Search

Genotype-phenotype association is found in many biological systems such as brains and brain diseases. However, predicting phenotypes from genotypes remains challenging, primarily due to complex underlying molecular and cellular mechanisms. Emerging multi-modal data enables studying such mechanisms at different scales. However, most of these approaches fail to incorporate biology into the machine learning models. Due to the black-box nature of many machine learning techniques, it is challenging to integrate these multi-modalities and interpret the results for biological insights, especially when some modality is missing. To this end, we developed DeepGAMI, an interpretable deep learning model to improve genotype-phenotype prediction from multi-modal data. DeepGAMI uses prior biological knowledge to define the neural network architecture. Notably, it embeds an auxiliary-learning layer for cross-modal imputation while training the model from multi-modal data. Using this pre-trained layer, we can impute latent features of additional modalities and thus enable predicting phenotypes from a single modality only. Finally, the model uses integrated gradient approach to prioritize multi-modal features and links for phenotypes. We applied DeepGAMI to (1) population-level bulk and cell-type-specific genotype and gene expression data for Schizophrenia (SCZ) cohort, (2) genotype and gene expression data for Alzheimer's Disease (AD) cohort, and (3) recent single-cell multi-modal data comprising transcriptomics and electrophysiology for neuronal cells in the mouse visual cortex. We found that DeepGAMI outperforms existing state-of-the-art methods and provides a profound understanding of gene regulatory mechanisms at cellular resolution from genotype to phenotype. DeepGAMI is an open-source tool and is available at https://github.com/daifengwanglab/DeepGAMI.

Additional file 6. Supplementary figures and analysis of PEER factors to mitigate confounding effects of inter-sample cellular heterogeneity.

BACKGROUND: Genome-wide association studies have found many genetic risk variants associated with Alzheimer’s disease (AD). However, how these risk variants affect deeper phenotypes such as disease progression and immune response remains elusive. Also, our understanding of cellular and molecular mechanisms from disease risk variants to various phenotypes is still limited. To address these problems, we performed integrative multi-omics analysis from genotype, transcriptomics, and epigenomics for revealing gene regulatory mechanisms from disease variants to AD phenotypes. METHOD: First, we cluster gene co-expression networks and identify gene modules for various AD phenotypes given population gene expression data. Next, we predict the transcription factors (TFs) that significantly regulate the genes in each module and the AD risk variants (e.g., SNPs) interrupting the TF binding sites on the regulatory elements. Finally, we construct a full gene regulatory network linking SNPs, interrupted TFs, and regulatory elements to target genes for each phenotype. This network thus provides mechanistic insights of gene regulation from disease risk variants to AD phenotypes. RESULTS: We applied our analysis to predict the gene regulatory networks in three major AD-relevant regions: hippocampus, dorsolateral prefrontal cortex (DLPFC), and lateral temporal lobe (LTL). These region networks provide a comprehensive functional genomic map linking AD SNPs to TFs and regulatory elements to target genes for various AD phenotypes. Comparative analyses further revealed cross-region-conserved and region-specific regulatory networks. For instance, AD SNPs rs13404184 and rs61068452 disrupt the bindings of TF SPI1 that regulates AD gene INPP5D in the hippocampus and lateral temporal lobe. However, SNP rs117863556 interrupts the bindings of TF REST to regulate GAB2 in the DLPFC only. Furthermore, driven by recent discoveries between AD and Covid-19, we found that many genes from our networks regulating Covid-19 pathways are also significantly differentially expressed in severe Covid patients (ICU), suggesting potential regulatory connections between AD and Covid. Thus, we used the machine learning models to predict severe Covid and prioritized highly predictive genes as AD-Covid genes. We also used Decision Curve Analysis to show that our AD-Covid genes outperform known Covid-19 genes for predicting Covid severity and deciding to send patients to ICU or not. In short, our results provide a deeper understanding of the interplay among multi-omics, brain regions, and AD phenotypes, including disease progression and Covid response. Our analysis is open-source available at https://github.com/daifengwanglab/ADSNPheno .

Transcription factor (TF) coordination plays a key role in target gene (TG) regulation via protein-protein interactions (PPIs) and DNA co-binding to regulatory elements. Single-cell technologies facilitate gene expression measurement for individual cells and cell-type identification, yet the connection between TF coordination and TG regulation of various cell types remains unclear. To address this, we have developed a novel computational approach, Network Regression Embeddings (NetREm), to reveal cell-type TF-TF coordination activities for TG regulation. NetREm leverages network-constrained regularization using prior knowledge of direct and/or indirect PPIs among TFs to analyze single-cell gene expression data. We test NetREm by simulation data and benchmark its performance in 4 real-world applications that have gold standard TF-TG networks available: mouse (mESCs) and simulated human (hESCs) embryonic stem (ESCs), human hematopoietic stem (HSCs), and mouse dendritic (mDCs) cells. Further, we use NetREm to prioritize valid novel TF-TF coordination links in human Peripheral Blood Mononuclear cell (PBMC) sub-types. We apply NetREm to analyze various cell types in both central (CNS) and peripheral (PNS) nerve system (NS) (e.g. neuronal, glial, Schwann cells (SCs)) as well as in Alzheimers disease (AD). Our findings uncover cell-type coordinating TFs and identify new TF-TG candidate links. We validate our top predictions using Cut&Run and knockout loss-of-function expression data in rat/mouse models and compare results with additional functional genomic data, including expression quantitative trait loci (eQTL) and Genome-Wide Association Studies (GWAS) to link genetic variants (single nucleotide polymorphisms (SNPs)) to TF coordination.

