Deep learning of left atrial structure and function provides link to atrial fibrillation risk

Increased left atrial volume and decreased left atrial function have long been associated with atrial fibrillation. The availability of large-scale cardiac magnetic resonance imaging data paired with genetic data provides a unique opportunity to assess the genetic contributions to left atrial structure and function, and understand their relationship with risk for atrial fibrillation. Here, we use deep learning and surface reconstruction models to measure left atrial minimum volume, maximum volume, stroke volume, and emptying fraction in 40,558 UK Biobank participants. In a genome-wide association study of 35,049 participants without pre-existing cardiovascular disease, we identify 20 common genetic loci associated with left atrial structure and function. We find that polygenic contributions to increased left atrial volume are associated with atrial fibrillation and its downstream consequences, including stroke. Through Mendelian randomization, we find evidence supporting a causal role for left atrial enlargement and dysfunction on atrial fibrillation risk.

High-throughput RNA isoform sequencing using programmed cDNA concatenation

Full-length RNA-sequencing methods using long-read technologies can capture complete transcript isoforms, but their throughput is limited. We introduce multiplexed arrays isoform sequencing (MAS-ISO-seq), a technique for programmably concatenating complementary DNAs (cDNAs) into molecules optimal for long-read sequencing, increasing the throughput >15-fold to nearly 40 million cDNA reads per run on the Sequel IIe sequencer. When applied to single-cell RNA sequencing of tumor-infiltrating T cells, MAS-ISO-seq demonstrated a 12- to 32-fold increase in the discovery of differentially spliced genes.

Demonstrating paths for unlocking the value of cloud genomics through cross cohort analysis

Recently, large scale genomic projects such as All of Us and the UK Biobank have introduced a new research paradigm where data are stored centrally in cloud-based Trusted Research Environments (TREs). To characterize the advantages and drawbacks of different TRE attributes in facilitating cross-cohort analysis, we conduct a Genome-Wide Association Study of standard lipid measures using two approaches: meta-analysis and pooled analysis. Comparison of full summary data from both approaches with an external study shows strong correlation of known loci with lipid levels (R2 ~ 83-97%). Importantly, 90 variants meet the significance threshold only in the meta-analysis and 64 variants are significant only in pooled analysis, with approximately 20% of variants in each of those groups being most prevalent in non-European, non-Asian ancestry individuals. These findings have important implications, as technical and policy choices lead to cross-cohort analyses generating similar, but not identical results, particularly for non-European ancestral populations.

Unsupervised removal of systematic background noise from droplet-based single-cell experiments using CellBender

Droplet-based single-cell assays, including single-cell RNA sequencing (scRNA-seq), single-nucleus RNA sequencing (snRNA-seq) and cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq), generate considerable background noise counts, the hallmark of which is nonzero counts in cell-free droplets and off-target gene expression in unexpected cell types. Such systematic background noise can lead to batch effects and spurious differential gene expression results. Here we develop a deep generative model based on the phenomenology of noise generation in droplet-based assays. The proposed model accurately distinguishes cell-containing droplets from cell-free droplets, learns the background noise profile and provides noise-free quantification in an end-to-end fashion. We implement this approach in the scalable and robust open-source software package CellBender. Analysis of simulated data demonstrates that CellBender operates near the theoretically optimal denoising limit. Extensive evaluations using real datasets and experimental benchmarks highlight enhanced concordance between droplet-based single-cell data and established gene expression patterns, while the learned background noise profile provides evidence of degraded or uncaptured cell types.

The All of Us Data and Research Center: Creating a Secure, Scalable, and Sustainable Ecosystem for Biomedical Research

The All of Us Research Program's Data and Research Center (DRC) was established to help acquire, curate, and provide access to one of the world's largest and most diverse datasets for precision medicine research. Already, over 500,000 participants are enrolled in All of Us, 80% of whom are underrepresented in biomedical research, and data are being analyzed by a community of over 2,300 researchers. The DRC created this thriving data ecosystem by collaborating with engaged participants, innovative program partners, and empowered researchers. In this review, we first describe how the DRC is organized to meet the needs of this broad group of stakeholders. We then outline guiding principles, common challenges, and innovative approaches used to build the All of Us data ecosystem. Finally, we share lessons learned to help others navigate important decisions and trade-offs in building a modern biomedical data platform.

A framework for individualized splice-switching oligonucleotide therapy

Splice-switching antisense oligonucleotides (ASOs) could be used to treat a subset of individuals with genetic diseases1, but the systematic identification of such individuals remains a challenge. Here we performed whole-genome sequencing analyses to characterize genetic variation in 235 individuals (from 209 families) with ataxia-telangiectasia, a severely debilitating and life-threatening recessive genetic disorder2,3, yielding a complete molecular diagnosis in almost all individuals. We developed a predictive taxonomy to assess the amenability of each individual to splice-switching ASO intervention; 9% and 6% of the individuals had variants that were 'probably' or 'possibly' amenable to ASO splice modulation, respectively. Most amenable variants were in deep intronic regions that are inaccessible to exon-targeted sequencing. We developed ASOs that successfully rescued mis-splicing and ATM cellular signalling in patient fibroblasts for two recurrent variants. In a pilot clinical study, one of these ASOs was used to treat a child who had been diagnosed with ataxia-telangiectasia soon after birth, and showed good tolerability without serious adverse events for three years. Our study provides a framework for the prospective identification of individuals with genetic diseases who might benefit from a therapeutic approach involving splice-switching ASOs.

Genetics of myocardial interstitial fibrosis in the human heart and association with disease

Myocardial interstitial fibrosis is associated with cardiovascular disease and adverse prognosis. Here, to investigate the biological pathways that underlie fibrosis in the human heart, we developed a machine learning model to measure native myocardial T1 time, a marker of myocardial fibrosis, in 41,505 UK Biobank participants who underwent cardiac magnetic resonance imaging. Greater T1 time was associated with diabetes mellitus, renal disease, aortic stenosis, cardiomyopathy, heart failure, atrial fibrillation, conduction disease and rheumatoid arthritis. Genome-wide association analysis identified 11 independent loci associated with T1 time. The identified loci implicated genes involved in glucose transport (SLC2A12), iron homeostasis (HFE, TMPRSS6), tissue repair (ADAMTSL1, VEGFC), oxidative stress (SOD2), cardiac hypertrophy (MYH7B) and calcium signaling (CAMK2D). Using a transforming growth factor β1-mediated cardiac fibroblast activation assay, we found that 9 of the 11 loci consisted of genes that exhibited temporal changes in expression or open chromatin conformation supporting their biological relevance to myofibroblast cell state acquisition. By harnessing machine learning to perform large-scale quantification of myocardial interstitial fibrosis using cardiac imaging, we validate associations between cardiac fibrosis and disease, and identify new biologically relevant pathways underlying fibrosis.

