Causal network perturbations for instance-specific analysis of single cell and disease samples

Discovering root causal genes with high-throughput perturbations

Root causal gene expression levels - or root causal genes for short - correspond to the initial changes to gene expression that generate patient symptoms as a downstream effect. Identifying root causal genes is critical towards developing treatments that modify disease near its onset, but no existing algorithms attempt to identify root causal genes from data. RNA-sequencing (RNA-seq) data introduces challenges such as measurement error, high dimensionality and non-linearity that compromise accurate estimation of root causal effects even with state-of-the-art approaches. We therefore instead leverage Perturb-seq, or high-throughput perturbations with single-cell RNA-seq readout, to learn the causal order between the genes. We then transfer the causal order to bulk RNA-seq and identify root causal genes specific to a given patient for the first time using a novel statistic. Experiments demonstrate large improvements in performance. Applications to macular degeneration and multiple sclerosis also reveal root causal genes that lie on known pathogenic pathways, delineate patient subgroups and implicate a newly defined omnigenic root causal model.

Development and validation of a mortality risk prediction model for chronic obstructive pulmonary disease: a cross-sectional study using probabilistic graphical modelling

Background

Chronic Obstructive Pulmonary Disease (COPD) is a leading cause of mortality. Predicting mortality risk in patients with COPD can be important for disease management strategies. Although all-cause mortality predictors have been developed previously, limited research exists on factors directly affecting COPD-specific mortality.

Methods

In a retrospective study, we used probabilistic graphs to analyse clinical cross-sectional data (COPDGene cohort), including demographics, spirometry, quantitative chest imaging, and symptom features, as well as gene expression data. COPDGene recruited current and former smokers, aged 45-80 years with >10 pack-years smoking history, from across the USA (Phase 1, 11/2007-4/2011) and invited them for a follow-up visit (Phase 2, 7/2013-7/2017). ECLIPSE cohort recruited current and former smokers (COPD patients and controls from USA and Europe), aged 45-80 with smoking history >10 pack-years (12/2005-11/2007). We applied graphical models on multi-modal data COPDGene Phase 1 participants to identify factors directly affecting all-cause and COPD-specific mortality (primary outcomes); and on Phase 2 follow-up cohort to identify additional molecular and social factors affecting mortality. We used penalized Cox regression with features selected by the causal graph to build VAPORED, a mortality risk prediction model. VAPORED was compared to existing scores (BODE: BMI, airflow obstruction, dyspnoea, exercise capacity; ADO: age, dyspnoea, airflow obstruction) on the ability to rank individuals by mortality risk, using four evaluation metrics (concordance, concordance probability estimate (CPE), cumulative/dynamic (C/D) area under the receiver operating characteristic curve (AUC), and integrated C/D AUC). The results were validated in ECLIPSE.

Findings

Graphical models, applied on the COPDGene Phase 1 samples (n = 8610), identified 11 and 7 variables directly linked to all-cause and COPD-specific mortality, respectively. Although many appear in both models, non-lung comorbidities appear only in the all-cause model, while forced vital capacity (FVC %predicted) appears in COPD-specific mortality model only. Additionally, the graph model of Phase 2 data (n = 3182) identified internet access, CD4 T cells and platelets to be linked to lower mortality risk. Furthermore, using the 7 variables linked to COPD-specific mortality (forced expiratory volume in 1 s/forced vital capacity (FEV1/FVC) ration, FVC %predicted, age, history of pneumonia, oxygen saturation, 6-min walk distance, dyspnoea) we developed VAPORED mortality risk score, which we validated on the ECLIPSE cohort (3-yr all-cause mortality data, n = 2312). VAPORED performed significantly better than ADO, BODE, and updated BODE indices in predicting all-cause mortality in ECLIPSE in terms of concordance (VAPORED [0.719] vs ADO [0.693; FDR p-value 0.014], BODE [0.695; FDR p-value 0.020], and updated BODE [0.694; FDR p-value 0.021]); CPE (VAPORED [0.714] vs ADO [0.673; FDR p-value <0.0001], BODE [0.662; FDR p-value <0.0001], and updated BODE [0.646; FDR p-value <0.0001]); 3-year C/D AUC (VAPORED [0.728] vs ADO [0.702; FDR p-value 0.017], BODE [0.704; FDR p-value 0.021], and updated BODE [0.703; FDR p-value 0.024]); integrated C/D AUC (VAPORED [0.723] vs ADO [0.698; FDR p-value 0.047], BODE [0.695; FDR p-value 0.024], and updated BODE [0.690; FDR p-value 0.021]). Finally, we developed a web tool to help clinicians calculate VAPORED mortality risk and compare it to ADO and BODE predictions.

