An atlas-scale generative model for unified representation learning of bulk RNA-seq data

Benchmarking gene expression reconstruction from single-cell latent representations

Single-cell transcriptomics is typically modeled in low-dimensional latent representations that improve the signal-to-noise ratio of the data. Such representations underpin data integration, cell state discovery, and perturbation prediction, with applications ranging from large-scale organ atlases to latent trajectory modeling. Recent virtual cell approaches further leverage these representations to predict cellular responses as distributional shifts in latent space. Each of these applications ultimately requires faithful gene expression reconstruction from latent spaces for biological interpretation, enabling gene-level analysis of predicted perturbed or batch-corrected cells. Yet representation choice is typically treated as an implementation detail rather than a primary modeling decision, with no systematic evaluation of how well latent representations support gene expression reconstruction. Here, we introduce ReconEval, a benchmark for evaluating gene expression reconstruction from single-cell latent spaces. We benchmark two classes of latent representations: end-to-end trained models such as PCA, autoencoders, and variational autoencoders, and pretrained single-cell foundation model embeddings coupled to newly trained decoders. Reconstruction is evaluated both directly and after latent-space perturbation prediction. Across perturbational and observational datasets totaling over 100 million cells, our metric suite quantifies statistical fidelity; biological signal preservation, including differential expression, coexpression, cell-cycle structure, cytokine response and pathway activity; and perturbation-specific effects. We find that autoencoders achieve the strongest stand-alone reconstruction at low dimensionality, while variational regularization does not improve generalization in reconstruction. Frozen foundation model embeddings retain recoverable gene-level information, with reconstruction quality depending strongly on decoder architecture and pretraining objective. In latent perturbation modeling, high-dimensional PCA matches foundation model embeddings, while low-dimensional AE embeddings are optimal for flow-based generative models. Overall, reconstruction depends critically on the interplay between representation and downstream model, and simpler representations can outperform complex alternatives given appropriate capacity. Our benchmark establishes reconstruction as a critical evaluation axis for single-cell foundation models. We envision it improving the biological interpretability of latent-space modeling, a prerequisite for future virtual cell models to be validated by domain experts and grounded in biology. ### Competing Interest Statement F.J.T. consults for Immunai, CytoReason, Valinor Industries, Bioturing and Phylo Inc., and has ownership interest in RN.AI Therapeutics, Dermagnostix, and Cellarity. The remaining authors declare no competing interests.

Realistic in silico generation and augmentation of single-cell RNA-seq data using generative adversarial networks

A fundamental problem in biomedical research is the low number of observations available, mostly due to a lack of available biosamples, prohibitive costs, or ethical reasons. Augmenting few real observations with generated in silico samples could lead to more robust analysis results and a higher reproducibility rate. Here, we propose the use of conditional single-cell generative adversarial neural networks (cscGAN) for the realistic generation of single-cell RNA-seq data. cscGAN learns non-linear gene–gene dependencies from complex, multiple cell type samples and uses this information to generate realistic cells of defined types. Augmenting sparse cell populations with cscGAN generated cells improves downstream analyses such as the detection of marker genes, the robustness and reliability of classifiers, the assessment of novel analysis algorithms, and might reduce the number of animal experiments and costs in consequence. cscGAN outperforms existing methods for single-cell RNA-seq data generation in quality and hold great promise for the realistic generation and augmentation of other biomedical data types.

Realistic in silico generation and augmentation of single cell RNA-seq data using Generative Adversarial Neural Networks

A fundamental problem in biomedical research is the low number of observations available, mostly due to a lack of available biosamples, prohibitive costs, or ethical reasons. Augmenting few real observations with generated in silico samples could lead to more robust analysis results and a higher reproducibility rate. Here we propose the use of conditional single cell Generative Adversarial Neural Networks (cscGANs) for the realistic generation of single cell RNA-seq data. cscGANs learn non-linear gene-gene dependencies from complex, multi cell type samples and use this information to generate realistic cells of defined types. Augmenting sparse cell populations with cscGAN generated cells improves downstream analyses such as the detection of marker genes, the robustness and reliability of classifiers, the assessment of novel analysis algorithms, and might reduce the number of animal experiments and costs in consequence. cscGANs outperform existing methods for single cell RNA-seq data generation in quality and hold great promise for the realistic generation and augmentation of other biomedical data types. *

Rectangle: robust and scalable multiscale deconvolution informed by single-cell RNA sequencing data

Bulk RNA-seq enables effective profiling of large cohorts and complex experimental designs, but current single-cell-informed deconvolution methods incompletely resolve closely related cell phenotypes, do not scale efficiently to large single-cell datasets, or fail to account for cellular content not represented in the reference. Here, we present Rectangle, an scverse Python framework for single-cell-informed deconvolution of bulk RNA-seq data.

