Weekly Paper Feed: March 11, 2026
Your weekly digest of the most interesting preprints and papers in genomics, single-cell biology, spatial transcriptomics, and deep learning for bioinformatics.
🧬 This Week's Top 5
1. GPU-Accelerated kallisto: RNA-seq Quantification in Seconds
Paper: RNA-seq analysis in seconds using GPUs
Source: bioRxiv · March 4, 2026
DOI: 10.64898/2026.03.04.709526
If you've ever waited 40 minutes for kallisto to finish on a large dataset, this paper is for you. Researchers have reimplemented kallisto's core algorithms — pseudoalignment, equivalence class intersection, and the EM algorithm — for massively parallel execution on GPUs. The result: a 30–50× speedup over multithreaded CPU kallisto.
On 100 Geuvadis samples, the GPU version processes 3.6 million paired-end reads per second, completing a typical sample in seconds. A benchmark dataset of 295 million reads went from 40 minutes to just 50 seconds.
The key insight is that this wasn't a naïve GPU port — it required careful algorithmic redesign of each component to actually exploit GPU parallelism. As single-cell and spatial datasets keep growing, GPU-native quantification tools like this will go from "nice to have" to essential infrastructure.
Why it matters: RNA-seq quantification is still a computational bottleneck in large-scale studies. This makes the first step of nearly every transcriptomics pipeline dramatically faster.
2. FX-seq: Single-Nucleus RNA-seq Finally Works on FFPE Tissue
Paper: Fixative eXchange (FX)-seq: Scalable Single-nucleus RNA Sequencing Analysis of PFA-fixed or FFPE Tissue
Source: bioRxiv · March 5, 2026
DOI: 10.64898/2026.03.05.709668
Single-nucleus RNA-seq from formalin-fixed, paraffin-embedded (FFPE) clinical samples has been frustratingly limited due to poor reverse transcription yields from crosslinked RNA. FX-seq solves this with a clever two-part chemistry: an organocatalyst removes PFA crosslinks (boosting RT yield), while a regiospecific Pt(II) crosslinker prevents RNA leakage.
Applied to 321,710 nuclei across mouse brain and human cancer specimens — including gastrointestinal stromal tumor and colorectal cancer — FX-seq reveals cellular heterogeneity that was previously inaccessible. It even works on H&E-stained sections, enabling integrated analysis with pathologist annotations to distinguish tumor from non-tumor regions.
The clinical implications are significant: large FFPE biobanks accumulated over decades suddenly become single-cell-accessible.
Why it matters: FFPE is the standard for clinical archiving. Unlocking snRNA-seq from these samples brings precision medicine and translational research closer together.
3. Spatial Transcriptomics Maps the Gut Microbiome at Single-Cell Resolution
Paper: Spatial transcriptomics maps host–gut microbiome biogeography at high resolution
Source: Nature Microbiology · March 2026
Link: nature.com/articles/s41564-026-02286-7
The gut microbiome has long been described as an "organ," but we've lacked the tools to study its tissue-level geography. This Nature Microbiology paper changes that: the authors adapted spatially resolved RNA-seq to simultaneously capture both host gene expression and microbial RNA in intact intestinal tissue.
Previous approaches either captured A-rich microbial RNAs with poly(dT) primers or used custom capture arrays, but with limited multiplexity. This work demonstrates high-resolution mapping of host–microbiome interactions across intact gut tissue, revealing how specific microbial communities co-localize with distinct host cell populations and trigger localized transcriptional responses.
Why it matters: Understanding where microbes live and what they're doing to neighboring host cells is key to untangling microbiome–disease relationships. This spatial approach makes that possible at scale.
4. CellTypeAI: Local LLM-Powered Cell Annotation for scRNA-seq
Paper: CellTypeAI: Automated cell identification for scRNA-seq using local generative-AI
Source: bioRxiv · March 3, 2026
DOI: 10.64898/2026.03.03.709253
Cell annotation remains one of the most tedious and error-prone steps in single-cell analysis. CellTypeAI addresses this with a streamlined, scalable tool that uses locally-hosted generative AI models enhanced by retrieval-augmented generation (RAG) to provide tissue-context-aware cell type annotation.
Unlike cloud-based tools (ChatGPT, DeepSeek, Claude), CellTypeAI runs entirely locally — making it suitable for sensitive clinical datasets where data cannot leave the institution. The authors show improved annotation accuracy over conventional marker-based methods and nascent cloud-based approaches, with direct integration into standard scRNA-seq analysis pipelines.
The RAG enhancement is particularly interesting: it allows the model to pull from curated cell-type databases dynamically, improving robustness to variation in marker gene expression across patients, samples, and pathologies.
Why it matters: Reliable, scalable cell annotation is a prerequisite for meaningful single-cell studies — and doing it locally removes a major barrier for clinical and sensitive-data contexts.
5. What Do Biological Foundation Models Actually Compute?
Paper: What Do Biological Foundation Models Compute? Sparse Autoencoders from Feature Recovery to Mechanistic Interpretability
Source: bioRxiv · March 4, 2026
DOI: 10.64898/2026.03.04.709491
Biological foundation models (protein language models, genomic LMs, single-cell foundation models) are increasingly used for variant interpretation and drug design — but we largely don't know what they've learned internally. This paper tackles that with a systematic review and framework for applying sparse autoencoders (SAEs) to decompose model activations into interpretable, biologically meaningful features.
The findings are striking: across independent studies using different architectures and evaluation strategies, SAEs consistently recover features spanning biological scales — from secondary structure elements in proteins to transcription factor binding sites in genomes. But the authors identify a critical validation gap: most studies validate features against the same sequence databases used for model training, risking circularity.
They propose a three-level interpretability framework (representational → computational → causal mechanistic) and argue that the real test lies in experimental validation via deep mutational scanning and massively parallel reporter assays.
Why it matters: As foundation models become infrastructure for genomics, understanding what they actually compute — not just what they predict — is essential for trust, safety, and biological insight.
📎 Quick Links
| # | Paper | Source | Link |
|---|---|---|---|
| 1 | GPU kallisto (RNA-seq in seconds) | bioRxiv | → |
| 2 | FX-seq (FFPE snRNA-seq) | bioRxiv | → |
| 3 | Spatial transcriptomics + gut microbiome | Nature Microbiology | → |
| 4 | CellTypeAI (local LLM annotation) | bioRxiv | → |
| 5 | Biological Foundation Models + SAEs | bioRxiv | → |
Have a paper we should cover next week? Reach out via the contact page or drop it in the community Discord.