Yang Joon Kim

Predicting how cells respond to genetic and chemical perturbations is a central challenge in drug discovery and functional genomics. A growing ecosystem of specialized single-cell foundation models has been developed to address this problem, yet their practical advantage over domain-agnostic approaches remains unclear. Here we evaluate the power of Tabular Foundation Models such as TabICL and TabPFN, general-purpose pre-trained regression models, against domain-specific architectures including PRESAGE, scGPT, scLAMBDA, STACK and Prophet across four complementary evaluation settings: cell-level in-context cross-cell-type prediction, pseudobulk perturbation prediction on five Perturb-seq datasets of cell-lines, a genome-wide CRISPR screen in primary human CD4 ⁺ T cells, and embryo-level cell-type composition prediction in a zebrafish developmental perturbation atlas. In the cell-level cross-cell type perturbation prediction, Tabular Foundation Models perform on par or better than specialized models. On pseudobulk perturbation prediction, Tabular Foundation Models consistently outperform specialized baselines across multiple evaluation metrics and datasets. On whole-emrbryo cell-type composition prediction, Tabular Foundation Models are competitive with specialized baselines. These results demonstrate that general-purpose tabular in-context learning provides a strong and scalable alternative to bespoke biological architectures for perturbation response modeling across cell systems and scales.

Zebrafish RNA-seq studies span diverse developmental, physiological, and disease contexts, yet most analyses remain confined to individual experiments and disregard the non-zebrafish component of the data. We present RNAquarium, a scalable framework for joint transcriptomic and metatranscriptomic analysis of RNA-seq data and apply it to all publicly available zebrafish RNA-seq datasets in the Sequence Read Archive. RNAquarium captures transcriptomic structure across development and tissues, reveals diverse microbial and viral associations, and identifies previously undescribed zebrafish viruses including a close relative of human influenza B virus linked to distinct host transcriptional states. We further demonstrate that archive-scale transcriptomes can support foundation-model training and prediction of infection-associated transcriptomic signatures. RNAquarium provides an open framework and interactive portal for exploring the breadth of zebrafish gene expression patterns and associated taxa profiled across a large research community. Moreover, our framework establishes a generalizable strategy for integrating transcriptomic and metatranscriptomic analyses across the diversity of life represented in public sequencing archives.

Deep learning DNA sequence-to-function models offer the promise of gaining mechanistic insights into genome regulation, however their performance is often limited by data scarcity in the species of interest. We present DanioDecima, a zebrafish-specific model leveraging transfer learning from human and mouse-trained models to predict tissue- and cell-type-specific gene expression during zebrafish embryogenesis. Initializing DanioDecima with pretrained human and mouse Borzoi and Decima weights raises the median pseudobulk Pearson r substantially across cell-types and improves gene-level correlations of test set genes. An in silico directed-evolution loop guided by DanioDecima scoring generated synthetic promoters whose motif architectures cluster by the expected target lineage. These findings exemplify a cross-species transfer learning methodology for sequence-to-function models, and position DanioDecima as a practical resource for zebrafish regulatory engineering.

Different glioblastoma (GBM) subtypes have been identified based on the tumor microenvironment (TME). The discovery of new therapies for these hard-to-treat tumors requires a thorough characterization of preclinical models, including their TME, to apply preclinical results to the most similar GBM subtype. Using single-nucleus RNA sequencing (snRNA-seq), we characterized the tumor and TME in an immunocompetent mouse model with intracranially implanted GBM stem cells at different stages and treatments. Visium spatial transcriptomics confirmed the location of annotated cells. This model exhibits GBM targets related to integration into neural circuits - Grik2, Nlgn3, Gap43 or Kcnn4-, immunoevasion - Nt5e, Cd274 or Irf8- and immunosuppression - Csf1r, Arg1, Mrc1 and Tgfb1. The landscape of cytokines, checkpoint ligands and receptors uncovered Mrc1, PD-L1, TIM-3 or B7-H3, among the immunotherapy targets that can be addressed in this model. The comparison with human GBMs unveiled crucial similarities with TME^Med GBM, the most frequent subtype.

During embryonic development, gene regulatory networks (GRNs) drive molecular differentiation of cell types. However, the temporal dynamics of these networks remain poorly understood. Here, we present Zebrahub-Multiome, a single-cell multiomic atlas that captures chromatin accessibility and gene expression from 94,562 cells across six stages of zebrafish embryogenesis (10-24 hours post-fertilization), capturing key developmental stages from the end of gastrulation to the onset of organogenesis. By measuring regulatory element activity alongside transcriptional output from the same cells, we identify 640,000 cis-regulatory elements organized into 402 hierarchically structured modules corresponding to specific developmental pathways. Early embryonic stages employ broadly shared regulatory programs that progressively fragment into lineage-specific modules. Timeresolved gene regulatory network inference reveals that transcription factors undergo functional transitions – from multilineage regulators to specialized, lineage-committed factors. These quantitative measurements reveal the regulatory network rewiring that drives cell fate specification. Our interactive web portal ( zebrahub.org/epigenomics ) enables exploration of gene dynamics, regulatory networks, and perturbation predictions, providing a quantitative framework for understanding vertebrate developmental regulation.

