Current challenges in developing foundational models for volumetric imaging data, such as magnetic resonance imaging (MRI), stem from the computational complexity of state-of-the-art architectures in high dimensions and curating sufficiently large datasets of volumes. To address these challenges, we introduce Raptor (Random Planar Tensor Reduction), a train-free method for generating semantically rich embeddings for volumetric data. Raptor leverages a frozen 2D foundation model, pretrained on natural images, to extract visual tokens from individual cross-sections of medical volumes. These tokens are then spatially compressed using random projections, significantly reducing computational complexity while retaining rich semantic information. Extensive experiments on 10 diverse medical volume tasks verify the superior performance of Raptor over state-of-the-art methods, including those pretrained exclusively on medical volumes (+3% SuPreM, +6% MISFM, +10% Merlin, +13% VoCo, and +14% SLIViT), while entirely bypassing the need for costly training. Our results highlight Raptor’s effectiveness and versatility as a foundation for advancing deep learning-based methods for medical volumes (code: github.com/sriramlab/raptor).

We present CACTI, a masked autoencoding approach for imputing tabular data that leverages the structure in missingness patterns and contextual information. Our approach employs a novel median truncated copy masking training strategy that encourages the model to learn from empirical patterns of missingness while incorporating semantic relationships between features — captured by column names and text descriptions — to better represent feature dependence. These dual sources of inductive bias enable CACTIto outperform state-of-the-art methods — an average $R^2$ gain of 7.8\% over the next best method (13.4%, 6.1%, and 5.3% under missing not at random, at random and completely at random, respectively) — across a diverse range of datasets and missingness conditions. Our results highlight the value of leveraging dataset-specific contextual information and missingness patterns to enhance imputation performance.

Multivariate analysis is becoming central in studies investigating high-throughput molecular data, yet, some important features of these data are seldom explored. Here, we present MANOCCA (Multivariate Analysis of Conditional CovAriance), a powerful method to test for the effect of a predictor on the covariance matrix of a multivariate outcome. The proposed test is by construction orthogonal to tests based on the mean and variance and is able to capture effects that are missed by both approaches. We first compare the performances of MANOCCA with existing correlation-based methods and show that MANOCCA is the only test correctly calibrated in simulation mimicking omics data. We then investigate the impact of reducing the dimensionality of the data using principal component analysis when the sample size is smaller than the number of pairwise covariance terms analysed. We show that, in many realistic scenarios, the maximum power can be achieved with a limited number of components. Finally, we apply MANOCCA to 1000 healthy individuals from the Milieu Interieur cohort, to assess the effect of health, lifestyle and genetic factors on the covariance of two sets of phenotypes, blood biomarkers and flow cytometry-based immune phenotypes. Our analyses identify significant associations between multiple factors and the covariance of both omics data.

Structural variation (SV) refers to insertions, deletions, inversions, and duplications in human genomes. SVs are present in approximately 1.5% of the human genome. Still, this small subset of genetic variation has been implicated in the pathogenesis of psoriasis, Crohn's disease and other autoimmune disorders, autism spectrum and other neurodevelopmental disorders, and schizophrenia. Since identifying structural variants is an important problem in genetics, several specialized computational techniques have been developed to detect structural variants directly from sequencing data. With advances in whole-genome sequencing (WGS) technologies, a plethora of SV detection methods have been developed. However, dissecting SVs from WGS data remains a challenge, with the majority of SV detection methods prone to a high false-positive rate, and no existing method able to precisely detect a full range of SVs present in a sample. Previous studies have shown that none of the existing SV callers can maintain high accuracy across various SV lengths and genomic coverages. Here, we report an integrated structural variant calling framework, Variant Identification and Structural Variant Analysis (VISTA), that leverages the results of individual callers using a novel and robust filtering and merging algorithm. In contrast to existing consensus-based tools which ignore the length and coverage, VISTA overcomes this limitation by executing various combinations of top-performing callers based on variant length and genomic coverage to generate SV events with high accuracy. We evaluated the performance of VISTA on comprehensive gold-standard datasets across varying organisms and coverage. We benchmarked VISTA using the Genome-in-a-Bottle gold standard SV set, haplotype-resolved de novo assemblies from the Human Pangenome Reference Consortium, along with an in-house polymerase chain reaction (PCR)-validated mouse gold standard set. VISTA maintained the highest F1 score among top consensus-based tools measured using a comprehensive gold standard across both mouse and human genomes. VISTA also has an optimized mode, where the calls can be optimized for precision or recall. VISTA-optimized can attain 100% precision and the highest sensitivity among other variant callers. In conclusion, VISTA represents a significant advancement in structural variant calling, offering a robust and accurate framework that outperforms existing consensus-based tools and sets a new standard for SV detection in genomic research.

