Causal disease effect sizes of proximal single-nucleotide polymorphisms (SNPs) are widely assumed to be independent but could be correlated. Here we introduce a new method, linkage disequilibrium SNP-pair effect correlation regression (LDSPEC), to estimate the correlation of causal disease effect sizes of derived alleles between proximal SNPs; LDSPEC produced robust estimates in simulations. Analyzing 70 UK Biobank diseases and traits (average N = 305,646), we detected significantly non-zero SNP-pair effect correlations (for example, -0.37 ± 0.09 for low-frequency positive linkage disequilibrium 0-100-bp SNP pairs) that decayed with distance and varied with allele frequency and linkage disequilibrium between SNPs. SNP pairs with shared functions had stronger effect correlations that spanned longer genomic distances. Consequently, SNP heritability estimates were smaller than estimates of the sum of causal effect size variances across SNPs, particularly for certain functional annotations. We recapitulated our findings via forward simulations involving stabilizing selection, implicating the action of linkage masking, whereby haplotypes containing linked SNPs with opposite effects on disease have reduced effects on fitness and escape negative selection.

Leveraging multi-ancestry data can improve fine-mapping power. We propose MultiSuSiE, an extension of Sum of Single Effects (SuSiE), to multiple ancestries that allows causal effect sizes to vary across ancestries. We evaluated MultiSuSiE using whole-genome sequencing data from 47,000 African-ancestry, 36,000 Latino-ancestry and 116,000 European-ancestry individuals from All of Us. In simulations, MultiSuSiE applied to Afr36k + Lat36k + Eur36k was well-calibrated and attained higher power than SuSiE applied to Eur109k; compared to recent multi-ancestry methods (SuSiEx and MESuSiE), MultiSuSiE attained higher power and lower computational cost. In analyses of 14 quantitative traits, MultiSuSiE applied to Afr47k + Lat36k + Eur116k identified 348 fine-mapped variants with posterior inclusion probability (PIP) > 0.9, and MultiSuSiE applied to Afr36k + Lat36k + Eur36k identified 59

Myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS) is a complex, heterogeneous, and systemic disease defined by a suite of symptoms, including unexplained persistent fatigue, post-exertional malaise (PEM), cognitive impairment, myalgia, orthostatic intolerance, and unrefreshing sleep. The disease mechanism of ME/CFS is unknown, with no effective curative treatments. In this study, we present a multi-site ME/CFS whole-genome analysis, which is powered by a novel deep learning framework, HEAL2. We show that HEAL2 not only has predictive value for ME/CFS based on personal rare variants, but also links genetic risk to various ME/CFS-associated symptoms. Model interpretation of HEAL2 identifies 115 ME/CFS-risk genes that exhibit significant intolerance to loss-of-function (LoF) mutations. Transcriptome and network analyses highlight the functional importance of these genes across a wide range of tissues and cell types, including the central nervous system (CNS) and immune cells. Patient-derived multi-omics data implicate reduced expression of ME/CFS risk genes within ME/CFS patients, including in the plasma proteome, and the transcriptomes of B and T cells, especially cytotoxic CD4 T cells, supporting their disease relevance. Pan-phenotype analysis of ME/CFS genes further reveals the genetic correlation between ME/CFS and other complex diseases and traits, including depression and long COVID-19. Overall, HEAL2 provides a candidate genetic-based diagnostic tool for ME/CFS, and our findings contribute to a comprehensive understanding of the genetic, molecular, and cellular basis of ME/CFS, yielding novel insights into therapeutic targets. Our deep learning model also offers a potent, broadly applicable framework for parallel rare variant analysis and genetic prediction for other complex diseases and traits.

Heritable diseases often manifest in a highly tissue-specific manner, with different disease loci mediated by genes in distinct tissues or cell types. We propose Tissue-Gene Fine-Mapping (TGFM), a fine-mapping method that infers the posterior probability (PIP) for each gene-tissue pair to mediate a disease locus by analyzing GWAS summary statistics (and in-sample LD) and leveraging eQTL data from diverse tissues to build cis-predicted expression models; TGFM also assigns PIPs to causal variants that are not mediated by gene expression in assayed genes and tissues. TGFM accounts for both co-regulation across genes and tissues and LD between SNPs (generalizing existing fine-mapping methods), and incorporates genome-wide estimates of each tissue's contribution to disease as tissue-level priors. TGFM was well-calibrated and moderately well-powered in simulations; unlike previous methods, TGFM was able to attain correct calibration by modeling uncertainty in cis-predicted expression models. We applied TGFM to 45 UK Biobank diseases/traits (average N = 316K) using eQTL data from 38 GTEx tissues. TGFM identified an average of 147 PIP > 0.5 causal genetic elements per disease/trait, of which 11% were gene-tissue pairs. Implicated gene-tissue pairs were concentrated in known disease-critical tissues, and causal genes were strongly enriched in disease-relevant gene sets. Causal gene-tissue pairs identified by TGFM recapitulated known biology (e.g., TPO-thyroid for Hypothyroidism), but also included biologically plausible novel findings (e.g., SLC20A2-artery aorta for Diastolic blood pressure). Further application of TGFM to single-cell eQTL data from 9 cell types in peripheral blood mononuclear cells (PBMC), analyzed jointly with GTEx tissues, identified 30 additional causal gene-PBMC cell type pairs at PIP > 0.5-primarily for autoimmune disease and blood cell traits, including the biologically plausible example of CD52 in classical monocyte cells for Monocyte count. In conclusion, TGFM is a robust and powerful method for fine-mapping causal tissues and genes at disease-associated loci.

