Research Group Fabian Theis

Learning disentangled representations from unlabelled data is a fundamental challenge in machine learning. Solving it may unlock other problems, such as generalization, interpretability, or fairness. Although remarkably challenging to solve in theory, disentanglement is often achieved in practice through prior matching. Furthermore, recent works have shown that prior matching approaches can be enhanced by leveraging geometrical considerations, e.g., by learning representations that preserve geometric features of the data, such as distances or angles between points. However, matching the prior while preserving geometric features is challenging, as a mapping that fully preserves these features while aligning the data distribution with the prior does not exist in general. To address these challenges, we introduce a novel approach to disentangled representation learning based on quadratic optimal transport. We formulate the problem using Gromov-Monge maps that transport one distribution onto another with minimal distortion of predefined geometric features, preserving them as much as can be achieved. To compute such maps, we propose the Gromov-Monge-Gap (GMG), a regularizer quantifying whether a map moves a reference distribution with minimal geometry distortion. We demonstrate the effectiveness of our approach for disentanglement across four standard benchmarks, outperforming other methods leveraging geometric considerations.

Advances in single-cell ‘-omics’ allow unprecedented insights into the transcriptional profiles of individual cells and, when combined with large-scale perturbation screens, enable measuring of the effect of targeted perturbations on the whole transcriptome. These advances provide an opportunity to better understand the causative role of genes in complex biological processes. In this Perspective, we delineate the application of causal machine learning to single-cell genomics and its associated challenges. We first present the causal model that is most commonly applied to single-cell biology and then identify and discuss potential approaches to three open problems: the lack of generalization of models to novel experimental conditions, the complexity of interpreting learned models, and the difficulty of learning cell dynamics.

Large language models based on the transformer deep learning architecture have revolutionized natural language processing. Motivated by the analogy between human language and the genome’s biological code, researchers have begun to develop genome language models (gLMs) based on transformers and related architectures. This Review explores the use of transformers and language models in genomics. We survey open questions in genomics amenable to the use of gLMs, and motivate the use of gLMs and the transformer architecture for these problems. We discuss the potential of gLMs for modelling the genome using unsupervised pretraining tasks, specifically focusing on the power of zero- and few-shot learning. We explore the strengths and limitations of the transformer architecture, as well as the strengths and limitations of current gLMs more broadly. Additionally, we contemplate the future of genomic modelling beyond the transformer architecture, based on current trends in research. This Review serves as a guide for computational biologists and computer scientists interested in transformers and language models for genomic data.

Single-cell data provide high-dimensional measurements of the transcriptional states of cells, but extracting insights into the regulatory functions of genes, particularly identifying transcriptional mechanisms affected by biological perturbations, remains a challenge. Many perturbations induce compensatory cellular responses, making it difficult to distinguish direct from indirect effects on gene regulation. Modeling how gene regulatory functions shape the temporal dynamics of these responses is key to improving our understanding of biological perturbations. Dynamical models based on differential equations offer a principled way to capture transcriptional dynamics, but their application to single-cell data has been hindered by computational constraints, stochasticity, sparsity, and noise. Existing methods either rely on low-dimensional representations or make strong simplifying assumptions, limiting their ability to model transcriptional dynamics at scale. We introduce a Functional and Learnable model of Cell dynamicS, FLeCS, that incorporates gene network structure into coupled differential equations to model gene regulatory functions. Given (pseudo)time-series single-cell data, FLeCS accurately infers cell dynamics at scale, provides improved functional insights into transcriptional mechanisms perturbed by gene knockouts, both in myeloid differentiation and K562 Perturb-seq experiments, and simulates single-cell trajectories of A549 cells following small-molecule perturbations.

Targeted spatial transcriptomic methods capture the topology of cell types and states in tissues at single-cell and subcellular resolution by measuring the expression of a predefined set of genes. The selection of an optimal set of probed genes is crucial for capturing the spatial signals present in a tissue. This requires selecting the most informative, yet minimal, set of genes to profile (gene set selection) for which it is possible to build probes (probe design). However, current selections often rely on marker genes, precluding them from detecting continuous spatial signals or new states. We present Spapros, an end-to-end probe set selection pipeline that optimizes both gene set specificity for cell type identification and within-cell type expression variation to resolve spatially distinct populations while considering prior knowledge as well as probe design and expression constraints. We evaluated Spapros and show that it outperforms other selection approaches in both cell type recovery and recovering expression variation beyond cell types. Furthermore, we used Spapros to design a single-cell resolution in situ hybridization on tissues (SCRINSHOT) experiment of adult lung tissue to demonstrate how probes selected with Spapros identify cell types of interest and detect spatial variation even within cell types.

Recent efforts to construct reference maps of cellular phenotypes have expanded the volume and diversity of single-cell omics data, providing an unprecedented resource for studying cell properties. Despite the availability of rich datasets and their continued growth, current single-cell models are unable to fully capitalize on the information they contain. Transformers have become the architecture of choice for foundation models in other domains owing to their ability to generalize to heterogeneous, large-scale datasets. Thus, the question arises of whether transformers could set off a similar shift in the field of single-cell modeling. Here we first describe the transformer architecture and its single-cell adaptations and then present a comprehensive review of the existing applications of transformers in single-cell analysis and critically discuss their future potential for single-cell biology. By studying limitations and technical challenges, we aim to provide a structured outlook for future research directions at the intersection of machine learning and single-cell biology.