Cross-modal autoencoder framework learns holistic representations of cardiovascular state

A fundamental challenge in diagnostics is integrating multiple modalities to develop a joint characterization of physiological state. Using the heart as a model system, we develop a cross-modal autoencoder framework for integrating distinct data modalities and constructing a holistic representation of cardiovascular state. In particular, we use our framework to construct such cross-modal representations from cardiac magnetic resonance images (MRIs), containing structural information, and electrocardiograms (ECGs), containing myoelectric information. We leverage the learned cross-modal representation to (1) improve phenotype prediction from a single, accessible phenotype such as ECGs; (2) enable imputation of hard-to-acquire cardiac MRIs from easy-to-acquire ECGs; and (3) develop a framework for performing genome-wide association studies in an unsupervised manner. Our results systematically integrate distinct diagnostic modalities into a common representation that better characterizes physiologic state.

The Genetic Determinants of Aortic Distention

BACKGROUND

As the largest conduit vessel, the aorta is responsible for the conversion of phasic systolic inflow from ventricular ejection into more continuous peripheral blood delivery. Systolic distention and diastolic recoil conserve energy and are enabled by the specialized composition of the aortic extracellular matrix. Aortic distensibility decreases with age and vascular disease.

OBJECTIVES

In this study, we sought to discover epidemiologic correlates and genetic determinants of aortic distensibility and strain.

METHODS

We trained a deep learning model to quantify thoracic aortic area throughout the cardiac cycle from cardiac magnetic resonance images and calculated aortic distensibility and strain in 42,342 UK Biobank participants.

RESULTS

Descending aortic distensibility was inversely associated with future incidence of cardiovascular diseases, such as stroke (HR: 0.59 per SD; P = 0.00031). The heritabilities of aortic distensibility and strain were 22% to 25% and 30% to 33%, respectively. Common variant analyses identified 12 and 26 loci for ascending and 11 and 21 loci for descending aortic distensibility and strain, respectively. Of the newly identified loci, 22 were not significantly associated with thoracic aortic diameter. Nearby genes were involved in elastogenesis and atherosclerosis. Aortic strain and distensibility polygenic scores had modest effect sizes for predicting cardiovascular outcomes (delaying or accelerating disease onset by 2%-18% per SD change in scores) and remained statistically significant predictors after accounting for aortic diameter polygenic scores.

CONCLUSIONS

Genetic determinants of aortic function influence risk for stroke and coronary artery disease and may lead to novel targets for medical intervention.

Clinical and genetic associations of deep learning-derived cardiac magnetic resonance-based left ventricular mass

Left ventricular mass is a risk marker for cardiovascular events, and may indicate an underlying cardiomyopathy. Cardiac magnetic resonance is the gold-standard for left ventricular mass estimation, but is challenging to obtain at scale. Here, we use deep learning to enable genome-wide association study of cardiac magnetic resonance-derived left ventricular mass indexed to body surface area within 43,230 UK Biobank participants. We identify 12 genome-wide associations (1 known at TTN and 11 novel for left ventricular mass), implicating genes previously associated with cardiac contractility and cardiomyopathy. Cardiac magnetic resonance-derived indexed left ventricular mass is associated with incident dilated and hypertrophic cardiomyopathies, and implantable cardioverter-defibrillator implant. An indexed left ventricular mass polygenic risk score ≥90th percentile is also associated with incident implantable cardioverter-defibrillator implant in separate UK Biobank (hazard ratio 1.22, 95% CI 1.05-1.44) and Mass General Brigham (hazard ratio 1.75, 95% CI 1.12-2.74) samples. Here, we perform a genome-wide association study of cardiac magnetic resonance-derived indexed left ventricular mass to identify 11 novel variants and demonstrate that cardiac magnetic resonance-derived and genetically predicted indexed left ventricular mass are associated with incident cardiomyopathy.

Machine Learning to Understand Genetic and Clinical Factors Associated With the Pulse Waveform Dicrotic Notch

BACKGROUND

Absence of a dicrotic notch on finger photoplethysmography is an easily ascertainable and inexpensive trait that has been associated with age and prevalent cardiovascular disease. However, the trait exists along a continuum, and little is known about its genetic underpinnings or prognostic value for incident cardiovascular disease.

METHODS

In 169 787 participants in the UK Biobank, we identified absent dicrotic notch on photoplethysmography and created a novel continuous trait reflecting notch smoothness using machine learning. Next, we determined the heritability, genetic basis, polygenic risk, and clinical relations for the binary absent notch trait and the newly derived continuous notch smoothness trait.

RESULTS

Heritability of the continuous notch smoothness trait was 7.5%, compared with 5.6% for the binary absent notch trait. A genome-wide association study of notch smoothness identified 15 significant loci, implicating genes including NT5C2 (P=1.2×10-26), IGFBP3 (P=4.8×10-18), and PHACTR1 (P=1.4×10-13), compared with 6 loci for the binary absent notch trait. Notch smoothness stratified risk of incident myocardial infarction or coronary artery disease, stroke, heart failure, and aortic stenosis. A polygenic risk score for notch smoothness was associated with incident cardiovascular disease and all-cause death in UK Biobank participants without available photoplethysmography data.

CONCLUSIONS

We found that a machine learning derived continuous trait reflecting dicrotic notch smoothness on photoplethysmography was heritable and associated with genes involved in vascular stiffness. Greater notch smoothness was associated with greater risk of incident cardiovascular disease. Raw digital phenotyping may identify individuals at risk for disease via specific genetic pathways.

Systematic single-variant and gene-based association testing of thousands of phenotypes in 394,841 UK Biobank exomes

Genome-wide association studies have successfully discovered thousands of common variants associated with human diseases and traits, but the landscape of rare variations in human disease has not been explored at scale. Exome-sequencing studies of population biobanks provide an opportunity to systematically evaluate the impact of rare coding variations across a wide range of phenotypes to discover genes and allelic series relevant to human health and disease. Here, we present results from systematic association analyses of 4,529 phenotypes using single-variant and gene tests of 394,841 individuals in the UK Biobank with exome-sequence data. We find that the discovery of genetic associations is tightly linked to frequency and is correlated with metrics of deleteriousness and natural selection. We highlight biological findings elucidated by these data and release the dataset as a public resource alongside the Genebass browser for rapidly exploring rare-variant association results.

Spatially Distinct Genetic Determinants of Aortic Dimensions Influence Risks of Aneurysm and Stenosis

BACKGROUND

The left ventricular outflow tract (LVOT) and ascending aorta are spatially complex, with distinct pathologies and embryologic origins. Prior work examined the genetics of thoracic aortic diameter in a single plane.

