Research Radar — 2026-05-20
Methods & AI
Computational
HESTIA: Scalable Multimodal Integration of Histology and High-Resolution Spatial Transcriptomics
bioRxiv Published 2026-05-18 preprint DOI: 10.1101/2026.05.14.723098
spatial transcriptomics deep learning histology multimodal integration high-resolution scalable
Summary: Introduces HESTIA, a scalable deep learning framework for multimodal integration of histology images and high-resolution spatial transcriptomics data. HESTIA addresses the fundamental challenge of aligning morphological features in tissue sections with spatially resolved gene expression at subcellular resolution, enabling joint representation learning across modalities. The framework is designed for scalability to large tissue sections and whole-slide images while preserving fine-grained spatial information, opening the door to systematic multimodal tissue atlasing.
Why it matters: Multimodal integration of histology and spatial transcriptomics is one of the most pressing computational challenges in spatial biology. HESTIA's scalable approach could become a foundational tool for the growing number of spatial atlasing initiatives, enabling researchers to extract richer biological insights from the combination of tissue morphology and molecular profiles.
Why for Yiru: Spatial transcriptomics and histology integration are directly relevant to understanding the tumour microenvironment — where tissue morphology and molecular states together define immune cell infiltration patterns, tumour architecture, and treatment response. HESTIA's multimodal framework could enable more comprehensive spatial analysis of TME samples.
SpatialArtifacts: Tissue Artifact Detection in Spatial Transcriptomics
bioRxiv Published 2026-05-18 preprint DOI: 10.1101/2026.05.15.725260
spatial transcriptomics quality control artifact detection tissue processing data preprocessing deep learning
Summary: Presents SpatialArtifacts, a computational method for automated detection and classification of tissue artifacts in spatial transcriptomics data. Tissue artifacts — including folding, tearing, and processing-induced damage — are pervasive in spatial transcriptomics experiments and can confound downstream analyses if not properly identified. SpatialArtifacts provides a systematic framework for flagging artifact-affected regions, enabling researchers to filter or account for technical artifacts before biological interpretation.
Why it matters: Quality control is the unsung hero of spatial transcriptomics — artifacts can masquerade as biological signals and lead to false discoveries. A dedicated artifact detection tool fills a critical gap in the spatial analysis toolkit, improving the reliability of all downstream analyses from differential expression to spatial domain identification.
Why for Yiru: Spatial TME analysis relies on accurate tissue quality assessment. Artifact detection is particularly important when analyzing clinical tumour samples, which often have variable tissue quality. A robust QC tool would improve confidence in spatial analyses of the TME.
Deep Learning for Cross-Domain Spatial Transcriptomic Modeling of Tissue Repair
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.13.724803
spatial transcriptomics deep learning tissue repair cross-domain modeling transfer learning wound healing
Summary: Develops a deep learning framework for cross-domain spatial transcriptomic modeling of tissue repair, enabling knowledge transfer across different tissue types and experimental platforms. The model learns domain-invariant representations of spatially resolved gene expression during wound healing and tissue regeneration, capturing conserved repair programs across tissues while accommodating tissue-specific features. Demonstrates that cross-domain modeling improves the detection of repair-associated spatial gene expression patterns compared to single-domain approaches.
Why it matters: Tissue repair is a universal biological process with relevance to wound healing, fibrosis, and cancer. Cross-domain modeling that identifies conserved repair programs across tissues could reveal fundamental principles of tissue regeneration and its dysregulation in disease.
Why for Yiru: Tissue repair and wound healing share key cellular programs with the tumour microenvironment — including fibroblast activation, immune recruitment, and extracellular matrix remodeling. Understanding conserved repair programs across tissues could inform TME biology and reveal parallels between regenerative and neoplastic processes.
HESpotEx: Spot-Level Gene Expression Prediction from Histology Images
Nature Computational Science Published 2026-05-15 research article DOI: 10.1038/s43588-026-00992-0
spatial transcriptomics deep learning histopathology gene expression prediction computational pathology H&E
Summary: Presents HESpotEx, a deep learning model that predicts spot-level gene expression directly from histology images (H&E-stained sections) without requiring paired spatial transcriptomics data for inference. HESpotEx learns the relationship between tissue morphology and local gene expression patterns from training datasets, enabling virtual spatial transcriptomics from routine histology slides. Published in Nature Computational Science, this represents a significant step toward making spatial gene expression inference accessible from widely available histopathology images.
Why it matters: The ability to predict gene expression from routine histology slides could democratize spatial transcriptomics by eliminating the need for expensive and specialized spatial assays. In clinical settings where H&E slides are routinely collected, HESpotEx could enable retrospective spatial gene expression analysis at scale.
