Introduction: Heterogeneity has been found to be a hallmark of cancer. Patient response to treatment, disease progression, and cells’ gene expression profiles are just a few of these variable characteristics. Identifying patterns in this heterogeneity is crucial to understanding cancer progression and developing better treatments. Additionally, kinase dysfunction is a well-documented phenomenon present in cancer. Examining kinase activity profiles has already provided some insight into cancerous cells’ heterogeneity and has the potential to provide much information.
Advances in phosphoproteomic techniques allow for an improved ability to directly measure phosphorylated products. Various algorithms exist that perform kinase activity quantification. Kinase Substrate Transfer to Activity Relationships (KSTAR) has proven to be particularly adept at quantifying kinase activity, especially in the tyrosine space (Crowl et al., 2022). The algorithm probabilistically selects edges from a weighted substrate-kinase network. Then, under the assumption that more active kinases will produce more phosphorylated products, it calculates the enrichment of a specific kinase’s substrates found in a phosphoproteomic experiment to determine that kinase’s activity score. In this work, we seek to extend KSTAR’s ability to quantify a kinases activity by adjusting the underlying substrate-kinase network. First, we will expand the coverage of the kinome for which KSTAR can produce activity scores. Second, we are investigating ways in which to address the statistical saturation that we observe when using KSTAR to measure the activity of kinases known to be highly active in cancer cells.
Materials and
Methods: KSTAR uses an underlying substrate-kinase network generated by NetworKIN (Linding et al., 2007). In order to increase the kinome coverage of KSTAR, we are growing the underlying substrate-kinase network by incorporating networks generated by other algorithms including Group-based Prediction System (GPS) (Wang et al. 2020) and Phosphorylation in a Protein Interaction Context for Kinases (PhosphoPICK) (Patrick et al., 2015). Additionally, we are testing using a substrate-kinase network made from position-specific scoring matrices (PSSMs) published by Johnson et al., (2023).
Benchmarking is done with various datasets from phosphoproteomic experiments measuring samples from cancer cell lines, tumor biopsies, and xenograft studies. There are three main types of experiments to test KSTAR’s performance: inhibition, stimulation, and clustering. In inhibition and stimulation experiments, kinase activity is increased/decrease using known inhibitors/activators. KSTAR’s ability to detect this expected change is tested in these experiments. In clustering experiments, the assumption is made that cells from the cell line should cluster together, regardless of lab in which they were cultured or where the phosphoproteomic experiment was run. We generate kinase activity profiles using KSTAR and determine whether in fact cells from the same line cluster together.
Results, Conclusions, and Discussions: The advantage of incorporating different networks into KSTAR is that they cover different portions of the kinome, but still have some overlap, which allows for comparing the networks, and presents the opportunity of using an ensemble approach to predicting activity of certain kinases. However, combining these networks is not entirely straightforward. For example, different networks use different metrics as edge weights, e.g. false discovery rate or p-values for relationships as measured by different statistical methods. We are currently investigating the relationship between different kinase-substrate networks to determine how best to combine them.
First, we investigated pairwise correlation in common kinase-substrate edges between networks generated by 3 algorithms: NetworKIN, PhosphoPICK and GPS. Preliminary results show that there is a moderate positive correlation between the raw edge weights from different networks for some of the common kinases (Fig. 1). This suggests there is some agreement between the networks, though not a general global agreement across all common kinases. Additionally, running KSTAR on various datasets using different underlying networks yields kinase activity scores with differing levels of agreement for different kinases (Fig. 2). Work in progress includes leveraging these differences to make stronger predictions with KSTAR for common kinases as well as using the non-common kinases to expand the coverage of the kinome.
