Weekly reads 19/1/26
Multimodal integration, interpretability, and robustness in spatial biology
This week's papers advance single-cell and spatial biology into a more integrated, interpretable, and sample-aware future. At the experimental frontier, DBiTplus closes a long-standing multimodal gap by combining whole-transcriptome sequencing and high-plex protein imaging on the same tissue section, allowing for true gene-protein co-mapping at single-cell resolution, even in clinical FFPE samples. On the regulatory side, TF-MINDl addresses the black box problem of deep sequence models by extracting reproducible, interpretable enhancer grammar that explains, and even engineers, cell-type-specific gene regulation in human neural development. Several methods then reconsider how we analyse complex datasets at scale: scSLIDE elevates samples (rather than cells) as the unit of inference to reconstruct continuous disease and developmental trajectories; BatchSVG uses a statistically grounded filter to remove batch-confounded spatially variable genes before they mislead downstream analyses, whereas SpaHDmap combines histology and spatial transcriptomics to produce high-resolution, interpretable embeddings that recover fine tissue architecture across technologies. Together, these studies reflect a broader shift in the field: from isolated modalities and black-box predictions to integrated measurements, interpretable models, and analysis frameworks that take into account spatial context, batch structure, and biological continuity.
Preprints/articles that I managed to read this week
Integration of imaging-based and sequencing-based spatial omics mapping on the same tissue section via DBiTplus
Enninful et al. Nat Methods (2026). https://doi.org/10.1038/s41592-025-02948-0
The paper in one sentence
DBiTplus is a multimodal spatial omics platform that combines whole-transcriptome sequencing and multiplexed protein imaging on the same tissue section, enabling spatially resolved, single-cell atlas construction and mechanistic insight into tissue biology and disease.
Summary
DBiTplus (Deterministic Barcoding in Tissue sequencing plus) introduces a unified workflow for spatially mapping both the transcriptome and proteome from the same tissue section. By leveraging spatial barcoding and an optimized RNase H–based cDNA retrieval method, it preserves tissue architecture for high-plex immunofluorescence imaging (e.g., CODEX, CellScape) while allowing transcriptome-wide sequencing. A computational pipeline aligns imaging and sequencing data, guiding spot deconvolution into pure cell-type sub-spots to generate single-cell–resolution spatial transcriptomes. Validated on mouse embryos and human lymphoma specimens, DBiTplus reveals spatial gene–protein concordance, maps lymphoma progression and transformation, and profiles non-coding RNAs—all within the same tissue context.
Personal highlights
Dual-modality mapping on the same tissue section: DBiTplus simultaneously captures whole-transcriptome spatial sequencing and multiplexed protein imaging from a single tissue slice, eliminating alignment errors from adjacent sections and ensuring true multimodal integration.
Enzymatic cDNA retrieval that preserves tissue integrity: the use of thermostable RNase H, optimized with Triton X-100 enables efficient cDNA release while maintaining tissue morphology and antigenicity for subsequent high-quality immunofluorescence imaging.
Imaging-guided spot deconvolution to single-cell resolution: by co-registering protein imaging with transcriptomic spots, DBiTplus splits each spot into pure cell-type sub-spots, enabling spatially resolved single-cell transcriptomics beyond the limits of traditional spot-based methods.
Compatibility with challenging clinical FFPE samples: DBiTplus works robustly on formalin-fixed paraffin-embedded tissues making it directly applicable to archival clinical specimens for retrospective studies.
Discovery of spatially resolved non-coding RNA biology: in addition to mRNAs and proteins, DBiTplus captures small non-coding RNAs (e.g., miRNAs) in situ, enabling investigation of their spatial roles in disease mechanisms such as Richter’s transformation.
Why should we care?
DBiTplus bridges a critical gap in spatial biology by allowing researchers to measure both the full transcriptome and dozens of proteins in precisely the same cells and tissue architecture. This matters because many biological and disease processes like cancer progression, immune response, and tissue development are driven by complex interactions between genes and proteins that cannot be fully understood from one modality alone.
Decoding the Enhancer Grammar of Human Neural Development with TF-MINDl
De Winter et al. bioRxiv (2026). doi:10.64898/2026.01.14.699402
The paper in one sentence
Researchers developed TF-MINDl, a computational framework that extracts interpretable cis-regulatory rules, such as transcription factor binding site combinations, from black-box deep learning models, revealing the enhancer code underlying early human neural development and enabling the design of synthetic, cell-type-specific enhancers.
