Weekly Paper Feed: March 25, 2026
5 papers this week · RNA foundation models, reference-free spatial annotation, subcellular transcriptomics, isoform-resolved spatial sequencing, and a landmark PDAC single-cell atlas
EVA: A Long-Context Generative Foundation Model for RNA Design
📄 A Long-Context Generative Foundation Model Deciphers RNA Design Principles
Lv, Cheng, Xie et al. · bioRxiv · March 17, 2026
What they did: RNA engineering is still largely empirical — designing sequences with desired functions requires extensive wet-lab iteration. EVA (Evolutionary Versatile Architect) changes that. It's a generative RNA foundation model trained on over 114 million full-length RNA sequences across evolutionary diversity, using a Mixture-of-Experts architecture with an 8,192-token context window. EVA captures both global transcript structure and fine-grained functional elements in a single unified representation.
Why it matters: Most RNA design tools are specialized — one model for aptamers, another for guide RNAs, another for mRNA optimization. EVA unifies all of this in a single architecture with controllable generation, enabling tasks ranging from mutational fitness prediction to de novo design of tRNAs, CRISPR gRNAs, mRNA therapeutics, and circular RNAs.
Key result: EVA achieves state-of-the-art performance on 7 of 9 established RNA benchmarks, improving structural modeling accuracy by up to an order of magnitude compared to existing approaches. It is fully open-source.
Binary-SPA: Reference-Free Cell Annotation for High-Resolution Spatial Transcriptomics
📄 Binary-SPA: A Reference-Free Method for Cell Annotation in High-Resolution Spatial Transcriptomics
Cai, Wang, Ren et al. · bioRxiv · March 17, 2026
What they did: Annotating cell types in high-resolution spatial transcriptomics (e.g. Visium HD, Xenium, CosMx) typically requires either a matched scRNA-seq reference or marker-based clustering — both have significant limitations. Binary-SPA is a two-stage computational framework that sidesteps external references entirely. First, it uses binary classification on predefined marker sets to identify high-confidence cells. Those become the internal reference for anchor-based label transfer in stage two, achieving 100% annotation coverage without any external dataset.
Why it matters: scRNA-seq reference panels often don't exist for clinical tissues, rare cell types, or novel experimental systems. Reference-free annotation removes this bottleneck entirely, making high-res ST analysis feasible in any context.
Key result: Binary-SPA matches or outperforms label transfer methods (which require same-tissue scRNA-seq references) while eliminating the need for external data. It performs robustly on challenging specimens including bone marrow biopsies, and was validated against matched COMET protein expression data with strong concordance.
scAPEX-seq: Mapping Subcellular Transcriptomes at Single-Cell Resolution
📄 Subcellular transcriptome sequencing with single cell APEX-seq identifies regulators of cell-cell interactions
bioRxiv · March 17, 2026
What they did: Standard scRNA-seq captures what genes are expressed — but not where the RNA is in the cell. RNA localization governs splicing, translation, and signaling, yet it's invisible to conventional droplet-based methods. scAPEX-seq combines proximity labeling (APEX-based) with droplet scRNA-seq to map the endoplasmic reticulum–associated transcriptome in thousands of individual cells. Applied to tumor-macrophage and CAR T cell co-cultures, it selectively enriches cell-surface and secretory transcripts that conventional scRNA-seq misses.
Why it matters: Cell-cell interaction biology is dominated by secreted and surface proteins — exactly the transcripts that are ER-enriched and underrepresented in standard single-cell data. scAPEX-seq is a direct solution, and it's scalable to existing droplet platforms.
Key result: In HER2+ tumor / CAR T cell co-cultures, scAPEX-seq resolved distinct activated CAR T cell states including a CTSW-high population. Overexpression of CTSW (a cathepsin protease) promotes stem-like CAR T phenotypes, long-term proliferation, and sustained tumor killing — a finding missed entirely by conventional scRNA-seq.
Isoform-Resolved Spatial Transcriptomics via a Low-Cost NGS+TGS Dual Workflow
📄 Isoform-resolved spatial transcriptomics on a lab-made high-density array via a single-chip NGS-TGS workflow
Liu, Liu, Lu et al. · bioRxiv · March 18, 2026
What they did: Commercial spatial transcriptomics platforms (Visium, Xenium) use short-read sequencing and cannot resolve RNA isoforms. Long-read spatial methods exist but remain expensive and inaccessible. This paper introduces a lab-made, high-density bead-in-microwell chip assembled with a standard centrifuge and decoded with a tri-part combinatorial long-barcode strategy (5.6×10⁷ barcode combinations across 4.1×10⁴ capture sites). A single chip splits full-length cDNA for both NGS and TGS sequencing at matched spatial coordinates — isoform resolution at short-read cost.
Why it matters: Transcript isoforms encode critical regulatory and proteomic diversity invisible to gene-level quantification. This platform democratizes isoform-resolved spatial transcriptomics — making it available to any lab without commercial instrument investment.
Key result: In a tomato-pepper graft interface, the platform identified thousands of unannotated isoforms and spatially restricted splicing events including novel CaGRP1 isoforms and SlPIP2 intron retention. In mouse embryos, it resolved unannotated Col1a2 isoforms in oligodendrocyte progenitors and Dalrd3 intron retention patterns.
A 1.6-Million-Cell Cross-Species Atlas of Pancreatic Cancer Progression
📄 Cross-species single-cell atlases chart progression, therapy-driven remodelling and immune evasion in pancreatic cancer
Parikh, Jimenez, Schneeweis et al. · bioRxiv · March 19, 2026
What they did: Pancreatic ductal adenocarcinoma (PDAC) is almost always diagnosed late, yet single-cell datasets capturing advanced, metastatic, and treated disease have been sparse. This study integrates 16 human PDAC studies (257 donors) with representative mouse models (101 tumors) using scANVI harmonization, compiling over 1.6 million cells into a four-level hierarchical taxonomy of 60+ distinct cell states. It spans the full disease arc: early → advanced → metastatic → post-treatment.
Why it matters: A cross-species atlas at this scale enables systematic identification of conserved vs. human-specific biology, therapy-induced immune remodeling, and rare cell states that are statistically invisible in individual studies. PDAC remains one of the deadliest cancers precisely because its biology at late stages is poorly understood.
Key result: The atlas resolves ten malignant transcriptional programs linked to disease progression, including an EMT-persistent state enriched after radiotherapy exposure. It also uncovers rare CD4⁺CD8⁺ double-positive T cells and documents how post-treatment disease reshapes the tumor microenvironment — a resource for both biomarker discovery and therapeutic target identification.
📎 Quick Links
| # | Paper | Source | Link |
|---|---|---|---|
| 1 | EVA (RNA foundation model) | bioRxiv | → |
| 2 | Binary-SPA (reference-free spatial annotation) | bioRxiv | → |
| 3 | scAPEX-seq (subcellular transcriptomics) | bioRxiv | → |
| 4 | Isoform-resolved spatial ST (NGS+TGS chip) | bioRxiv | → |
| 5 | Cross-species PDAC single-cell atlas | bioRxiv | → |
Papers sourced from bioRxiv (week of March 18–25, 2026).