Search

Abstract A prominent trend in single-cell transcriptomics is providing spatial context alongside a characterization of each cell’s molecular state. This typically requires targeting an a priori selection of genes, often covering less than 1% of the genome, and a key question is how to optimally determine the small gene panel. We address this challenge by introducing a flexible deep learning framework, PERSIST, to identify informative gene targets for spatial transcriptomics studies by leveraging reference scRNA-seq data. Using datasets spanning different brain regions, species, and scRNA-seq technologies, we show that PERSIST reliably identifies panels that provide more accurate prediction of the genome-wide expression profile, thereby capturing more information with fewer genes. PERSIST can be adapted to specific biological goals, and we demonstrate that PERSIST’s binarization of gene expression levels enables models trained on scRNA-seq data to generalize with to spatial transcriptomics data, despite the complex shift between these technologies.

Abstract As interest in using unsupervised deep learning models to analyze gene expression data has grown, an increasing number of methods have been developed to make these models more interpretable. These methods can be separated into two groups: post hoc analyses of black box models through feature attribution methods and approaches to build inherently interpretable models through biologically-constrained architectures. We argue that these approaches are not mutually exclusive, but can in fact be usefully combined. We propose PAUSE ( https://github.com/suinleelab/PAUSE ), an unsupervised pathway attribution method that identifies major sources of transcriptomic variation when combined with biologically-constrained neural network models.

Qui et al. present a new approach, IMPACT, that uses explainable artificial intelligence to analyze all-cause mortality. IMPACT provides insights into the individualized mortality risk scores, while maintaining high model accuracy and the expressive power to capture complex, non-linear relationships between mortality and individuals’ features.

Series of machine learning models, relevant for tasks in biology, medicine, and finance, usually involve complex feature attribution techniques. The authors introduce a tractable method to compute local feature attributions for a series of machine learning models inspired by connections to the Shapley value.

Abstract Hundreds of millions of surgical procedures take place annually across the world, which generate a prevalent type of electronic health record (EHR) data comprising time series physiological signals. Here, we present a transferable embedding method (i.e., a method to transform time series signals into input features for predictive machine learning models) named PHASE (PHysiologicAl Signal Embeddings) that enables us to more accurately forecast adverse surgical outcomes based on physiological signals. We evaluate PHASE on minute-by-minute EHR data of more than 50,000 surgeries from two operating room (OR) datasets and patient stays in an intensive care unit (ICU) dataset. PHASE outperforms other state-of-the-art approaches, such as long-short term memory networks trained on raw data and gradient boosted trees trained on handcrafted features, in predicting six distinct outcomes: hypoxemia, hypocapnia, hypotension, hypertension, phenylephrine, and epinephrine. In a transfer learning setting where we train embedding models in one dataset then embed signals and predict adverse events in unseen data, PHASE achieves significantly higher prediction accuracy at lower computational cost compared to conventional approaches. Finally, given the importance of understanding models in clinical applications we demonstrate that PHASE is explainable and validate our predictive models using local feature attribution methods.

The molecular basis of Alzheimer’s Disease has been obscured by heterogeneity and scarcity of brain gene expression data, which limit effectiveness in complex models. Here, the authors introduce a multi-task deep learning framework to learn generalizable and nuanced relationships between gene expression and neuropathology.

Air pollution might affect atherosclerosis through DNA methylation changes in cells crucial to atherosclerosis, such as monocytes. We conducted an epigenome-wide study of DNA methylation in CD14+ monocytes and long-term ambient air pollution exposure in adults participating in the Multi-Ethnic Study of Atherosclerosis (MESA). We also assessed the association between differentially methylated signals and cis-gene expression. Using spatiotemporal models, one-year average concentrations of outdoor fine particulate matter (PM2.5) and oxides of nitrogen (NOX) were estimated at participants’ homes. We assessed DNA methylation and gene expression using Illumina 450k and HumanHT-12 v4 Expression BeadChips, respectively (n = 1,207). We used bump hunting and site-specific approaches to identify differentially methylated signals (false discovery rate of 0.05) and used linear models to assess associations between differentially methylated signals and cis-gene expression. Four differentially methylated regions (DMRs) located on chromosomes 5, 6, 7, and 16 (within or near SDHAP3, ZFP57, HOXA5, and PRM1, respectively) were associated with PM2.5. The DMRs on chromosomes 5 and 6 also associated with NOX. The DMR on chromosome 5 had the smallest p-value for both PM2.5 (p = 1.4×10−6) and NOX (p = 7.7×10−6). Three differentially methylated CpGs were identified for PM2.5, and cg05926640 (near TOMM20) had the smallest p-value (p = 5.6×10−8). NOX significantly associated with cg11756214 within ZNF347 (p = 5.6×10−8). Several differentially methylated signals were also associated with cis-gene expression. The DMR located on chromosome 7 was associated with the expression of HOXA5, HOXA9, and HOXA10. The DMRs located on chromosomes 5 and 16 were associated with expression of MRPL36 and DEXI, respectively. The CpG cg05926640 was associated with expression of ARID4B, IRF2BP2, and TOMM20. We identified differential DNA methylation in monocytes associated with long-term air pollution exposure. Methylation signals associated with gene expression might help explain how air pollution contributes to cardiovascular disease.

