Molecular signatures that can be transferred across different omics platforms

Cross-Platform Omics Prediction procedure: a statistical machine learning framework for wider implementation of precision medicine

In this modern era of precision medicine, molecular signatures identified from advanced omics technologies hold great promise to better guide clinical decisions. However, current approaches are often location-specific due to the inherent differences between platforms and across multiple centres, thus limiting the transferability of molecular signatures. We present Cross-Platform Omics Prediction (CPOP), a penalised regression model that can use omics data to predict patient outcomes in a platform-independent manner and across time and experiments. CPOP improves on the traditional prediction framework of using gene-based features by selecting ratio-based features with similar estimated effect sizes. These components gave CPOP the ability to have a stable performance across datasets of similar biology, minimising the effect of technical noise often generated by omics platforms. We present a comprehensive evaluation using melanoma transcriptomics data to demonstrate its potential to be used as a critical part of a clinical screening framework for precision medicine. Additional assessment of generalisation was demonstrated with ovarian cancer and inflammatory bowel disease studies.

Construction of Prediction Model for Atrial Fibrillation with Valvular Heart Disease Based on Machine Learning

Background

Valvular heart disease (VHD) is a major precipitating factor of atrial fibrillation (AF) that contributes to decreased cardiac function, heart failure, and stroke. Stroke induced by VHD combined with atrial fibrillation (AF-VHD) is a much more serious condition in comparison to VHD alone. The aim of this study was to explore the molecular mechanism governing VHD progression and to provide candidate treatment targets for AF-VHD.

Methods

Four public mRNA microarray datasets were downloaded and differentially expressed genes (DEGs) screening was performed. Weighted gene correlation network analysis was carried out to detect key modules and explore their relationships and disease status. Candidate hub signature genes were then screened within the key module using machine learning methods. The receiver operating characteristic curve and nomogram model analysis were used to determine the potential clinical significance of the hub genes. Subsequently, target gene protein levels in independent human atrial tissue samples were detected using western blotting. Specific expression analysis of the hub genes in the tissue and cell samples was performed using single-cell sequencing analysis in the Human Protein Atlas tool.

Results

A total of 819 common DEGs in combined datasets were screened. Fourteen modules were identified using the cut tree dynamic function. The cyan and purple modules were considered the most clinically significant for AF-VHD. Then, 25 hub genes in the cyan and purple modules were selected for further analysis. The pathways related to dilated cardiomyopathy, hypertrophic cardiomyopathy, and heart contraction were concentrated in the purple and cyan modules of the AF-VHD. Genes of importance (CSRP3, MCOLN3, SLC25A5, and FIBP) were then identified based on machine learning. Of these, CSRP3 had a potential clinical significance and was specifically expressed in the heart tissue.

Conclusions

The identified genes may play critical roles in the pathophysiological process of AF-VHD, providing new insights into VHD development to AF and helping to determine potential biomarkers and therapeutic targets for treating AF-VHD.

Breast tumor stiffness instructs bone metastasis via maintenance of mechanical conditioning

While the immediate and transitory response of breast cancer cells to pathological stiffness in their native microenvironment has been well explored, it remains unclear how stiffness-induced phenotypes are maintained over time after cancer cell dissemination in vivo. Here, we show that fibrotic-like matrix stiffness promotes distinct metastatic phenotypes in cancer cells, which are preserved after transition to softer microenvironments, such as bone marrow. Using differential gene expression analysis of stiffness-responsive breast cancer cells, we establish a multigenic score of mechanical conditioning (MeCo) and find that it is associated with bone metastasis in patients with breast cancer. The maintenance of mechanical conditioning is regulated by RUNX2, an osteogenic transcription factor, established driver of bone metastasis, and mitotic bookmarker that preserves chromatin accessibility at target gene loci. Using genetic and functional approaches, we demonstrate that mechanical conditioning maintenance can be simulated, repressed, or extended, with corresponding changes in bone metastatic potential.

Watson et al. demonstrate that mechanical conditioning by stiff microenvironments in breast tumors is maintained in cancer cells after dissemination to softer microenvironments, including bone marrow. They show that mechanical conditioning promotes invasion and osteolysis and establish a mechanical conditioning (MeCo) score, associated with bone metastasis in patients.

Integrative computational epigenomics to build data-driven gene regulation hypotheses

BACKGROUND

Diseases are complex phenotypes often arising as an emergent property of a non-linear network of genetic and epigenetic interactions. To translate this resulting state into a causal relationship with a subset of regulatory features, many experiments deploy an array of laboratory assays from multiple modalities. Often, each of these resulting datasets is large, heterogeneous, and noisy. Thus, it is non-trivial to unify these complex datasets into an interpretable phenotype. Although recent methods address this problem with varying degrees of success, they are constrained by their scopes or limitations. Therefore, an important gap in the field is the lack of a universal data harmonizer with the capability to arbitrarily integrate multi-modal datasets.