Maximum Inner Product Search (MIPS) is a ubiquitous task in machine learning applications such as recommendation systems. Given a query vector and $n$ atom vectors in $d$-dimensional space, the goal of MIPS is to find the atom that has the highest inner product with the query vector. Existing MIPS algorithms scale at least as $O(\sqrt{d})$, which becomes computationally prohibitive in high-dimensional settings. In this work, we present BanditMIPS, a novel randomized MIPS algorithm whose complexity is independent of $d$. BanditMIPS estimates the inner product for each atom by subsampling coordinates and adaptively evaluates more coordinates for more promising atoms. The specific adaptive sampling strategy is motivated by multi-armed bandits. We provide theoretical guarantees that BanditMIPS returns the correct answer with high probability, while improving the complexity in $d$ from $O(\sqrt{d})$ to $O(1)$. We also perform experiments on four synthetic and real-world datasets and demonstrate that BanditMIPS outperforms prior state-of-the-art algorithms. For example, in the Movie Lens dataset ($n$=4,000, $d$=6,000), BanditMIPS is 20$\times$ faster than the next best algorithm while returning the same answer. BanditMIPS requires no preprocessing of the data and includes a hyperparameter that practitioners may use to trade off accuracy and runtime. We also propose a variant of our algorithm, named BanditMIPS-$\alpha$, which achieves further speedups by employing non-uniform sampling across coordinates. Finally, we demonstrate how known preprocessing techniques can be used to further accelerate BanditMIPS, and discuss applications to Matching Pursuit and Fourier analysis.

Genome-wide association studies (GWASs) have identified tens of thousands of disease associated variants and provided critical insights into developing effective treatments. However, limited sample sizes have hindered the discovery of variants for uncommon and rare diseases. Here, we introduce KGWAS, a novel geometric deep learning method that leverages a massive functional knowledge graph across variants and genes to improve detection power in small-cohort GWASs significantly. KGWAS assesses the strength of a variant's association to disease based on the aggregate GWAS evidence across molecular elements interacting with the variant within the knowledge graph. Comprehensive simulations and replication experiments showed that, for small sample sizes ( N =1-10K), KGWAS identified up to 100% more statistically significant associations than state-of-the-art GWAS methods and achieved the same statistical power with up to 2.67× fewer samples. We applied KGWAS to 554 uncommon UK Biobank diseases ( N case <5K) and identified 183 more associations (46.9% improvement) than the original GWAS, where the gain further increases to 79.8% for 141 rare diseases (N case <300). The KGWAS-only discoveries are supported by abundant functional evidence, such as rs2155219 (on 11q13) associated with ulcerative colitis potentially via regulating LRRC32 expression in CD4+ regulatory T cells, and rs7312765 (on 12q12) associated with the rare disease myasthenia gravis potentially via regulating PPHLN1 expression in neuron-related cell types. Furthermore, KGWAS consistently improves downstream analyses such as identifying disease-specific network links for interpreting GWAS variants, identifying disease-associated genes, and identifying disease-relevant cell populations. Overall, KGWAS is a flexible and powerful AI model that integrates growing functional genomics data to discover novel variants, genes, cells, and networks, especially valuable for small cohort diseases.

