Weekly Paper Feed: February 12–18, 2026
Five standout preprints from this week spanning single-cell genomics, spatial transcriptomics, deep learning, and epigenomics.
1. VIPerturb-seq: Genome-Wide Perturb-seq Goes Mainstream
📄 Paper: Genome-wide single-cell perturbation screens with VIPerturb-seq
👥 Authors: (New York Genome Center / NYU)
📅 Posted: February 12, 2026 | bioRxiv
Key Findings
- Introduces a split-probe detection strategy for fixed cells that enables genome-wide CRISPR library detection in Perturb-seq experiments
- Achieves a 50-fold throughput improvement through combinatorial indexing workflows
- Demonstrates both unbiased and phenotypically enriched genome-wide screens using a CRISPRi library (GuEST-List)
- Enables "Very Important Perturbation" (VIP) enrichment — prioritizing phenotypically interesting hits before single-cell profiling
Why It Matters
Perturb-seq has been transformative for mapping gene function, but scaling to genome-wide libraries has been prohibitively expensive for most labs. VIPerturb-seq shatters that barrier with probe-based detection and combinatorial indexing, making routine genome-wide functional screens accessible to individual labs — not just large consortia. The phenotypic enrichment option is particularly clever: screen broadly, then focus single-cell profiling resources on the perturbations that actually matter.
Quick Take
🔥 A genuine democratization of genome-wide Perturb-seq. The 50x throughput gain and optional enrichment step make this immediately practical for labs asking targeted biological questions at genome scale.
2. Do We Actually Need Foundation Models for Single-Cell?
📄 Paper: Parameter-free representations outperform single-cell foundation models on downstream benchmarks
👥 Authors: Boston University / Chan Zuckerberg Initiative
📅 Posted: February 11, 2026 | bioRxiv
Key Findings
- Simple, parameter-free linear pipelines (careful normalization + linear methods) match or exceed foundation model performance on standard scRNA-seq benchmarks
- Outperforms transformer-based models like TranscriptFormer on out-of-distribution tasks — novel cell types and organisms not in training data
- Argues that cell identity biology can be captured by simple linear representations of gene expression
Why It Matters
This is the kind of paper the field needs. As billion-parameter foundation models become the default approach, it's critical to ask: are we over-engineering the solution? The authors show that for the most common downstream tasks (cell-type classification, disease-state prediction, cross-species learning), the signal in scRNA-seq data is strong enough that interpretable linear methods do just fine — and generalize better. A wake-up call for rigorous benchmarking practices.
Quick Take
🎯 A healthy reality check. Foundation models are exciting, but if PCA + normalization beats them on out-of-distribution tasks, the community should think carefully about what these models are actually learning.
3. The Most Comprehensive Basal Ganglia Epigenomic Atlas Yet
📄 Paper: A Multimodal Single-Cell Epigenomic and 3D Genome Atlas of the Human Basal Ganglia
👥 Authors: Ecker Lab (Salk Institute) + BRAIN Initiative consortium
📅 Posted: February 12, 2026 | bioRxiv
Key Findings
- First multimodal single-cell epigenomic atlas of the human basal ganglia — jointly profiling DNA methylation and 3D chromatin conformation via snm3C-seq
- 261,331 cells across 8 basal ganglia subregions, 31 subclasses, 59 cell groups
- Reveals cell-type- and region-specific differential methylation validated by MERFISH spatial transcriptomics
- Non-neuronal cells show distinct 3D genome organization: smaller chromatin compartments, more long-range contacts, shorter loops
- Identified medium spiny neuron subtypes with distinct epigenomic signatures; 3D genome embeddings highlight dorsal, ventral, and hybrid populations
Why It Matters
The basal ganglia are central to motor control and reward processing, and their dysfunction underlies Parkinson's, Huntington's, and addiction. This atlas provides the epigenomic reference needed to interpret GWAS risk loci and understand gene regulation in these critical brain structures. The integration of methylation, 3D chromatin, accessibility, and spatial transcriptomics in one cohesive atlas sets a new standard for brain region characterization.
Quick Take
🧠 A monumental resource. The depth of multi-omic integration here — methylation, 3D genome, MERFISH validation — gives this atlas staying power as a community reference for years to come.
4. SpaDecoder: 3D-Aware Spatial Transcriptomics Deconvolution
📄 Paper: Spatiotemporal cell type deconvolution leveraging tissue structure
👥 Authors: Zhang Lab (Georgia Tech)
📅 Posted: February 10, 2026 | bioRxiv
Key Findings
- Presents SpaDecoder, a parallelized matrix factorization method for per-spot deconvolution across multiple 3D spatial or temporal tissue slices
- Uses an adaptively inferred 3D neighborhood Gaussian kernel — going beyond naive 2D proximity assumptions that fail at tissue boundaries
- Handles batch effects and variability in single-cell reference profiles
- Enables downstream analyses: anteroposterior variability, gene expression imputation, key tissue region identification, colocalized cell types, and spatio-temporal cell location prediction
Why It Matters
Most spatial transcriptomics deconvolution tools treat each tissue section in isolation or assume nearby spots are always similar. SpaDecoder's 3D-aware approach is more biologically realistic — tissue is 3-dimensional, and leveraging that structure improves accuracy. The adaptive kernel elegantly handles the boundary problem where neighboring spots can belong to completely different tissue compartments.
Quick Take
⚡ Spatial deconvolution that actually thinks in 3D. Open-source and ready to use: github.com/ZhangLabGT/spadecoder
5. Tissue-Scale Cell Typing from Nuclear Architecture Alone
📄 Paper: Machine learning-guided spatial omics for tissue-scale discovery of cell-type-specific architectures
👥 Authors: (Howard Hughes Medical Institute)
📅 Posted: February 12, 2026 | bioRxiv
Key Findings
- Integrates highly multiplexed protein and RNA imaging (up to 46 proteins + 79 RNAs) with explainable ML to map cell-type-specific molecular and structural architectures
- Profiled ~370,000 cells in intact mouse brain tissue at ~260 nm subcellular resolution
- Nuclear protein and chromatin features alone are sufficient to accurately classify brain cell types and their spatial organization
- Explainable deep learning identifies human-interpretable subnuclear structural features that distinguish cell types
Why It Matters
The idea that you can classify cell types just from nuclear architecture — without relying on transcript abundance — opens up exciting possibilities for high-content screening and pathology. The explainability angle is equally important: rather than treating the classifier as a black box, the model reveals what structural features matter, which could point to new biology about how nuclear organization differs across cell types.
Quick Take
🔬 When your nuclei are as distinctive as fingerprints. The combination of spatial multi-omics at scale + explainable ML makes this both a powerful atlas and a method development milestone.
Wrapping Up
This week's preprints highlight two converging themes: scalability (VIPerturb-seq's 50x throughput, SpaDecoder's 3D parallelization, tissue-scale imaging of 370K cells) and interpretability (parameter-free methods beating black-box models, explainable ML revealing nuclear features). The field is maturing — not just doing more, but doing it more thoughtfully.
The basal ganglia atlas stands out as a landmark resource that will be referenced for years. And the foundation model reality check? Essential reading for anyone training large models on scRNA-seq data.
See you next week! 📚
Tags: #single-cell #spatial-transcriptomics #deep-learning #genomics #perturb-seq #epigenomics #foundation-models