Why for Yiru: Predicting molecular features from histology images is directly relevant to computational pathology approaches for TME analysis. The ability to infer gene expression from H&E slides could enable large-scale spatial analysis of archived tumour specimens without requiring fresh tissue for spatial sequencing.
scMAGCA: Interpretable Single-Cell Multiomics Integration with Multi-Scale Attention Graph Convolutional Autoencoders
Nature Communications Published 2026-05-15 research article DOI: 10.1038/s41467-026-73055-7
single-cell multiomics gene regulation interpretability graph neural network batch correction data integration
Summary: Introduces scMAGCA, an interpretable framework for single-cell multiomics data integration using multi-scale attention graph convolutional autoencoders. scMAGCA jointly embeds multiple omics modalities (e.g., RNA, ATAC, protein) from the same cells into a shared latent space while preserving modality-specific features. The multi-scale attention mechanism provides interpretability by identifying which genomic regions and genes drive the integrated representation, and the graph-based architecture enables effective batch correction and noise reduction.
Why it matters: Single-cell multiomics is the next frontier in cellular profiling, but integrating disparate data modalities without losing biological signal is technically challenging. scMAGCA's interpretable approach addresses the black-box problem that plagues many integration methods, enabling researchers to understand which molecular features drive the integrated representations.
Why for Yiru: Multiomics integration is essential for comprehensive TME characterization — combining gene expression, chromatin accessibility, and protein measurements from the same cells could reveal regulatory mechanisms underlying immune cell states and tumour cell plasticity.
Deep Learning Models for Chemical Perturbation Prediction Do Not Yet Use Drug Features
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.13.724458
deep learning drug discovery benchmarking chemical perturbation model evaluation feature engineering
Summary: Systematically evaluates the extent to which deep learning models for chemical perturbation prediction actually utilize drug-specific features versus relying primarily on cell-line and gene-expression context. Through rigorous benchmarking, the authors find that many state-of-the-art models perform nearly as well when drug features are randomized or removed, suggesting they are driven by cellular context rather than chemical structure. This raises concerns about the true generalizability of these models to novel compounds.
Why it matters: Chemical perturbation prediction is central to drug discovery and toxicity screening. If models do not actually use drug features, they cannot generalize to novel compounds — undermining their utility for virtual screening. This benchmarking study highlights critical evaluation gaps in the field.
Why for Yiru: Rigorous benchmarking and critical evaluation of computational models are essential for ensuring that methods deployed in biomedical research are fit for purpose. The finding that models may exploit dataset biases rather than learn true chemical biology relationships has parallels across computational biology.
Biomedical discoveries
Biomedicine
Spatial Proteomic Atlas of Tertiary Lymphoid Structures in Non-Small Cell Lung Cancer
bioRxiv Published 2026-05-18 preprint DOI: 10.1101/2026.05.14.723890
spatial proteomics TLS NSCLC immunotherapy tumour microenvironment B cells
Summary: Generates a comprehensive spatial proteomic atlas of tertiary lymphoid structures (TLS) in non-small cell lung cancer, mapping the protein-level organization of immune cell subsets, stromal components, and signalling molecules within and around TLS. The atlas reveals distinct TLS maturation states defined by spatial protein expression patterns, identifies molecular features associated with TLS functionality, and correlates TLS proteomic architecture with immunotherapy response. This represents one of the most detailed spatial proteomic characterizations of TLS to date.
Why it matters: TLS are lymphoid aggregates that form in tumours and are strongly associated with immunotherapy response, yet their molecular architecture and functional heterogeneity remain poorly understood at the protein level. A spatial proteomic atlas provides actionable insights into which TLS features predict treatment benefit and how TLS function might be therapeutically enhanced.
Why for Yiru: TLS biology and its relationship to immunotherapy response are directly relevant to TME research. Understanding the spatial proteomic organization of TLS — including B cell, T cell, and stromal compartmentalization — connects to interests in immune cell organization and anti-tumour immunity.
Integrated Multi-Omics Identifies Distinct Macrophage Alterations During Progression of Metabolic Dysfunction-Associated Steatohepatitis
Nature Genetics Published 2026-05-18 research article DOI: 10.1038/s41588-026-02600-3
macrophage multi-omics MASH spatial biology GPNMB liver disease single-cell
Summary: Integrates single-nucleus transcriptomics, spatial multi-omics, and proteomics on human liver samples to delineate the evolving landscape of hepatic macrophages across the MASLD-to-MASH disease spectrum. Reveals progressive depletion of Kupffer cells accompanied by emergence of diverse, phenotypically distinct macrophage subsets. Spatial multi-omics demonstrates that disease progression toward MASH is marked by accumulation of antigen-presenting, phagocytic GPNMB+ macrophages supported by IL32-producing hepatocytes. Identified macrophage markers enable patient stratification by disease activity and stage across independent clinical cohorts.
