Logic-based mechanistic machine learning on high-content images reveals how drugs differentially regulate cardiac fibroblasts.

Fibroblasts are essential regulators of extracellular matrix deposition following cardiac injury. These cells exhibit highly plastic responses in phenotype during fibrosis in response to environmental stimuli. Here, we test whether and how candidate anti-fibrotic drugs differentially regulate measures of cardiac fibroblast phenotype, which may help identify treatments for cardiac fibrosis. We conducted a high-content microscopy screen of human cardiac fibroblasts treated with 13 clinically relevant drugs in the context of TGFß and/or IL-1ß, measuring phenotype across 137 single-cell features. We used the phenotypic data from our high-content imaging to train a logic-based mechanistic machine learning model (LogiMML) for fibroblast signaling. The model predicted how pirfenidone and Src inhibitor WH-4-023 reduce actin filament assembly and actin-myosin stress fiber formation, respectively. Validating the LogiMML model prediction that PI3K partially mediates the effects of Src inhibition, we found that PI3K inhibition reduces actin-myosin stress fiber formation and procollagen I production in human cardiac fibroblasts. In this study, we establish a modeling approach combining the strengths of logic-based network models and regularized regression models. We apply this approach to predict mechanisms that mediate the differential effects of drugs on fibroblasts, revealing Src inhibition acting via PI3K as a potential therapy for cardiac fibrosis.

A systematic analysis of the effects of splicing on the diversity of post-translational modifications in protein isoforms.

Post-translational modifications (PTMs) and splicing are known to be important regulatory processes for controlling protein function and activity. However, there have been limitations in analyzing the interplay of alternative splicing and PTMs, which stems from the deep differences in genomic and proteomic databases. In this work, we bridged the protein- and genome-centric world views to map PTMs to genomic locations for subsequent projection of PTMs onto alternative isoforms. We then performed a systematic analysis of the diversification of PTMs by alternative splicing, including exploration of the modification-specific rates of inclusion across isoforms and how often the regulatory sequences directly flanking a PTM are impacted by splicing, which might indicate altered regulatory or binding interactions in the alternatively spliced isoform. We found that 6-51% of PTMs are excluded from at least one isoform, depending on the modification type. Further, approximately 2% of prospective PTM sites exhibited altered regulatory sequences surrounding the modification site, suggesting that regulatory or binding interactions might be diversified in these proteoforms. Lastly, we applied this PTM-to-isoform mapping approach to explore the impacts of disease-related splicing in prostate cancer, identifying possible new hypotheses explaining the variable consequences of ESRP1 expression in different cancers. As a part of this work, we have provided an easily implementable tool for annotating splice events identified from RNA-sequencing with PTMs and their functional consequences, called PTM-POSE.

Logic-based mechanistic machine learning on high-content images reveals how drugs differentially regulate cardiac fibroblasts.

Fibroblasts are essential regulators of extracellular matrix deposition following cardiac injury. These cells exhibit highly plastic responses in phenotype during fibrosis in response to environmental stimuli. Here, we test whether and how candidate anti-fibrotic drugs differentially regulate measures of cardiac fibroblast phenotype, which may help identify treatments for cardiac fibrosis. We conducted a high content microscopy screen of human cardiac fibroblasts treated with 13 clinically relevant drugs in the context of TGFß and/or IL-1ß, measuring phenotype across 137 single-cell features. We used the phenotypic data from our high content imaging to train a logic-based mechanistic machine learning model (LogiMML) for fibroblast signaling. The model predicted how pirfenidone and Src inhibitor WH-4-023 reduce actin filament assembly and actin-myosin stress fiber formation, respectively. Validating the LogiMML model prediction that PI3K partially mediates the effects of Src inhibition, we found that PI3K inhibition reduces actin-myosin stress fiber formation and procollagen I production in human cardiac fibroblasts. In this study, we establish a modeling approach combining the strengths of logic-based network models and regularized regression models, apply this approach to predict mechanisms that mediate the differential effects of drugs on fibroblasts, revealing Src inhibition acting via PI3K as a potential therapy for cardiac fibrosis.

A path to translation: How 3D patient tumor avatars enable next generation precision oncology.

3D patient tumor avatars (3D-PTAs) hold promise for next-generation precision medicine. Here, we describe the benefits and challenges of 3D-PTA technologies and necessary future steps to realize their potential for clinical decision making. 3D-PTAs require standardization criteria and prospective trials to establish clinical benefits. Innovative trial designs that combine omics and 3D-PTA readouts may lead to more accurate clinical predictors, and an integrated platform that combines diagnostic and therapeutic development will accelerate new treatments for patients with refractory disease.

