Paper Outlook — August 17-23, 2026

· 5 min read · Paper Feed

Paper Outlook — August 17-23, 2026

TL;DR

  • This week's picks split into two camps: new biology about the cellular neighborhood of tumors (neutrophils and fibroblasts), and new computational tools that make single-cell and spatial gene-expression data easier to work with.
  • Two papers tackle spatial transcriptomics, the technology that maps gene activity across a tissue. One introduces a method for finding genes whose expression smoothly tracks with features visible under the microscope; the other compares 18 cell-type-mixing tools across 50 datasets and offers practical guidance.
  • Deep learning keeps inching toward interpretability: a Transformer-based cell-typing tool (TOSICA) uses biological concepts like pathways as its vocabulary, and a multi-task network (UnitedNet) explains how different molecular layers inside the same cell relate to each other.
  • For cancer researchers, the takeaway is that the cells surrounding a tumor — even short-lived immune cells like neutrophils — can settle into stable, disease-driving states with clear links to how patients fare.

Deterministic reprogramming of neutrophils within tumors

This study shows that neutrophils, short-lived immune cells, enter tumors and undergo permanent rewiring of their genes and proteins, settling into a single terminal state that promotes blood vessel growth. These reprogrammed cells cluster in low-oxygen, sugar-hungry regions at the tumor core. The work recasts neutrophils from bystanders into active contributors to cancer progression, and a possible handle for therapy.

Conceptual illustration for: Deterministic reprogramming of neutrophils within tumors

Science (2024) · 335 citations

Read the original → https://doi.org/10.1126/science.adf6493

Cancer-associated fibroblast phenotypes are associated with patient outcome in non-small cell lung cancer

Using a high-resolution tissue-imaging approach on more than 1,000 lung cancer patients, the authors catalogue 11 subtypes of fibroblast support cells and group them into 4 patterns that predict how long patients survive. One pattern, rich in 'tumor-like' fibroblasts, was strongly tied to poor prognosis. The work positions these support cells as independent risk factors and a plausible angle for personalized treatment.

Overview of experimental and analytical set-up (A) Experimental set-up. NSCLC samples (2070 cores, 1,070 patients) and controls were stained with metal-tagged antibodies and imaged using IMC. Cell and tumor-stroma masks were generated. (B) Analytical setup. Single-cell data were used to identify cell phenotypes, and to evaluate correlations between cell types and clinical data. CAF based patient stratifications were used for survival and TMA composition analyses. Finally, spatial properties of CAFs and immune cell types were analyzed and integrated with clinical data. (C) Markers in the antibody panel.

Figure 1 from L. Cords et al., Cancer Cell (2024). Licensed CC BY-NC-ND. Source: 10.1016/j.ccell.2023.12.021.

Cancer Cell (2024) · 264 citations

Read the original → https://doi.org/10.1016/j.ccell.2023.12.021

Inferring histology-associated gene expression gradients in spatial transcriptomic studies

The authors present Spatial Gradient Screening, a tool that scans spatial gene expression maps to find genes whose activity smoothly tracks with features visible in tissue histology, without forcing the data into pre-defined groups. They demonstrate it on injured mouse brain and brain tumor samples. The method helps researchers pick up subtle, location-driven gene patterns that traditional group-based approaches tend to miss.

Histologically defined group-based analysis. a – c SPATA’s manual annotation tool is employed to delineate borders, facilitating the grouping of spots based on histology. d Gene expression-driven UMAP projection of spots. e A dot plot showcases the 18 most statistically significant unique marker genes, ranked by their average log2-fold change, in accordance with histological areas. f and g Surface plots are color-coded to highlight inferred copy number alterations that are characteristic of glioblastoma. h and i Statistical analyses examine copy number alterations across histological areas (two-sided test, no adjustments for multiple comparisons, Tumor ( n = 1307), Transition ( n = 478), Infiltrated Cortex ( n = 1428). The minima represent the smallest and the maxima represent the largest value within 1.5 times the interquartile range (IQR) below or above the first or third quartile (Q1, Q3), respectively. The median is shown as a line inside the box. The box bounds are the first quartile (Q1, 25th percentile) and the third quartile (Q3, 75th percentile). The whiskers extend from Q1 and Q3 to the minima and maxima. j A heatmap provides a comprehensive view of alterations across all chromosomes in relation to histological areas, corroborating the tumor area’s classification with multiple alterations, the nearly unaltered cortex infiltrated by the tumor, and the transition zone exhibiting intermediate levels of alterations. k A volcano plot from the DEA analysis across histological areas highlights marker genes for the Tumor and Transition areas, characterized by an adjusted p -value of 0 (infinite −log10).

Fig. 1 from J. Kueckelhaus et al., Nature Communications (2024). Licensed CC BY. Source: 10.1038/s41467-024-50904-x.

Nature Communications (2024) · 74 citations

Read the original → https://doi.org/10.1038/s41467-024-50904-x

A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics

This benchmark study systematically compares 18 computational methods for estimating which cell types are mixed inside each spot of a spatial transcriptomics experiment, across 50 real and simulated datasets. It evaluates accuracy, robustness, and ease of use, and produces practical guidelines for choosing a tool. A handy reference for anyone planning a spatial transcriptomics project.

