• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Tuesday, November 17, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Biology

Contrastive Graph Learning Framework Improves Spatial Domain Detection Across Tissue Platforms

by
November 17, 2026
in Biology
Reading Time: 6 mins read
0
Contrastive Graph Learning Framework Improves Spatial Domain Detection Across Tissue Platforms

Contrastive Graph Learning Framework Improves Spatial Domain Detection Across Tissue Platforms

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A computational framework named CASTLE, developed by researchers Lai Wei, Aoqi Xie, and Yuehua Cui from the Department of Statistics and Probability at Michigan State University, improves the identification of spatial domains in spatially resolved transcriptomics by integrating physical cell locations with local cell-type composition. Published in PLOS Genetics, the method, titled Cell-type Aware SpaTial domain detection via contrastive Learning Embedding, couples a biologically informed neighborhood graph with a self-supervised contrastive learning objective. In extensive benchmarking across multiple tissue types, spatial resolutions, and technological platforms, the method demonstrated competitive or superior clustering accuracy, refined domain boundaries, and enhanced across-sample stability compared to both established and emerging spatial clustering baselines. The authors reported that explicitly accounting for local cell-type mixtures or expression profiles allows the model to capture fine-grained tissue structures and functional subdomains that conventional spatial clustering workflows often obscure.

Spatially resolved transcriptomics platforms have transformed biological research by measuring comprehensive gene-expression profiles while preserving the physical microenvironment of tissue sections. A foundational computational task in analyzing such data is spatial domain delineation, which partitions tissue into coherent anatomical, functional, or pathological regions. Accurate domain identification is essential for discovering domain-specific marker genes, evaluating spatial microenvironments, and charting cell-to-cell communication networks. Traditional computational workflows have frequently relied on non-spatial clustering algorithms, such as the Louvain or Leiden graph-partitioning techniques implemented in standard single-cell analysis toolkits, or spatial probabilistic formulations including hidden Markov random fields and Bayesian spatial priors. In recent years, graph neural network architectures have emerged as prominent solutions, learning low-dimensional representations by convolving expression measurements across spatial adjacency graphs. However, precisely resolving boundaries between neighboring functional zones remains a significant computational hurdle across complex and heterogeneous tissues.

A critical limitation of existing graph neural network methods is their reliance on physical proximity alone or on potentially circular pre-clustering steps. While spatial distance establishes basic adjacency, many biological tissues display sharp domain borders dictated primarily by shifts in cell-type composition rather than physical distance alone. In structures such as the human dorsolateral prefrontal cortex, the mouse olfactory bulb, and the hippocampus, distinct cellular mosaics and laminar distributions define functional zones. Although some previous models have attempted to integrate cell-type awareness, they frequently introduce circular label feedback. For instance, some pipelines run Louvain clustering on gene expression principal components to generate pre-clustering labels that subsequently reshape the neighborhood graph for downstream clustering. This feedback mechanism risks reinforcing initial classification errors and over-smoothing true transitional boundaries in heterogeneous regions. CASTLE addresses this conceptual challenge by directly incorporating independently derived cell-type proportions or single-cell annotations into graph construction without iterative circular feedback.

The CASTLE framework begins by pre-processing spatially resolved transcriptomics data, removing non-tissue background pixels, normalizing raw library sizes per spot, applying a logarithmic transformation, scaling features to zero mean and unit variance, and selecting the top 3,000 highly variable genes. To represent spatial context, CASTLE first constructs an undirected k-nearest neighbor spatial graph over the measured tissue coordinates. The method then refines the connectivity and edge weights of this graph using biological similarity metrics tailored to the data platform. For multicellular spot-level data such as 10x Visium, CASTLE computes the cosine similarity between cell-type composition vectors estimated by deconvolution algorithms such as RCTD or gwSPADE. For single-cell resolution platforms like MERFISH, categorical cell-type identities set similarity to one for identical cell classes and zero otherwise. When annotations are unavailable, noisy, or derived from subcellular platforms, CASTLE utilizes a label-free fallback called CASTLE-ES, computing cosine similarities across low-dimensional principal component embeddings.

To learn robust low-dimensional representations, CASTLE pairs its biologically informed graph with a single-layer graph convolutional network encoder and a mirrored decoder regularized by a self-supervised contrastive learning objective. The contrastive component adopts a local-context strategy inspired by GraphST, comparing representations against an augmented negative view. CASTLE generates a corrupted graph by randomly permuting gene-expression features across spatial coordinates while preserving the underlying geometric neighborhoods. Edges are reweighted according to the permuted feature similarities, producing a hard negative counterpart that maintains spatial topology while disrupting local biological coherence. The encoder maps both the original and corrupted graphs into latent embeddings. For every spot, a local-context summary vector is calculated by averaging the latent vectors of its immediate topological neighbors and applying a nonlinear sigmoid function. A discriminator then scores agreement between spots and their context vectors, maximizing mutual information in the true view while penalizing spurious correlations derived from the permuted graph.