KSTAR: An algorithm to predict patient-specific kinase activities from phosphoproteomic data.

Kinase inhibitors as targeted therapies have played an important role in improving cancer outcomes. However, there are still considerable challenges, such as resistance, non-response, patient stratification, polypharmacology, and identifying combination therapy where understanding a tumor kinase activity profile could be transformative. Here, we develop a graph- and statistics-based algorithm, called KSTAR, to convert phosphoproteomic measurements of cells and tissues into a kinase activity score that is generalizable and useful for clinical pipelines, requiring no quantification of the phosphorylation sites. In this work, we demonstrate that KSTAR reliably captures expected kinase activity differences across different tissues and stimulation contexts, allows for the direct comparison of samples from independent experiments, and is robust across a wide range of dataset sizes. Finally, we apply KSTAR to clinical breast cancer phosphoproteomic data and find that there is potential for kinase activity inference from KSTAR to complement the current clinical diagnosis of HER2 status in breast cancer patients.

KinPred: A unified and sustainable approach for harnessing proteome-level human kinase-substrate predictions.

Tyrosine and serine/threonine kinases are essential regulators of cell processes and are important targets for human therapies. Unfortunately, very little is known about specific kinase-substrate relationships, making it difficult to infer meaning from dysregulated phosphoproteomic datasets or for researchers to identify possible kinases that regulate specific or novel phosphorylation sites. The last two decades have seen an explosion in algorithms to extrapolate from what little is known into the larger unknown-predicting kinase relationships with site-specific substrates using a variety of approaches that include the sequence-specificity of kinase catalytic domains and various other factors, such as evolutionary relationships, co-expression, and protein-protein interaction networks. Unfortunately, a number of limitations prevent researchers from easily harnessing these resources, such as loss of resource accessibility, limited information in publishing that results in a poor mapping to a human reference, and not being updated to match the growth of the human phosphoproteome. Here, we propose a methodological framework for publishing predictions in a unified way, which entails ensuring predictions have been run on a current reference proteome, mapping the same substrates and kinases across resources to a common reference, filtering for the human phosphoproteome, and providing methods for updating the resource easily in the future. We applied this framework on three currently available resources, published in the last decade, which provide kinase-specific predictions in the human proteome. Using the unified datasets, we then explore the role of study bias, the emergent network properties of these predictive algorithms, and comparisons within and between predictive algorithms. The combination of the code for unification and analysis, as well as the unified predictions are available under the resource we named KinPred. We believe this resource will be useful for a wide range of applications and establishes best practices for long-term usability and sustainability for new and existing predictive algorithms.

Defining phenotypic and functional heterogeneity of glioblastoma stem cells by mass cytometry.

Most patients with glioblastoma (GBM) die within 2 years. A major therapeutic goal is to target GBM stem cells (GSCs), a subpopulation of cells that contribute to treatment resistance and recurrence. Since their discovery in 2003, GSCs have been isolated using single-surface markers, such as CD15, CD44, CD133, and α6 integrin. It remains unknown how these single-surface marker-defined GSC populations compare with each other in terms of signaling and function and whether expression of different combinations of these markers is associated with different functional capacity. Using mass cytometry and fresh operating room specimens, we found 15 distinct GSC subpopulations in patients, and they differed in their MEK/ERK, WNT, and AKT pathway activation status. Once in culture, some subpopulations were lost and previously undetectable ones materialized. GSCs that highly expressed all 4 surface markers had the greatest self-renewal capacity, WNT inhibitor sensitivity, and in vivo tumorigenicity. This work highlights the potential signaling and phenotypic diversity of GSCs. Larger patient sample sizes and antibody panels are required to confirm these findings.

New analysis pipeline for high-throughput domain-peptide affinity experiments improves SH2 interaction data.