The human hippocampus and prefrontal cortex play critical roles in learning and cognition1,2, yet the dynamic molecular characteristics of their development remain enigmatic. Here we investigated the epigenomic and three-dimensional chromatin conformational reorganization during the development of the hippocampus and prefrontal cortex, using more than 53,000 joint single-nucleus profiles of chromatin conformation and DNA methylation generated by single-nucleus methyl-3C sequencing (snm3C-seq3)3. The remodelling of DNA methylation is temporally separated from chromatin conformation dynamics. Using single-cell profiling and multimodal single-molecule imaging approaches, we have found that short-range chromatin interactions are enriched in neurons, whereas long-range interactions are enriched in glial cells and non-brain tissues. We reconstructed the regulatory programs of cell-type development and differentiation, finding putatively causal common variants for schizophrenia strongly overlapping with chromatin loop-connected, cell-type-specific regulatory regions. Our data provide multimodal resources for studying gene regulatory dynamics in brain development and demonstrate that single-cell three-dimensional multi-omics is a powerful approach for dissecting neuropsychiatric risk loci. Using a single-nucleus multi-omics approach, a study jointly profiles the reorganization of the epigenome and the three-dimensional chromatin conformation during the development of the human hippocampus and prefrontal cortex.

Myasthenia gravis (MG) is etiologically associated with thymus abnormalities, but its pathology in the thymus remains unclear. In this study, we attempt to narrow down the features associated with MG using spatial transcriptome analysis of thymoma and thymic hyperplasia samples. We find that the majority of thymomas are constituted by the cortical region. However, the small medullary region is enlarged in seropositive thymomas and contains polygenic enrichment and MG-specific germinal center structures. Neuromuscular medullary thymic epithelial cells, previously identified as MG-specific autoantigen-producing cells, are enriched in the cortico-medullary junction. The medulla is characterized by a specific chemokine pattern and immune cell composition, including migratory dendritic cells and effector regulatory T cells. Similar germinal center structures and immune microenvironments are also observed in the thymic hyperplasia medulla. This study shows that the medulla and junction areas are linked to MG pathology and provides insights into future MG research.

Clustering is a fundamental task in data science with wide-ranging applications. In $k$-medoids clustering, cluster centers must be actual datapoints and arbitrary distance metrics may be used; these features allow for greater interpretability of the cluster centers and the clustering of exotic objects in $k$-medoids clustering, respectively. $k$-medoids clustering has recently grown in popularity due to the discovery of more efficient $k$-medoids algorithms. In particular, recent research has proposed BanditPAM, a randomized $k$-medoids algorithm with state-of-the-art complexity and clustering accuracy. In this paper, we present BanditPAM++, which accelerates BanditPAM via two algorithmic improvements, and is $O(k)$ faster than BanditPAM in complexity and substantially faster than BanditPAM in wall-clock runtime. First, we demonstrate that BanditPAM has a special structure that allows the reuse of clustering information $\textit{within}$ each iteration. Second, we demonstrate that BanditPAM has additional structure that permits the reuse of information $\textit{across}$ different iterations. These observations inspire our proposed algorithm, BanditPAM++, which returns the same clustering solutions as BanditPAM but often several times faster. For example, on the CIFAR10 dataset, BanditPAM++ returns the same results as BanditPAM but runs over 10$\times$ faster. Finally, we provide a high-performance C++ implementation of BanditPAM++, callable from Python and R, that may be of interest to practitioners at https://github.com/motiwari/BanditPAM. Auxiliary code to reproduce all of our experiments via a one-line script is available at https://github.com/ThrunGroup/BanditPAM_plusplus_experiments.

The genetic architecture of human diseases and complex traits has been extensively studied, but little is known about the relationship of causal disease effect sizes between proximal SNPs, which have largely been assumed to be independent. We introduce a new method, LD SNP-pair effect correlation regression (LDSPEC), to estimate the correlation of causal disease effect sizes of derived alleles between proximal SNPs, depending on their allele frequencies, LD, and functional annotations; LDSPEC produced robust estimates in simulations across various genetic architectures. We applied LDSPEC to 70 diseases and complex traits from the UK Biobank (average N=306K), meta-analyzing results across diseases/traits. We detected significantly nonzero effect correlations for proximal SNP pairs (e.g., -0.37±0.09 for low-frequency positive-LD 0-100bp SNP pairs) that decayed with distance (e.g., -0.07±0.01 for low-frequency positive-LD 1-10kb), varied with allele frequency (e.g., -0.15±0.04 for common positive-LD 0-100bp), and varied with LD between SNPs (e.g., +0.12±0.05 for common negative-LD 0-100bp) (because we consider derived alleles, positive-LD and negative-LD SNP pairs may yield very different results). We further determined that SNP pairs with shared functions had stronger effect correlations that spanned longer genomic distances, e.g., -0.37±0.08 for low-frequency positive-LD same-gene promoter SNP pairs (average genomic distance of 47kb (due to alternative splicing)) and -0.32±0.04 for low-frequency positive-LD H3K27ac 0-1kb SNP pairs. Consequently, SNP-heritability estimates were substantially smaller than estimates of the sum of causal effect size variances across all SNPs (ratio of 0.87±0.02 across diseases/traits), particularly for certain functional annotations (e.g., 0.78±0.01 for common Super enhancer SNPs)-even though these quantities are widely assumed to be equal. We recapitulated our findings via forward simulations with an evolutionary model involving stabilizing selection, implicating the action of linkage masking, whereby haplotypes containing linked SNPs with opposite effects on disease have reduced effects on fitness and escape negative selection.

