Weekly Paper Feed: February 11, 2026
TL;DR: This week's highlights span spatial transcriptomics breakthroughs, genomic language models, and deep learning tools for single-cell analysis β five papers that push the boundaries of how we see and interpret biological data.
1. ovrlpy: Untangling 3D Signal Overlaps in Spatial Transcriptomics
π Paper: Identifying 3D signal overlaps in spatial transcriptomics data with ovrlpy
π₯ Authors: Tiesmeyer S, MΓΌller-BΓΆtticher N, Malt A et al.
π
Published: 2026 | Nature Biotechnology
Key Findings
- Imaging-based spatial transcriptomics captures transcripts in 3D, but cell segmentation is typically performed in 2D β leading to "spatial doublets" where overlapping cells in the z-axis contaminate profiles
- ovrlpy is a new computational tool that detects overlapping cells, tissue folds, and segmentation errors by analyzing transcript localization in three dimensions
- The tool identifies spatial doublets that conventional 2D pipelines miss entirely
Why It Matters
As spatial transcriptomics data becomes richer in z-resolution, the assumption that each segmented "cell" contains transcripts from a single source breaks down. ovrlpy provides a critical quality control layer that will become standard in spatial analysis workflows.
Quick Take
A deceptively simple but essential tool β published in Nature Biotech for good reason. If you do any imaging-based spatial transcriptomics, this should be in your pipeline yesterday.
2. Seq-Scope-X: Spatial Omics Beyond Optical Resolution
π Paper: Seq-Scope-eXpanded: spatial omics beyond optical resolution
π₯ Authors: Anacleto A, Cheng W, Feng Q et al.
π
Published: 2026 | Nature Communications
Key Findings
- Seq-Scope-X combines submicrometer-resolution Seq-Scope with tissue expansion to surpass the 200-300 nm optical resolution limit
- In liver tissue, the method resolves nuclear vs. cytoplasmic compartments within nearly every single cell, revealing widespread differences in transcriptome patterns between compartments
- Individual hepatocytes can dynamically switch metabolic roles β independently confirmed by imaging-based methods
- Works in brain and colon tissue, and adapts for spatial proteomics with hundreds of antibody stains
Why It Matters
Breaking the optical resolution barrier for sequencing-based spatial transcriptomics opens entirely new questions about subcellular RNA localization and its functional significance. The nuclear/cytoplasmic transcriptome differences they found suggest post-transcriptional regulation is even more dynamic than previously appreciated.
Quick Take
This is the kind of platform paper that spawns dozens of follow-up studies. The expansion + sequencing combo is elegant, and the biological insights (subcellular compartmentalization at scale) are genuinely surprising.
3. Nucleotide GPT: What Do Genomic Language Models Actually Learn?
π Paper: Probing genomic language models: Nucleotide GPT and the role of pretraining
π₯ Authors: McLaughlin SM, Lim DA
π
Published: 2026 | Briefings in Bioinformatics
Key Findings
- Introduces Nucleotide GPT, a decoder-only transformer with single-nucleotide tokenization, to dissect what genomic language models learn during pretraining
- Repetitive elements (REs), which make up 30-60% of mammalian genomes, dominate the pretraining objective β models spend most effort learning repeat patterns
- Models with moderate RE downweighting (0.5) consistently achieved optimal performance across seven genomic classification tasks
- Sparse autoencoder (SAE) interpretability analysis reveals RE-associated features dominate internal representations
Why It Matters
This is a crucial reality check for the genomic language model field. If models are mostly learning to recognize transposable elements and other repeats, their utility for predicting functional genomic elements is limited. The paper provides concrete evidence for why pretraining strategies need rethinking.
Quick Take
Refreshingly honest work. Instead of claiming SOTA on benchmarks, they ask what the model actually learned β and the answer (mostly repeats) should give the field pause. The RE downweighting solution is simple but effective.
4. SAFAARI: Deep Learning for Single-Cell Integration Without Losing Biology
π Paper: SAFAARI: Contrastive Adversarial Open-set Domain Adaptation for Single-cell Integration & Annotation
π₯ Authors: Aminzadeh F, Wu J, He J et al.
π
Published: 2026 | Genomics, Proteomics & Bioinformatics
Key Findings
- SAFAARI is a unified deep learning framework for cell annotation, batch correction, and multi-omics integration
- Uses supervised contrastive learning + adversarial domain adaptation for domain-invariant embeddings
- Key innovation: identifies novel/unknown cell types and handles class imbalance for rare cell detection
- Outperformed existing methods in both qualitative and quantitative benchmarks across simulated and real-world datasets, including cross-omics integration
Why It Matters
Single-cell integration remains one of the biggest practical headaches in the field. Most tools either over-correct (removing real biology along with batch effects) or under-correct. SAFAARI's explicit handling of unknown cell types is particularly important β in real experiments, you should expect populations not in your reference.
Quick Take
Solid engineering with practical impact. The open-set domain adaptation approach is well-suited for exploratory single-cell studies where you genuinely don't know what cell types you'll find. Code is open source.
5. AGED: Reference-Free Cell Type Deconvolution for Spatial Transcriptomics
π Paper: Attention-guided Enhanced Deconvolution enables reference-free cell type estimation in spatial transcriptomics
π₯ Authors: Yang X, Wang Y, Chen X
π
Published: 2026 | Scientific Reports
Key Findings
- AGED is a two-stage framework combining probabilistic modeling with neural attention architectures for reference-free spatial transcriptomics deconvolution
- Automatically determines the optimal number of cell types β no manual specification needed
- Uses Performer-based networks with linear-complexity attention, plus a novel "Attention-Guide" refinement stage
- Achieved superior reconstruction (r=0.86) on Mouse Olfactory Bulb and mapped cell types in human pancreatic cancer and thymus tissue
Why It Matters
Most deconvolution methods require matched single-cell reference atlases β which often don't exist for rare tissues or novel disease contexts. AGED removes this dependency entirely while also solving the annoying problem of having to guess how many cell types to look for.
Quick Take
Reference-free deconvolution is a hard problem, and this is a meaningful step forward. The automatic cell type number selection is especially practical. Not yet at the accuracy of reference-based methods, but for tissues without good references, it's a game-changer.
Tags: #spatial-transcriptomics #deep-learning #genomics #single-cell #RNA-seq #weekly-paper-feed