OBJECTIVES

We sought to elucidate the genetic basis for the diameter of the LVOT, aortic root, and ascending aorta.

METHODS

Using deep learning, we analyzed 2.3 million cardiac magnetic resonance images from 43,317 UK Biobank participants. We computed the diameters of the LVOT, the aortic root, and at 6 locations of ascending aorta. For each diameter, we conducted a genome-wide association study and generated a polygenic score. Finally, we investigated associations between these scores and disease incidence.

RESULTS

A total of 79 loci were significantly associated with at least 1 diameter. Of these, 35 were novel, and most were associated with 1 or 2 diameters. A polygenic score of aortic diameter approximately 13 mm from the sinotubular junction most strongly predicted thoracic aortic aneurysm (n = 427,016; mean HR: 1.42 per SD; 95% CI: 1.34-1.50; P = 6.67 × 10^-21). A polygenic score predicting a smaller aortic root was predictive of aortic stenosis (n = 426,502; mean HR: 1.08 per SD; 95% CI: 1.03-1.12; P = 5 × 10^-6).

CONCLUSIONS

We detected distinct genetic loci underpinning the diameters of the LVOT, aortic root, and at several segments of ascending aorta. We spatially defined a region of aorta whose genetics may be most relevant to predicting thoracic aortic aneurysm. We further described a genetic signature that may predispose to aortic stenosis. Understanding genetic contributions to proximal aortic diameter may enable identification of individuals at risk for aortic disease and facilitate prioritization of therapeutic targets.

Clinical Implementation of Combined Monogenic and Polygenic Risk Disclosure for Coronary Artery Disease

BACKGROUND

State-of-the-art genetic risk interpretation for a common complex disease such as coronary artery disease (CAD) requires assessment for both monogenic variants—such as those related to familial hypercholesterolemia—as well as the cumulative impact of many common variants, as quantified by a polygenic score.

OBJECTIVES

The objective of the study was to describe a combined monogenic and polygenic CAD risk assessment program and examine its impact on patient understanding and changes to clinical management.

METHODS

Study participants attended an initial visit in a preventive genomics clinic and a disclosure visit to discuss results and recommendations, primarily via telemedicine. Digital postdisclosure surveys and chart review evaluated the impact of disclosure.

RESULTS

There were 60 participants (mean age 51 years, 37% women, 72% with no known CAD), including 30 (50%) referred by their cardiologists and 30 (50%) self-referred. Two (3%) participants had a monogenic variant pathogenic for familial hypercholesterolemia, and 19 (32%) had a high polygenic score in the top quintile of the population distribution. In a postdisclosure survey, both the genetic test report (in 80% of participants) and the discussion with the clinician (in 89% of participants) were ranked as very or extremely helpful in understanding the result. Of the 42 participants without CAD, 17 or 40% had a change in management, including statin initiation, statin intensification, or coronary imaging.

CONCLUSIONS

Combined monogenic and polygenic assessments for CAD risk provided by preventive genomics clinics are beneficial for patients and result in changes in management in a significant portion of patients.

Deep learning on resting electrocardiogram to identify impaired heart rate recovery

Background and Objective

Postexercise heart rate recovery (HRR) is an important indicator of cardiac autonomic function and abnormal HRR is associated with adverse outcomes. We hypothesized that deep learning on resting electrocardiogram (ECG) tracings may identify individuals with impaired HRR.

Methods

We trained a deep learning model (convolutional neural network) to infer HRR based on resting ECG waveforms (HRR_pred) among UK Biobank participants who had undergone exercise testing. We examined the association of HRR_pred with incident cardiovascular disease using Cox models, and investigated the genetic architecture of HRR_pred in genome-wide association analysis.

Results

Among 56,793 individuals (mean age 57 years, 51% women), the HRR_pred model was moderately correlated with actual HRR (r = 0.48, 95% confidence interval [CI] 0.47-0.48). Over a median follow-up of 10 years, we observed 2060 incident diabetes mellitus (DM) events, 862 heart failure events, and 2065 deaths. Higher HRR_pred was associated with lower risk of DM (hazard ratio [HR] 0.79 per 1 standard deviation change, 95% CI 0.76-0.83), heart failure (HR 0.89, 95% CI 0.83-0.95), and death (HR 0.83, 95% CI 0.79-0.86). After accounting for resting heart rate, the association of HRR_pred with incident DM and all-cause mortality were similar. Genetic determinants of HRR_pred included known heart rate, cardiac conduction system, cardiomyopathy, and metabolic trait loci.

Conclusion

Deep learning-derived estimates of HRR using resting ECG independently associated with future clinical outcomes, including new-onset DM and all-cause mortality. Inferring postexercise heart rate response from a resting ECG may have potential clinical implications and impact on preventive strategies warrants future study.

Genetic analysis of right heart structure and function in 40,000 people

Congenital heart diseases often involve maldevelopment of the evolutionarily recent right heart chamber. To gain insight into right heart structure and function, we fine-tuned deep learning models to recognize the right atrium, right ventricle and pulmonary artery, measuring right heart structures in 40,000 individuals from the UK Biobank with magnetic resonance imaging. Genome-wide association studies identified 130 distinct loci associated with at least one right heart measurement, of which 72 were not associated with left heart structures. Loci were found near genes previously linked with congenital heart disease, including NKX2-5, TBX5/TBX3, WNT9B and GATA4. A genome-wide polygenic predictor of right ventricular ejection fraction was associated with incident dilated cardiomyopathy (hazard ratio, 1.33 per standard deviation; P = 7.1 × 10-13) and remained significant after accounting for a left ventricular polygenic score. Harnessing deep learning to perform large-scale cardiac phenotyping, our results yield insights into the genetic determinants of right heart structure and function.

Cohort design and natural language processing to reduce bias in electronic health records research

Electronic health record (EHR) datasets are statistically powerful but are subject to ascertainment bias and missingness. Using the Mass General Brigham multi-institutional EHR, we approximated a community-based cohort by sampling patients receiving longitudinal primary care between 2001-2018 (Community Care Cohort Project [C3PO], n = 520,868). We utilized natural language processing (NLP) to recover vital signs from unstructured notes. We assessed the validity of C3PO by deploying established risk models for myocardial infarction/stroke and atrial fibrillation. We then compared C3PO to Convenience Samples including all individuals from the same EHR with complete data, but without a longitudinal primary care requirement. NLP reduced the missingness of vital signs by 31%. NLP-recovered vital signs were highly correlated with values derived from structured fields (Pearson r range 0.95-0.99). Atrial fibrillation and myocardial infarction/stroke incidence were lower and risk models were better calibrated in C3PO as opposed to the Convenience Samples (calibration error range for myocardial infarction/stroke: 0.012-0.030 in C3PO vs. 0.028-0.046 in Convenience Samples; calibration error for atrial fibrillation 0.028 in C3PO vs. 0.036 in Convenience Samples). Sampling patients receiving regular primary care and using NLP to recover missing data may reduce bias and maximize generalizability of EHR research.