The training objective of CASTLE combines self-supervised contrastive learning on local microenvironments with an expression self-reconstruction loss. The encoder produces a latent spot representation that the decoder maps back to the original feature dimension to reconstruct the normalized expression matrix. The total loss balances the mean squared reconstruction error and the symmetric binary cross-entropy contrastive losses using weighting hyper-parameters alpha and beta, which were fixed at 1.0 and 0.1, respectively, across benchmarks. Optimization is conducted using the Adam algorithm with a learning rate of 0.001 over 600 training epochs. Following convergence, the reconstructed expression matrix is subjected to Gaussian mixture modeling via the mclust algorithm to delineate final spatial domains. For unannotated datasets, cluster counts are determined by maximizing the average Silhouette Coefficient. An optional post-processing refinement reassigns isolated spots to the majority domain within their immediate neighborhood, utilizing radii of 20 spots for multicellular assays and 50 cells for single-cell platforms.

Wei and colleagues comprehensively validated CASTLE against nine leading spatial clustering methods, including GraphST, STAGATE, stAA, CellCharter, NicheCompass, MENDER, IRIS, BASS, and BANKSY. On the 10x Visium human dorsolateral prefrontal cortex benchmark consisting of 12 annotated slices, CASTLE achieved the highest median Adjusted Rand Index and the lowest variance across sections among the primary representative methods. For sample 151670, CASTLE attained an Adjusted Rand Index of 0.6485, whereas competing representative methods scored between 0.43 and 0.47, with CASTLE reliably capturing laminar borders and white matter where others suffered over-segmentation. In sensitivity evaluations across different deconvolution tools, CASTLE achieved Adjusted Rand Indexes of 0.6485 with RCTD, 0.5836 with CARD, and 0.5835 with cell2location on slice 151670. On mouse olfactory bulb 100-micrometer data, CASTLE attained an Adjusted Rand Index of 0.5372 against STdeconvolve reference labels, outperforming GraphST at 0.4875, STAGATE at 0.2373, and stAA at 0.1468 while preserving laminar geometry.

The research team further demonstrated CASTLE’s capacity to dissect complex pathological tissue using a 10x Genomics Human Breast Cancer Visium section comprising 3,798 spots and 36,601 genes, categorized into 20 pathologist-annotated regions across ductal carcinoma in situ, lobular carcinoma in situ, invasive ductal carcinoma, tumor edges, and healthy tissue. CASTLE achieved the highest overall Adjusted Rand Index of 0.5688 against the 20 reference regions, while revealing functional subdomains within coarsely annotated sections. In the Healthy_1 region, CASTLE resolved Cluster 4, marked by ACTA2, DCN, and VWF expression alongside extracellular matrix and hemostasis pathways, separate from Cluster 18, which expressed CD34 and exhibited strong enrichment for oxidative phosphorylation and mitochondrial electron transport. Within invasive ductal carcinoma region IDC_2, CASTLE separated three distinct transcriptional programs: Cluster 8, an estrogen-receptor luminal state enriched for AP-2 transcription factor regulation and marker AC093001.1; Cluster 10, characterized by extracellular matrix breakdown; and Cluster 19, marked by cap-dependent translation initiation and mitotic cell division.

CASTLE was also evaluated on cellular-resolution imaging platforms, including a MERFISH dataset of the mouse hypothalamic preoptic area spanning five sections, 155 genes, 15 cell types, and eight anatomical regions. Across the five sections, CASTLE achieved a mean Adjusted Rand Index of 0.4402, and its label-free expression variant CASTLE-ES achieved 0.4264, remaining close to BASS at 0.4434 and outperforming GraphST, STAGATE, and stAA, though BANKSY recorded a higher mean score of 0.5006. On three mouse medial prefrontal cortex sections profiled by STARmap, CASTLE-ES demonstrated lower variability across slices than stAA, outperforming GraphST and STAGATE. In a single-cell 10x Genomics Xenium human breast cancer dataset analyzed across six major biological compartments, CASTLE achieved the highest Adjusted Rand Index among nine evaluated computational methods, maintaining robust clustering across sensitivity checks involving the reassignment of rare mast cells and coarser 16- and 13-category annotation schemes.