Protein domain interactions with short linear peptides, such as those of the Src homology 2 (SH2) domain with phosphotyrosine-containing peptide motifs (pTyr), are ubiquitous and important to many biochemical processes of the cell. The desire to map and quantify these interactions has resulted in the development of high-throughput (HTP) quantitative measurement techniques, such as microarray or fluorescence polarization assays. For example, in the last 15 years, experiments have progressed from measuring single interactions to covering 500,000 of the 5.5 million possible SH2-pTyr interactions in the human proteome. However, high variability in affinity measurements and disagreements about positive interactions between published data sets led us here to reevaluate the analysis methods and raw data of published SH2-pTyr HTP experiments. We identified several opportunities for improving the identification of positive and negative interactions and the accuracy of affinity measurements. We implemented model-fitting techniques that are more statistically appropriate for the nonlinear SH2-pTyr interaction data. We also developed a method to account for protein concentration errors due to impurities and degradation or protein inactivity and aggregation. Our revised analysis increases the reported affinity accuracy, reduces the false-negative rate, and increases the amount of useful data by adding reliable true-negative results. We demonstrate improvement in classification of binding versus nonbinding when using machine-learning techniques, suggesting improved coherence in the reanalyzed data sets. We present revised SH2-pTyr affinity results and propose a new analysis pipeline for future HTP measurements of domain-peptide interactions.

ProteoClade: A taxonomic toolkit for multi-species and metaproteomic analysis.

We present ProteoClade, a Python toolkit that performs taxa-specific peptide assignment, protein inference, and quantitation for multi-species proteomics experiments. ProteoClade scales to hundreds of millions of protein sequences, requires minimal computational resources, and is open source, multi-platform, and accessible to non-programmers. We demonstrate its utility for processing quantitative proteomic data derived from patient-derived xenografts and its speed and scalability enable a novel de novo proteomic workflow for complex microbiota samples.

ASPEN, a methodology for reconstructing protein evolution with improved accuracy using ensemble models.

Evolutionary reconstruction algorithms produce models of the evolutionary history of proteins or species. Such algorithms are highly sensitive to their inputs: the sequences used and their alignments. Here, we asked whether the variance introduced by selecting different input sequences could be used to better identify accurate evolutionary models. We subsampled from available ortholog sequences and measured the distribution of observed relationships between paralogs produced across hundreds of models inferred from the subsamples. We observed two important phenomena. First, the reproducibility of an all-sequence, single-alignment reconstruction, measured by comparing topologies inferred from 90% subsamples, directly correlates with the accuracy of that single-alignment reconstruction, producing a measurable value for something that has been traditionally unknowable. Second, topologies that are most consistent with the observations made in the ensemble are more accurate and we present a meta algorithm that exploits this property to improve model accuracy.

Predicting Patient Response to the Antiarrhythmic Mexiletine Based on Genetic Variation.

RATIONALE: Mutations in the SCN5A gene, encoding the α subunit of the Nav1.5 channel, cause a life-threatening form of cardiac arrhythmia, long QT syndrome type 3 (LQT3). Mexiletine, which is structurally related to the Na+ channel-blocking anesthetic lidocaine, is used to treat LQT3 patients. However, the patient response is variable, depending on the genetic mutation in SCN5A. OBJECTIVE: The goal of this study is to understand the molecular basis of patients' variable responses and build a predictive statistical model that can be used to personalize mexiletine treatment based on patient's genetic variant. METHODS AND RESULTS: We monitored the cardiac Na+ channel voltage-sensing domain (VSD) conformational dynamics simultaneously with other gating properties for the LQT3 variants. To systematically identify the relationship between mexiletine block and channel biophysical properties, we used a system-based statistical modeling approach to connect the multivariate properties to patient phenotype. We found that mexiletine altered the conformation of the Domain III VSD, which is the same VSD that many tested LQT3 mutations affect. Analysis of 15 LQT3 variants showed a strong correlation between the activation of the Domain III-VSD and the strength of the inhibition of the channel by mexiletine. Based on this improved molecular-level understanding, we generated a systems-based model based on a dataset of 32 LQT3 patients, which then successfully predicted the response of 7 out of 8 patients to mexiletine in a blinded, retrospective trial. CONCLUSIONS: Our results imply that the modulated receptor theory of local anesthetic action, which confines local anesthetic binding effects to the channel pore, should be revised to include drug interaction with the Domain III-VSD. Using an algorithm that incorporates this mode of action, we can predict patient-specific responses to mexiletine, improving therapeutic decision making.

Proteome-Level Analysis Indicates Global Mechanisms for Post-Translational Regulation of RRM Domains.