The ability to predict human phenotypes and identify biomarkers of disease from metagenomic data is crucial for the development of therapeutics for microbiome-associated diseases. However, metagenomic data is commonly affected by technical variables unrelated to the phenotype of interest, such as sequencing protocol, which can make it difficult to predict phenotype and find biomarkers of disease. Supervised methods to correct for background noise, originally designed for gene expression and RNA-seq data, are commonly applied to microbiome data but may be limited because they cannot account for unmeasured sources of variation. Unsupervised approaches address this issue, but current methods are limited because they are ill-equipped to deal with the unique aspects of microbiome data, which is compositional, highly skewed, and sparse. We perform a comparative analysis of the ability of different denoising transformations in combination with supervised correction methods as well as an unsupervised principal component correction approach that is presently used in other domains but has not been applied to microbiome data to date. We find that the unsupervised principal component correction approach has comparable ability in reducing false discovery of biomarkers as the supervised approaches, with the added benefit of not needing to know the sources of variation apriori. However, in prediction tasks, it appears to only improve prediction when technical variables contribute to the majority of variance in the data. As new and larger metagenomic datasets become increasingly available, background noise correction will become essential for generating reproducible microbiome analyses.

The holdout randomization test (HRT) discovers a set of covariates most predictive of a response. Given the covariate distribution, HRTs can explicitly control the false discovery rate (FDR). However, if this distribution is unknown and must be estimated from data, HRTs can inflate the FDR. To alleviate the inflation of FDR, we propose the contrarian randomization test (CONTRA), which is designed explicitly for scenarios where the covariate distribution must be estimated from data and may even be misspecified. Our key insight is to use an equal mixture of two “contrarian” probabilistic models in determining the importance of a covariate. One model is fit with the real data, while the other is fit using the same data, but with the covariate being tested replaced with samples from an estimate of the covariate distribution. CONTRA is flexible enough to achieve a power of 1 asymptotically, can reduce the FDR compared to state-of-the-art CVS methods when the covariate distribution is misspecified, and is computationally efficient in high dimensions and large sample sizes. We further demonstrate the effectiveness of CONTRA on numerous synthetic benchmarks, and highlight its capabilities on a genetic dataset.

In recent years, methods were proposed for assigning feature importance scores to measure the contribution of individual features. While in some cases the goal is to understand a specific model, in many cases the goal is to understand the contribution of certain properties (features) to a real-world phenomenon. Thus, a distinction has been made between feature importance scores that explain a model and scores that explain the data. When explaining the data, machine learning models are used as proxies in settings where conducting many real-world experiments is expensive or prohibited. While existing feature importance scores show great success in explaining models, we demonstrate their limitations when explaining the data, especially in the presence of correlations between features. Therefore, we develop a set of axioms to capture properties expected from a feature importance score when explaining data and prove that there exists only one score that satisfies all of them, the Marginal Contribution Feature Importance (MCI). We analyze the theoretical properties of this score function and demonstrate its merits empirically.

Linear mixed models (LMMs) can be applied in the meta-analyses of responses from individuals across multiple contexts, increasing power to detect associations while accounting for confounding effects arising from within-individual variation. However, traditional approaches to fitting these models can be computationally intractable. Here, we describe an efficient and exact method for fitting a multiple-context linear mixed model. Whereas existing exact methods may be cubic in their time complexity with respect to the number of individuals, our approach for multiple-context LMMs (mcLMM) is linear. These improvements allow for large-scale analyses requiring computing time and memory magnitudes of order less than existing methods. As examples, we apply our approach to identify expression quantitative trait loci from large-scale gene expression data measured across multiple tissues as well as joint analyses of multiple phenotypes in genome-wide association studies at biobank scale.