Inverting the model of genomics data sharing with the NHGRI Genomic Data Science Analysis, Visualization, and Informatics Lab-space

The NHGRI Genomic Data Science Analysis, Visualization, and Informatics Lab-space (AnVIL; https://anvilproject.org) was developed to address a widespread community need for a unified computing environment for genomics data storage, management, and analysis. In this perspective, we present AnVIL, describe its ecosystem and interoperability with other platforms, and highlight how this platform and associated initiatives contribute to improved genomic data sharing efforts. The AnVIL is a federated cloud platform designed to manage and store genomics and related data, enable population-scale analysis, and facilitate collaboration through the sharing of data, code, and analysis results. By inverting the traditional model of data sharing, the AnVIL eliminates the need for data movement while also adding security measures for active threat detection and monitoring and provides scalable, shared computing resources for any researcher. We describe the core data management and analysis components of the AnVIL, which currently consists of Terra, Gen3, Galaxy, RStudio/Bioconductor, Dockstore, and Jupyter, and describe several flagship genomics datasets available within the AnVIL. We continue to extend and innovate the AnVIL ecosystem by implementing new capabilities, including mechanisms for interoperability and responsible data sharing, while streamlining access management. The AnVIL opens many new opportunities for analysis, collaboration, and data sharing that are needed to drive research and to make discoveries through the joint analysis of hundreds of thousands to millions of genomes along with associated clinical and molecular data types.

ECG-Based Deep Learning and Clinical Risk Factors to Predict Atrial Fibrillation

BACKGROUND

Artificial intelligence (AI)-enabled analysis of 12-lead ECGs may facilitate efficient estimation of incident atrial fibrillation (AF) risk. However, it remains unclear whether AI provides meaningful and generalizable improvement in predictive accuracy beyond clinical risk factors for AF.

METHODS

We trained a convolutional neural network (ECG-AI) to infer 5-year incident AF risk using 12-lead ECGs in patients receiving longitudinal primary care at Massachusetts General Hospital (MGH). We then fit 3 Cox proportional hazards models, composed of ECG-AI 5-year AF probability, CHARGE-AF clinical risk score (Cohorts for Heart and Aging in Genomic Epidemiology-Atrial Fibrillation), and terms for both ECG-AI and CHARGE-AF (CH-AI), respectively. We assessed model performance by calculating discrimination (area under the receiver operating characteristic curve) and calibration in an internal test set and 2 external test sets (Brigham and Women's Hospital [BWH] and UK Biobank). Models were recalibrated to estimate 2-year AF risk in the UK Biobank given limited available follow-up. We used saliency mapping to identify ECG features most influential on ECG-AI risk predictions and assessed correlation between ECG-AI and CHARGE-AF linear predictors.

RESULTS

The training set comprised 45 770 individuals (age 55±17 years, 53% women, 2171 AF events) and the test sets comprised 83 162 individuals (age 59±13 years, 56% women, 2424 AF events). Area under the receiver operating characteristic curve was comparable using CHARGE-AF (MGH, 0.802 [95% CI, 0.767-0.836]; BWH, 0.752 [95% CI, 0.741-0.763]; UK Biobank, 0.732 [95% CI, 0.704-0.759]) and ECG-AI (MGH, 0.823 [95% CI, 0.790-0.856]; BWH, 0.747 [95% CI, 0.736-0.759]; UK Biobank, 0.705 [95% CI, 0.673-0.737]). Area under the receiver operating characteristic curve was highest using CH-AI (MGH, 0.838 [95% CI, 0.807 to 0.869]; BWH, 0.777 [95% CI, 0.766 to 0.788]; UK Biobank, 0.746 [95% CI, 0.716 to 0.776]). Calibration error was low using ECG-AI (MGH, 0.0212; BWH, 0.0129; UK Biobank, 0.0035) and CH-AI (MGH, 0.012; BWH, 0.0108; UK Biobank, 0.0001). In saliency analyses, the ECG P-wave had the greatest influence on AI model predictions. ECG-AI and CHARGE-AF linear predictors were correlated (Pearson r: MGH, 0.61; BWH, 0.66; UK Biobank, 0.41).

CONCLUSIONS

AI-based analysis of 12-lead ECGs has similar predictive usefulness to a clinical risk factor model for incident AF and the approaches are complementary. ECG-AI may enable efficient quantification of future AF risk.

Deep learning enables genetic analysis of the human thoracic aorta

Enlargement or aneurysm of the aorta predisposes to dissection, an important cause of sudden death. We trained a deep learning model to evaluate the dimensions of the ascending and descending thoracic aorta in 4.6 million cardiac magnetic resonance images from the UK Biobank. We then conducted genome-wide association studies in 39,688 individuals, identifying 82 loci associated with ascending and 47 with descending thoracic aortic diameter, of which 14 loci overlapped. Transcriptome-wide analyses, rare-variant burden tests and human aortic single nucleus RNA sequencing prioritized genes including SVIL, which was strongly associated with descending aortic diameter. A polygenic score for ascending aortic diameter was associated with thoracic aortic aneurysm in 385,621 UK Biobank participants (hazard ratio = 1.43 per s.d., confidence interval 1.32-1.54, P = 3.3 × 10-20). Our results illustrate the potential for rapidly defining quantitative traits with deep learning, an approach that can be broadly applied to biomedical images.

Empirical validation of an automated approach to data use oversight

The current paradigm for data use oversight of biomedical datasets is onerous, extending the timescale and resources needed to obtain access for secondary analyses, thus hindering scientific discovery. For a researcher to utilize a controlled-access dataset, a data access committee must review her research plans to determine whether they are consistent with the data use limitations (DULs) specified by the informed consent form. The newly created GA4GH data use ontology (DUO) holds the potential to streamline this process by making data use oversight computable. Here, we describe an open-source software platform, the Data Use Oversight System (DUOS), that connects with DUO terminology to enable automated data use oversight. We analyze dbGaP data acquired since 2006, finding an exponential increase in data access requests, which will not be sustainable with current manual oversight review. We perform an empirical evaluation of DUOS and DUO on selected datasets from the Broad Institute's data repository. We were able to structure 118/123 of the evaluated DULs (96%) and 52/52 (100%) of research proposals using DUO terminology, and we find that DUOS' automated data access adjudication in all cases agreed with the DAC manual review. This first empirical evaluation of the feasibility of automated data use oversight demonstrates comparable accuracy to human-based data access oversight in real-world data governance.