Identification of markers of drug response is essential for precision therapy. Here the authors introduce an algorithm that uses prior information about each gene’s importance in AML to identify the most predictive gene-drug associations from transcriptome and drug response data from 30 AML samples.

Human facial asymmetry is due to a complex interaction of genetic and environmental factors. To identify genetic influences on facial asymmetry, we developed a method for automated scoring that summarizes local morphology features and their spatial distribution. A genome-wide association study using asymmetry scores from two local symmetry features was conducted and significant genetic associations were identified for one asymmetry feature, including genes thought to play a role in craniofacial disorders and development: NFATC1, SOX5, NBAS, and TCF7L1. These results provide evidence that normal variation in facial asymmetry may be impacted by common genetic variants and further motivate the development of automated summaries of complex phenotypes.

Triple-negative breast cancer is a heterogeneous disease characterized by poor clinical outcomes and a shortage of targeted treatment options. To discover molecular features of triple-negative breast cancer, we performed quantitative proteomics analysis of twenty human-derived breast cell lines and four primary breast tumors to a depth of more than 12,000 distinct proteins. We used this data to identify breast cancer subtypes at the protein level and demonstrate the precise quantification of biomarkers, signaling proteins, and biological pathways by mass spectrometry. We integrated proteomics data with exome sequence resources to identify genomic aberrations that affect protein expression. We performed a high-throughput drug screen to identify protein markers of drug sensitivity and understand the mechanisms of drug resistance. The genome and proteome provide complementary information that, when combined, yield a powerful engine for therapeutic discovery. This resource is available to the cancer research community to catalyze further analysis and investigation.

We present a computational framework, called DISCERN (DIfferential SparsE Regulatory Network), to identify informative topological changes in gene-regulator dependence networks inferred on the basis of mRNA expression datasets within distinct biological states. DISCERN takes two expression datasets as input: an expression dataset of diseased tissues from patients with a disease of interest and another expression dataset from matching normal tissues. DISCERN estimates the extent to which each gene is perturbed-having distinct regulator connectivity in the inferred gene-regulator dependencies between the disease and normal conditions. This approach has distinct advantages over existing methods. First, DISCERN infers conditional dependencies between candidate regulators and genes, where conditional dependence relationships discriminate the evidence for direct interactions from indirect interactions more precisely than pairwise correlation. Second, DISCERN uses a new likelihood-based scoring function to alleviate concerns about accuracy of the specific edges inferred in a particular network. DISCERN identifies perturbed genes more accurately in synthetic data than existing methods to identify perturbed genes between distinct states. In expression datasets from patients with acute myeloid leukemia (AML), breast cancer and lung cancer, genes with high DISCERN scores in each cancer are enriched for known tumor drivers, genes associated with the biological processes known to be important in the disease, and genes associated with patient prognosis, in the respective cancer. Finally, we show that DISCERN can uncover potential mechanisms underlying network perturbation by explaining observed epigenomic activity patterns in cancer and normal tissue types more accurately than alternative methods, based on the available epigenomic data from the ENCODE project.

Genome-wide RNA expression data provide a detailed view of an organism's biological state; hence, a dataset measuring expression variation between genetically diverse individuals (eQTL data) may provide important insights into the genetics of complex traits. However, with data from a relatively small number of individuals, it is difficult to distinguish true causal polymorphisms from the large number of possibilities. The problem is particularly challenging in populations with significant linkage disequilibrium, where traits are often linked to large chromosomal regions containing many genes. Here, we present a novel method, Lirnet, that automatically learns a regulatory potential for each sequence polymorphism, estimating how likely it is to have a significant effect on gene expression. This regulatory potential is defined in terms of "regulatory features"-including the function of the gene and the conservation, type, and position of genetic polymorphisms-that are available for any organism. The extent to which the different features influence the regulatory potential is learned automatically, making Lirnet readily applicable to different datasets, organisms, and feature sets. We apply Lirnet both to the human HapMap eQTL dataset and to a yeast eQTL dataset and provide statistical and biological results demonstrating that Lirnet produces significantly better regulatory programs than other recent approaches. We demonstrate in the yeast data that Lirnet can correctly suggest a specific causal sequence variation within a large, linked chromosomal region. In one example, Lirnet uncovered a novel, experimentally validated connection between Puf3-a sequence-specific RNA binding protein-and P-bodies-cytoplasmic structures that regulate translation and RNA stability-as well as the particular causative polymorphism, a SNP in Mkt1, that induces the variation in the pathway.