Searching reads from unknown origins in a reference database and finding evolutionarily similar genomes is central to many applications. Quantifying the similarity by estimating the distance between each read and matching references could further help downstream analyses such as taxonomic characterization or even placing the read on a reference phylogeny. Such distances can be computed using alignment. Since alignment becomes impractical for ultra-large reference databases, many use k -mer-based tools for downstream tasks, but these methods do not compute distances. We introduce krepp to compute the distance between a read and a genome using a maximum likelihood framework built around k -mers . We further introduce algorithms to place reads on a reference tree. Empirical evaluations show that krepp accurately estimates distances, scales well with large databases, and can place short reads coming from any part of the genome on a reference phylogeny.
By folding into particular 3D structures, proteins play a key role in living beings. To learn meaningful representation from a protein structure for downstream tasks, not only the global backbone topology but the local fine-grained orientational relations between amino acids should also be considered. In this work, we propose the Orientation-Aware Graph Neural Networks (OAGNNs) to better sense the geometric characteristics in protein structure (e.g. inner-residue torsion angles, inter-residue orientations). Extending a single weight from a scalar to a 3D vector, we construct a rich set of geometric-meaningful operations to process both the classical and SO(3) representations of a given structure. To plug our designed perceptron unit into existing Graph Neural Networks, we further introduce an equivariant message passing paradigm, showing superior versatility in maintaining SO(3)-equivariance at the global scale. Experiments have shown that our OAGNNs have a remarkable ability to sense geometric orientational features compared to classical networks. OAGNNs have also achieved state-of-the-art performance on various computational biology applications related to protein 3D structures. The code is available at https://github.com/Ced3-han/OAGNN/tree/main .
Protein function prediction remains a critical and challenging task in the field of bioinformatics. Despite various existing computational methods, most focus on predicting functions for novel proteins, often overlooking those with incomplete annotations. Predicting missing Gene Ontology (GO) terms for these partially annotated proteins is also important for gaining a comprehensive understanding of their biological roles. To address this gap, we introduce DualGOFiller, a dual-channel graph neural network (GNN) model, which is a novel method specifically focused on predicting missing GO terms for partially annotated proteins. DualGOFiller integrates heterogeneous (protein-GO bipartite graph) and homogeneous (protein-protein interaction network and GO directed acyclic graph) network information, enriched with attribute data from protein sequences and GO term descriptions. It further employs graph contrastive learning (GCL) to strengthen the alignment across different views. Experiments demonstrate the superiority of DualGOFiller over state-of-the-art methods, including traditional recommendation system approaches and several protein function prediction methods. The ablation study further confirms the significant contributions of key component in DualGOFiller for boosting predictive performance. The source code is available at https://github.com/ZhuLab-Fudan/DualGOFiller .
Clonal evolution of cancer results in intratumor heterogeneity (ITH), making treatment and cure challenging. Single-cell sequencing has advanced our understanding of ITH, but tracing subclonal evolution using mutational profiles of cells is limited by scale and noise. Moreover, available tumor progression tree inference methods usually offer a single tree to explain the progression of a tumor, and do not inform about alternative evolutionary scenarios. In this paper we introduce the partition function for a tumor progression tree, to assess the reliability of any proposed subclonal structure in a single-cell sequenced tumor. Given that tumor progression can be represented as a perfect phylogeny, the partition function for a given set of cells, R, and a given mutation, ρ , is defined as the likelihood that ρ is present exclusively in R. As such, the partition function evaluates whether R forms a subclone with ρ as a possible subclonal driver, which is especially useful if the cells of R are biologically or clinically significant, e.g., have aggressive growth, therapy resistance, or metastatic potential. We also present an algorithm to estimate the partition function and provide theoretical proofs that our algorithm asymptotically approaches the ground truth. In addition, we present empirical evidence of its accuracy on simulated data, as well as its practical value in analyzing the progression history of real tumors. The implementation is available at: https://github.com/algo-cancer/Partition-Function/ .
Advancements in sequencing technology have expanded data availability, capturing diverse phenotypic traits and biological perturbations. However, increased resolution also raises complexity, as studies now examine multiple dimensions, including donor phenotypes, anatomical regions, cell types, and time points. Integrating datasets across studies promises insights into health and disease beyond the scope of individual experiments, but this requires methods that can separate technical artifacts from meaningful biological signals while providing interpretable insights into condition-related genetic factors. Existing approaches tend to focus on either integration or interpretability, rarely addressing both simultaneously. To overcome these challenges, we introduce ALPINE, a joint supervised-unsupervised non-negative matrix factorization framework that disentangles technical and biological variation while directly identifying condition-associated genes.