Interpretation

Our work is an important step towards improving our identification of high-risk patients and generating hypotheses of potential biological mechanisms and social factors driving mortality in patients with COPD at the population level. The main limitation of our study is the fact that the analysed datasets consist of older people with extensive smoking history and limited racial diversity. Thus, the results are relevant to high-risk individuals or those diagnosed with COPD and the VAPORED score is validated for them.

Funding

This research was supported by NIH [NHLBI, NLM]. The COPDGene study is supported by the COPD Foundation, through grants from AstraZeneca, Bayer Pharmaceuticals, Boehringer Ingelheim, Genentech, GlaxoSmithKline, Novartis, Pfizer and Sunovion.

Transcriptome data are insufficient to control false discoveries in regulatory network inference

Inference of causal transcriptional regulatory networks (TRNs) from transcriptomic data suffers notoriously from false positives. Approaches to control the false discovery rate (FDR), for example, via permutation, bootstrapping, or multivariate Gaussian distributions, suffer from several complications: difficulty in distinguishing direct from indirect regulation, nonlinear effects, and causal structure inference requiring "causal sufficiency," meaning experiments that are free of any unmeasured, confounding variables. Here, we use a recently developed statistical framework, model-X knockoffs, to control the FDR while accounting for indirect effects, nonlinear dose-response, and user-provided covariates. We adjust the procedure to estimate the FDR correctly even when measured against incomplete gold standards. However, benchmarking against chromatin immunoprecipitation (ChIP) and other gold standards reveals higher observed than reported FDR. This indicates that unmeasured confounding is a major driver of FDR in TRN inference. A record of this paper's transparent peer review process is included in the supplemental information.

Causal network perturbation analysis identifies known and novel type-2 diabetes driver genes

The molecular pathogenesis of diabetes is multifactorial, involving genetic predisposition and environmental factors that are not yet fully understood. However, pancreatic β-cell failure remains among the primary reasons underlying the progression of type-2 diabetes (T2D) making targeting β-cell dysfunction an attractive pathway for diabetes treatment. To identify genetic contributors to β-cell dysfunction, we investigated single-cell gene expression changes in β-cells from healthy (C57BL/6J) and diabetic (NZO/HlLtJ) mice fed with normal or high-fat, high-sugar diet (HFHS). Our study presents an innovative integration of the causal network perturbation assessment (ssNPA) framework with meta-cell transcriptome analysis to explore the genetic underpinnings of type-2 diabetes (T2D). By generating a reference causal network and in silico perturbation, we identified novel genes implicated in T2D and validated our candidates using the Knockout Mouse Phenotyping (KOMP) Project database.