Learning disentangled representations from unlabelled data is a fundamental challenge in machine learning. Solving it may unlock other problems, such as generalization, interpretability, or fairness. Although remarkably challenging to solve in theory, disentanglement is often achieved in practice through prior matching. Furthermore, recent works have shown that prior matching approaches can be enhanced by leveraging geometrical considerations, e.g., by learning representations that preserve geometric features of the data, such as distances or angles between points. However, matching the prior while preserving geometric features is challenging, as a mapping that fully preserves these features while aligning the data distribution with the prior does not exist in general. To address these challenges, we introduce a novel approach to disentangled representation learning based on quadratic optimal transport. We formulate the problem using Gromov-Monge maps that transport one distribution onto another with minimal distortion of predefined geometric features, preserving them as much as can be achieved. To compute such maps, we propose the Gromov-Monge-Gap (GMG), a regularizer quantifying whether a map moves a reference distribution with minimal geometry distortion. We demonstrate the effectiveness of our approach for disentanglement across four standard benchmarks, outperforming other methods leveraging geometric considerations.

Spatial omics technologies are increasingly leveraged to characterize how disease disrupts tissue organization and cellular niches. While multiple methods to analyze spatial variation within a sample have been published, statistical and computational approaches to compare cell spatial organization across samples or conditions are mostly lacking. We present GraphCompass, a comprehensive set of omics-adapted graph analysis methods to quantitatively evaluate and compare the spatial arrangement of cells in samples representing diverse biological conditions. GraphCompass builds upon the Squidpy spatial omics toolbox and encompasses various statistical approaches to perform cross-condition analyses at the level of individual cell types, niches, and samples. Additionally, GraphCompass provides custom visualization functions that enable effective communication of results. We demonstrate how GraphCompass can be used to address key biological questions, such as how cellular organization and tissue architecture differ across various disease states and which spatial patterns correlate with a given pathological condition. GraphCompass can be applied to various popular omics techniques, including, but not limited to, spatial proteomics (e.g. MIBI-TOF), spot-based transcriptomics (e.g. 10× Genomics Visium), and single-cell resolved transcriptomics (e.g. Stereo-seq). In this work, we showcase the capabilities of GraphCompass through its application to three different studies that may also serve as benchmark datasets for further method development. With its easy-to-use implementation, extensive documentation, and comprehensive tutorials, GraphCompass is accessible to biologists with varying levels of computational expertise. By facilitating comparative analyses of cell spatial organization, GraphCompass promises to be a valuable asset in advancing our understanding of tissue function in health and disease.

Although recent advances in higher-order Graph Neural Networks (GNNs) improve the theoretical expressiveness and molecular property predictive performance, they often fall short of the empirical performance of models that explicitly use fragment information as inductive bias. However, for these approaches, there exists no theoretic expressivity study. In this work, we propose the Fragment-WL test, an extension to the well-known Weisfeiler & Leman (WL) test, which enables the theoretic analysis of these fragment-biased GNNs. Building on the insights gained from the Fragment-WL test, we develop a new GNN architecture and a fragmentation with infinite vocabulary that significantly boosts expressiveness. We show the effectiveness of our model on synthetic and real-world data where we outperform all GNNs on Peptides and have 12% lower error than all GNNs on ZINC and 34% lower error than other fragment-biased models. Furthermore, we show that our model exhibits superior generalization capabilities compared to the latest transformer-based architectures, positioning it as a robust solution for a range of molecular modeling tasks.

In optimal transport (OT), a Monge map is known as a mapping that transports a source distribution to a target distribution in the most cost-efficient way. Recently, multiple neural estimators for Monge maps have been developed and applied in diverse unpaired domain translation tasks, e.g. in single-cell biology and computer vision. However, the classic OT framework enforces mass conservation, which makes it prone to outliers and limits its applicability in real-world scenarios. The latter can be particularly harmful in OT domain translation tasks, where the relative position of a sample within a distribution is explicitly taken into account. While unbalanced OT tackles this challenge in the discrete setting, its integration into neural Monge map estimators has received limited attention. We propose a theoretically grounded method to incorporate unbalancedness into any Monge map estimator. We improve existing estimators to model cell trajectories over time and to predict cellular responses to perturbations. Moreover, our approach seamlessly integrates with the OT flow matching (OT-FM) framework. While we show that OT-FM performs competitively in image translation, we further improve performance by incorporating unbalancedness (UOT-FM), which better preserves relevant features. We hence establish UOT-FM as a principled method for unpaired image translation.

RNA velocity has been rapidly adopted to guide interpretation of transcriptional dynamics in snapshot single-cell data; however, current approaches for estimating RNA velocity lack effective strategies for quantifying uncertainty and determining the overall applicability to the system of interest. Here, we present veloVI (velocity variational inference), a deep generative modeling framework for estimating RNA velocity. veloVI learns a gene-specific dynamical model of RNA metabolism and provides a transcriptome-wide quantification of velocity uncertainty. We show that veloVI compares favorably to previous approaches with respect to goodness of fit, consistency across transcriptionally similar cells and stability across preprocessing pipelines for quantifying RNA abundance. Further, we demonstrate that veloVI’s posterior velocity uncertainty can be used to assess whether velocity analysis is appropriate for a given dataset. Finally, we highlight veloVI as a flexible framework for modeling transcriptional dynamics by adapting the underlying dynamical model to use time-dependent transcription rates.

Recent advances in multiplexed single‐cell transcriptomics experiments facilitate the high‐throughput study of drug and genetic perturbations. However, an exhaustive exploration of the combinatorial perturbation space is experimentally unfeasible. Therefore, computational methods are needed to predict, interpret, and prioritize perturbations. Here, we present the compositional perturbation autoencoder (CPA), which combines the interpretability of linear models with the flexibility of deep‐learning approaches for single‐cell response modeling. CPA learns to in silico predict transcriptional perturbation response at the single‐cell level for unseen dosages, cell types, time points, and species. Using newly generated single‐cell drug combination data, we validate that CPA can predict unseen drug combinations while outperforming baseline models. Additionally, the architecture’s modularity enables incorporating the chemical representation of the drugs, allowing the prediction of cellular response to completely unseen drugs. Furthermore, CPA is also applicable to genetic combinatorial screens. We demonstrate this by imputing in silico 5,329 missing combinations (97.6% of all possibilities) in a single‐cell Perturb‐seq experiment with diverse genetic interactions. We envision CPA will facilitate efficient experimental design and hypothesis generation by enabling in silico response prediction at the single‐cell level and thus accelerate therapeutic applications using single‐cell technologies.