Antimicrobial resistance is a global health challenge, complicating the treatment of bacterial infections and leading to higher patient morbidity and mortality. Rapid and reliable identification of resistant pathogens is crucial for timely and effective therapeutic interventions, but traditional culture-based methods are time-consuming. Machine learning models using bacterial MALDI-TOF mass spectrometry data have shown promising results in early resistance prediction—with predictions available at least 24 h earlier than conventional phenotypical test results. However, their clinical adoption has been hindered by subpar predictive performance and the lack of interpretable, statistically valid uncertainty estimates. In this work, we introduce a novel antimicrobial resistance prediction framework that addresses this gap with a novel knowledge-graph-enhanced conformal predictor. Conformal prediction (CP) constructs prediction sets with statistical coverage guarantees, ensuring that bacterial resistance to a certain antibiotic is detected with a specified error rate. Our proposed conformal predictor constructs improved prediction sets over standard CP approaches using a knowledge graph capturing the interdependencies in antimicrobial resistance patterns. In addition, we introduce a novel classifier framework that overcomes the limitations of previous efforts by incorporating multigraph-based antibiotic representations. Evaluating our approach on Klebsiella pneumoniae, we demonstrate state-of-the-art predictive performance across most clinically highly-relevant antibiotics. Furthermore, we show that our knowledge-graph-enhanced conformal predictor reduces false discovery rates while maintaining reliable coverage guarantees compared to standard CP approaches. All code required to reproduce the presented results can be found under https://github.com/BorgwardtLab/ConformalAMR .
Single-cell RNA sequencing provides detailed insights into cellular heterogeneity and responses to external stimuli. However, distinguishing inherent cellular variation from extrinsic effects induced by external stimuli remains a major analytical challenge. Here, we present scCausalVI, a causality-aware generative model designed to disentangle these sources of variation. scCausalVI decouples intrinsic cellular states from treatment effects through a deep structural causal network that explicitly models the causal mechanisms governing cell-state-specific responses to external perturbations while accounting for technical variations. Our model integrates structural causal modeling with cross-condition in silico prediction to infer gene expression profiles under hypothetical scenarios. Comprehensive benchmarking demonstrates that scCausalVI outperforms existing methods in disentangling causal relationships, quantifying treatment effects, generalizing to unseen cell types, and separating biological signals from technical variation in multi-source data integration. Applied to COVID-19 datasets, scCausalVI effectively identifies treatment-responsive populations and delineates molecular signatures of cellular susceptibility.
Spatial-omics technologies profile cells in their native spatial context within tissues, enabling more complete understanding of cellular properties. However, a key computational challenge remains: identifying cellular interactions that underlie cell types and states, which are essential for spatial organization and provide a biologically grounded framework for understanding cell identities and spatial patterns. These interactions over different distances require multiscale modeling, which represents a major gap in existing methods. Here, we introduce Steamboat, an interpretable machine learning framework leveraging a self-supervised, multi-head attention model that uniquely decomposes gene expression of a cell into multiple key factors: intrinsic cell programs, neighboring cell communication, and long-range interactions. By applying Steamboat to diverse tissues in health and disease across various spatial-omics technologies, we demonstrate its ability to uncover critical multiscale cellular interactions, capturing classical contact signaling and revealing previously unrecognized patterns of cellular communication. Steamboat provides a powerful approach for spatial-omics analysis, offering new insights into the multiscale spatial organization of cells and their communication across a broad range of biological contexts.
Spatial transcriptomics (ST) has transformed our understanding of tissue architecture and cellular interactions, but integrating ST data across platforms remains challenging due to differences in gene panels, data sparsity, and technical variability. Here, we introduce Lloki, a novel framework for integrating imaging-based ST data from diverse platforms without requiring shared gene panels. Lloki addresses ST integration through two key alignment tasks: feature alignment across technologies and batch alignment across datasets. Optimal transport-guided feature propagation adjusts data sparsity to match scRNA-seq references through graph-based imputation, enabling single-cell foundation models such as scGPT to generate unified features. Batch alignment then refines scGPT-transformed embeddings, mitigating batch effects while preserving biological variability. Evaluations on mouse brain samples from five different technologies demonstrate that Lloki outperforms existing methods and is effective for cross-technology spatial gene program identification and tissue slice alignment. Applying Lloki to five ovarian cancer datasets, we identify an integrated gene program indicative of tumor-infiltrating T cells across gene panels. Together, Lloki provides a robust foundation for cross-platform ST studies, with the potential to scale to large atlas datasets, enabling deeper insights into cellular organization and tissue environments.