Organism-scale annotation with Pan-human Azimuth

Single-cell atlases now span many human tissues, but inconsistent annotations across studies limit their utility as a unified reference. We introduce Pan-human Azimuth, a supervised neural network that maps human cells from diverse tissues and datasets onto a single hierarchical organism-scale typology.

VAPOR: A variational autoencoder with transport operators to disentangle cellular gene expression dynamics of co-occurring biological processes in time and space

Abstract Single-cell and spatial transcriptomics enable the analysis of cellular states and dynamics in gene expression, revealing how diverse biological processes relate to these states over time and space. To study these dynamics, trajectory inference methods order cells along computationally inferred paths to reconstruct gradual transitions in cell states. However, by encouraging smooth and continuous trajectories, these approaches tend to conflate co-occurring processes-such as proliferation, maturation, and spatial organization-that are jointly reflected in gene expression, potentially overlooking process-specific gene expression dynamics. To address this, we developed VAPOR, which integrates a variational autoencoder with transport operators to model and disentangle cellular gene expression dynamics for potentially co-occurring biological processes. VAPOR inputs single-cell (or spatial) gene expression data into a variational autoencoder (VAE) to learn the latent states of cells and then models their latent dynamics as an ordinary differential equation. The latent dynamics are further decomposed into process-specific components parameterized by transport operators (TOs) and their corresponding process weights. Each TO defines a process-specific dynamics, and its weight for each cell quantifies the process's contribution to the cell dynamics. After assessment by simulation studies, we applied VAPOR with benchmarking to real data, including time-course scRNA-seq from postconceptual human brain development, spatial transcriptomics of the mouse hippocampus, and cross-species scRNA-seq spanning human and macaque first-trimester forebrain development. In these applications, VAPOR has identified a variety of temporal and spatial co-occurring processes, such as cell cycle, gliogenesis, neurogenesis, and neuronal migration, along with associated dynamic genes, including those species-specific to human and macaque development. VAPOR is available as an open-source tool for general-purpose use. ### Competing Interest Statement The authors have declared no competing interest.

GAN-Based Data Augmentation for Prediction Improvement Using Gene Expression Data in Cancer

Within the area of bioinformatics, Deep Learning (DL) models have shown exceptional results in applications in which histological images, scans and tomographies are used. However, when gene expression data is under analysis, the performance is often limited, further hampered by the complexity of these models that require several instances, in the order of thousands, to provide good results. Due to the difficulty and the costs involved in the collection of medical data, the application of Data Augmentation (DA) techniques to alleviate the lack of samples is a topic of great relevance. State-of-the-art models based on Conditional Generative Adversarial Networks (CGAN) and some introduced modifications are used in this work to investigate the effect of DA for prediction of the vital status of patients from RNA-Seq gene expression data. Experimental results on several real-world data sets demonstrate the effectiveness and efficiency of the proposed models. The application of DA methods significantly increase prediction accuracy, leading by 12% with respect to benchmark data sets and 3.15% with respect to data processed with feature selection. Results based on CGAN models outperform in most cases, alternative methods like the SMOTE or noise injection techniques.