Machine learning for molecules holds great potential for efficiently exploring the vast chemical space and thus streamlining the drug discovery process by facilitating the design of new therapeutic molecules. Deep generative models have shown promising results for molecule generation, but the benefits of specific inductive biases for learning distributions over small graphs are unclear. Our study aims to investigate the impact of subgraph structures and vocabulary design on distribution learning, using small drug molecules as a case study. To this end, we introduce Subcover, a new subgraph-based fragmentation scheme, and evaluate it through a two-step variational auto-encoder. Our results show that Subcover’s improved identification of chemically meaningful subgraphs leads to a relative improvement of the FCD score by 30%, outperforming previous methods. Our findings highlight the potential of Subcover to enhance the performance and scalability of existing methods, contributing to the advancement of drug discovery.

Models of intercellular communication in tissues are based on molecular profiles of dissociated cells, are limited to receptor–ligand signaling and ignore spatial proximity in situ. We present node-centric expression modeling, a method based on graph neural networks that estimates the effects of niche composition on gene expression in an unbiased manner from spatial molecular profiling data. We recover signatures of molecular processes known to underlie cell communication.

Recent advances in single-cell technologies have enabled high-throughput molecular profiling of cells across modalities and locations. Single-cell transcriptomics data can now be complemented by chromatin accessibility, surface protein expression, adaptive immune receptor repertoire profiling and spatial information. The increasing availability of single-cell data across modalities has motivated the development of novel computational methods to help analysts derive biological insights. As the field grows, it becomes increasingly difficult to navigate the vast landscape of tools and analysis steps. Here, we summarize independent benchmarking studies of unimodal and multimodal single-cell analysis across modalities to suggest comprehensive best-practice workflows for the most common analysis steps. Where independent benchmarks are not available, we review and contrast popular methods. Our article serves as an entry point for novices in the field of single-cell (multi-)omic analysis and guides advanced users to the most recent best practices.

Neural Ordinary Differential Equations (NODEs) have proven successful in learning dynamical systems in terms of accurately recovering the observed trajectories. While different types of sparsity have been proposed to improve robustness, the generalization properties of NODEs for dynamical systems beyond the observed data are underexplored. We systematically study the influence of weight and feature sparsity on forecasting as well as on identifying the underlying dynamical laws. Besides assessing existing methods, we propose a regularization technique to sparsify input-output connections’’ and extract relevant features during training. Moreover, we curate real-world datasets including human motion capture and human hematopoiesis single-cell RNA-seq data to realistically analyze different levels of out-of-distribution (OOD) generalization in forecasting and dynamics identification respectively. Our extensive empirical evaluation on these challenging benchmarks suggests that weight sparsity improves generalization in the presence of noise or irregular sampling. However, it does not prevent learning spurious feature dependencies in the inferred dynamics, rendering them impractical for predictions under interventions, or for inferring the true underlying dynamics. Instead, feature sparsity can indeed help with recovering sparse ground-truth dynamics compared to unregularized NODEs.

Single-cell transcriptomics enabled the study of cellular heterogeneity in response to perturbations at the resolution of individual cells. However, scaling high-throughput screens (HTSs) to measure cellular responses for many drugs remains a challenge due to technical limitations and, more importantly, the cost of such multiplexed experiments. Thus, transferring information from routinely performed bulk RNA HTS is required to enrich single-cell data meaningfully.We introduce chemCPA, a new encoder-decoder architecture to study the perturbational effects of unseen drugs. We combine the model with an architecture surgery for transfer learning and demonstrate how training on existing bulk RNA HTS datasets can improve generalisation performance. Better generalisation reduces the need for extensive and costly screens at single-cell resolution. We envision that our proposed method will facilitate more efficient experiment designs through its ability to generate in-silico hypotheses, ultimately accelerating drug discovery.

Tissue phenotypes such as metabolic states, inflammation, and tumor properties are functions of molecular states of cells that constitute the tissue. Recent spatial molecular profiling assays measure tissue architecture motifs in a molecular and often unbiased way and thus can explain some aspects of emergence of these phenotypes. Here, we characterize the ability of graph neural networks to model tissue-level emergent phenotypes based on spatial data by evaluating phenotype prediction across model complexities. First, we show that immune cell dispersion in colorectal tumors, which is known to be predictive of disease outcome, can be captured by graph neural networks. Second, we show that breast cancer tumor classes can be predicted from gene expression alone without spatial information and are thus too simplistic a phenotype to require a complex model of emergence. Third, we show that representation learning approaches for spatial graphs of molecular profiles are limited by overfitting in the prevalent regime of up to 100s of images per study. We address overfitting with within-graph self-supervision and illustrate its promise for tissue representation learning as a constraint for node representations.

Despite the therapeutic promise of direct reprogramming, basic principles concerning fate erasure and the mechanisms to resolve cell identity conflicts remain unclear. To tackle these fundamental questions, we established a single‐cell protocol for the simultaneous analysis of multiple cell fate conversion events based on combinatorial and traceable reprogramming factor expression: Collide‐seq. Collide‐seq revealed the lack of a common mechanism through which fibroblast‐specific gene expression loss is initiated. Moreover, we found that the transcriptome of converting cells abruptly changes when a critical level of each reprogramming factor is attained, with higher or lower levels not contributing to major changes. By simultaneously inducing multiple competing reprogramming factors, we also found a deterministic system, in which titration of fates against each other yields dominant or colliding fates. By investigating one collision in detail, we show that reprogramming factors can disturb cell identity programs independent of their ability to bind their target genes. Taken together, Collide‐seq has shed light on several fundamental principles of fate conversion that may aid in improving current reprogramming paradigms.