Accurate protein structure prediction is challenging especially for families and types that are not well-represented in current training data. Active learning selects candidates for labeling with the aim of most rapidly improving model performance. In a general, many labeling strategies have been proposed. However, with protein structure prediction, most of these strategies don’t apply or are difficult due to the high-dimensional, variable-dimensional regression target and the inherent complexity of the models involved. We applied a novel active learning strategy, DEWDROP, to protein structure prediction on two different protein datasets: VHH-only antibodies (Nanobodies™), and Mycobacterial proteins. We introduce a domain-specific fine-tuned Equifold model for VHH structures and apply DEWDROP to generate ensembles of predictions using Monte Carlo dropout. Using the statistics of these we select batches with high information content for labeling. We show that DEWDROP (1) improves model training efficiency through batch optimization outperforming baselines, and (2) selects data with relevant high information content.
Interpreting the functional significance of non-coding GWAS variants remains challenging. While co-localizing variants with cell-type specific cis-regulatory elements (CREs) has improved our understanding, many variants remain unassociated. In this study, we propose GEM-Finder (Genomic Element Mapping for Fine Discovery of Promoter-Linked Variants), a novel analytical framework that integrates transcriptomic, epigenomic (H3K27ac ChIP-seq), and chromatin interaction data. GEM-Finder utilizes long-range chromatin interactions to identify CREs that connect differentially expressed genes of specific cell types. When we apply GEM-Finder to endothelial differentiation, unlike conventional methods primarily focused on cell-type specific CREs, GEM-Finder identifies 7.6 times more disease/trait associations. Our enrichment analyses reveal both shared and unique links across 53 human diseases/traits, uncovering novel associations between endothelial differentiation and various human diseases/traits. These findings underscore the importance of incorporating long-range chromatin interactions for a more precise identification of disease-associated CREs and functional characterization of non-coding GWAS variants.
Genomic language models (gLMs) have shown mostly modest success in identifying evolutionarily constrained elements in mammalian genomes. To address this issue, we introduce a novel framework for training gLMs that explicitly models nucleotide evolution on phylogenetic trees using multispecies whole-genome alignments. Our approach integrates an alignment into the loss function during training but does not require it for making predictions, thereby enhancing the model's applicability. We applied this framework to train PhyloGPN, a model that excels at predicting functionally disruptive variants from a single sequence alone and demonstrates strong transfer learning capabilities.
Gene-based rare variant association tests (RVATs) are essential for uncovering disease mechanisms and identifying candidate drug targets, yet existing frameworks lack flexibility in integrating multiple variant annotations. Here, we introduce BayesRVAT, a Bayesian framework for RVAT which models variant effects using priors informed by multiple annotations. We show that BayesRVAT outperforms state-of-the-art burden test strategies in both simulations and an analysis of 12 blood traits from the UK Biobank.
Tandem repeats (TRs) are sequences of DNA in which ≥2 bp are repeated back-to-back at specific locations in the genome. TR expansions, in which the number of repeat units exceeds the normal range, have been implicated in more than 50 conditions. However, accurately measuring the copy number of TRs is challenging, especially when their expansions are larger than the fragment sizes used in standard short-read genome sequencing. Here, we introduce ScatTR, a novel computational method that leverages a maximum likelihood framework to estimate the copy number of large TR expansions from short-read sequencing data. ScatTR calculates the likelihood of different alignments between sequencing reads and reference sequences that represent various TR lengths and employs a Monte Carlo technique to find the best match. In simulated data, ScatTR outperforms state-of-the-art methods, particularly for TRs with longer motifs and those with lengths that greatly exceed typical sequencing fragment sizes. When applied to data from the 1000 Genomes Project, ScatTR detects potential large TR expansions that other methods missed, highlighting its ability to better characterize genome-wide TR variation.
Haplotype assembly is the problem of reconstructing the combination of alleles on the maternally and paternally inherited chromosome copies. Individual haplotypes are essential to our understanding of how combinations of different variants impact phenotype. In this work, we focus on read-based haplotype assembly of individual diploid genomes, which reconstructs the two haplotypes directly from read alignments at variant loci. We introduce ralphi, a novel deep reinforcement learning framework for haplotype assembly, which integrates the representational power of deep learning with reinforcement learning to accurately partition read fragments into their respective haplotype sets. To set the reward objective for reinforcement learning, our approach uses the classic reduction of the problem to the maximum fragment cut formulation on fragment graphs, where nodes correspond to reads and edge weights capture the conflict or agreement of the reads at shared variant sites. We trained ralphi on a diverse dataset of fragment graph topologies derived from genomes of different populations included in the 1000 Genomes Project. We show that ralphi consistently achieves lower error rates at comparable or longer haplotype block lengths over the state of the art for short and long ONT reads at varying coverage in standard human genome benchmarks. ralphi is available at https://github.com/PopicLab/ralphi .