Hierarchical classification of immune cell transcriptomes at population-scale

Accurate immune cell classification is essential for interpreting single-cell RNA sequencing (scRNA-seq) data. However, progress is constrained by the lack of independent, high-resolution benchmarks, as the routine integration of datasets introduces statistical dependencies that artificially inflate model generalizability. Here, we present the single-cell universal classification omnibus (Suco), a resource of independent, uniform expert annotations, and Compocyte, a modular hierarchical classifier. Together, they establish a framework designed for the scale of human population immunology. This approach substantially outperforms existing classifiers while facilitating expert review of ambiguous annotations. Applying Compocyte across 50 studies, including three newly generated datasets, we classified 15.6 million leukocytes from 3,965 patients. Within this expansive cohort, we identified a new tumor-associated resorptive macrophage phenotype, a non-canonical monocyte subtype in subclinical cytokine release syndrome, and the programmatic erosion of T cell memory stemness across metastatic sites. Suco and Compocyte thus provide a generalizable architecture and benchmark capable of sustaining high-resolution annotation across massive clinical cohorts. ### Competing Interest Statement CMR has consulted regarding oncology drug development with Amgen, AstraZeneca, Daiichi Sankyo, Genentech, Merck, and Novartis, and has received licensing and royalty payments for DLL3-directed therapeutics. T.W. reports stock ownership for Roche, Astra Zeneca, Bayer, Innate Pharma, Kyntra, Illumina, 10x Genomics, and Merck KGaA as well as research funding from Atrandi Biosciences, Vilnius, Lithuania; CanVirex AG, Basel Switzerland; and Institut fuer Klinische Krebsforschung GmbH, Frankfurt, Germany, and travel funding from Roche, Basel, Switzerland. S.Z. reports advisory board membership and honoraria from Amgen, Astellas, AstraZeneca, Bayer, Bristol-Myers Squibb, Daiichi Sankyo, Eisai, EUSA, Gilead, Ipsen, Johnson&Johnson, Lilly, MedSir, Medtoday, Merck, MSD, Novartis, Pfizer, Roche, Sanofi Aventis, StreamedUp, Urotrials, Urotube, Zentiva and resarch funding from Eisai. S.Z. reports clinical trial support from Amgen, AstraZeneca, AVEO, Bayer, Biontech, Bristol-Myers Squibb, Calithera, Exelixis, Gilead, Lilly, MSD, Novartis, Pfizer, Roche, Seagen/Astellas, Urotrials and travels & conference support from Amgen, Astellas, AstraZeneca, Bayer, EISAI, Ipsen, Johnson&Johnson, Merck, MSD, Pfizer. All remaining authors declare no relevant competing interests. Spanish Association Against Cancer, PI049999 Federal Ministry of Research, Technology and Space, 001001KT2322 National Cancer Institute, R35 CA263816 National Cancer Institute, U24 CA213274 National Cancer Institute, P30 CA008748 Research Council of Lithuania, P-MIP-24-93

Gene regulatory networks: from correlative models to causal explanations

Gene regulatory networks (GRNs) explain how the genome controls cellular behaviour and tissue morphogenesis, serving to connect molecular mechanism to functional output. Single-cell technologies now provide descriptions of these networks with unprecedented detail, but this advance has also revealed gene regulatory systems that are too complex for our existing conceptual frameworks. GRNs, which should provide mechanistic explanations, are increasingly reduced to statistical correlations — ‘hairballs’ that fail to capture molecular causation. Here, we explore why this dilemma exists and propose a path forward. We argue that methods in ‘representation learning’ can be used to model GRNs, without needing to capture every molecular detail. For this framework, we advocate three linked principles: models must be inherently mechanistic, with structures grounded in cellular and evolutionary biology; molecular principles and constraints must be used to reduce the solution space for learning GRN models; and more sophisticated forms of experimental perturbation and synthetic biological engineering are needed to train models and test predictions. By reimagining GRNs through these principles, we can bridge the gap from data abundance to new conceptual understanding.

EmbeddingGemma: Powerful and Lightweight Text Representations

We introduce EmbeddingGemma, a new lightweight, open text embedding model based on the Gemma 3 language model family. Our innovative training recipe strategically captures knowledge from larger models via encoder-decoder initialization and geometric embedding distillation. We improve model robustness and expressiveness with a spread-out regularizer, and ensure generalizability by merging checkpoints from varied, optimized mixtures. Evaluated on the Massive Text Embedding Benchmark (MTEB) across multilingual, English, and code domains, EmbeddingGemma (300M) achieves state-of-the-art results. Notably, it outperforms prior top models, both proprietary and open, with fewer than 500M parameters, and provides performance comparable to models double its size, offering an exceptional performance-to-cost ratio. Remarkably, this lead persists when quantizing model weights or truncating embedding outputs. This makes EmbeddingGemma particularly well-suited for low-latency and high-throughput use cases such as on-device applications. We provide ablation studies exploring our key design choices. We release EmbeddingGemma to the community to promote further research.

Embryo-scale reverse genetics at single-cell resolution

The maturation of single-cell transcriptomic technologies has facilitated the generation of comprehensive cellular atlases from whole embryos1–4. A majority of these data, however, has been collected from wild-type embryos without an appreciation for the latent variation that is present in development. Here we present the ‘zebrafish single-cell atlas of perturbed embryos’: single-cell transcriptomic data from 1,812 individually resolved developing zebrafish embryos, encompassing 19 timepoints, 23 genetic perturbations and a total of 3.2 million cells. The high degree of replication in our study (eight or more embryos per condition) enables us to estimate the variance in cell type abundance organism-wide and to detect perturbation-dependent deviance in cell type composition relative to wild-type embryos. Our approach is sensitive to rare cell types, resolving developmental trajectories and genetic dependencies in the cranial ganglia neurons, a cell population that comprises less than 1% of the embryo. Additionally, time-series profiling of individual mutants identified a group of brachyury-independent cells with strikingly similar transcriptomes to notochord sheath cells, leading to new hypotheses about early origins of the skull. We anticipate that standardized collection of high-resolution, organism-scale single-cell data from large numbers of individual embryos will enable mapping of the genetic dependencies of zebrafish cell types, while also addressing longstanding challenges in developmental genetics, including the cellular and transcriptional plasticity underlying phenotypic diversity across individuals.