Injury to adult mammalian central nervous system (CNS) axons results in limited regeneration. Rodent studies have revealed a developmental switch in CNS axon regenerative ability, yet whether this is conserved in humans is unknown. Using human fibroblasts from 8 gestational-weeks to 72 years-old, we performed direct reprogramming to transdifferentiate fibroblasts into induced neurons (Fib-iNs), avoiding pluripotency which restores cells to an embryonic state. We found that early gestational Fib-iNs grew longer neurites than all other ages, mirroring the developmental switch in regenerative ability in rodents. RNA-sequencing and screening revealed ARID1A as a developmentally-regulated modifier of neurite growth in human neurons. These data suggest that age-specific epigenetic changes may drive the intrinsic loss of neurite growth ability in human CNS neurons during development. One-Sentence Summary: Directly-reprogrammed human neurons demonstrate a developmental decrease in neurite growth ability.

Background: Genome-wide association studies have found many genetic risk variants associated with Alzheimer's disease (AD). However, how these risk variants affect deeper phenotypes such as disease progression and immune response remains elusive. Also, our understanding of cellular and molecular mechanisms from disease risk variants to various phenotypes is still limited. To address these problems, we performed an integrative multi-omics analysis of genotype, transcriptomics, and epigenomics for revealing gene regulatory mechanisms from disease variants to AD phenotypes., Method: First, given the population gene expression data of a cohort, we construct and cluster its gene co-expression network to identify gene co-expression modules for various AD phenotypes. Next, we predict transcription factors (TFs) regulating co-expressed genes and AD risk SNPs that interrupt TF binding sites on regulatory elements. Finally, we construct a gene regulatory network (GRN) linking SNPs, interrupted TFs, and regulatory elements to target genes and gene modules for each phenotype in the cohort. This network thus provides systematic insights into gene regulatory mechanisms from risk variants to AD phenotypes., Results: Our analysis predicted GRNs in three major AD-relevant regions: Hippocampus, Dorsolateral Prefrontal Cortex (DLPFC), Lateral Temporal Lobe (LTL). Comparative analyses revealed cross-region-conserved and region-specific GRNs, in which many immunological genes are present. For instance, SNPs rs13404184 and rs61068452 disrupt SPI1 binding and regulation of AD gene INPP5D in the Hippocampus and LTL. However, SNP rs117863556 interrupts bindings of REST to regulate GAB2 in DLPFC only. Driven by emerging neuroinflammation in AD, we used Covid-19 as a proxy to identify possible regulatory mechanisms for neuroimmunology in AD. To this end, we looked at the GRN subnetworks relating to genes from shared AD-Covid pathways. From those subnetworks, our machine learning analysis prioritized the AD-Covid genes for predicting Covid-19 severity. Decision Curve Analysis also validated our AD-Covid genes outperform known Covid-19 genes for classifying severe Covid-19 patients. This suggests AD-Covid genes along with linked SNPs can be potential novel biomarkers for neuroimmunology in AD. Finally, our results are open-source available as a comprehensive functional genomic map for AD, providing a deeper mechanistic understanding of the interplay among multi-omics, brain regions, gene functions like neuroimmunology, and phenotypes.

The search for potential antibody-based diagnostics, vaccines, and therapeutics for pandemic severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) has focused almost exclusively on the spike (S) and nucleocapsid (N) proteins. Coronavirus membrane (M), ORF3a, and ORF8 proteins are humoral immunogens in other coronaviruses (CoVs) but remain largely uninvestigated for SARS-CoV-2. Here we use ultradense peptide microarray mapping to show that SARS-CoV-2 infection induces robust antibody responses to epitopes throughout the SARS-CoV-2 proteome, particularly in M, in which one epitope achieved excellent diagnostic accuracy. We map 79 B cell epitopes throughout the SARS-CoV-2 proteome and demonstrate that antibodies that develop in response to SARS-CoV-2 infection bind homologous peptide sequences in the six other known human CoVs. We also confirm reactivity against four of our top-ranking epitopes by enzyme-linked immunosorbent assay (ELISA). Illness severity correlated with increased reactivity to nine SARS-CoV-2 epitopes in S, M, N, and ORF3a in our population. Our results demonstrate previously unknown, highly reactive B cell epitopes throughout the full proteome of SARS-CoV-2 and other CoV proteins.