Large variation in response to inhaled corticosteroids (ICS) has been reported in both asthma and chronic obstructive pulmonary disease (COPD), which may partly be explained by genetic factors. The transcriptome of the airways changes following ICS treatment,1 which may be directed by single nucleotide polymorphisms (SNPs), that affect deoxyribonucleic acid (DNA) methylation (methylation-Quantitative Trait Loci, meQTL).
A strong and consistent response of the airways to ICS in both asthma and COPD patients1, 2 has been found, and severe childhood asthma has been associated with increased odds of COPD development in later life,3 showing that overlap between the diseases may exist. We hypothesised that preselection of steroid-inducible meQTL that affect DNA methylation upon ICS treatment may increase power to find SNPs that also clinically affect response to ICS and that these genetic variants might overlap between asthma and COPD. The aim of this study was to identify SNPs that affect change in DNA methylation in the airway wall upon ICS treatment, and to investigate whether these SNPs are associated with asthma exacerbations in children despite treatment with ICS.
For the identification of meQTLs, we investigated 43 Dutch COPD patients from the Groningen and Leiden Universities study of Corticosteroids in Obstructive Lung Disease (GLUCOLD) study (Table S1).1 Longitudinal airway wall DNA methylation (EPIC 850 K array) and gene expression (ribonucleic acid-sequencing, RNA-seq) was collected from these patients pre- and post-6 months of fluticasone ± salmeterol (500/50 μg twice daily) treatment (Figure S1). We focused on methylation sites that previously were shown to be altered during ICS treatment (1049 CpG sites).4 This analysis identified 76 inducible meQTL caused by 71 independent SNPs with an false discovery rate (FDR) < 0.05 (Table S2). The most significant association was between cg13086983 and rs10917023, where the G allele (minor allele frequency: 7.7%) induced higher methylation (Beta: 0.849, p value: 4.21 × 10−06). Of these 76 CpG sites, 24 were associated with 24 gene transcripts (Table S3). The most significant association was found between the Cytosine-phosphate-Guanine (CpG) site cg08570199 and the CCDC80 gene (Beta coefficient: −1.249, p-value: 2.05 × 10−4; Figure 1A–D).

Large single-cell atlases are now routinely generated to serve as references for analysis of smaller-scale studies. Yet learning from reference data is complicated by batch effects between datasets, limited availability of computational resources and sharing restrictions on raw data. Here we introduce a deep learning strategy for mapping query datasets on top of a reference called single-cell architectural surgery (scArches). scArches uses transfer learning and parameter optimization to enable efficient, decentralized, iterative reference building and contextualization of new datasets with existing references without sharing raw data. Using examples from mouse brain, pancreas, immune and whole-organism atlases, we show that scArches preserves biological state information while removing batch effects, despite using four orders of magnitude fewer parameters than de novo integration. scArches generalizes to multimodal reference mapping, allowing imputation of missing modalities. Finally, scArches retains coronavirus disease 2019 (COVID-19) disease variation when mapping to a healthy reference, enabling the discovery of disease-specific cell states. scArches will facilitate collaborative projects by enabling iterative construction, updating, sharing and efficient use of reference atlases.

Single-cell transcriptomics enabled the study of cellular heterogeneity in response to perturbations at the resolution of individual cells. However, scaling high-throughput screens (HTSs) to measure cellular responses for many drugs remains a challenge due to technical limitations and, more importantly, the cost of such multiplexed experiments. Thus, transferring information from routinely performed bulk RNA-seq HTS is required to enrich single-cell data meaningfully. We introduce a new encoder-decoder architecture to study the perturbational effects of unseen drugs. We combine the model with a transfer learning scheme and demonstrate how training on existing bulk RNA-seq HTS datasets can improve generalisation performance. Better generalisation reduces the need for extensive and costly screens at single-cell resolution. We envision that our proposed method will facilitate more efficient experiment designs through its ability to generate in-silico hypotheses, ultimately accelerating targeted drug discovery.

Despite its high prevalence, the cellular and molecular mechanisms of chronic obstructive pulmonary disease (COPD) are far from being understood. Here, we determine disease-related changes in cellular and molecular compositions within the alveolar space and peripheral blood of a cohort of COPD patients and controls. Myeloid cells were the largest cellular compartment in the alveolar space with invading monocytes and proliferating macrophages elevated in COPD. Modeling cell-to-cell communication, signaling pathway usage, and transcription factor binding predicts TGF-β1 to be a major upstream regulator of transcriptional changes in alveolar macrophages of COPD patients. Functionally, macrophages in COPD showed reduced antigen presentation capacity, accumulation of cholesteryl ester, reduced cellular chemotaxis, and mitochondrial dysfunction, reminiscent of impaired immune activation.

Spatial omics data are advancing the study of tissue organization and cellular communication at an unprecedented scale. Flexible tools are required to store, integrate and visualize the large diversity of spatial omics data. Here, we present Squidpy, a Python framework that brings together tools from omics and image analysis to enable scalable description of spatial molecular data, such as transcriptome or multivariate proteins. Squidpy provides efficient infrastructure and numerous analysis methods that allow to efficiently store, manipulate and interactively visualize spatial omics data. Squidpy is extensible and can be interfaced with a variety of already existing libraries for the scalable analysis of spatial omics data.

Computational trajectory inference enables the reconstruction of cell state dynamics from single-cell RNA sequencing experiments. However, trajectory inference requires that the direction of a biological process is known, largely limiting its application to differentiating systems in normal development. Here, we present CellRank (https://cellrank.org) for single-cell fate mapping in diverse scenarios, including regeneration, reprogramming and disease, for which direction is unknown. Our approach combines the robustness of trajectory inference with directional information from RNA velocity, taking into account the gradual and stochastic nature of cellular fate decisions, as well as uncertainty in velocity vectors. On pancreas development data, CellRank automatically detects initial, intermediate and terminal populations, predicts fate potentials and visualizes continuous gene expression trends along individual lineages. Applied to lineage-traced cellular reprogramming data, predicted fate probabilities correctly recover reprogramming outcomes. CellRank also predicts a new dedifferentiation trajectory during postinjury lung regeneration, including previously unknown intermediate cell states, which we confirm experimentally.