Evaluation of single-sample network inference methods for precision oncology

A major challenge in precision oncology is to detect targetable cancer vulnerabilities in individual patients. Modeling high-throughput omics data in biological networks allows identifying key molecules and processes of tumorigenesis. Traditionally, network inference methods rely on many samples to contain sufficient information for learning, resulting in aggregate networks. However, to implement patient-tailored approaches in precision oncology, we need to interpret omics data at the level of individual patients. Several single-sample network inference methods have been developed that infer biological networks for an individual sample from bulk RNA-seq data. However, only a limited comparison of these methods has been made and many methods rely on 'normal tissue' samples as reference, which are not always available. Here, we conducted an evaluation of the single-sample network inference methods SSN, LIONESS, SWEET, iENA, CSN and SSPGI using transcriptomic profiles of lung and brain cancer cell lines from the CCLE database. The methods constructed functional gene networks with distinct network characteristics. Hub gene analyses revealed different degrees of subtype-specificity across methods. Single-sample networks were able to distinguish between tumor subtypes, as exemplified by node strength clustering, enrichment of known subtype-specific driver genes among hubs and differential node strength. We also showed that single-sample networks correlated better to other omics data from the same cell line as compared to aggregate networks. We conclude that single-sample network inference methods can reflect sample-specific biology when 'normal tissue' samples are absent and we point out peculiarities of each method.

Distinct COPD subtypes in former smokers revealed by gene network perturbation analysis

BACKGROUND

Chronic obstructive pulmonary disease (COPD) varies significantly in symptomatic and physiologic presentation. Identifying disease subtypes from molecular data, collected from easily accessible blood samples, can help stratify patients and guide disease management and treatment.

METHODS

Blood gene expression measured by RNA-sequencing in the COPDGene Study was analyzed using a network perturbation analysis method. Each COPD sample was compared against a learned reference gene network to determine the part that is deregulated. Gene deregulation values were used to cluster the disease samples.

RESULTS

The discovery set included 617 former smokers from COPDGene. Four distinct gene network subtypes are identified with significant differences in symptoms, exercise capacity and mortality. These clusters do not necessarily correspond with the levels of lung function impairment and are independently validated in two external cohorts: 769 former smokers from COPDGene and 431 former smokers in the Multi-Ethnic Study of Atherosclerosis (MESA). Additionally, we identify several genes that are significantly deregulated across these subtypes, including DSP and GSTM1, which have been previously associated with COPD through genome-wide association study (GWAS).

CONCLUSIONS

The identified subtypes differ in mortality and in their clinical and functional characteristics, underlining the need for multi-dimensional assessment potentially supplemented by selected markers of gene expression. The subtypes were consistent across cohorts and could be used for new patient stratification and disease prognosis.

Causal Discovery in High-dimensional, Multicollinear Datasets

As the cost of high-throughput genomic sequencing technology declines, its application in clinical research becomes increasingly popular. The collected datasets often contain tens or hundreds of thousands of biological features that need to be mined to extract meaningful information. One area of particular interest is discovering underlying causal mechanisms of disease outcomes. Over the past few decades, causal discovery algorithms have been developed and expanded to infer such relationships. However, these algorithms suffer from the curse of dimensionality and multicollinearity. A recently introduced, non-orthogonal, general empirical Bayes approach to matrix factorization has been demonstrated to successfully infer latent factors with interpretable structures from observed variables. We hypothesize that applying this strategy to causal discovery algorithms can solve both the high dimensionality and collinearity problems, inherent to most biomedical datasets. We evaluate this strategy on simulated data and apply it to two real-world datasets. In a breast cancer dataset, we identified important survival-associated latent factors and biologically meaningful enriched pathways within factors related to important clinical features. In a SARS-CoV-2 dataset, we were able to predict whether a patient (1) had Covid-19 and (2) would enter the ICU. Furthermore, we were able to associate factors with known Covid-19 related biological pathways.