Summary
This study addresses a major challenge in genomics: while sequence-to-function (S2F) deep learning models can accurately predict enhancer activity, they operate as “black boxes,” making it difficult to extract human-interpretable rules. To bridge this gap, the authors introduced TF-MINDl, a tool that uses contribution scores from S2F models to identify, cluster, and annotate transcription factor binding site (TFBS) instances across genomic regions. Using two independent single-cell multioome atlases—one from human neural tube organoids and another from a human embryo—the team trained S2F models (DeepNeuralTube) and applied TF-MINDl to uncover consistent enhancer codes for dorsal-ventral progenitors, neural crest, mesenchyme, and neurons. They validated TF-MINDl predictions with ChIP-seq data, cross-species comparisons, and topic modeling, and demonstrated its utility by designing synthetic enhancers for facial mesenchyme cells that functioned as predicted in luciferase reporter assays. The work shows that TF-MINDl can reproducibly extract biologically meaningful cis-regulatory rules, offering a step toward formalizing the genomic regulatory code and enabling the principled design of synthetic regulatory elements.
Personal highlights
From black-box to rule-based enhancer decoding: TF-MINDl transforms S2F model contribution scores into annotated TF binding site instances, enabling the systematic extraction of cis-regulatory grammar, such as motif combinations, affinities, and co-occurrence patterns, across cell types.
Cross-system validation using organoids and human embryos: by applying TF-MINDl to both neural tube organoids and a 4-week human embryo, the study demonstrates robust, concordant enhancer codes, validating organoids as a faithful model for early human neural development.
Experimental validation via synthetic enhancer design: extracted rules were used to engineer synthetic enhancers in facial mesenchyme; designs with two high-affinity “coordinator” motifs (or one plus accessory sites) sufficed to drive cell-type-specific activity, confirming the sufficiency of the decoded logic.
Integration of topic modeling for cross-species prediction: TF-MINDl’s topic modeling of TFBS co-occurrences achieved predictive accuracy comparable to deep learning models in cross-species (human–zebrafish) comparisons, highlighting the conservation and generalizability of the extracted rules.
Scalable, annotation-ready framework for regulatory genomics: built on AnnData and scverse ecosystems, TF-MINDl provides an accessible pipeline for genome-wide TFBS instance annotation, clustering, and rule extraction, supporting future efforts in non-coding variant interpretation and mechanistic modeling.
Why should we care?
TF-MINDl moves beyond simply predicting where enhancers are active to explaining why, unpacking the combinatorial TF binding logic that drives cell-type-specific gene expression. For developmental biologists, it offers a validated rulebook for human neural enhancers, bridging organoid and in vivo models. For disease researchers, it provides a framework to interpret non-coding mutations in neurodevelopmental disorders. For synthetic biologists, it enables the principled design of cell-type-specific enhancers. And for the AI-in-science community, it represents a meaningful advance in interpretable deep learning, turning black-box predictors into transparent, mechanistic models of genomic regulation.
scSLIDE: Reconstructing Developmental and Disease Progression with Sample‑Level Embeddings
Jiang et al. bioRxiv (2025). https://doi.org/10.64898/2025.12.10.693462
The paper in one sentence
scSLIDE is a computational framework that transforms single‑cell data into sample‑level density profiles, enabling the discovery of continuous trajectories in disease and development beyond traditional case‑control comparisons.
Summary
scSLIDE (single‑cell Sample‑Level Integration using Density Estimation) shifts the focus of single‑cell analysis from individual cells to whole samples. It constructs a semi‑supervised cell embedding that captures both cell‑type and cell‑state variation, then represents each sample as a density profile across cellular landmarks. This sample‑level representation allows for clustering, trajectory inference, and identification of molecular programs associated with sample heterogeneity. Applied to COVID‑19, Alzheimer’s disease, and zebrafish embryogenesis, scSLIDE uncovers continuous axes of variation, such as infection severity, temporal progression, and developmental pseudostages, that are obscured by conventional binary case‑control analyses.
Personal highlights
Sample‑centric density profiling over cellular landmarks: scSLIDE represents each sample as a distribution of its cells across a landmark‑defined state space, transforming single‑cell measurements into compact, comparable sample‑level profiles that capture compositional and state shifts.
Semi‑supervised embedding balancing cell‑type resolution and phenotype‑associated variation: by integrating unsupervised (cell‑type‑oriented) and supervised (PLS‑based) embeddings via weighted nearest neighbors, scSLIDE retains high‑resolution cell‑type information while prioritizing subtle phenotype‑driven signals.
Continuous trajectory inference disentangles multiple axes of sample heterogeneity: in COVID‑19 data, scSLIDE separates infection status, time‑since‑onset, and disease severity into distinct diffusion components; in Alzheimer’s disease, it reconstructs a continuous severity trajectory that aligns with independent neuropathology scores.
Enhanced reproducibility and power for trajectory‑based differential expression: Modeling sample variation along continuous axes, rather than binary labels, yields more reproducible gene‑expression signatures across independent cohorts and identifies gradual, biologically meaningful expression changes.
Why should we care?
scSLIDE reframes how we analyze single‑cell data by placing the sample, not the cell, at the center of interpretation. This shift is crucial for translational research, where questions about patient stratification, disease progression, and therapeutic response depend on sample‑level variation. By moving beyond binary case‑control comparisons, scSLIDE reveals continuous biological trajectories that better reflect the graded, heterogeneous nature of development and disease.