Why it matters: MASLD affects over 30% of the global population yet the immune mechanisms driving progression from benign steatosis to inflammatory MASH remain poorly understood. This study provides the most comprehensive macrophage atlas of human MASH progression to date and identifies GPNMB+ macrophages as a spatially defined, stage-specific population with potential as both biomarkers and therapeutic targets.
Why for Yiru: Macrophage heterogeneity and spatial organization in chronic inflammatory disease directly parallel questions in the tumour microenvironment. The multi-omics integration approach and identification of spatially coordinated hepatocyte-macrophage interactions are methodologically and conceptually relevant to TME research.
Immunotherapy Drug Target Identification Using Machine Learning and Patient-Derived Tumour Explant Validation
Nature Machine Intelligence Published 2026-05-18 research article DOI: 10.1038/s42256-026-01201-3
immunotherapy drug target discovery graph neural network machine learning patient-derived explants TME
Summary: Introduces MIDAS (Mining Immunotherapy Drug tArgetS), a multimodal graph neural network system for immuno-oncology target discovery that integrates gene interactions, multi-omic patient profiles, immune cell biology, antigen processing, disease associations, and genetic perturbation phenotypes. MIDAS generalizes to time-sliced data, outcompetes state-of-the-art baselines including OpenTargets, ranks approved targets above those in clinical development, and recovers immunotherapy-response-associated genes in unseen patients. Functional perturbation of oncostatin M–oncostatin M receptor signaling in TRACERx melanoma patient-derived explants reduced dysfunctional CD8+ T cells and CCL4 levels, providing clinical validation.
Why it matters: Current immunotherapy benefits only a minority of patients, and novel target discovery remains slow and expensive. MIDAS represents a systematic computational framework that not only predicts targets but validates them in clinically relevant patient-derived explants — bridging the gap between computational prediction and translational immunology with direct therapeutic implications.
Why for Yiru: Immuno-oncology target discovery and TME modulation are directly aligned with Boss's research interests. The patient-derived explant validation approach and the identification of oncostatin M as a TME modulator connect to interests in spatial TME biology and immunotherapy resistance mechanisms.
Reprogramming Tumour-Associated Macrophages from Immune Suppressive to Inflammatory State by Checkpoint Kinase 1 Inhibitor Combination Treatment
bioRxiv Published 2026-05-17 preprint DOI: 10.1101/2026.05.13.724422
tumour-associated macrophages CHK1 inhibitor TAM reprogramming immunotherapy tumour microenvironment drug combination
Summary: Demonstrates that Checkpoint kinase 1 (Chk1) inhibitor combination treatment reprograms tumour-associated macrophages (TAMs) from an immune-suppressive M2-like state to an inflammatory, anti-tumour M1-like state. The study characterizes the molecular mechanism by which Chk1 inhibition reshapes the TAM transcriptional and functional landscape, and shows that TAM reprogramming contributes to the anti-tumour efficacy of Chk1 inhibitor combinations in preclinical models, adding an immune component to what was previously considered a purely tumour-cell-intrinsic therapeutic strategy.
Why it matters: Chk1 inhibitors are in clinical development primarily for their tumour-cell-intrinsic effects on DNA damage response. The discovery that they also reprogram TAMs toward anti-tumour phenotypes adds an unexpected immunological dimension and suggests combination strategies with immunotherapy that leverage both tumour-cell and immune-cell effects.
Why for Yiru: TAM reprogramming is a major therapeutic axis in the TME. Understanding how existing clinical-stage drugs like Chk1 inhibitors reshape macrophage states could inform rational combination immunotherapy design and spatial analysis of treatment response.
Overweight Status Drives Early Tumor Microenvironment Reprogramming in Pancreatic Ductal Adenocarcinoma: A Cell-Type-Resolved Bayesian Hierarchical Modeling and Interactome Analysis
bioRxiv Published 2026-05-17 preprint DOI: 10.1101/2026.05.14.721695
TME pancreatic cancer obesity Bayesian single-cell interactome cell-type resolved
Summary: Uses cell-type-resolved Bayesian hierarchical modeling and interactome analysis to investigate how overweight status drives early tumour microenvironment reprogramming in pancreatic ductal adenocarcinoma (PDAC). Reveals that overweight-associated metabolic and inflammatory signals reshape cell-cell communication networks in the pre-malignant and early tumour TME, altering fibroblast, immune, and epithelial cell states in ways that may accelerate PDAC progression. Identifies specific ligand-receptor interactions and signaling pathways through which systemic metabolic status is transduced into local TME changes.
Why it matters: Obesity is a major risk factor for pancreatic cancer and is associated with worse outcomes, but the mechanisms linking systemic metabolic status to local TME reprogramming are poorly understood. This study provides a computational framework and specific molecular hypotheses for how overweight status primes the TME for cancer progression.
