation state of practically all proteins in a cell at any moment. Starting from known cases of phosphorylation by a kinase, many computational methods have been developed to predict if the kinase might tag a certain spot on another protein or if an observed tag was attached by the kinase, with different models for each kinase. I have developed a new method that instead uses a single model to assign one or more kinases to each observed tag, built from the latest large-scale experimental data. This change in focus and unbiased training data allows my method to be significantly more accurate than past methods. I also explored useful applications for my method. For example, I used it to show that much of our knowledge about which kinase is responsible for each tag is probably inaccurately biased towards the commonly studied ones.

Systematic identification of ALK substrates by integrated phosphoproteome and interactome analysis

Integrated analysis of the phosphoproteome and interactome of anaplastic lymphoma kinase (ALK)-overexpressing HEK 293 cells revealed 37 ALK substrate candidates, contributing to the improvement of kinase activity prediction.

The sensitivity of phosphorylation site identification by mass spectrometry has improved markedly. However, the lack of kinase–substrate relationship (KSR) data hinders the improvement of the range and accuracy of kinase activity prediction. In this study, we aimed to develop a method for acquiring systematic KSR data on anaplastic lymphoma kinase (ALK) using mass spectrometry and to apply this method to the prediction of kinase activity. Thirty-seven ALK substrate candidates, including 34 phosphorylation sites not annotated in the PhosphoSitePlus database, were identified by integrated analysis of the phosphoproteome and crosslinking interactome of HEK 293 cells with doxycycline-induced ALK overexpression. Furthermore, KSRs of ALK were validated by an in vitro kinase assay. Finally, using phosphoproteomic data from ALK mutant cell lines and patient-derived cells treated with ALK inhibitors, we found that the prediction of ALK activity was improved when the KSRs identified in this study were used instead of the public KSR dataset. Our approach is applicable to other kinases, and future identification of KSRs will facilitate more accurate estimations of kinase activity and elucidation of phosphorylation signals.

Learning in Medicine: The Importance of Statistical Thinking

In many fields, including medicine and biology, there has been in the last years an increasing diffusion and availability of complex data from different sources. Examples include biological experiments or data from health care providers. These data encompass information that can potentially enhance therapeutic advancement and constitute the core of modern system medicine. When analyzing these complex data, it is important to appropriately quantify uncertainty, avoiding using only algorithmic and automated approaches, which are not always appropriate. Improper application of algorithmic approaches, which ignore domain knowledge, could result in filling the literature with imprecise and/or misleading conclusions. In this chapter, we highlight the importance of statistical thinking when leveraging complex data and models to enhance science progress. In particular, we discuss the reproducibility and replicability issues, the importance of uncertainty quantification, and some common pitfalls in the analysis of big data.