Single-cell RNA sequencing measures gene expression at an unprecedented resolution and scale and allows the analysis of cellular phenotypes which was not possible before. In this context, graphs occur as a natural representation of the system —both as gene-centric and cell-centric. However, many advances in machine learning on graphs are not yet harnessed in models on single-cell data. Taking the inference of cell types or gene interactions as examples, graph representation learning has a wide applicability to both cell and gene graphs. Recent advances in spatial molecular profiling additionally put graph learning in the focus of attention because of the innate resemblance of spatial information to spatial graphs. We argue that graph embedding techniques have great potential for various applications across single-cell biology. Here, we discuss how graph representation learning maps to current models and concepts used in single-cell biology and formalise overlaps to developments in graph-based deep learning.

Single cell RNA-seq has revolutionized transcriptomics by providing cell type resolution for differential gene expression and expression quantitative trait loci (eQTL) analyses. However, efficient power analysis methods for single cell data and inter-individual comparisons are lacking. Here, we present scPower; a statistical framework for the design and power analysis of multi-sample single cell transcriptomic experiments. We modelled the relationship between sample size, the number of cells per individual, sequencing depth, and the power of detecting differentially expressed genes within cell types. We systematically evaluated these optimal parameter combinations for several single cell profiling platforms, and generated broad recommendations. In general, shallow sequencing of high numbers of cells leads to higher overall power than deep sequencing of fewer cells. The model, including priors, is implemented as an R package and is accessible as a web tool. scPower is a highly customizable tool that experimentalists can use to quickly compare a multitude of experimental designs and optimize for a limited budget.

Background: Due to the ongoing COVID-19 pandemic, demand for diagnostic testing has increased drastically, resulting in shortages of necessary materials to conduct the tests and overwhelming the capacity of testing laboratories. The supply scarcity and capacity limits affect test administration: priority must be given to hospitalized patients and symptomatic individuals, which can prevent the identification of asymptomatic and presymptomatic individuals and hence effective tracking and tracing policies. We describe optimized group testing strategies applicable to SARS-CoV-2 tests in scenarios tailored to the current COVID-19 pandemic and assess significant gains compared to individual testing.
Methods: We account for biochemically realistic scenarios in the context of dilution effects on SARS-CoV-2 samples and consider evidence on specificity and sensitivity of PCR-based tests for the novel coronavirus. Because of the current uncertainty and the temporal and spatial changes in the prevalence regime, we provide analysis for several realistic scenarios and propose fast and reliable strategies for massive testing procedures.
Key Findings: We find significant efficiency gaps between different group testing strategies in realistic scenarios for SARS-CoV-2 testing, highlighting the need for an informed decision of the pooling protocol depending on estimated prevalence, target specificity, and high- vs. low-risk population. For example, using one of the presented methods, all 1.47 million inhabitants of Munich, Germany, could be tested using only around 141 thousand tests if the infection rate is below 0.4% is assumed. Using 1 million tests, the 6.69 million inhabitants from the city of Rio de Janeiro, Brazil, could be tested as long as the infection rate does not exceed 1%. Moreover, we provide an interactive web application, available at www.group-testing.com, for visualizing the different strategies and designing pooling schemes according to specific prevalence scenarios and test configurations.
Interpretation: Altogether, this work may help provide a basis for an efficient upscaling of current testing procedures, which takes the population heterogeneity into account and is fine-grained towards the desired study populations, e.g., mild/asymptomatic individuals vs. symptomatic ones but also mixtures thereof.

Single-cell RNA-seq datasets are often first analyzed independently without harnessing model fits from previous studies, and are then contextualized with public data sets, requiring time-consuming data wrangling. We address these issues with sfaira, a single-cell data zoo for public data sets paired with a model zoo for executable pre-trained models. The data zoo is designed to facilitate contribution of data sets using ontologies for metadata. We propose an adaption of cross-entropy loss for cell type classification tailored to datasets annotated at different levels of coarseness. We demonstrate the utility of sfaira by training models across anatomic data partitions on 8 million cells.

RNA velocity has enabled the recovery of directed dynamic information from single‐cell transcriptomics by connecting measurements to the underlying kinetics of gene expression. This approach has opened up new ways of studying cellular dynamics. Here, we review the current state of RNA velocity modeling approaches, discuss various examples illustrating limitations and potential pitfalls, and provide guidance on how the ensuing challenges may be addressed. We then outline future directions on how to generalize the concept of RNA velocity to a wider variety of biological systems and modalities.

The in vivo phenotypic profile of T cells reactive to severe acute respiratory syndrome (SARS)-CoV-2 antigens remains poorly understood. Conventional methods to detect antigen-reactive T cells require in vitro antigenic re-stimulation or highly individualized peptide-human leukocyte antigen (pHLA) multimers. Here, we use single-cell RNA sequencing to identify and profile SARS-CoV-2-reactive T cells from Coronavirus Disease 2019 (COVID-19) patients. To do so, we induce transcriptional shifts by antigenic stimulation in vitro and take advantage of natural T cell receptor (TCR) sequences of clonally expanded T cells as barcodes for ‘reverse phenotyping’. This allows identification of SARS-CoV-2-reactive TCRs and reveals phenotypic effects introduced by antigen-specific stimulation. We characterize transcriptional signatures of currently and previously activated SARS-CoV-2-reactive T cells, and show correspondence with phenotypes of T cells from the respiratory tract of patients with severe disease in the presence or absence of virus in independent cohorts. Reverse phenotyping is a powerful tool to provide an integrated insight into cellular states of SARS-CoV-2-reactive T cells across tissues and activation states.