RRM, or RNA-recognition motif, domains are the largest class of single-stranded RNA binding domains in the human proteome and play important roles in RNA processing, splicing, export, stability, packaging, and degradation. Using a current database of post-translational modifications (PTMs), ProteomeScout, we found that RRM domains are also one of the most heavily modified domains in the human proteome. Here, we present two interesting findings about RRM domain modifications, found by mapping known PTMs onto RRM domain alignments and structures. First, we find significant overlap of ubiquitination and acetylation within RRM domains, suggesting the possibility for ubiquitination-acetylation crosstalk. Additionally, an analysis of quantitative study of ubiquitination changes in response to proteasome inhibition highlights the uniqueness of RRM domain ubiquitination - RRM domain ubiquitination decreases in response to proteasome inhibition, whereas the majority of sites increase. Second, we found conservation of tyrosine phosphorylation within the RNP1 and RNP2 consensus sequences, which coordinate RNA binding - suggesting a possible role for regulation of RNA binding by tyrosine kinase signaling. These observations suggest there are unique regulatory mechanisms of RRM function that have yet to be uncovered and that the RRM domain represents a model system for further studies on understanding PTM crosstalk.

Using ProteomeScout: A Resource of Post-Translational Modifications, Their Experiments, and the Proteins That They Annotate.

Post-translational modifications (PTMs) of protein amino acids are ubiquitous and important to protein function, localization, degradation, and more. In recent years, there has been an explosion in the discovery of PTMs as a result of improvements in PTM measurement techniques, including quantitative measurements of PTMs across multiple conditions. ProteomeScout is a repository for such discovery and quantitative experiments and provides tools for visualizing PTMs within proteins, including where they are relative to other PTMS, domains, mutations, and structure. ProteomeScout additionally provides analysis tools for identifying statistically significant relationships in experimental datasets. This unit describes four basic protocols for working with the ProteomeScout Web interface or programmatically with the database download. © 2017 by John Wiley & Sons, Inc.

Ensemble clustering of phosphoproteomic data identifies differences in protein interactions and cell-cell junction integrity of HER2-overexpressing cells.

Overexpression of HER2, a receptor tyrosine kinase of the ERBB family, in breast cancer is related to increased cancer progression and aggressiveness. A breast epithelial cell model with the single perturbation of HER2 overexpression is capable of replicating the increased aggressiveness of HER2 overexpressing cancers. In previous work, Wolf-Yadlin and colleagues (Wolf-Yadlin et al., Mol. Syst. Biol., 2006, 2) measured the proximal tyrosine phosphorylation dynamics of the parental and HER2 overexpressing cells (24H) in response to EGF. Here, we apply an ensemble clustering approach to dynamic phosphorylation measurements of the two cell models in order to identify signaling events that explain the increased migratory potential of HER2 overexpressing cells. The use of an ensemble approach for identifying relationships within a dataset and how these relationships change across datasets uncovers relationships that cannot be found by the direct comparison of dynamic responses in the two conditions. Of particular note is a drastic change in the clustering of SHC1 phosphorylation (on site Y349) from an EGFR-MAPK module in parental cells to a module consisting of an E-cadherin junction protein phosphorylation site, catenin delta-1 Y228, in HER2 overexpressing (24H) cells. Given the importance of E-cadherin junctions in healthy epithelial wound healing and migration, we chose to test the computationally-derived identification of altered cell junctions and CTNND1:SHC1 relationships. Our cell and molecular biology experiments demonstrate that SHC and CTNND1 interact in an EGF- and HER2-dependent manner and that the cell junctions are phenotypically affected by HER2, breaking down in response to EGF and yet avoiding apoptosis as a result of cell junction loss. The results suggest a mechanism by which HER2 alters the localization of the SHC-MAPK signaling axis and a phenotypic effect on cell junction integrity.

High-Resolution Identification of Specificity Determining Positions in the LacI Protein Family Using Ensembles of Sub-Sampled Alignments.

Since the advent of large-scale genomic sequencing, and the consequent availability of large numbers of homologous protein sequences, there has been burgeoning development of methods for extracting functional information from multiple sequence alignments (MSAs). One type of analysis seeks to identify specificity determining positions (SDPs) based on the assumption that such positions are highly conserved within groups of sequences sharing functional specificity, but conserved to different amino acids in different specificity groups. This unsupervised approach to utilizing evolutionary information may elucidate mechanisms of specificity in protein-protein interactions, catalytic activity of enzymes, sensitivity to allosteric regulation, and other types of protein functionality. We present an analysis of SDPs in the LacI family of transcriptional regulators in which we 1) relax the constraint that all specificity groups must contribute to SDP signal, and 2) use a novel approach to robust treatment of sequence alignment uncertainty based on sub-sampling. We find that the vast majority of SDP signal occurs at positions with a conservation pattern that significantly complicates detection by previously described methods. This pattern, which we term "partial SDP", consists of the commonly accepted SDP conservation pattern among a subset of specificity groups and strong degeneracy among the rest. An upshot of this fact is that the SDP complement of every specificity group appears to be unique. Additionally, sub-sampling gives us the ability to assign a confidence interval to the SDP score, as well as increase fidelity, as compared to analysis of a single, comprehensive alignment-the current standard in multiple sequence alignment methodologies.