The analysis of longitudinal data from electronic health records (EHRs) has the potential to improve clinical diagnoses and enable personalized medicine, motivating efforts to identify disease subtypes from patient comorbidity information. Here we introduce an age-dependent topic modeling (ATM) method that provides a low-rank representation of longitudinal records of hundreds of distinct diseases in large EHR datasets. We applied ATM to 282,957 UK Biobank samples, identifying 52 diseases with heterogeneous comorbidity profiles; analyses of 211,908 All of Us samples produced concordant results. We defined subtypes of the 52 heterogeneous diseases based on their comorbidity profiles and compared genetic risk across disease subtypes using polygenic risk scores (PRSs), identifying 18 disease subtypes whose PRS differed significantly from other subtypes of the same disease. We further identified specific genetic variants with subtype-dependent effects on disease risk. In conclusion, ATM identifies disease subtypes with differential genome-wide and locus-specific genetic risk profiles.

In autoimmune diseases such as rheumatoid arthritis, the immune system attacks the body's own cells. Developing a precise understanding of the cell states where noncoding autoimmune risk variants impart causal mechanisms is critical to developing curative therapies. Here, to identify noncoding regions with accessible chromatin that associate with cell-state-defining gene expression patterns, we leveraged multimodal single-nucleus RNA and assay for transposase-accessible chromatin (ATAC) sequencing data across 28,674 cells from the inflamed synovial tissue of 12 donors. Specifically, we used a multivariate Poisson model to predict peak accessibility from single-nucleus RNA sequencing principal components. For 14 autoimmune diseases, we discovered that cell-state-dependent ('dynamic') chromatin accessibility peaks in immune cell types were enriched for heritability, compared with cell-state-invariant ('cs-invariant') peaks. These dynamic peaks marked regulatory elements associated with T peripheral helper, regulatory T, dendritic and STAT1+CXCL10+ myeloid cell states. We argue that dynamic regulatory elements can help identify precise cell states enriched for disease-critical genetic variation.

The human frontal cortex and hippocampus play critical roles in learning and cognition. We investigated the epigenomic and 3D chromatin conformational reorganization during the development of the frontal cortex and hippocampus, using more than 53,000 joint single-nucleus profiles of chromatin conformation and DNA methylation (sn-m3C-seq). The remodeling of DNA methylation predominantly occurs during late-gestational to early-infant development and is temporally separated from chromatin conformation dynamics. Neurons have a unique Domain-Dominant chromatin conformation that is different from the Compartment-Dominant conformation of glial cells and non-brain tissues. We reconstructed the regulatory programs of cell-type differentiation and found putatively causal common variants for schizophrenia strongly overlap with chromatin loop-connected, cell-type-specific regulatory regions. Our data demonstrate that single-cell 3D-regulome is an effective approach for dissecting neuropsychiatric risk loci.

Attempts to identify and prioritize functional DNA elements in coding and non-coding regions, particularly through use of in silico functional annotation data, continue to increase in popularity. However, specific functional roles can vary widely from one variant to another, making it challenging to summarize different aspects of variant function with a one-dimensional rating. Here we propose multi-dimensional annotation-class integrative estimation (MACIE), an unsupervised multivariate mixed-model framework capable of integrating annotations of diverse origin to assess multi-dimensional functional roles for both coding and non-coding variants. Unlike existing one-dimensional scoring methods, MACIE views variant functionality as a composite attribute encompassing multiple characteristics and estimates the joint posterior functional probabilities of each genomic position. This estimate offers more comprehensive and interpretable information in the presence of multiple aspects of functionality. Applied to a variety of independent coding and non-coding datasets, MACIE demonstrates powerful and robust performance in discriminating between functional and non-functional variants. We also show an application of MACIE to fine-mapping and heritability enrichment analysis by using the lipids GWAS summary statistics data from the European Network for Genetic and Genomic Epidemiology Consortium.