Cell biology is fundamentally limited in its ability to collect complete data on cellular phenotypes and the wide range of responses to perturbation. Areas such as computer vision and speech recognition have addressed this problem of characterizing unseen or unlabeled conditions with the combined advances of big data, deep learning, and computing resources in the past 5 years. Similarly, recent advances in machine learning approaches enabled by single-cell data start to address prediction tasks in perturbation response modeling. We first define objectives in learning perturbation response in single-cell omics; survey existing approaches, resources, and datasets (https://github.com/theislab/sc-pert); and discuss how a perturbation atlas can enable deep learning models to construct an informative perturbation latent space. We then examine future avenues toward more powerful and explainable modeling using deep neural networks, which enable the integration of disparate information sources and an understanding of heterogeneous, complex, and unseen systems.

Tissue niches are sources of cellular variation and key to understanding both single-cell and tissue phenotypes. The interaction of a cell with its niche can be described through cell communication events. These events cannot be directly observed in molecular profiling assays of single cells and have to be inferred. However, computational models of cell communication and variance attribution defined on data from dissociated tissues suffer from multiple limitations with respect to their ability to define and to identify communication events. We address these limitations using spatial molecular profiling data with node-centric expression modeling (NCEM), a computational method based on graph neural networks which reconciles variance attribution and communication modeling in a single model of tissue niches. We use these models in varying complexity across spatial assays, such as immunohistochemistry and MERFISH, and biological systems to demonstrate that the statistical cell–cell dependencies discovered by NCEM are plausible signatures of known molecular processes underlying cell communication. We identify principles of tissue organisation as cell communication events across multiple datasets using interpretation mechanisms. In the primary motor cortex, we found gene expression variation that is due to niche composition variation across cortical depth. Using the same approach, we also identified niche-dependent cell state variation in CD8 T cells from inflamed colon and colorectal cancer. Finally, we show that NCEMs can be extended to mixed models of explicit cell communication events and latent intrinsic sources of variation in conditional variational autoencoders to yield holistic models of cellular variation in spatial molecular profiling data. Altogether, this graphical model of cellular niches is a step towards understanding emergent tissue phenotypes.

Recent advances in multiplexed single-cell transcriptomics experiments are facilitating the high-throughput study of drug and genetic perturbations. However, an exhaustive exploration of the combinatorial perturbation space is experimentally unfeasible, so computational methods are needed to predict, interpret, and prioritize perturbations. Here, we present the compositional perturbation autoencoder (CPA), which combines the interpretability of linear models with the flexibility of deep-learning approaches for single-cell response modeling. CPA encodes and learns transcriptional drug responses across different cell type, dose, and drug combinations. The model produces easy-to-interpret embeddings for drugs and cell types, which enables drug similarity analysis and predictions for unseen dosage and drug combinations. We show that CPA accurately models single-cell perturbations across compounds, doses, species, and time. We further demonstrate that CPA predicts combinatorial genetic interactions of several types, implying that it captures features that distinguish different interaction programs. Finally, we demonstrate that CPA can generate in-silico 5,329 missing genetic combination perturbations (97.6% of all possibilities) with diverse genetic interactions. We envision our model will facilitate efficient experimental design and hypothesis generation by enabling in-silico response prediction at the single-cell level, and thus accelerate therapeutic applications using single-cell technologies.

Spatial omics data are advancing the study of tissue organization and cellular communication at an unprecedented scale. Here, we present Squidpy, a Python framework that brings together tools from omics and image analysis to enable scalable description of spatial molecular data, such as transcriptome or multivariate proteins. Squidpy provides both infrastructure and numerous analysis methods that allow to efficiently store, manipulate and interactively visualize spatial omics data.

The size and shape of peptide ions in the gas phase are an under-explored dimension for mass spectrometry-based proteomics. To investigate the nature and utility of the peptide collisional cross section (CCS) space, we measure more than a million data points from whole-proteome digests of five organisms with trapped ion mobility spectrometry (TIMS) and parallel accumulation-serial fragmentation (PASEF). The scale and precision (CV < 1%) of our data is sufficient to train a deep recurrent neural network that accurately predicts CCS values solely based on the peptide sequence. Cross section predictions for the synthetic ProteomeTools peptides validate the model within a 1.4% median relative error (R > 0.99). Hydrophobicity, proportion of prolines and position of histidines are main determinants of the cross sections in addition to sequence-specific interactions. CCS values can now be predicted for any peptide and organism, forming a basis for advanced proteomics workflows that make full use of the additional information.

Chronic activation and dysregulation of the neuroendocrine stress response have severe physiological and psychological consequences, including the development of metabolic and stress-related psychiatric disorders. We provide the first unbiased, cell type–specific, molecular characterization of all three components of the hypothalamic-pituitary-adrenal axis, under baseline and chronic stress conditions. Among others, we identified a previously unreported subpopulation of Abcb1b+ cells involved in stress adaptation in the adrenal gland. We validated our findings in a mouse stress model, adrenal tissues from patients with Cushing’s syndrome, adrenocortical cell lines, and peripheral cortisol and genotyping data from depressed patients. This extensive dataset provides a valuable resource for researchers and clinicians interested in the organism’s nervous and endocrine responses to stress and the interplay between these tissues. Our findings raise the possibility that modulating ABCB1 function may be important in the development of treatment strategies for patients suffering from metabolic and stress-related psychiatric disorders.

Motivation: While generative models have shown great success in sampling high-dimensional samples conditional on low-dimensional descriptors (stroke thickness in MNIST, hair color in CelebA, speaker identity in WaveNet), their generation out-of-distribution poses fundamental problems due to the difficulty of learning compact joint distribution across conditions. The canonical example of the conditional variational autoencoder (CVAE), for instance, does not explicitly relate conditions during training and, hence, has no explicit incentive of learning such a compact representation.