BatchSVG: Identifying Batch‑Biased Genes in Spatially Variable Gene Detection
Shah et al. bioRxiv (2025). https://doi.org/10.64898/2025.12.09.693192
The paper in one sentence
BatchSVG is a statistical tool that identifies and removes spatially variable genes (SVGs) confounded by batch effects, thereby improving the biological consistency of downstream spatial analyses in multi‑sample datasets.
Summary
BatchSVG addresses a key challenge in spatial transcriptomics: distinguishing true biological spatial variation from technical batch artifacts in multi‑sample experiments. The method fits a binomial deviance model per gene, once with and once without a batch covariate (e.g., slide or sample ID), and quantifies the change in residual deviance and rank for each gene. Genes whose deviance drops substantially or whose rank decreases when batch is accounted for are flagged as batch‑biased. A data‑driven threshold, based on standard deviations of the relative change in deviance (RCD) and rank deviance (RD), is used to select genes for removal. Applied to spatial datasets, BatchSVG refines SVG lists, leading to more biologically coherent spatial domains in downstream clustering (e.g., with PRECAST) without compromising quality‑control metrics.
Personal highlights
Deviance‑ and rank‑based detection of batch‑biased SVGs: BatchSVG fits binomial models with/without batch covariates and quantifies changes in residual deviance and gene rank, pinpointing SVGs whose apparent spatial signal is largely explained by technical variation.
Data‑driven thresholding adapts to dataset‑specific variability: instead of fixed cut‑offs, BatchSVG uses the number of standard deviations of the relative change in deviance (RCD) and rank deviance (RD) to establish adaptive, dataset‑specific thresholds for batch‑bias identification.
Improved biological interpretability of spatial domains: Removing batch‑biased SVGs before clustering yields spatial domains that align better with known anatomy and marker genes, as quantified by increased normalized mutual information (NMI) with manual annotations.
Methodology generalizable beyond spatial data: Although designed for SRT, BatchSVG’s core approach, comparing model fits with/without a batch covariate, can be applied to other omics data (e.g., single‑nucleus RNA‑seq) where batch‑confounded features need identification.
Why should we care?
BatchSVG tackles a subtle but critical problem in the age of large‑scale spatial atlases: batch‑confounded spatial signals can masquerade as biology and mislead downstream interpretation. As spatial studies scale across dozens of slides, donors, or sequencing runs, traditional SVG detection methods, which treat each section independently, fail to disentangle technical artifacts from true spatial patterning. BatchSVG provides a simple, statistically grounded filter to clean SVG lists before clustering or domain detection, ensuring that the features driving analysis reflect tissue architecture rather than technical noise.
The interpretable multimodal dimension reduction framework SpaHDmap enhances resolution in spatial transcriptomics
Tang et al. Nature Cell Biology (2026). https://doi.org/10.1038/s41556-025-01838-z
The paper in one sentence
SpaHDmap is a deep learning–based framework that integrates histology images with spatial transcriptomics data to produce interpretable, high-resolution embeddings, enabling fine-grained spatial analysis across multiple samples and tissue types.
Summary
Spatial transcriptomics (ST) technologies map gene expression across tissue sections, but data are often noisy, sparse, and limited in resolution. SpaHDmap addresses these challenges by combining non-negative matrix factorization (NMF) with a multimodal encoder–decoder architecture that fuses gene expression with high-resolution histology images (e.g., H&E, IHC). This integration allows SpaHDmap to generate pixel-level embeddings that are both interpretable and biologically meaningful. The framework supports multi-sample analysis, works with various image types, and outperforms existing methods in recovering fine spatial structures, as demonstrated in brain, cancer, and developmental datasets.
Personal highlights
Multimodal fusion of histology and transcriptomics: SpaHDmap integrates high-resolution histology images with ST data through a unified encoder–decoder framework, leveraging visual context to enhance spatial resolution and denoise expression signals.
Interpretable, high-resolution embeddings via NMF: by embedding non-negative matrix factorization into a deep learning pipeline, SpaHDmap produces spatially coherent, parts-based representations that link embedding dimensions to known biological regions and gene programs.
Flexible support for multiple samples and image types: the framework can jointly analyze datasets from different conditions, remove batch effects, and handle diverse histology modalities, from H&E to IHC and even pseudo-images generated from subcellular ST data.
Fine-grained spatial domain discovery: SpaHDmap identifies sub-spot and pixel-level tissue architectures, such as hippocampal sublayers, meningeal layers, and tumor–stroma interfaces, that are missed by spot-level methods.
Robust and scalable across technologies: evaluated on 30+ datasets from platforms like 10x Visium, Xenium, and Stereo-seq, SpaHDmap consistently outperforms state-of-the-art methods in embedding accuracy, clustering resolution, and biological interpretability.
Other papers that peeked my interest and were added to the purgatory of my “to read” pile
Thanks for reading.
Cheers,
Seb.