Weekly Paper Feed: March 18, 2026
5 papers this week · 3D spatial reconstruction, single-cell long-read sequencing, spatial benchmarking, mtDNA multiomics, and a next-gen full-length scRNA-seq protocol
UniST: Reconstructing Full 3D Tissue Architecture from Sparse Spatial Slices
📄 UniST: A Unified Computational Framework for 3D Spatial Transcriptomics Reconstruction
Liu et al. · bioRxiv · March 16, 2026
What they did: Most spatial transcriptomics datasets are 2D sections — you lose the full three-dimensional architecture of the tissue in the process. UniST is a generative AI framework that reconstructs dense, continuous 3D ST landscapes from sparse serial sections without modifying the underlying experimental protocol. It integrates three complementary modules: kernel point convolution with cross-attention for point cloud upsampling, optical flow-based interpolation for continuous slice reconstruction, and a graph autoencoder with implicit neural representations for gene expression imputation.
Why it matters: 3D tissue organization is not just a nice-to-have — spatial domains, tumor-immune boundaries, and tertiary lymphoid structures often span multiple z-planes and are fragmented or entirely missed in 2D analyses. UniST fills those gaps computationally.
Key result: In a mouse embryo dataset, UniST reconstructed a dense 3D heart architecture from sparsely sampled slices. In two human cancer tissues (including PDAC and colorectal cancer), UniST recovered tumor-immune boundaries and tertiary lymphoid structures that were fragmented in the original sparse data.
BenchDrop-seq: Long-Read Single-Cell RNA-seq Without Microfluidics
📄 BenchDrop-seq: a microfluidics-free platform for benchtop single-cell long-read RNA sequencing
bioRxiv · March 12, 2026
What they did: Single-cell long-read RNA-seq has been constrained by a fundamental access problem: it requires microfluidic barcoding hardware, specialized instruments, and high per-cell costs that put it out of reach for most labs. BenchDrop-seq solves this using particle-templated partitioning for barcoding — no microfluidics required — coupled to Oxford Nanopore sequencing with an open-source pipeline for barcode recovery, alignment, and transcript quantification.
Why it matters: Long-read scRNA-seq enables isoform-resolved measurements — things short-read 10x simply can't see. Democratizing access to this technology without requiring Drop-seq or 10x Chromium hardware changes who can do this science.
Key result: Validated in both a homogeneous cell line and heterogeneous primary tissue; high barcode recovery, accurate gene-level quantification, and reproducible detection of cell-type-specific transcript isoforms inaccessible to short-read assays.
ESPeR-seq: Eliminating Phantom UMIs and Finally Capturing True Transcript Ends
📄 ESPeR-seq: Extremely Sensitive and Pure, End-to-end, RNA-seq library preparation
bioRxiv · March 15, 2026
What they did: Smart-seq methods have long been the gold standard for full-length, high-sensitivity single-cell RNA-seq — but they suffer from three persistent problems: non-specific PCR products that reduce sensitivity, inability to precisely capture Transcription End Sites (TES), and "phantom UMIs" — artificial molecular barcodes generated during PCR that systematically inflate molecule counts. ESPeR-seq fixes all three. The key innovations are an "Omega-dT" primer that bypasses synthetic poly-T tracts to restore accurate TES capture, and a biochemical "multi-lock" mechanism using uracil-containing TSOs plus a uracil-intolerant DNA polymerase to eliminate PCR artifacts and phantom UMIs. A new metric — the logQ-slope — diagnoses UMI fidelity quantitatively.
Why it matters: Phantom UMIs are a quantification bug that every Smart-seq-style experiment inherits and almost no one is correcting for. ESPeR-seq both characterizes and eliminates this artifact, with implications for any full-length scRNA-seq data you've already generated.
Key result: Benchmarking shows that state-of-the-art methods (including existing Smart-seq variants) still exhibit UMI inflation detected by the logQ-slope metric; ESPeR-seq strictly prevents it. It also enables de novo reconstruction of gene models from aggregated single-cell populations, identifying novel multi-exon genes, unannotated 3' UTR extensions, and candidate eRNAs.
Single-Cell Multi-omics Tracks Mitochondrial DNA Mutations Across Cell States
📄 Single-cell multi-omic analysis of mitochondrial mutational mosaicism and dynamics
Hsieh, Kautz, Nitsch et al. · Nature Communications · March 2026
What they did: Somatic mitochondrial DNA (mtDNA) mutations occur at >200× the nuclear rate in human blood cells, but understanding how they propagate across cell types and states has been limited by bulk sequencing masking heteroplasmy variation. This study uses single-cell multi-omics to simultaneously profile mtDNA variant allele frequencies (VAFs), chromatin accessibility, and gene expression in thousands of individual cells — resolving how mtDNA mutations distribute across hematopoietic cell types, clonally expand, and associate with transcriptional and epigenetic states.
Why it matters: mtDNA mutations are increasingly implicated in aging, cancer, and rare mitochondriopathies. Single-cell resolution of mtDNA heteroplasmy finally allows genotype-to-phenotype mapping at the cell level — a problem that bulk sequencing structurally cannot solve.
Key result: The study maps cell-type-specific mtDNA variant dynamics in human hematopoiesis, demonstrating that mtDNA VAFs are far more heterogeneous across cell states than bulk data implies, with distinct clonal expansion patterns in different lineages.
Spatial Domain Detection Benchmark: It's the Preprocessing, Not the Architecture
📄 An explanatory benchmark of spatial domain detection reveals key drivers of method performance
Prusina, Schmidt et al. · bioRxiv · March 16, 2026
What they did: Dozens of computational methods exist for detecting spatial domains in spatial transcriptomics data, but comparative evaluations have led to contradictory conclusions because they test narrow subsets of datasets. This benchmark evaluates 26 spatial domain detection methods across 63 real tissue sections from six ST technologies, supplemented by >1,000 semi-synthetic datasets that systematically vary resolution, gene panel size, and tissue architecture. They also introduce a modular, plug-and-play benchmarking framework that enables ablation studies.
Why it matters: If you're picking a spatial domain detection method for your next project, this is required reading. The benchmark gives a principled, technology-aware basis for method selection — and the conclusion is surprising.
Key result: An ablation study of neural network-based methods shows that preprocessing and clustering choices often have a larger impact on performance than architectural differences between methods. Most spatial methods outperform non-spatial baselines, but performance depends strongly on data resolution and cellular heterogeneity — not just method choice.
📎 Quick Links
| # | Paper | Source | Link |
|---|---|---|---|
| 1 | UniST (3D spatial reconstruction) | bioRxiv | → |
| 2 | BenchDrop-seq (microfluidics-free long-read scRNA-seq) | bioRxiv | → |
| 3 | ESPeR-seq (phantom UMI elimination) | bioRxiv | → |
| 4 | Single-cell mtDNA multi-omics | Nature Communications | → |
| 5 | Spatial domain detection benchmark | bioRxiv | → |
Papers sourced from bioRxiv and Nature Communications (week of March 11–18, 2026).