Machine learning for perturbational single-cell omics

Cell biology is fundamentally limited in its ability to collect complete data on cellular phenotypes and the wide range of responses to perturbation. Areas such as computer vision and speech recognition have addressed this problem of characterizing unseen or unlabeled conditions with the combined advances of big data, deep learning, and computing resources in the past 5 years. Similarly, recent advances in machine learning approaches enabled by single-cell data start to address prediction tasks in perturbation response modeling. We first define objectives in learning perturbation response in single-cell omics; survey existing approaches, resources, and datasets (https://github.com/theislab/sc-pert); and discuss how a perturbation atlas can enable deep learning models to construct an informative perturbation latent space. We then examine future avenues toward more powerful and explainable modeling using deep neural networks, which enable the integration of disparate information sources and an understanding of heterogeneous, complex, and unseen systems.

Detection of differentially abundant cell subpopulations in scRNA-seq data

Comprehensive and accurate comparisons of transcriptomic distributions of cells from samples taken from two different biological states, such as healthy versus diseased individuals, are an emerging challenge in single-cell RNA sequencing (scRNA-seq) analysis. Current methods for detecting differentially abundant (DA) subpopulations between samples rely heavily on initial clustering of all cells in both samples. Often, this clustering step is inadequate since the DA subpopulations may not align with a clear cluster structure, and important differences between the two biological states can be missed. Here, we introduce DA-seq, a targeted approach for identifying DA subpopulations not restricted to clusters. DA-seq is a multiscale method that quantifies a local DA measure for each cell, which is computed from its k nearest neighboring cells across a range of k values. Based on this measure, DA-seq delineates contiguous significant DA subpopulations in the transcriptomic space. We apply DA-seq to several scRNA-seq datasets and highlight its improved ability to detect differences between distinct phenotypes in severe versus mildly ill COVID-19 patients, melanomas subjected to immune checkpoint therapy comparing responders to nonresponders, embryonic development at two time points, and young versus aging brain tissue. DA-seq enabled us to detect differences between these phenotypes. Importantly, we find that DA-seq not only recovers the DA cell types as discovered in the original studies but also reveals additional DA subpopulations that were not described before. Analysis of these subpopulations yields biological insights that would otherwise be undetected using conventional computational approaches.

Evaluation of Single Sample Network Inference Methods for Metabolomics-Based Systems Medicine

Networks and network analyses are fundamental tools of systems biology. Networks are built by inferring pair-wise relationships among biological entities from a large number of samples such that subject-specific information is lost. The possibility of constructing these sample (individual)-specific networks from single molecular profiles might offer new insights in systems and personalized medicine and as a consequence is attracting more and more research interest. In this study, we evaluated and compared LIONESS (Linear Interpolation to Obtain Network Estimates for Single Samples) and ssPCC (single sample network based on Pearson correlation) in the metabolomics context of metabolite–metabolite association networks. We illustrated and explored the characteristics of these two methods on (i) simulated data, (ii) data generated from a dynamic metabolic model to simulate real-life observed metabolite concentration profiles, and (iii) 22 metabolomic data sets and (iv) we applied single sample network inference to a study case pertaining to the investigation of necrotizing soft tissue infections to show how these methods can be applied in metabolomics. We also proposed some adaptations of the methods that can be used for data exploration. Overall, despite some limitations, we found single sample networks to be a promising tool for the analysis of metabolomics data.

Elucidation of Biological Networks across Complex Diseases Using Single-Cell Omics

Single-cell multimodal omics (scMulti-omics) technologies have made it possible to trace cellular lineages during differentiation and to identify new cell types in heterogeneous cell populations. The derived information is especially promising for computing cell-type-specific biological networks encoded in complex diseases and improving our understanding of the underlying gene regulatory mechanisms. The integration of these networks could, therefore, give rise to a heterogeneous regulatory landscape (HRL) in support of disease diagnosis and drug therapeutics. In this review, we provide an overview of this field and pay particular attention to how diverse biological networks can be inferred in a specific cell type based on integrative methods. Then, we discuss how HRL can advance our understanding of regulatory mechanisms underlying complex diseases and aid in the prediction of prognosis and therapeutic responses. Finally, we outline challenges and future trends that will be central to bringing the field of HRL in complex diseases forward.