MOTIVATION: While gene-environment (GxE) interactions contribute importantly to many different phenotypes, detecting such interactions requires well-powered studies and has proven difficult. To address this, we combine two approaches to improve GxE power: simultaneously evaluating multiple phenotypes and using a two-step analysis approach. Previous work shows that the power to identify a main genetic effect can be improved by simultaneously analyzing multiple related phenotypes. For a univariate phenotype, two-step methods produce higher power for detecting a GxE interaction compared to single step analysis. Therefore, we propose a two-step approach to test for an overall GxE effect for multiple phenotypes. RESULTS: Using simulations we demonstrate that, when more than one phenotype has GxE effect (i.e. GxE pleiotropy), our approach offers substantial gain in power (18-43%) to detect an aggregate-level GxE effect for a multivariate phenotype compared to an analogous two-step method to identify GxE effect for a univariate phenotype. We applied the proposed approach to simultaneously analyze three lipids, LDL, HDL and Triglyceride with the frequency of alcohol consumption as environmental factor in the UK Biobank. The method identified two loci with an overall GxE effect on the vector of lipids, one of which was missed by the competing approaches. AVAILABILITY AND IMPLEMENTATION: We provide an R package MPGE implementing the proposed approach which is available from CRAN: https://cran.r-project.org/web/packages/MPGE/index.html. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Admixture, the interbreeding between previously distinct populations, is a pervasive force in evolution. The evolutionary history of populations in the presence of admixture can be modeled by augmenting phylogenetic trees with additional nodes that represent admixture events. While enabling a more faithful representation of evolutionary history, admixture graphs present formidable inferential challenges, and there is an increasing need for methods that are accurate, fully automated and computationally efficient. One key challenge arises from the size of the space of admixture graphs. Given that exhaustively evaluating all admixture graphs can be prohibitively expensive, heuristics have been developed to enable efficient search over this space. One heuristic, implemented in the popular method TreeMix, consists of adding edges to a starting tree while optimizing a suitable objective function. RESULTS: Here, we present a demographic model (with one admixed population incident to a leaf) where TreeMix and any other starting-tree-based maximum likelihood heuristic using its likelihood function is guaranteed to get stuck in a local optimum and return an incorrect network topology. To address this issue, we propose a new search strategy that we term maximum likelihood network orientation (MLNO). We augment TreeMix with an exhaustive search for an MLNO, referring to this approach as OrientAGraph. In evaluations including previously published admixture graphs, OrientAGraph outperformed TreeMix on 4/8 models (there are no differences in the other cases). Overall, OrientAGraph found graphs with higher likelihood scores and topological accuracy while remaining computationally efficient. Lastly, our study reveals several directions for improving maximum likelihood admixture graph estimation. AVAILABILITY AND IMPLEMENTATION: OrientAGraph is available on Github (https://github.com/sriramlab/OrientAGraph) under the GNU General Public License v3.0. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

The number of variants that have a non-zero effect on a trait (i.e. polygenicity) is a fundamental parameter in the study of the genetic architecture of a complex trait. Although many previous studies have investigated polygenicity at a genome-wide scale, a detailed understanding of how polygenicity varies across genomic regions is currently lacking. In this work, we propose an accurate and scalable statistical framework to estimate regional polygenicity for a complex trait. We show that our approach yields approximately unbiased estimates of regional polygenicity in simulations across a wide-range of various genetic architectures. We then partition the polygenicity of anthropometric and blood pressure traits across 6-Mb genomic regions (N = 290K, UK Biobank) and observe that all analyzed traits are highly polygenic: over one-third of regions harbor at least one causal variant for each of the traits analyzed. Additionally, we observe wide variation in regional polygenicity: on average across all traits, 48.9% of regions contain at least 5 causal SNPs, 5.44% of regions contain at least 50 causal SNPs. Finally, we find that heritability is proportional to polygenicity at the regional level, which is consistent with the hypothesis that heritability enrichments are largely driven by the variation in the number of causal SNPs.