The Data Use Ontology to streamline responsible access to human biomedical datasets

Human biomedical datasets that are critical for research and clinical studies to benefit human health also often contain sensitive or potentially identifying information of individual participants. Thus, care must be taken when they are processed and made available to comply with ethical and regulatory frameworks and informed consent data conditions. To enable and streamline data access for these biomedical datasets, the Global Alliance for Genomics and Health (GA4GH) Data Use and Researcher Identities (DURI) work stream developed and approved the Data Use Ontology (DUO) standard. DUO is a hierarchical vocabulary of human and machine-readable data use terms that consistently and unambiguously represents a dataset's allowable data uses. DUO has been implemented by major international stakeholders such as the Broad and Sanger Institutes and is currently used in annotation of over 200,000 datasets worldwide. Using DUO in data management and access facilitates researchers' discovery and access of relevant datasets. DUO annotations increase the FAIRness of datasets and support data linkages using common data use profiles when integrating the data for secondary analyses. DUO is implemented in the Web Ontology Language (OWL) and, to increase community awareness and engagement, hosted in an open, centralized GitHub repository. DUO, together with the GA4GH Passport standard, offers a new, efficient, and streamlined data authorization and access framework that has enabled increased sharing of biomedical datasets worldwide.

GA4GH: International policies and standards for data sharing across genomic research and healthcare

The Global Alliance for Genomics and Health (GA4GH) aims to accelerate biomedical advances by enabling the responsible sharing of clinical and genomic data through both harmonized data aggregation and federated approaches. The decreasing cost of genomic sequencing (along with other genome-wide molecular assays) and increasing evidence of its clinical utility will soon drive the generation of sequence data from tens of millions of humans, with increasing levels of diversity. In this perspective, we present the GA4GH strategies for addressing the major challenges of this data revolution. We describe the GA4GH organization, which is fueled by the development efforts of eight Work Streams and informed by the needs of 24 Driver Projects and other key stakeholders. We present the GA4GH suite of secure, interoperable technical standards and policy frameworks and review the current status of standards, their relevance to key domains of research and clinical care, and future plans of GA4GH. Broad international participation in building, adopting, and deploying GA4GH standards and frameworks will catalyze an unprecedented effort in data sharing that will be critical to advancing genomic medicine and ensuring that all populations can access its benefits.

GA4GH Passport standard for digital identity and access permissions

The Global Alliance for Genomics and Health (GA4GH) supports international standards that enable a federated data sharing model for the research community while respecting data security, ethical and regulatory frameworks, and data authorization and access processes for sensitive data. The GA4GH Passport standard (Passport) defines a machine-readable digital identity that conveys roles and data access permissions (called "visas") for individual users. Visas are issued by data stewards, including data access committees (DACs) working with public databases, the entities responsible for the quality, integrity, and access arrangements for the datasets in the management of human biomedical data. Passports streamline management of data access rights across data systems by using visas that present a data user's digital identity and permissions across organizations, tools, environments, and services. We describe real-world implementations of the GA4GH Passport standard in use cases from ELIXIR Europe, National Institutes of Health, and the Autism Sharing Initiative. These implementations demonstrate that the Passport standard has provided transparent mechanisms for establishing permissions and authorizing data access across platforms.

Deep Learning to Predict Cardiac Magnetic Resonance-Derived Left Ventricular Mass and Hypertrophy From 12-Lead ECGs

BACKGROUND

Classical methods for detecting left ventricular (LV) hypertrophy (LVH) using 12-lead ECGs are insensitive. Deep learning models using ECG to infer cardiac magnetic resonance (CMR)-derived LV mass may improve LVH detection.

METHODS

Within 32 239 individuals of the UK Biobank prospective cohort who underwent CMR and 12-lead ECG, we trained a convolutional neural network to predict CMR-derived LV mass using 12-lead ECGs (left ventricular mass-artificial intelligence [LVM-AI]). In independent test sets (UK Biobank [n=4903] and Mass General Brigham [MGB, n=1371]), we assessed correlation between LVM-AI predicted and CMR-derived LV mass and compared LVH discrimination using LVM-AI versus traditional ECG-based rules (ie, Sokolow-Lyon, Cornell, lead aVL rule, or any ECG rule). In the UK Biobank and an ambulatory MGB cohort (MGB outcomes, n=28 612), we assessed associations between LVM-AI predicted LVH and incident cardiovascular outcomes using age- and sex-adjusted Cox regression.

RESULTS

LVM-AI predicted LV mass correlated with CMR-derived LV mass in both test sets, although correlation was greater in the UK Biobank (r=0.79) versus MGB (r=0.60, P<0.001 for both). When compared with any ECG rule, LVM-AI demonstrated similar LVH discrimination in the UK Biobank (LVM-AI c-statistic 0.653 [95% CI, 0.608 -0.698] versus any ECG rule c-statistic 0.618 [95% CI, 0.574 -0.663], P=0.11) and superior discrimination in MGB (0.621; 95% CI, 0.592 -0.649 versus 0.588; 95% CI, 0.564 -0.611, P=0.02). LVM-AI-predicted LVH was associated with incident atrial fibrillation, myocardial infarction, heart failure, and ventricular arrhythmias.

CONCLUSIONS

Deep learning-inferred LV mass estimates from 12-lead ECGs correlate with CMR-derived LV mass, associate with incident cardiovascular disease, and may improve LVH discrimination compared to traditional ECG rules.

Deep learning to estimate cardiac magnetic resonance-derived left ventricular mass

Background

Cardiac magnetic resonance (CMR) is the gold standard for left ventricular hypertrophy (LVH) diagnosis. CMR-derived LV mass can be estimated using proprietary algorithms (eg, InlineVF), but their accuracy and availability may be limited.

Objective

To develop an open-source deep learning model to estimate CMR-derived LV mass.

Methods

Within participants of the UK Biobank prospective cohort undergoing CMR, we trained 2 convolutional neural networks to estimate LV mass. The first (ML4Hreg) performed regression informed by manually labeled LV mass (available in 5065 individuals), while the second (ML4Hseg) performed LV segmentation informed by InlineVF (version D13A) contours. We compared ML4Hreg, ML4Hseg, and InlineVF against manually labeled LV mass within an independent holdout set using Pearson correlation and mean absolute error (MAE). We assessed associations between CMR-derived LVH and prevalent cardiovascular disease using logistic regression adjusted for age and sex.