Deriving the sequence of transitions between cell types, or differentiation events, that occur during organismal development is one of the fundamental challenges in developmental biology. Single-cell and spatial sequencing of samples from different developmental timepoints provide data to investigate differentiation but inferring a sequence of differentiation events requires: (1) finding trajectories, or ancestor:descendant relationships, between cells from consecutive timepoints; (2) coarse-graining these trajectories into a differentiation map, or collection of transitions between cell types, rather than individual cells. We introduce Hidden-Markov Optimal Transport (HM-OT), an algorithm that simultaneously groups cells into cell types and learns transitions between these cell types from single-cell or spatial developmental time series. HM-OT uses low-rank optimal transport to simultaneously align samples in a time series and learn a sequence of clusterings and a differentiation map with minimal total transport cost.
Spatially resolved transcriptomics (SRT) technologies measure gene expression across thousands of spatial locations within a tissue slice. Multiple SRT technologies are currently available and others are in active development, with each technology having varying spatial resolution (subcellular, single-cell, or multicellular regions), gene coverage (targeted vs. whole-transcriptome), and sequencing depth per location. For example, the widely used 10x Genomics Visium platform measures whole transcriptomes from multiple-cell-sized spots, whereas the 10x Genomics Xenium platform measures a few hundred genes at subcellular resolution. A number of studies apply multiple SRT technologies to slices that originate from the same biological tissue. Integration of data from different SRT technologies can overcome limitations of the individual technologies, enabling the imputation of expression from unmeasured genes in targeted technologies and/or the deconvolution of admixed expression from technologies with lower spatial resolution. Here, we introduce Spatial Integration for Imputation and Deconvolution (SIID), an algorithm to reconstruct a latent spatial gene expression matrix from a pair of observations from different SRT technologies. SIID leverages a spatial alignment and uses a joint nonnegative factorization model to accurately impute missing gene expression and infer gene expression signatures of cell types from admixed SRT data. In simulations involving paired SRT data sets from different technologies (e.g., Xenium and Visium), SIID shows superior performance in reconstructing spot-to-cell-type assignments, recovering cell type-specific gene expression and imputing missing data compared to contemporary tools. When applied to real-world 10x Xenium-Visium pairs from human breast and colon cancer tissues, SIID achieves highest performance in imputing holdout gene expression.
A common problem in phylogeny is to reconstruct the ancestral states of a feature measured at the present time. The classic Fitch-Hartigan and Sankoff algorithms compute the most parsimonious or most likely reconstruction. However, these approaches do not readily extend to structured ancestral reconstruction problems, such as those encountered when inferring the routes of metastases in cancer, deriving the transmission history of viruses, or detecting horizontal gene transfer in phylogenetic networks. We develop a combinatorial optimization approach to ancestral reconstruction problems based on the tree-labeling polytope, a geometric object whose vertices represent the ancestral labelings of a tree. We derive algorithms for three structured ancestral reconstruction problems: parsimonious migration history, softwired small parsimony, and convex recoloring. We apply these algorithms to analyze routes of metastasis in a mouse model of lung adenocarcinoma using lineage-tracing data from thousands of single cells.
Predicting the effects of amino acid substitutions on protein thermostability is critical for understanding diseases and engineering proteins. Recent protein generative models excel in predicting properties from structural or evolutionary contexts but fall short in stability prediction. We introduce SPURS, a deep learning framework integrating a protein language model (ESM) and an inverse folding model (ProteinMPNN). SPURS rewires ProteinMPNN’s structural representations into ESM’s attention layers, combining structural and sequence information for mutation prediction. Trained on a mega-scale thermostability dataset, SPURS achieves state-of-the-art accuracy and scalability across benchmarks. Beyond stability, it identifies functional sites unsupervised and enhances low-N fitness models, establishing itself as a versatile tool for protein engineering and analysis.
Protein function prediction is an essential problem in the field of computational biology and has a significant influence on many applications like drug discovery and antibody design. A long-lasting challenge for this problem is the data imbalance issue in existing protein function data, known as the long-tail problem: a small set of function classes are associated with most of the annotated proteins, while the majority of classes only have very few labeled proteins. This leads to the biased performance of current machine learning-based protein function prediction models, which focus more on optimizing prediction accuracy for well-represented function classes while performing poorly for underrepresented classes. In this work, we developed MSRep, a deep learning-based framework for protein function annotation aimed to mitigate the data imbalance issue. The inspiration for MSRep comes from an interesting phenomenon, called neural collapse (NC), observed in high-accuracy deep classification models. MSRep leveraged an NC-inspired loss to refine a pretrained protein language model embedding into a new space with NC properties, where both the well- and under-represented classes are equally represented in the embedding space. Evaluations across four protein function annotation tasks suggested that MSRep provides accurate predictions for both well- and underrepresented classes, outperforming several state-of-the-art annotation tools.