Avoiding common pitfalls when clustering biological data.

Clustering is an unsupervised learning method, which groups data points based on similarity, and is used to reveal the underlying structure of data. This computational approach is essential to understanding and visualizing the complex data that are acquired in high-throughput multidimensional biological experiments. Clustering enables researchers to make biological inferences for further experiments. Although a powerful technique, inappropriate application can lead biological researchers to waste resources and time in experimental follow-up. We review common pitfalls identified from the published molecular biology literature and present methods to avoid them. Commonly encountered pitfalls relate to the high-dimensional nature of biological data from high-throughput experiments, the failure to consider more than one clustering method for a given problem, and the difficulty in determining whether clustering has produced meaningful results. We present concrete examples of problems and solutions (clustering results) in the form of toy problems and real biological data for these issues. We also discuss ensemble clustering as an easy-to-implement method that enables the exploration of multiple clustering solutions and improves robustness of clustering solutions. Increased awareness of common clustering pitfalls will help researchers avoid overinterpreting or misinterpreting the results and missing valuable insights when clustering biological data.

Different Epidermal Growth Factor Receptor (EGFR) Agonists Produce Unique Signatures for the Recruitment of Downstream Signaling Proteins.

The EGF receptor can bind seven different agonist ligands. Although each agonist appears to stimulate the same suite of downstream signaling proteins, different agonists are capable of inducing distinct responses in the same cell. To determine the basis for these differences, we used luciferase fragment complementation imaging to monitor the recruitment of Cbl, CrkL, Gab1, Grb2, PI3K, p52 Shc, p66 Shc, and Shp2 to the EGF receptor when stimulated by the seven EGF receptor ligands. Recruitment of all eight proteins was rapid, dose-dependent, and inhibited by erlotinib and lapatinib, although to differing extents. Comparison of the time course of recruitment of the eight proteins in response to a fixed concentration of each growth factor revealed differences among the growth factors that could contribute to their differing biological effects. Principal component analysis of the resulting data set confirmed that the recruitment of these proteins differed between agonists and also between different doses of the same agonist. Ensemble clustering of the overall response to the different growth factors suggests that these EGF receptor ligands fall into two major groups as follows: (i) EGF, amphiregulin, and EPR; and (ii) betacellulin, TGFα, and epigen. Heparin-binding EGF is distantly related to both clusters. Our data identify differences in network utilization by different EGF receptor agonists and highlight the need to characterize network interactions under conditions other than high dose EGF.

Reproducible Analysis of Post-Translational Modifications in Proteomes--Application to Human Mutations.

BACKGROUND: Protein post-translational modifications (PTMs) are an important aspect of protein regulation. The number of PTMs discovered within the human proteome, and other proteomes, has been rapidly expanding in recent years. As a consequence of the rate in which new PTMs are identified, analysis done in one year may result in different conclusions when repeated in subsequent years. Among the various functional questions pertaining to PTMs, one important relationship to address is the interplay between modifications and mutations. Specifically, because the linear sequence surrounding a modification site often determines molecular recognition, it is hypothesized that mutations near sites of PTMs may be more likely to result in a detrimental effect on protein function, resulting in the development of disease. METHODS AND RESULTS: We wrote an application programming interface (API) to make analysis of ProteomeScout, a comprehensive database of PTMs and protein information, easy and reproducible. We used this API to analyze the relationship between PTMs and human mutations associated with disease (based on the 'Clinical Significance' annotation from dbSNP). Proteins containing pathogenic mutations demonstrated a significant study bias which was controlled for by analyzing only well-studied proteins, based on their having at least one pathogenic mutation. We found that pathogenic mutations are significantly more likely to lie within eight amino acids of a phosphoserine, phosphotyrosine or ubiquitination site when compared to mutations in general, based on a Fisher's Exact test. Despite the skew of pathogenic mutations occurring on positively charged arginines, we could not account for this relationship based only on residue type. Finally, we hypothesize a potential mechanism for a pathogenic mutation on RAF1, based on its proximity to a phosphorylation site, which represents a subtle regulation difference that may explain why its biochemical effect has failed to be uncovered previously. The combination of the API and a dynamically expanding PTM database will make the reanalysis of this question and other systems-level questions easier in the future.