Results

We generated CMR-derived LV mass estimates within 38,574 individuals. Among 891 individuals in the holdout set, ML4Hseg reproduced manually labeled LV mass more accurately (r = 0.864, 95% confidence interval [CI] 0.847-0.880; MAE 10.41 g, 95% CI 9.82-10.99) than ML4Hreg (r = 0.843, 95% CI 0.823-0.861; MAE 10.51, 95% CI 9.86-11.15, P = .01) and InlineVF (r = 0.795, 95% CI 0.770-0.818; MAE 14.30, 95% CI 13.46-11.01, P < .01). LVH defined using ML4Hseg demonstrated the strongest associations with hypertension (odds ratio 2.76, 95% CI 2.51-3.04), atrial fibrillation (1.75, 95% CI 1.37-2.20), and heart failure (4.67, 95% CI 3.28-6.49).

Conclusions

ML4Hseg is an open-source deep learning model providing automated quantification of CMR-derived LV mass. Deep learning models characterizing cardiac structure may facilitate broad cardiovascular discovery.

How Will Machine Learning Inform the Clinical Care of Atrial Fibrillation?

Machine learning applications in cardiology have rapidly evolved in the past decade. With the availability of machine learning tools coupled with vast data sources, the management of atrial fibrillation (AF), a common chronic disease with significant associated morbidity and socioeconomic impact, is undergoing a knowledge and practice transformation in the increasingly complex healthcare environment. Among other advances, deep-learning machine learning methods, including convolutional neural networks, have enabled the development of AF screening pathways using the ubiquitous 12-lead ECG to detect asymptomatic paroxysmal AF in at-risk populations (such as those with cryptogenic stroke), the refinement of AF and stroke prediction schemes through comprehensive digital phenotyping using structured and unstructured data abstraction from the electronic health record or wearable monitoring technologies, and the optimization of treatment strategies, ranging from stroke prophylaxis to monitoring of antiarrhythmic drug (AAD) therapy. Although the clinical and population-wide impact of these tools continues to be elucidated, such transformative progress does not come without challenges, such as the concerns about adopting black box technologies, assessing input data quality for training such models, and the risk of perpetuating rather than alleviating health disparities. This review critically appraises the advances of machine learning related to the care of AF thus far, their potential future directions, and its potential limitations and challenges.

Registered access: authorizing data access

The Global Alliance for Genomics and Health (GA4GH) proposes a data access policy model-"registered access"-to increase and improve access to data requiring an agreement to basic terms and conditions, such as the use of DNA sequence and health data in research. A registered access policy would enable a range of categories of users to gain access, starting with researchers and clinical care professionals. It would also facilitate general use and reuse of data but within the bounds of consent restrictions and other ethical obligations. In piloting registered access with the Scientific Demonstration data sharing projects of GA4GH, we provide additional ethics, policy and technical guidance to facilitate the implementation of this access model in an international setting.

matchbox: An open-source tool for patient matching via the Matchmaker Exchange

Rare disease investigators constantly face challenges in identifying additional cases to build evidence for gene-disease causality. The Matchmaker Exchange (MME) addresses this limitation by providing a mechanism for matching patients across genomic centers via a federated network. The MME has revolutionized searching for additional cases by making it possible to query across institutional boundaries, so that what was once a laborious and manual process of contacting researchers is now automated and computable. However, while the MME network is beginning to scale, the growth of additional nodes is limited by the lack of easy-to-use solutions that can be implemented by any rare disease database owner, even one without significant software engineering resources. Here, we describe matchbox, which is an open-source, platform-independent, portable bridge between any given rare disease genomic center and the MME network, which has already led to novel gene discoveries. We also describe how matchbox greatly reduces the barrier to participation by overcoming challenges for new databases to join the MME.

"Matching" consent to purpose: The example of the Matchmaker Exchange

The Matchmaker Exchange (MME) connects rare disease clinicians and researchers to facilitate the sharing of data from undiagnosed patients for the purpose of novel gene discovery. Such sharing raises the odds that two or more similar patients with candidate genes in common may be found, thereby allowing their condition to be more readily studied and understood. Consent considerations for data sharing in MME included both the ethical and legal differences between clinical and research settings and the level of privacy risk involved in sharing varying amounts of rare disease patient data to enable patient matches. In this commentary, we discuss these consent considerations and the resulting MME Consent Policy as they may be relevant to other international data sharing initiatives.

International Cooperation to Enable the Diagnosis of All Rare Genetic Diseases

Provision of a molecularly confirmed diagnosis in a timely manner for children and adults with rare genetic diseases shortens their "diagnostic odyssey," improves disease management, and fosters genetic counseling with respect to recurrence risks while assuring reproductive choices. In a general clinical genetics setting, the current diagnostic rate is approximately 50%, but for those who do not receive a molecular diagnosis after the initial genetics evaluation, that rate is much lower. Diagnostic success for these more challenging affected individuals depends to a large extent on progress in the discovery of genes associated with, and mechanisms underlying, rare diseases. Thus, continued research is required for moving toward a more complete catalog of disease-related genes and variants. The International Rare Diseases Research Consortium (IRDiRC) was established in 2011 to bring together researchers and organizations invested in rare disease research to develop a means of achieving molecular diagnosis for all rare diseases. Here, we review the current and future bottlenecks to gene discovery and suggest strategies for enabling progress in this regard. Each successful discovery will define potential diagnostic, preventive, and therapeutic opportunities for the corresponding rare disease, enabling precision medicine for this patient population.

The Matchmaker Exchange: a platform for rare disease gene discovery

There are few better examples of the need for data sharing than in the rare disease community, where patients, physicians, and researchers must search for "the needle in a haystack" to uncover rare, novel causes of disease within the genome. Impeding the pace of discovery has been the existence of many small siloed datasets within individual research or clinical laboratory databases and/or disease-specific organizations, hoping for serendipitous occasions when two distant investigators happen to learn they have a rare phenotype in common and can "match" these cases to build evidence for causality. However, serendipity has never proven to be a reliable or scalable approach in science. As such, the Matchmaker Exchange (MME) was launched to provide a robust and systematic approach to rare disease gene discovery through the creation of a federated network connecting databases of genotypes and rare phenotypes using a common application programming interface (API). The core building blocks of the MME have been defined and assembled. Three MME services have now been connected through the API and are available for community use. Additional databases that support internal matching are anticipated to join the MME network as it continues to grow.

Consent Codes: Upholding Standard Data Use Conditions

A systematic way of recording data use conditions that are based on consent permissions as found in the datasets of the main public genome archives (NCBI dbGaP and EMBL-EBI/CRG EGA).