RESULTS

In this review, we perform a critical analysis of methods with the explicit aim of harmonizing data, as opposed to case-specific integration. This revealed that matrix factorization, latent variable analysis, and deep learning are potent strategies. Finally, we describe the properties of an ideal universal data harmonization framework.

CONCLUSIONS

A sufficiently advanced universal harmonizer has major medical implications, such as (i) identifying dysregulated biological pathways responsible for a disease is a powerful diagnostic tool; (2) investigating these pathways further allows the biological community to better understand a disease's mechanisms; and (3) precision medicine also benefits from developments in this area, particularly in the context of the growing field of selective epigenome editing, which can suppress or induce a desired phenotype.

Gaussian and Mixed Graphical Models as (multi-)omics data analysis tools

Gaussian Graphical Models (GGMs) are tools to infer dependencies between biological variables. Popular applications are the reconstruction of gene, protein, and metabolite association networks. GGMs are an exploratory research tool that can be useful to discover interesting relations between genes (functional clusters) or to identify therapeutically interesting genes, but do not necessarily infer a network in the mechanistic sense. Although GGMs are well investigated from a theoretical and applied perspective, important extensions are not well known within the biological community. GGMs assume, for instance, multivariate normal distributed data. If this assumption is violated Mixed Graphical Models (MGMs) can be the better choice. In this review, we provide the theoretical foundations of GGMs, present extensions such as MGMs or multi-class GGMs, and illustrate how those methods can provide insight in biological mechanisms. We summarize several applications and present user-friendly estimation software. This article is part of a Special Issue entitled: Transcriptional Profiles and Regulatory Gene Networks edited by Dr. Dr. Federico Manuel Giorgi and Dr. Shaun Mahony.

Platform independent protein-based cell-of-origin subtyping of diffuse large B-cell lymphoma in formalin-fixed paraffin-embedded tissue

Diffuse large B-cell lymphoma (DLBCL) is commonly classified by gene expression profiling according to its cell of origin (COO) into activated B-cell (ABC)-like and germinal center B-cell (GCB)-like subgroups. Here we report the application of label-free nano-liquid chromatography - Sequential Window Acquisition of all THeoretical fragment-ion spectra - mass spectrometry (nanoLC-SWATH-MS) to the COO classification of DLBCL in formalin-fixed paraffin-embedded (FFPE) tissue. To generate a protein signature capable of predicting Affymetrix-based GCB scores, the summed log₂-transformed fragment ion intensities of 780 proteins quantified in a training set of 42 DLBCL cases were used as independent variables in a penalized zero-sum elastic net regression model with variable selection. The eight-protein signature obtained showed an excellent correlation (r = 0.873) between predicted and true GCB scores and yielded only 9 (21.4%) minor discrepancies between the three classifications: ABC, GCB, and unclassified. The robustness of the model was validated successfully in two independent cohorts of 42 and 31 DLBCL cases, the latter cohort comprising only patients aged >75 years, with Pearson correlation coefficients of 0.846 and 0.815, respectively, between predicted and NanoString nCounter based GCB scores. We further show that the 8-protein signature is directly transferable to both a triple quadrupole and a Q Exactive quadrupole-Orbitrap mass spectrometer, thus obviating the need for proprietary instrumentation and reagents. This method may therefore be used for robust and competitive classification of DLBCLs on the protein level.

Comparison of GeneChip, nCounter, and Real-Time PCR-Based Gene Expressions Predicting Locoregional Tumor Control after Primary and Postoperative Radiochemotherapy in Head and Neck Squamous Cell Carcinoma

This article compares the expression and applicability of biomarkers, from single genes and gene signatures, identified in patients with locally advanced head and neck squamous cell carcinoma using the GeneChip Human Transcriptome Array 2.0, nCounter, and real-time PCR analyses. Two multicenter, retrospective cohorts of patients with head and neck squamous cell carcinoma from the German Cancer Consortium Radiation Oncology Group who received postoperative radiochemotherapy or primary radiochemotherapy were considered. Real-time PCR was performed for a limited number of 38 genes of the cohort who received postoperative radiochemotherapy only. Correlations between the methods were evaluated by the Spearman rank correlation coefficient. Patients were stratified based on the expression of putative cancer stem cell markers, hypoxia-associated gene signatures, and a previously developed seven-gene signature. Locoregional tumor control was compared between these patient subgroups using log-rank tests. Gene expressions obtained from nCounter analyses were moderately correlated to GeneChip analyses (median ρ = approximately 0.68). A higher correlation was obtained between nCounter analyses and real-time PCR (median ρ = 0.84). Significant associations with locoregional tumor control were observed for most of the considered biomarkers evaluated by GeneChip and nCounter analyses. In general, all applied biomarkers (single genes and gene signatures) classified approximately 70% to 85% of the patients similarly. Overall, gene signatures seem to be more robust and had a better transferability among different measurement methods.