A common workflow in data exploration is to learn a low-dimensional representation of the data, identify groups of points in that representation, and examine the differences between the groups to determine what they represent. We treat this workflow as an interpretable machine learning problem by leveraging the model that learned the low-dimensional representation to help identify the key differences between the groups. To solve this problem, we introduce a new type of explanation, a Global Counterfactual Explanation (GCE), and our algorithm, Transitive Global Translations (TGT), for computing GCEs. TGT identifies the differences between each pair of groups using compressed sensing but constrains those pairwise differences to be consistent among all of the groups. Empirically, we demonstrate that TGT is able to identify explanations that accurately explain the model while being relatively sparse, and that these explanations match real patterns in the data.

Motivation: A large proportion of risk regions identified by genome-wide association studies (GWAS) are shared across multiple diseases and traits. Understanding whether this clustering is due to sharing of causal variants or chance colocalization can provide insights into shared etiology of complex traits and diseases. Results: In this work, we propose a flexible, unifying framework to quantify the overlap between a pair of traits called UNITY (Unifying Non-Infinitesimal Trait analYsis). We formulate a Bayesian generative model that relates the overlap between pairs of traits to GWAS summary statistic data under a non-infinitesimal genetic architecture underlying each trait. We propose a Metropolis-Hastings sampler to compute the posterior density of the genetic overlap parameters in this model. We validate our method through comprehensive simulations and analyze summary statistics from height and body mass index GWAS to show that it produces estimates consistent with the known genetic makeup of both traits. Availability and implementation: The UNITY software is made freely available to the research community at: https://github.com/bogdanlab/UNITY. Supplementary information: Supplementary data are available at Bioinformatics online.

Motivation: Heritability, the proportion of variation in a trait that can be explained by genetic variation, is an important parameter in efforts to understand the genetic architecture of complex phenotypes as well as in the design and interpretation of genome-wide association studies. Attempts to understand the heritability of complex phenotypes attributable to genome-wide single nucleotide polymorphism (SNP) variation data has motivated the analysis of large datasets as well as the development of sophisticated tools to estimate heritability in these datasets. Linear mixed models (LMMs) have emerged as a key tool for heritability estimation where the parameters of the LMMs, i.e. the variance components, are related to the heritability attributable to the SNPs analyzed. Likelihood-based inference in LMMs, however, poses serious computational burdens. Results: We propose a scalable randomized algorithm for estimating variance components in LMMs. Our method is based on a method-of-moment estimator that has a runtime complexity O(NMB) for N individuals and M SNPs (where B is a parameter that controls the number of random matrix-vector multiplications). Further, by leveraging the structure of the genotype matrix, we can reduce the time complexity to O(NMBmax( log⁡3N, log⁡3M)). We demonstrate the scalability and accuracy of our method on simulated as well as on empirical data. On standard hardware, our method computes heritability on a dataset of 500 000 individuals and 100 000 SNPs in 38 min. Availability and implementation: The RHE-reg software is made freely available to the research community at: https://github.com/sriramlab/RHE-reg.

MOTIVATION: A basic problem of broad public and scientific interest is to use the DNA of an individual to infer the genomic ancestries of the parents. In particular, we are often interested in the fraction of each parent's genome that comes from specific ancestries (e.g. European, African, Native American, etc). This has many applications ranging from understanding the inheritance of ancestry-related risks and traits to quantifying human assortative mating patterns. RESULTS: We model the problem of parental genomic ancestry inference as a pooled semi-Markov process. We develop a general mathematical framework for pooled semi-Markov processes and construct efficient inference algorithms for these models. Applying our inference algorithm to genotype data from 231 Mexican trios and 258 Puerto Rican trios where we have the true genomic ancestry of each parent, we demonstrate that our method accurately infers parameters of the semi-Markov processes and parents' genomic ancestries. We additionally validated the method on simulations. Our model of pooled semi-Markov process and inference algorithms may be of independent interest in other settings in genomics and machine learning.