Assigning cell identity to clusters of single cells is an essential step towards extracting biological insights from many genomics datasets. Although annotation workflows for datasets built with a single modality are well established, limitations exist in annotating cell types in datasets with multiple modalities due to the need for a framework to exploit them jointly. While, in principle, different modalities could convey complementary information about cell identity, it is unclear to what extent they can be combined to improve the accuracy and resolution of cell type annotations. Here, we present a conceptual framework to examine and jointly interrogate distinct modalities to identify cell types. We integrated our framework into a series of vignettes, using immune cells as a well-studied example, and demonstrate cell type annotation workflows ranging from using single-cell RNA-seq datasets alone, to using multiple modalities such as single-cell Multiome (RNA and chromatin accessibility), CITE-seq (RNA and surface proteins). In some cases, one or other single modality is superior to the other for identification of specific cell types, in others combining the two modalities improves resolution and the ability to identify finer subpopulations. Finally, we use interactive software from CZ CELLxGENE community tools to visualize and integrate histological and spatial transcriptomic data.

Elucidating the developmental processes of organisms requires a comprehensive understanding of cellular lineages in the spatial, temporal, and molecular domains. In this study, we introduce Zebrahub, a dynamic atlas of zebrafish embryonic development that integrates single-cell sequencing time course data with lineage reconstructions facilitated by light-sheet microscopy. This atlas offers high-resolution and in-depth molecular insights into zebrafish development, achieved through the sequencing of individual embryos across ten developmental stages, complemented by trajectory reconstructions. Zebrahub also incorporates an interactive tool to navigate the complex cellular flows and lineages derived from light-sheet microscopy data, enabling in silico fate mapping experiments. To demonstrate the versatility of our multi-modal resource, we utilize Zebrahub to provide fresh insights into the pluripotency of Neuro-Mesodermal Progenitors (NMPs). Our publicly accessible web-based platform, Zebrahub, is a foundational resource for studying developmental processes at both transcriptional and spatiotemporal levels, providing researchers with an integrated approach to exploring and analyzing the complexities of cellular lineages during zebrafish embryogenesis.

Spatial transcriptomics extends single-cell RNA sequencing (scRNA-seq) by providing spatial context for cell type identification and analysis. Imaging-based spatial technologies such as multiplexed error-robust fluorescence in situ hybridization (MERFISH) can achieve single-cell resolution, directly mapping single-cell identities to spatial positions. MERFISH produces a different data type than scRNA-seq, and a technical comparison between the two modalities is necessary to ascertain how to best integrate them. We performed MERFISH on the mouse liver and kidney and compared the resulting bulk and single-cell RNA statistics with those from the Tabula Muris Senis cell atlas and from two Visium datasets. MERFISH quantitatively reproduced the bulk RNA-seq and scRNA-seq results with improvements in overall dropout rates and sensitivity. Finally, we found that MERFISH independently resolved distinct cell types and spatial structure in both the liver and kidney. Computational integration with the Tabula Muris Senis atlas did not enhance these results. We conclude that MERFISH provides a quantitatively comparable method for single-cell gene expression and can identify cell types without the need for computational integration with scRNA-seq atlases.

A challenge in quantitative biology is to predict output patterns of gene expression from knowledge of input transcription factor patterns and from the arrangement of binding sites for these transcription factors on regulatory DNA. We tested whether widespread thermodynamic models could be used to infer parameters describing simple regulatory architectures that inform parameter-free predictions of more complex enhancers in the context of transcriptional repression by Runt in the early fruit fly embryo. By modulating the number and placement of Runt binding sites within an enhancer, and quantifying the resulting transcriptional activity using live imaging, we discovered that thermodynamic models call for higher-order cooperativity between multiple molecular players. This higher-order cooperativity captures the combinatorial complexity underlying eukaryotic transcriptional regulation and cannot be determined from simpler regulatory architectures, highlighting the challenges in reaching a predictive understanding of transcriptional regulation in eukaryotes and calling for approaches that quantitatively dissect their molecular nature.

Transcriptional control is fundamental to cellular function. However, despite knowing that transcription factors can repress or activate specific genes, how these functions are implemented at the molecular level has remained elusive. Here we combine optogenetics, single-cell live-imaging, and mathematical modeling to study how a zinc-finger repressor, Knirps, induces switch-like transitions into long-lived quiescent states. Using optogenetics, we demonstrate that repression is rapidly reversible ( ∼ 1 minute) and memoryless. Furthermore, we show that the repressor acts by decreasing the frequency of transcriptional bursts in a manner consistent with an equilibrium binding model. Our results provide a quantitative framework for dissecting the in vivo biochemistry of eukaryotic transcriptional regulation.