Constraining classifiers in molecular analysis: invariance and robustness

Analysing molecular profiles requires the selection of classification models that can cope with the high dimensionality and variability of these data. Also, improper reference point choice and scaling pose additional challenges. Often model selection is somewhat guided by ad hoc simulations rather than by sophisticated considerations on the properties of a categorization model. Here, we derive and report four linked linear concept classes/models with distinct invariance properties for high-dimensional molecular classification. We can further show that these concept classes also form a half-order of complexity classes in terms of Vapnik-Chervonenkis dimensions, which also implies increased generalization abilities. We implemented support vector machines with these properties. Surprisingly, we were able to attain comparable or even superior generalization abilities to the standard linear one on the 27 investigated RNA-Seq and microarray datasets. Our results indicate that a priori chosen invariant models can replace ad hoc robustness analysis by interpretable and theoretically guaranteed properties in molecular categorization.

Expression Concordance of 325 Novel RNA Biomarkers between Data Generated by NanoString nCounter and Affymetrix GeneChip

Background

With the development of new drug combinations and targeted treatments for multiple types of cancer, the ability to stratify categories of patient populations and to develop companion diagnostics has become increasingly important. A panel of 325 RNA biomarkers was selected based on cancer-related biological processes of healthy cells and gene expression changes over time during nonmalignant epithelial cell organization. This "cancer in reverse" approach resulted in a panel of biomarkers relevant for at least 7 cancer types, providing gene expression profiles representing key cellular signaling pathways beyond mutations in "driver genes." Objective. To further investigate this biomarker panel, the objective of the current study is to (1) validate the assay reproducibility for the 325 RNA biomarkers and (2) compare gene expression profiles side by side using two technology platforms.

Methods and Results

We have mapped the 325 RNA transcripts and in a custom NanoString nCounter expression panel to be compared to all potential probe sets in the Affymetrix Human Genome U133 Plus 2.0. The experiments were conducted with 10 unique biological formalin-fixed paraffin-embedded (FFPE) breast tumor samples. Each site extracted RNA from four sections of 10-micron thick FFPE tissue over three different days by two different operators using an optimized standard operating procedure and quality control criteria. Samples were analyzed using mas5 in BioConductor and NanoStringNorm in R. Pearson correlation showed reproducibility between sites for all 60 samples with r = 0.995 for Affymetrix and r = 0.999 for NanoString. Correlation in multiple days and multiple users was for Affymetrix r = (0.962 - 0.999) and for NanoString r = (0.982 - 0.991).

Conclusion

The 325 RNA biomarkers showed reproducibility in two technology platforms with moderate to high concordance. Future directions include performing clinical validation studies and generating rationale for patient selection in clinical trials using the technically validated assay.

From hype to reality: data science enabling personalized medicine

BACKGROUND

Personalized, precision, P4, or stratified medicine is understood as a medical approach in which patients are stratified based on their disease subtype, risk, prognosis, or treatment response using specialized diagnostic tests. The key idea is to base medical decisions on individual patient characteristics, including molecular and behavioral biomarkers, rather than on population averages. Personalized medicine is deeply connected to and dependent on data science, specifically machine learning (often named Artificial Intelligence in the mainstream media). While during recent years there has been a lot of enthusiasm about the potential of 'big data' and machine learning-based solutions, there exist only few examples that impact current clinical practice. The lack of impact on clinical practice can largely be attributed to insufficient performance of predictive models, difficulties to interpret complex model predictions, and lack of validation via prospective clinical trials that demonstrate a clear benefit compared to the standard of care. In this paper, we review the potential of state-of-the-art data science approaches for personalized medicine, discuss open challenges, and highlight directions that may help to overcome them in the future.

CONCLUSIONS

There is a need for an interdisciplinary effort, including data scientists, physicians, patient advocates, regulatory agencies, and health insurance organizations. Partially unrealistic expectations and concerns about data science-based solutions need to be better managed. In parallel, computational methods must advance more to provide direct benefit to clinical practice.

Scale-Invariant Biomarker Discovery in Urine and Plasma Metabolite Fingerprints

Metabolomics data is typically scaled to a common reference like a constant volume of body fluid, a constant creatinine level, or a constant area under the spectrum. Such scaling of the data, however, may affect the selection of biomarkers and the biological interpretation of results in unforeseen ways. Here, we studied how both the outcome of hypothesis tests for differential metabolite concentration and the screening for multivariate metabolite signatures are affected by the choice of scale. To overcome this problem for metabolite signatures and to establish a scale-invariant biomarker discovery algorithm, we extended linear zero-sum regression to the logistic regression framework and showed in two applications to ¹H NMR-based metabolomics data how this approach overcomes the scaling problem. Logistic zero-sum regression is available as an R package as well as a high-performance computing implementation that can be downloaded at https://github.com/rehbergT/zeroSum .