Comparing strategies to fine-map the association of common SNPs at chromosome 9p21 with type 2 diabetes and myocardial infarction

Noncoding variants at human chromosome 9p21 near CDKN2A and CDKN2B are associated with type 2 diabetes, myocardial infarction, aneurysm, vertical cup disc ratio and at least five cancers. Here we compare approaches to more comprehensively assess genetic variation in the region. We carried out targeted sequencing at high coverage in 47 individuals and compared the results to pilot data from the 1000 Genomes Project. We imputed variants into type 2 diabetes and myocardial infarction cohorts directly from targeted sequencing, from a genotyped reference panel derived from sequencing and from 1000 Genomes Project low-coverage data. Polymorphisms with frequency >5% were captured well by all strategies. Imputation of intermediate-frequency polymorphisms required a higher density of tag SNPs in disease samples than is available on first-generation genome-wide association study (GWAS) arrays. Our association analyses identified more comprehensive sets of variants showing equivalent statistical association with type 2 diabetes or myocardial infarction, but did not identify stronger associations than the original GWAS signals.

A framework for variation discovery and genotyping using next-generation DNA sequencing data

Recent advances in sequencing technology make it possible to comprehensively catalog genetic variation in population samples, creating a foundation for understanding human disease, ancestry and evolution. The amounts of raw data produced are prodigious, and many computational steps are required to translate this output into high-quality variant calls. We present a unified analytic framework to discover and genotype variation among multiple samples simultaneously that achieves sensitive and specific results across five sequencing technologies and three distinct, canonical experimental designs. Our process includes (i) initial read mapping; (ii) local realignment around indels; (iii) base quality score recalibration; (iv) SNP discovery and genotyping to find all potential variants; and (v) machine learning to separate true segregating variation from machine artifacts common to next-generation sequencing technologies. We here discuss the application of these tools, instantiated in the Genome Analysis Toolkit, to deep whole-genome, whole-exome capture and multi-sample low-pass (∼4×) 1000 Genomes Project datasets.

Multi-step control of muscle diversity by Hox proteins in the Drosophila embryo

Hox transcription factors control many aspects of animal morphogenetic diversity. The segmental pattern of Drosophila larval muscles shows stereotyped variations along the anteroposterior body axis. Each muscle is seeded by a founder cell and the properties specific to each muscle reflect the expression by each founder cell of a specific combination of 'identity' transcription factors. Founder cells originate from asymmetric division of progenitor cells specified at fixed positions. Using the dorsal DA3 muscle lineage as a paradigm, we show here that Hox proteins play a decisive role in establishing the pattern of Drosophila muscles by controlling the expression of identity transcription factors, such as Nautilus and Collier (Col), at the progenitor stage. High-resolution analysis, using newly designed intron-containing reporter genes to detect primary transcripts, shows that the progenitor stage is the key step at which segment-specific information carried by Hox proteins is superimposed on intrasegmental positional information. Differential control of col transcription by the Antennapedia and Ultrabithorax/Abdominal-A paralogs is mediated by separate cis-regulatory modules (CRMs). Hox proteins also control the segment-specific number of myoblasts allocated to the DA3 muscle. We conclude that Hox proteins both regulate and contribute to the combinatorial code of transcription factors that specify muscle identity and act at several steps during the muscle-specification process to generate muscle diversity.

Diversity and complexity in DNA recognition by transcription factors

Sequence preferences of DNA binding proteins are a primary mechanism by which cells interpret the genome. Despite the central importance of these proteins in physiology, development, and evolution, comprehensive DNA binding specificities have been determined experimentally for only a few proteins. Here, we used microarrays containing all 10-base pair sequences to examine the binding specificities of 104 distinct mouse DNA binding proteins representing 22 structural classes. Our results reveal a complex landscape of binding, with virtually every protein analyzed possessing unique preferences. Roughly half of the proteins each recognized multiple distinctly different sequence motifs, challenging our molecular understanding of how proteins interact with their DNA binding sites. This complexity in DNA recognition may be important in gene regulation and in the evolution of transcriptional regulatory networks.

High-resolution DNA-binding specificity analysis of yeast transcription factors

Transcription factors (TFs) regulate the expression of genes through sequence-specific interactions with DNA-binding sites. However, despite recent progress in identifying in vivo TF binding sites by microarray readout of chromatin immunoprecipitation (ChIP-chip), nearly half of all known yeast TFs are of unknown DNA-binding specificities, and many additional predicted TFs remain uncharacterized. To address these gaps in our knowledge of yeast TFs and their cis regulatory sequences, we have determined high-resolution binding profiles for 89 known and predicted yeast TFs, over more than 2.3 million gapped and ungapped 8-bp sequences ("k-mers"). We report 50 new or significantly different direct DNA-binding site motifs for yeast DNA-binding proteins and motifs for eight proteins for which only a consensus sequence was previously known; in total, this corresponds to over a 50% increase in the number of yeast DNA-binding proteins with experimentally determined DNA-binding specificities. Among other novel regulators, we discovered proteins that bind the PAC (Polymerase A and C) motif (GATGAG) and regulate ribosomal RNA (rRNA) transcription and processing, core cellular processes that are constituent to ribosome biogenesis. In contrast to earlier data types, these comprehensive k-mer binding data permit us to consider the regulatory potential of genomic sequence at the individual word level. These k-mer data allowed us to reannotate in vivo TF binding targets as direct or indirect and to examine TFs' potential effects on gene expression in approximately 1,700 environmental and cellular conditions. These approaches could be adapted to identify TFs and cis regulatory elements in higher eukaryotes.

Predicting the binding preference of transcription factors to individual DNA k-mers

MOTIVATION

Recognition of specific DNA sequences is a central mechanism by which transcription factors (TFs) control gene expression. Many TF-binding preferences, however, are unknown or poorly characterized, in part due to the difficulty associated with determining their specificity experimentally, and an incomplete understanding of the mechanisms governing sequence specificity. New techniques that estimate the affinity of TFs to all possible k-mers provide a new opportunity to study DNA-protein interaction mechanisms, and may facilitate inference of binding preferences for members of a given TF family when such information is available for other family members.

RESULTS

We employed a new dataset consisting of the relative preferences of mouse homeodomains for all eight-base DNA sequences in order to ask how well we can predict the binding profiles of homeodomains when only their protein sequences are given. We evaluated a panel of standard statistical inference techniques, as well as variations of the protein features considered. Nearest neighbour among functionally important residues emerged among the most effective methods. Our results underscore the complexity of TF-DNA recognition, and suggest a rational approach for future analyses of TF families.