Post-deployment monitoring of ML systems is critical for ensuring reliability, especially as new user inputs can differ from the training distribution. Here we propose a novel approach, MLDemon, for ML DEployment MONitoring. MLDemon integrates both unlabeled data and a small amount of on-demand labels to produce a real-time estimate of the ML model's current performance on a given data stream. Subject to budget constraints, MLDemon decides when to acquire additional, potentially costly, expert supervised labels to verify the model. On temporal datasets with diverse distribution drifts and models, MLDemon outperforms existing approaches. Moreover, we provide theoretical analysis to show that MLDemon is minimax rate optimal for a broad class of distribution drifts.

Random forests are some of the most widely used machine learning models today, especially in domains that necessitate interpretability. We present an algorithm that accelerates the training of random forests and other popular tree-based learning methods. At the core of our algorithm is a novel node-splitting subroutine, dubbed MABSplit, used to efficiently find split points when constructing decision trees. Our algorithm borrows techniques from the multi-armed bandit literature to judiciously determine how to allocate samples and computational power across candidate split points. We provide theoretical guarantees that MABSplit improves the sample complexity of each node split from linear to logarithmic in the number of data points. In some settings, MABSplit leads to 100x faster training (an 99% reduction in training time) without any decrease in generalization performance. We demonstrate similar speedups when MABSplit is used across a variety of forest-based variants, such as Extremely Random Forests and Random Patches. We also show our algorithm can be used in both classification and regression tasks. Finally, we show that MABSplit outperforms existing methods in generalization performance and feature importance calculations under a fixed computational budget. All of our experimental results are reproducible via a one-line script at https://github.com/ThrunGroup/FastForest.

Dysbiosis of the oral microbiome mediates chronic periodontal disease. Realignment of microbial dysbiosis towards health may prevent disease. Treatment with antibiotics and probiotics can modulate the microbial, immunological, and clinical landscape of periodontal disease with some success. Antibacterial peptides or bacteriocins, such as nisin, and a nisin-producing probiotic, Lactococcus lactis, have not been examined in this context, yet warrant examination because of their biomedical benefits in eradicating biofilms and pathogenic bacteria, modulating immune mechanisms, and their safety profile in humans. This study’s goal was to examine the potential for nisin and a nisin-producing probiotic to abrogate periodontal bone loss, the host inflammatory response, and changes in oral microbiome composition in a polymicrobial mouse model of periodontal disease. Nisin and a nisin-producing Lactococcus lactis probiotic significantly decreased the levels of several periodontal pathogens, alveolar bone loss, and the oral and systemic inflammatory host response. Surprisingly, nisin and/or the nisin-producing L. lactis probiotic enhanced the population of fibroblasts and osteoblasts despite the polymicrobial infection. Nisin mediated human periodontal ligament cell proliferation dose-dependently by increasing the proliferation marker, Ki-67. Nisin and probiotic treatment significantly shifted the oral microbiome towards the healthy control state; health was associated with Proteobacteria, whereas 3 retroviruses were associated with disease. Disease-associated microbial species were correlated with IL-6 levels. Nisin or nisin-producing probiotic’s ability to shift the oral microbiome towards health, mitigate periodontal destruction and the host immune response, and promote a novel proliferative phenotype in reparative connective tissue cells, addresses key aspects of the pathogenesis of periodontal disease and reveals a new biomedical application for nisin in treatment of periodontitis and reparative medicine.

Gene expression at the individual cell-level resolution, as quantified by single-cell RNA-sequencing (scRNA-seq), can provide unique insights into the pathology and cellular origin of diseases and complex traits. Here, we introduce single-cell Disease Relevance Score (scDRS), an approach that links scRNA-seq with polygenic risk of disease at individual cell resolution; scDRS identifies individual cells that show excess expression levels for genes in a disease-specific gene set constructed from GWAS data. We determined via simulations that scDRS is well-calibrated and powerful in identifying individual cells associated to disease. We applied scDRS to GWAS data from 74 diseases and complex traits (average N=341K) in conjunction with 16 scRNA-seq data sets spanning 1.3 million cells from 31 tissues and organs. At the cell type level, scDRS broadly recapitulated known links between classical cell types and disease, and also produced novel biologically plausible findings. At the individual cell level, scDRS identified subpopulations of disease-associated cells that are not captured by existing cell type labels, including subpopulations of CD4+ T cells associated with inflammatory bowel disease, partially characterized by their effector-like states; subpopulations of hippocampal CA1 pyramidal neurons associated with schizophrenia, partially characterized by their spatial location at the proximal part of the hippocampal CA1 region; and subpopulations of hepatocytes associated with triglyceride levels, partially characterized by their higher ploidy levels. At the gene level, we determined that genes whose expression across individual cells was correlated with the scDRS score (thus reflecting co-expression with GWAS disease genes) were strongly enriched for gold-standard drug target and Mendelian disease genes.