Results: We overcome the limitation of the CVAE by matching distributions across conditions using maximum mean discrepancy in the decoder layer that follows the bottleneck. This introduces a strong regularization both for reconstructing samples within the same condition and for transforming samples across conditions, resulting in much improved generalization. As this amount to solving a style-transfer problem, we refer to the model as transfer VAE (trVAE). Benchmarking trVAE on high-dimensional image and single-cell RNA-seq, we demonstrate higher robustness and higher accuracy than existing approaches. We also show qualitatively improved predictions by tackling previously problematic minority classes and multiple conditions in the context of cellular perturbation response to treatment and disease based on high-dimensional single-cell gene expression data. For generic tasks, we improve Pearson correlations of high-dimensional estimated means and variances with their ground truths from 0.89 to 0.97 and 0.75 to 0.87, respectively. We further demonstrate that trVAE learns cell-type-specific responses after perturbation and improves the prediction of most cell-type-specific genes by 65%.

Multispectral Optoacoustic Tomography (MSOT) resolves oxy- (HbO2) and deoxy-hemoglobin (Hb) to perform vascular imaging. MSOT suffers from gradual signal attenuation with depth due to light-tissue interactions: an effect that hinders the precise manual segmentation of vessels. Furthermore, vascular assessment requires functional tests, which last several minutes and result in recording thousands of images. Here, we introduce a deep learning approach with a sparse-UNET (S-UNET) for automatic vascular segmentation in MSOT images to avoid the rigorous and time-consuming manual segmentation. We evaluated the S-UNET on a test-set of 33 images, achieving a median DICE score of 0.88. Apart from high segmentation performance, our method based its decision on two wavelengths with physical meaning for the task-at-hand: 850 nm (peak absorption of oxy-hemoglobin) and 810 nm (isosbestic point of oxy-and deoxy-hemoglobin). Thus, our approach achieves precise data-driven vascular segmentation for automated vascular assessment and may boost MSOT further towards its clinical translation.

High-content imaging and single-cell genomics are two of the most prominent high-throughput technologies for studying cellular properties and functions at scale. Recent studies have demonstrated that information in large imaging datasets can be used to estimate gene mutations and to predict the cell-cycle state and the cellular decision making directly from cellular morphology. Thus, high-throughput imaging methodologies, such as imaging flow cytometry can potentially aim beyond simple sorting of cell-populations. We introduce IFC-seq, a machine learning methodology for predicting the expression profile of every cell in an imaging flow cytometry experiment. Since it is to-date unfeasible to observe single-cell gene expression and morphology in flow, we integrate uncoupled imaging data with an independent transcriptomics dataset by leveraging common surface markers. We demonstrate that IFC-seq successfully models gene expression of a moderate number of key gene-markers for two independent imaging flow cytometry datasets: (i) human blood mononuclear cells and (ii) mouse myeloid progenitor cells. In the case of mouse myeloid progenitor cells IFC-seq can predict gene expression directly from brightfield images in a label-free manner, using a convolutional neural network. The proposed method promises to add gene expression information to existing and new imaging flow cytometry datasets, at no additional cost.

It has recently become possible to simultaneously assay T-cell specificity with respect to large sets of antigens and the T-cell receptor sequence in high-throughput single-cell experiments. Leveraging this new type of data, we propose and benchmark a collection of deep learning architectures to model T-cell specificity in single cells. In agreement with previous results, we found that models that treat antigens as categorical outcome variables outperform those that model the TCR and antigen sequence jointly. Moreover, we show that variability in single-cell immune repertoire screens can be mitigated by modeling cell-specific covariates. Lastly, we demonstrate that the number of bound pMHC complexes can be predicted in a continuous fashion providing a gateway to disentangle cell-to-dextramer binding strength and receptor-to-pMHC affinity. We provide these models in the Python package TcellMatch to allow imputation of antigen specificities in single-cell RNA-seq studies on T cells without the need for MHC staining.

RNA velocity has opened up new ways of studying cellular differentiation in single-cell RNA-sequencing data. It describes the rate of gene expression change for an individual gene at a given time point based on the ratio of its spliced and unspliced messenger RNA (mRNA). However, errors in velocity estimates arise if the central assumptions of a common splicing rate and the observation of the full splicing dynamics with steady-state mRNA levels are violated. Here we present scVelo, a method that overcomes these limitations by solving the full transcriptional dynamics of splicing kinetics using a likelihood-based dynamical model. This generalizes RNA velocity to systems with transient cell states, which are common in development and in response to perturbations. We apply scVelo to disentangling subpopulation kinetics in neurogenesis and pancreatic endocrinogenesis. We infer gene-specific rates of transcription, splicing and degradation, recover each cell’s position in the underlying differentiation processes and detect putative driver genes. scVelo will facilitate the study of lineage decisions and gene regulation.

The in vivo detection of dead cells remains a major challenge due to technical hurdles. Here, we present a novel method, where injection of fluorescent milk fat globule-EGF factor 8 protein (MFG-E8) in vivo combined with imaging flow cytometry and deep learning allows the identification of dead cells based on their surface exposure of phosphatidylserine (PS) and other image parameters. A convolutional autoencoder (CAE) was trained on defined pictures and successfully used to identify apoptotic cells in vivo. However, unexpectedly, these analyses also revealed that the great majority of PS+ cells were not apoptotic, but rather live cells associated with PS+ extracellular vesicles (EVs). During acute viral infection apoptotic cells increased slightly, while up to 30% of lymphocytes were decorated with PS+ EVs of antigen-presenting cell (APC) exosomal origin. The combination of recombinant fluorescent MFG-E8 and the CAE-method will greatly facilitate analyses of cell death and EVs in vivo.