Self-Driving Datasets: From 20 Million Papers to Nuanced Biomedical Knowledge at Scale

Manually curated biomedical repositories -- spanning bioactivity, genomics, and chemistry -- are expensive to maintain, lag behind primary literature, and discard experimental context, obscuring nuances needed to assess data correctness and coverage. We show that PubMed itself can be autonomously and cost-effectively turned into structured datasets that are larger, more nuanced, and more accurate than the curated databases they replace. We present three coupled contributions: (1) an LLM-based entity-tagging pipeline, grounded in nine biomedical ontologies, that tags 4.5B entities across 19 categories in a 22.5M-paper, 2.5T-token PubMed corpus; (2) hybrid sparse-dense retrieval supporting entity-filtered semantic queries over the tagged corpus; and (3) Starling, a multi-agent deep research system that, given only a natural-language task description, designs precision- and recall-targeted retrieval filters, induces an extraction schema, and emits structured records with nuance-rich fields and supporting passages. Across six tasks -- blood-brain barrier permeability, oral bioavailability, acute toxicity (LD50), gene-disease associations, protein subcellular localization, and chemical reactions -- Starling produces ~6.3M records (91K-3M per task); several are, to our knowledge, the largest public datasets for their property. Frontier-model rejection of our extractions is 0.6-7.7% across tasks, far below error rates we measure on widely used curated counterparts (e.g., 16.5% on BBB_Martins, 7.3% on Bioavailability_Ma). Beyond scale and accuracy, the supporting passages carry nuance tabular databases discard -- e.g., oral bioavailability may depend on fed vs. fasted state. Together, the corpus, retrieval, and agent establish a foundation for AI-driven therapeutic design. Code and datasets: https://github.com/starling-labs/starling.

Designing RNA sequencing experiments: A practical guide to reproducible gene expression analysis

RNA sequencing (RNA-seq) has become a cornerstone of modern biotechnology, offering a comprehensive and high-resolution view of gene expression that enables the discovery of novel transcripts across diverse biological systems. Its applications extend beyond basic transcriptomics, providing powerful tools for uncovering molecular mechanisms underlying disease, environmental responses, and chemical toxicity. In biotechnology and biomedical research, RNA-seq facilitates the identification of regulatory networks and biomarkers that inform therapeutic development, risk assessment, and precision medicine.

AI models collapse when trained on recursively generated data

Stable diffusion revolutionized image creation from descriptive text. GPT-2 (ref. 1), GPT-3(.5) (ref. 2) and GPT-4 (ref. 3) demonstrated high performance across a variety of language tasks. ChatGPT introduced such language models to the public. It is now clear that generative artificial intelligence (AI) such as large language models (LLMs) is here to stay and will substantially change the ecosystem of online text and images. Here we consider what may happen to GPT-{n} once LLMs contribute much of the text found online. We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear. We refer to this effect as ‘model collapse’ and show that it can occur in LLMs as well as in variational autoencoders (VAEs) and Gaussian mixture models (GMMs). We build theoretical intuition behind the phenomenon and portray its ubiquity among all learned generative models. We demonstrate that it must be taken seriously if we are to sustain the benefits of training from large-scale data scraped from the web. Indeed, the value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of LLM-generated content in data crawled from the Internet.

NeuroRVQ: Multi-Scale Biosignal Tokenization for Generative Foundation Models

NeuroRVQ: Multi-Scale Biosignal Tokenization for Generative Foundation Models

Quantifying cell-state densities in single-cell phenotypic landscapes using Mellon

Cell-state density characterizes the distribution of cells along phenotypic landscapes and is crucial for unraveling the mechanisms that drive diverse biological processes. Here, we present Mellon, an algorithm for estimation of cell-state densities from high-dimensional representations of single-cell data. We demonstrate Mellon’s efficacy by dissecting the density landscape of differentiating systems, revealing a consistent pattern of high-density regions corresponding to major cell types intertwined with low-density, rare transitory states. We present evidence implicating enhancer priming and the activation of master regulators in emergence of these transitory states. Mellon offers the flexibility to perform temporal interpolation of time-series data, providing a detailed view of cell-state dynamics during developmental processes. Mellon facilitates density estimation across various single-cell data modalities, scaling linearly with the number of cells. Our work underscores the importance of cell-state density in understanding the differentiation processes, and the potential of Mellon to provide insights into mechanisms guiding biological trajectories.