MOTIVATION: Local ancestry analysis of genotype data from recently admixed populations (e.g. Latinos, African Americans) provides key insights into population history and disease genetics. Although methods for local ancestry inference have been extensively validated in simulations (under many unrealistic assumptions), no empirical study of local ancestry accuracy in Latinos exists to date. Hence, interpreting findings that rely on local ancestry in Latinos is challenging. RESULTS: Here, we use 489 nuclear families from the mainland USA, Puerto Rico and Mexico in conjunction with 3204 unrelated Latinos from the Multiethnic Cohort study to provide the first empirical characterization of local ancestry inference accuracy in Latinos. Our approach for identifying errors does not rely on simulations but on the observation that local ancestry in families follows Mendelian inheritance. We measure the rate of local ancestry assignments that lead to Mendelian inconsistencies in local ancestry in trios (MILANC), which provides a lower bound on errors in the local ancestry estimates. We show that MILANC rates observed in simulations underestimate the rate observed in real data, and that MILANC varies substantially across the genome. Second, across a wide range of methods, we observe that loci with large deviations in local ancestry also show enrichment in MILANC rates. Therefore, local ancestry estimates at such loci should be interpreted with caution. Finally, we reconstruct ancestral haplotype panels to be used as reference panels in local ancestry inference and show that ancestry inference is significantly improved by incoroprating these reference panels. AVAILABILITY AND IMPLEMENTATION: We provide the reconstructed reference panels together with the maps of MILANC rates as a public resource for researchers analyzing local ancestry in Latinos at http://bogdanlab.pathology.ucla.edu SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: It is becoming increasingly evident that the analysis of genotype data from recently admixed populations is providing important insights into medical genetics and population history. Such analyses have been used to identify novel disease loci, to understand recombination rate variation and to detect recent selection events. The utility of such studies crucially depends on accurate and unbiased estimation of the ancestry at every genomic locus in recently admixed populations. Although various methods have been proposed and shown to be extremely accurate in two-way admixtures (e.g. African Americans), only a few approaches have been proposed and thoroughly benchmarked on multi-way admixtures (e.g. Latino populations of the Americas). RESULTS: To address these challenges we introduce here methods for local ancestry inference which leverage the structure of linkage disequilibrium in the ancestral population (LAMP-LD), and incorporate the constraint of Mendelian segregation when inferring local ancestry in nuclear family trios (LAMP-HAP). Our algorithms uniquely combine hidden Markov models (HMMs) of haplotype diversity within a novel window-based framework to achieve superior accuracy as compared with published methods. Further, unlike previous methods, the structure of our HMM does not depend on the number of reference haplotypes but on a fixed constant, and it is thereby capable of utilizing large datasets while remaining highly efficient and robust to over-fitting. Through simulations and analysis of real data from 489 nuclear trio families from the mainland US, Puerto Rico and Mexico, we demonstrate that our methods achieve superior accuracy compared with published methods for local ancestry inference in Latinos.

MOTIVATION: The identification of catalytic residues is a key step in understanding the function of enzymes. While a variety of computational methods have been developed for this task, accuracies have remained fairly low. The best existing method exploits information from sequence and structure to achieve a precision (the fraction of predicted catalytic residues that are catalytic) of 18.5% at a corresponding recall (the fraction of catalytic residues identified) of 57% on a standard benchmark. Here we present a new method, Discern, which provides a significant improvement over the state-of-the-art through the use of statistical techniques to derive a model with a small set of features that are jointly predictive of enzyme active sites. RESULTS: In cross-validation experiments on two benchmark datasets from the Catalytic Site Atlas and CATRES resources containing a total of 437 manually curated enzymes spanning 487 SCOP families, Discern increases catalytic site recall between 12% and 20% over methods that combine information from both sequence and structure, and by >or=50% over methods that make use of sequence conservation signal only. Controlled experiments show that Discern's improvement in catalytic residue prediction is derived from the combination of three ingredients: the use of the INTREPID phylogenomic method to extract conservation information; the use of 3D structure data, including features computed for residues that are proximal in the structure; and a statistical regularization procedure to prevent overfitting.