To explore ultra-high-density spatial transcriptomics, the investigators applied CASTLE to subcellular and micro-binned datasets, including mouse olfactory bulb sections profiled by Stereo-seq and mouse hippocampus sections profiled by Slide-seqV2. In the Stereo-seq analysis, where tissue was divided into seven laminar layers using DAPI staining as ground truth, CASTLE was the only tested framework that cleanly delineated the internal rostral migratory stream from the granule cell layer, while GraphST and stAA mixed the granule and internal plexiform layers. On Slide-seqV2 hippocampus data containing over 10,000 beads, CASTLE successfully mapped CA1, CA3, and dentate gyrus compartments alongside canonical marker genes Wfs1, Ly6e, and C1ql2. Additionally, when evaluated on a 10x Genomics Visium HD human colorectal cancer dataset comprising 137,048 spatial bins at 16-micrometer resolution across 18,085 genes, CASTLE-ES achieved a Silhouette Coefficient of 0.36, outscoring GraphST at 0.33, STAGATE at 0.29, and stAA at 0.24, accurately separating smooth muscle, connective tissue, and neoplasm.

While CASTLE achieved strong performance across diverse platforms and resolutions, the authors documented several methodological limitations inherent to its design. The framework assumes that spatial domains correlate at least partially with underlying cell-type composition or discrete transcriptional shifts; consequently, its advantage may be attenuated in tissue regions characterized by continuous biological gradients or uniform cell compositions where boundaries are governed by subtle transient states. Additionally, inaccuracies or technical artifacts arising from upstream cell-type deconvolution or single-cell labeling can propagate into initial graph topology, impacting final domain assignments. The current implementation processes individual tissue sections independently and does not incorporate explicit multi-sample batch integration or automated hyper-parameter tuning across extreme resolutions.

Subject of Research: Spatial domain identification in spatially resolved transcriptomics using cell-type-aware graph contrastive learning

Article Title: CASTLE: Cell-type aware SpaTial domain detection via contrastive learning embedding

Article References: Wei, L., Xie, A., & Cui, Y. (2026). CASTLE: Cell-type aware SpaTial domain detection via contrastive learning embedding. PLOS Genetics, 22(11), e1012365. https://doi.org/10.1371/journal.pgen.1012365

Image Credits: AI Generated

DOI: 10.1371/journal.pgen.1012365

Keywords: CASTLE framework, spatial transcriptomics, spatial domain detection, contrastive learning, graph neural networks, cell-type deconvolution, dorsolateral prefrontal cortex, breast cancer microenvironment, MERFISH, Stereo-seq, Slide-seqV2, Visium HD, Gaussian mixture modeling, tissue architecture, local context embedding, spatial clustering

News Source: Drew Townsend. (November 17, 2026). Contrastive Graph Learning Framework Improves Spatial Domain Detection Across Tissue Platforms. Scienmag.

Tags: Breast cancer microenvironmentCASTLE frameworkcell type deconvolutioncontrastive learningdorsolateral prefrontal cortexGaussian mixture modelinggraph neural networkslocal context embeddingMERFISHSlide-seqV2spatial clusteringspatial domain detectionSpatial transcriptomicsStereo-seqtissue architectureVisium HD
Share12Tweet7Share2ShareShareShare1

Related Posts

Local Regulatory T Cell Delivery Strategies Show Preclinical Promise in Bone Repair

Local Regulatory T Cell Delivery Strategies Show Preclinical Promise in Bone Repair

November 17, 2026
Oxfendazole Suppresses Bone Destruction in Preclinical Osteosarcoma Models by Inhibiting ANXA2

Oxfendazole Suppresses Bone Destruction in Preclinical Osteosarcoma Models by Inhibiting ANXA2

November 17, 2026

TMPRSS2 Protease Impairs Antibody Recognition of the SARS-CoV-2 Spike Protein

November 17, 2026

Veterinary Brucellosis Testing Requires Validated Decision Rules Across Animal Populations

November 17, 2026

POPULAR NEWS

  • Plane-Based Gonial Angle Segmentation Better Represents Mandibular Growth in Longitudinal Study

    Plane-Based Gonial Angle Segmentation Better Represents Mandibular Growth in Longitudinal Study

    29 shares
    Share 12 Tweet 7
  • Quality Improvement Project Raises Pre-Antibiotic Microbiological Analysis Rate to 79.2 Percent

    29 shares
    Share 12 Tweet 7
  • LLM-Driven Social Simulations Face Significant Epistemic Validity Challenges

    29 shares
    Share 12 Tweet 7
  • Veterinary Brucellosis Testing Requires Validated Decision Rules Across Animal Populations

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

High Patient Satisfaction Reported in Bladder Cancer Prehabilitation Program

Local Regulatory T Cell Delivery Strategies Show Preclinical Promise in Bone Repair

Study Finds Heart Valve Infection in Nearly One in Ten Spinal Infection Admissions

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.