The summarization of benchmarking pipeline. A Eighteen cellular deconvolution methods, classified based on their data requirements and computational techniques, were evaluated with 50 simulated and real-world spatial transcriptomics datasets. B Datasets from six spatial transcriptomics technologies were used for benchmarking, and the scatter plot shows the resolution of each spot and number of spots and genes in each technology. C The benchmarking results were measured according to the accuracy, robustness, and usability of the methods.

Fig. 1 from Haoyang Li et al., Nature Communications (2023). Licensed CC BY. Source: 10.1038/s41467-023-37168-7.

Nature Communications (2023) · 235 citations

Read the original → https://doi.org/10.1038/s41467-023-37168-7

Transformer for one stop interpretable cell type annotation

TOSICA is a deep learning model based on the Transformer architecture that automatically transfers cell-type labels from a labeled reference dataset to a new single-cell dataset, using biologically meaningful inputs such as gene pathways instead of raw gene lists. It is fast, accurate, and its attention weights reveal which pathways drive each label. That makes large-scale cell annotation both reproducible and more interpretable than many existing tools.

Algorithmic framework of TOSICA. a The model is trained on single-cell RNA sequencing data and cell type label for each cell. Based on databases or expert knowledge, masked learnable embeddings are used to convert the reference input data (n genes) to k input tokens representing each gene set (GS), to which class token (CLS) is added. In the attention function, query (Q), key (K), and value (V) matrix are linearly projected from these GSs and CLS combined tokens and the weights (attention, A) is computed by a compatibility function of the Q with the corresponding K, then assigned to each V for computing output (O). In each Multi-head Self-attention layer, the attention function is performed H times in parallel. The CLS of O, considered as latent space of each cell, is used as input of the whole conjunction neural network cell type classifier. Meanwhile, the attention of class (CLS) token to gene set (GS) tokens is referred as attention score and used for cell embedding. b hArtery and hBone datasets use healthy samples as training data and predict disease samples. hPancreas and mBrain datasets are split by data source. Training and test data in mPancreas and mAtlas come from different timepoints.

Fig. 1 from Jiawei Chen et al., Nature Communications (2023). Licensed CC BY. Source: 10.1038/s41467-023-35923-4.

Nature Communications (2023) · 202 citations

Read the original → https://doi.org/10.1038/s41467-023-35923-4

Explainable multi-task learning for multi-modality biological data analysis

UnitedNet is a neural network framework that jointly performs multiple analytical tasks on single cells measured with several molecular layers at once, such as RNA levels, DNA accessibility, and protein abundance. It integrates these layers and also explains the gene-regulatory links between them in a cell-type-specific way. The aim is to give biologists a single, interpretable model for understanding how different molecular programs interact inside the same cell.

Multi-task learning of multi-modality biological data by UnitedNet. a Schematics of representative multi-modality biological data: (i) simultaneously measured transcriptomics and intracellular electrophysiology (multi-sensing data), (ii) integratively profiled transcriptomics and DNA accessibility (multi-omics data), and (iii) spatially resolved transcriptomics and proteomics (spatial omics data). b Schematics of designation of the shared-latent space. Different modality measurements from the same cell can be projected to a shared latent space as latent codes by encoders. The latent codes can be projected back to the modality space by decoders. In the latent space, latent codes representing different modalities from the same cell can be integrated as a unimodal shared latent code used for joint group identification. c Schematics showing the network structure of UnitedNet based on the shared-latent space designed in ( b ). d , e Schematics showing the multi-task learning between joint group identification ( d ) and cross-modal prediction ( e ) by UnitedNet for analyzing multi-modality data. f , g Schematics showing the application of explainable learning methods to dissect the trained UnitedNet for identifying group-to-feature relevance ( f ) and the cross-modal feature-to-feature relevance ( g ). First, an explainable learning method dissects the encoders of a trained UnitedNet to identify the most relevant input features to each group ( f ). Then, within the grouped input features from ( f ), the explainable learning method further dissects the encoders and decoders of the trained UnitedNet to identify group-specific cross-modal feature-to-feature relevance ( g ). The drawings in panel a were created with BioRender.com.

Fig. 1 from Xin-Hui Tang et al., Nature Communications (2023). Licensed CC BY. Source: 10.1038/s41467-023-37477-x.

Nature Communications (2023) · 107 citations

Read the original → https://doi.org/10.1038/s41467-023-37477-x

Quick links

Paper Venue / Year Link
Deterministic reprogramming of neutrophils within tumors Science (2024) Read →
Cancer-associated fibroblast phenotypes are associated with patient outcome in non-small cell lung cancer Cancer Cell (2024) Read →
Inferring histology-associated gene expression gradients in spatial transcriptomic studies Nature Communications (2024) Read →
A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics Nature Communications (2023) Read →
Transformer for one stop interpretable cell type annotation Nature Communications (2023) Read →
Explainable multi-task learning for multi-modality biological data analysis Nature Communications (2023) Read →

Cancer Genomics Oncology Tumor Spatial Transcriptomics Spatial Omics Single-Cell Transcriptomics