Large single-cell atlases are now routinely generated with the aim of serving as reference to analyse future smaller-scale studies. Yet, learning from reference data is complicated by batch effects between datasets, limited availability of computational resources, and sharing restrictions on raw data. Leveraging advances in machine learning, we propose a deep learning strategy to map query datasets on top of a reference called single-cell architectural surgery (scArches, https://github.com/theislab/scarches). It uses transfer learning and parameter optimization to enable efficient, decentralized, iterative reference building, and the contextualization of new datasets with existing references without sharing raw data. Using examples from mouse brain, pancreas, and whole organism atlases, we showcase that scArches preserves nuanced biological state information while removing batch effects in the data, despite using four orders of magnitude fewer parameters compared to de novo integration. To demonstrate mapping disease variation, we show that scArches preserves detailed COVID-19 disease variation upon reference mapping, enabling discovery of new cell identities that are unseen during training. We envision our method to facilitate collaborative projects by enabling the iterative construction, updating, sharing, and efficient use of reference atlases.

Despite the epidemics of chronic obstructive pulmonary disease (COPD), the cellular and molecular mechanisms of this disease are far from being understood. Here, we characterize and classify the cellular composition within the alveolar space and peripheral blood of COPD patients and control donors using a clinically applicable single-cell RNA-seq technology corroborated by advanced computational approaches for: machine learning-based cell-type classification, identification of differentially expressed genes, prediction of metabolic changes, and modeling of cellular trajectories within a patient cohort. These high-resolution approaches revealed: massive transcriptional plasticity of macrophages in the alveolar space with increased levels of invading and proliferating cells, loss of MHC expression, reduced cellular motility, altered lipid metabolism, and a metabolic shift reminiscent of mitochondrial dysfunction in COPD patients. Collectively, single-cell omics of multi-tissue samples was used to build the first cellular and molecular framework for COPD pathophysiology as a prerequisite to develop molecular biomarkers and causal therapies against this deadly disease.

Dedifferentiation of insulin-secreting β cells in the islets of Langerhans has been proposed to be a major mechanism of β-cell dysfunction. Whether dedifferentiated β cells can be targeted by pharmacological intervention for diabetes remission, and ways in which this could be accomplished, are unknown as yet. Here we report the use of streptozotocin-induced diabetes to study β-cell dedifferentiation in mice. Single-cell RNA sequencing (scRNA-seq) of islets identified markers and pathways associated with β-cell dedifferentiation and dysfunction. Single and combinatorial pharmacology further show that insulin treatment triggers insulin receptor pathway activation in β cells and restores maturation and function for diabetes remission. Additional β-cell selective delivery of oestrogen by Glucagon-like peptide-1 (GLP-1–oestrogen conjugate) decreases daily insulin requirements by 60%, triggers oestrogen-specific activation of the endoplasmic-reticulum-associated protein degradation system, and further increases β-cell survival and regeneration. GLP-1–oestrogen also protects human β cells against cytokine-induced dysfunction. This study not only describes mechanisms of β-cell dedifferentiation and regeneration, but also reveals pharmacological entry points to target dedifferentiated β cells for diabetes remission.

Accurately modeling cellular response to perturbations is a central goal of computational biology. While such modeling has been based on statistical, mechanistic and machine learning models in specific settings, no generalization of predictions to phenomena absent from training data (out-of-sample) has yet been demonstrated. Here, we present scGen (https://github.com/theislab/scgen), a model combining variational autoencoders and latent space vector arithmetics for high-dimensional single-cell gene expression data. We show that scGen accurately models perturbation and infection response of cells across cell types, studies and species. In particular, we demonstrate that scGen learns cell-type and species-specific responses implying that it captures features that distinguish responding from non-responding genes and cells. With the upcoming availability of large-scale atlases of organs in a healthy state, we envision scGen to become a tool for experimental design through in silico screening of perturbation response in the context of disease and drug treatment.

Single-cell RNA sequencing (scRNA-seq) has highlighted the important role of intercellular heterogeneity in phenotype variability in both health and disease1. However, current scRNA-seq approaches provide only a snapshot of gene expression and convey little information on the true temporal dynamics and stochastic nature of transcription. A further key limitation of scRNA-seq analysis is that the RNA profile of each individual cell can be analysed only once. Here we introduce single-cell, thiol-(SH)-linked alkylation of RNA for metabolic labelling sequencing (scSLAM-seq), which integrates metabolic RNA labelling2, biochemical nucleoside conversion3 and scRNA-seq to record transcriptional activity directly by differentiating between new and old RNA for thousands of genes per single cell. We use scSLAM-seq to study the onset of infection with lytic cytomegalovirus in single mouse fibroblasts. The cell-cycle state and dose of infection deduced from old RNA enable dose–response analysis based on new RNA. scSLAM-seq thereby both visualizes and explains differences in transcriptional activity at the single-cell level. Furthermore, it depicts ‘on–off’ switches and transcriptional burst kinetics in host gene expression with extensive gene-specific differences that correlate with promoter-intrinsic features (TBP–TATA-box interactions and DNA methylation). Thus, gene-specific, and not cell-specific, features explain the heterogeneity in transcriptomes between individual cells and the transcriptional response to perturbations.

Single‐cell RNA‐seq has enabled gene expression to be studied at an unprecedented resolution. The promise of this technology is attracting a growing user base for single‐cell analysis methods. As more analysis tools are becoming available, it is becoming increasingly difficult to navigate this landscape and produce an up‐to‐date workflow to analyse one’s data. Here, we detail the steps of a typical single‐cell RNA‐seq analysis, including pre‐processing (quality control, normalization, data correction, feature selection, and dimensionality reduction) and cell‐ and gene‐level downstream analysis. We formulate current best‐practice recommendations for these steps based on independent comparison studies. We have integrated these best‐practice recommendations into a workflow, which we apply to a public dataset to further illustrate how these steps work in practice. Our documented case study can be found at https://www.github.com/theislab/single-cell-tutorial. This review will serve as a workflow tutorial for new entrants into the field, and help established users update their analysis pipelines.