UNLABELLED: A characterization of the genetic variation of recently admixed populations may reveal historical population events, and is useful for the detection of single nucleotide polymorphisms (SNPs) associated with diseases through association studies and admixture mapping. Inference of locus-specific ancestry is key to our understanding of the genetic variation of such populations. While a number of methods for the inference of locus-specific ancestry are accurate when the ancestral populations are quite distant (e.g. African-Americans), current methods incur a large error rate when inferring the locus-specific ancestry in admixed populations where the ancestral populations are closely related (e.g. Americans of European descent). RESULTS: In this work, we extend previous methods for the inference of locus-specific ancestry by the incorporation of a refined model of recombination events. We present an efficient dynamic programming algorithm to infer the locus-specific ancestries in this model, resulting in a method that attains improved accuracies; the improvement is most significant when the ancestral populations are closely related. An evaluation on a wide range of scenarios, including admixtures of the 52 population groups from the Human Genome Diversity Project demonstrates that locus-specific ancestry can indeed be accurately inferred in these admixtures using our method. Finally, we demonstrate that imputation methods can be improved by the incorporation of locus-specific ancestry, when applied to admixed populations. AVAILABILITY: The implementation of the WINPOP model is available as part of the LAMP package at http://lamp.icsi.berkeley.edu/lamp.

BACKGROUND: Identifying the catalytic residues in enzymes can aid in understanding the molecular basis of an enzyme's function and has significant implications for designing new drugs, identifying genetic disorders, and engineering proteins with novel functions. Since experimentally determining catalytic sites is expensive, better computational methods for identifying catalytic residues are needed. RESULTS: We propose ResBoost, a new computational method to learn characteristics of catalytic residues. The method effectively selects and combines rules of thumb into a simple, easily interpretable logical expression that can be used for prediction. We formally define the rules of thumb that are often used to narrow the list of candidate residues, including residue evolutionary conservation, 3D clustering, solvent accessibility, and hydrophilicity. ResBoost builds on two methods from machine learning, the AdaBoost algorithm and Alternating Decision Trees, and provides precise control over the inherent trade-off between sensitivity and specificity. We evaluated ResBoost using cross-validation on a dataset of 100 enzymes from the hand-curated Catalytic Site Atlas (CSA). CONCLUSION: ResBoost achieved 85% sensitivity for a 9.8% false positive rate and 73% sensitivity for a 5.7% false positive rate. ResBoost reduces the number of false positives by up to 56% compared to the use of evolutionary conservation scoring alone. We also illustrate the ability of ResBoost to identify recently validated catalytic residues not listed in the CSA.

MOTIVATION: Identification of functionally important residues in proteins plays a significant role in biological discovery. Here, we present INTREPID--an information-theoretic approach for functional site identification that exploits the information in large diverse multiple sequence alignments (MSAs). INTREPID uses a traversal of the phylogeny in combination with a positional conservation score, based on Jensen-Shannon divergence, to rank positions in an MSA. While knowledge of protein 3D structure can significantly improve the accuracy of functional site identification, since structural information is not available for a majority of proteins, INTREPID relies solely on sequence information. We evaluated INTREPID on two tasks: predicting catalytic residues and predicting specificity determinants. RESULTS: In catalytic residue prediction, INTREPID provides significant improvements over Evolutionary Trace, ConSurf as well as over a baseline global conservation method on a set of 100 manually curated enzymes from the Catalytic Site Atlas. In particular, INTREPID is able to better predict catalytic positions that are not globally conserved and hence, attains improved sensitivity at high values of specificity. We also investigated the performance of INTREPID as a function of the evolutionary divergence of the protein family. We found that INTREPID is better able to exploit the diversity in such families and that accuracy improves when homologs with very low sequence identity are included in an alignment. In specificity determinant prediction, when subtype information is known, INTREPID-SPEC, a variant of INTREPID, attains accuracies that are competitive with other approaches for this task. AVAILABILITY: INTREPID is available for 16919 families in the PhyloFacts resource (http://phylogenomics.berkeley.edu/phylofacts).