Design of compact, universal DNA microarrays for protein binding microarray experiments

Our group has recently developed a compact, universal protein binding microarray (PBM) that can be used to determine the binding preferences of transcription factors (TFs). This design represents all possible sequence variants of a given length k (i.e., all k-mers) on a single array, allowing a complete characterization of the binding specificities of a given TF. Here, we present the mathematical foundations of this design based on de Bruijn sequences generated by linear feedback shift registers. We show that these sequences represent the maximum number of variants for any given set of array dimensions (i.e., number of spots and spot lengths), while also exhibiting desirable pseudo-randomness properties. Moreover, de Bruijn sequences can be selected that represent gapped sequence patterns, further increasing the coverage of the array. This design yields a powerful experimental platform that allows the binding preferences of TFs to be determined with unprecedented resolution.

Variation in homeodomain DNA binding revealed by high-resolution analysis of sequence preferences

Most homeodomains are unique within a genome, yet many are highly conserved across vast evolutionary distances, implying strong selection on their precise DNA-binding specificities. We determined the binding preferences of the majority (168) of mouse homeodomains to all possible 8-base sequences, revealing rich and complex patterns of sequence specificity and showing that there are at least 65 distinct homeodomain DNA-binding activities. We developed a computational system that successfully predicts binding sites for homeodomain proteins as distant from mouse as Drosophila and C. elegans, and we infer full 8-mer binding profiles for the majority of known animal homeodomains. Our results provide an unprecedented level of resolution in the analysis of this simple domain structure and suggest that variation in sequence recognition may be a factor in its functional diversity and evolutionary success.

Inferring condition-specific transcription factor function from DNA binding and gene expression data

Numerous genomic and proteomic datasets are permitting the elucidation of transcriptional regulatory networks in the yeast Saccharomyces cerevisiae. However, predicting the condition dependence of regulatory network interactions has been challenging, because most protein–DNA interactions identified in vivo are from assays performed in one or a few cellular states. Here, we present a novel method to predict the condition-specific functions of S. cerevisiae transcription factors (TFs) by integrating 1327 microarray gene expression data sets and either comprehensive TF binding site data from protein binding microarrays (PBMs) or in silico motif data. Importantly, our method does not impose arbitrary thresholds for calling target regions ‘bound' or genes ‘differentially expressed', but rather allows all the information derived from a TF binding or gene expression experiment to be considered. We show that this method can identify environmental, physical, and genetic interactions, as well as distinct sets of genes that might be activated or repressed by a single TF under particular conditions. This approach can be used to suggest conditions for directed in vivo experimentation and to predict TF function.

Compact, universal DNA microarrays to comprehensively determine transcription-factor binding site specificities

Transcription factors (TFs) interact with specific DNA regulatory sequences to control gene expression throughout myriad cellular processes. However, the DNA binding specificities of only a small fraction of TFs are sufficiently characterized to predict the sequences that they can and cannot bind. We present a maximally compact, synthetic DNA sequence design for protein binding microarray (PBM) experiments that represents all possible DNA sequence variants of a given length k (that is, all 'k-mers') on a single, universal microarray. We constructed such all k-mer microarrays covering all 10-base pair (bp) binding sites by converting high-density single-stranded oligonucleotide arrays to double-stranded (ds) DNA arrays. Using these microarrays we comprehensively determined the binding specificities over a full range of affinities for five TFs of different structural classes from yeast, worm, mouse and human. The unbiased coverage of all k-mers permits high-throughput interrogation of binding site preferences, including nucleotide interdependencies, at unprecedented resolution.

Expression-guided in silico evaluation of candidate cis regulatory codes for Drosophila muscle founder cells

While combinatorial models of transcriptional regulation can be inferred for metazoan systems from a priori biological knowledge, validation requires extensive and time-consuming experimental work. Thus, there is a need for computational methods that can evaluate hypothesized cis regulatory codes before the difficult task of experimental verification is undertaken. We have developed a novel computational framework (termed “CodeFinder”) that integrates transcription factor binding site and gene expression information to evaluate whether a hypothesized transcriptional regulatory model (TRM; i.e., a set of co-regulating transcription factors) is likely to target a given set of co-expressed genes. Our basic approach is to simultaneously predict cis regulatory modules (CRMs) associated with a given gene set and quantify the enrichment for combinatorial subsets of transcription factor binding site motifs comprising the hypothesized TRM within these predicted CRMs. As a model system, we have examined a TRM experimentally demonstrated to drive the expression of two genes in a sub-population of cells in the developing Drosophila mesoderm, the somatic muscle founder cells. This TRM was previously hypothesized to be a general mode of regulation for genes expressed in this cell population. In contrast, the present analyses suggest that a modified form of this cis regulatory code applies to only a subset of founder cell genes, those whose gene expression responds to specific genetic perturbations in a similar manner to the gene on which the original model was based. We have confirmed this hypothesis by experimentally discovering six (out of 12 tested) new CRMs driving expression in the embryonic mesoderm, four of which drive expression in founder cells.

Although genome sequences and much gene expression data are readily available, the determination of sets of transcription factors regulating particular gene expression patterns remains a problem of fundamental importance. Tissue-specific gene expression in developing animals is regulated through the combinatorial interactions of transcription factors with DNA regulatory elements termed cis regulatory modules (CRMs). Although genetic and biochemical experiments allow the identification of transcription factors and CRMs, those experiments are laborious and time-consuming. Philippakis et al. introduce a new approach (termed “CodeFinder”) for quantifying the enrichment for particular combinations of transcription factor binding site motifs within predicted CRMs associated with a given gene set of interest, identified from gene expression data. The authors' analyses allowed them to discover a specific combination of transcription factor binding site motifs that constitute a core cis regulatory code for expression of a particular subset of genes in muscle founder cells, an embryonic cell population in the developing fruit fly (Drosophila melanogaster) mesoderm, and also led them to the discovery and subsequent experimental validation of novel, tissue-specific CRMs. Importantly, the CodeFinder approach is generally applicable, and thus could be used to support, refute, or refine a known or hypothesized cis regulatory code for any biological system or genome of interest.

Modulefinder: a tool for computational discovery of cis regulatory modules

Regulation of gene expression occurs largely through the binding of sequence-specific transcription factors (TFs) to genomic binding sites (BSs). We present a rigorous scoring scheme, implemented as a C program termed "ModuleFinder", that evaluates the likelihood that a given genomic region is a cis regulatory module (CRM) for an input set of TFs according to its degree of: (1) homotypic site clustering; (2) heterotypic site clustering; and (3) evolutionary conservation across multiple genomes. Importantly, ModuleFinder obtains all parameters needed to appropriately weight the relative contributions of these sequence features directly from the input sequences and TFBS motifs, and does not need to first be trained. Using two previously described collections of experimentally verified CRMs in mammals and in fly as validation datasets, we show that ModuleFinder is able to identify CRMs with great sensitivity and specificity.