Why for Yiru: The intersection of systemic metabolism and TME biology is highly relevant to understanding cancer risk and progression. The cell-type-resolved Bayesian modeling approach is methodologically relevant to spatial and single-cell TME analysis.
NGFR Identifies a Basal Subpopulation in Bladder Cancer Associated with Immunotherapy Resistance
bioRxiv Published 2026-05-18 preprint DOI: 10.1101/2026.05.14.725085
bladder cancer immunotherapy resistance NGFR biomarker basal subtype tumour heterogeneity
Summary: Identifies NGFR (nerve growth factor receptor) as a marker of a basal subpopulation in bladder cancer that is associated with resistance to immunotherapy. Using single-cell and spatial transcriptomic analyses of bladder cancer specimens, the study characterizes the molecular features of NGFR+ tumour cells, their spatial organization within the TME, and their relationship to immune exclusion and checkpoint blockade failure. NGFR emerges as both a biomarker of immunotherapy resistance and a potential therapeutic vulnerability.
Why it matters: Immunotherapy resistance remains the major barrier in bladder cancer treatment. Identifying specific tumour cell subpopulations that drive resistance — and their molecular markers — enables patient stratification and reveals new therapeutic targets for overcoming resistance.
Why for Yiru: Immunotherapy resistance mechanisms and tumour cell subpopulations that drive immune evasion are central to TME research. The identification of NGFR as a resistance-associated marker connects to interests in tumour heterogeneity and biomarker discovery.
Cross-disciplinary watchlist
Other Fields
Temporal Architecture of the Seminiferous Cycle Revealed by Spatial Transcriptomics
Cell Published 2026-05-15 research article DOI: 10.1016/j.cell.2026.04.032
spatial transcriptomics spermatogenesis developmental biology testis cell state transitions temporal dynamics
Summary: Uses spatial transcriptomics to reconstruct the temporal architecture of the seminiferous cycle in unprecedented resolution, mapping the spatial organization of germ cell differentiation stages along the seminiferous tubules. The study reveals how the physical arrangement of developing germ cells in space encodes the temporal progression of spermatogenesis, identifying spatially coordinated gene expression programs that govern each stage of sperm development. This Cell publication represents a landmark in applying spatial technologies to fundamental developmental biology questions.
Why it matters: The seminiferous cycle is a classic model of spatially organized developmental progression, but its molecular architecture has never been mapped at genome scale with spatial resolution. This study demonstrates the power of spatial transcriptomics to decode temporally organized biological processes from their physical arrangement in tissue.
Why for Yiru: The concept of temporal dynamics encoded in spatial organization has direct parallels in the TME, where tumour evolution, immune infiltration, and treatment response all have spatiotemporal dimensions. Methodological approaches for extracting temporal information from spatial data are broadly applicable.
EMReady2: Improving Cryo-EM Density Maps with Mamba-Based Deep Learning
Nature Communications Published 2026-05-16 research article DOI: 10.1038/s41467-026-71794-1
cryo-EM deep learning Mamba structural biology density map protein structure
Summary: Presents EMReady2, a Mamba-based deep learning framework for improving cryo-electron microscopy density maps. EMReady2 leverages the Mamba state-space model architecture — known for efficient long-range dependency modeling — to enhance the quality and interpretability of cryo-EM reconstructions. The method improves map resolution, reduces noise, and sharpens structural features across diverse macromolecular complexes, outperforming existing map improvement tools.
Why it matters: Cryo-EM has revolutionized structural biology but density map quality remains a bottleneck for many targets, especially flexible or low-abundance complexes. A state-of-the-art deep learning tool for map improvement expands the range of structures accessible to cryo-EM and accelerates structure-based drug design.
Why for Yiru: Structural biology and deep learning are increasingly intertwined. The application of state-space models (Mamba) to structural biology problems is a methodological development relevant to staying current with deep learning architectures in biomedical research.
Biological Foundation Models Enable Annotation and Structure Prediction Across Divergent Genomes
bioRxiv Published 2026-05-16 preprint DOI: 10.1101/2026.05.15.724572
foundation models genome annotation protein structure cross-species deep learning comparative genomics
Summary: Develops biological foundation models capable of genome annotation and protein structure prediction across evolutionarily divergent genomes, from model organisms to non-model species with limited experimental data. The models leverage pretraining on diverse genomic and proteomic data to generalize across the tree of life, enabling functional annotation and structure prediction in species where traditional methods fail due to lack of training data.
Why it matters: The vast majority of Earth's biodiversity remains genomically uncharacterized because annotation tools require species-specific training data. Foundation models that generalize across divergent genomes could unlock functional genomics for non-model organisms and accelerate discoveries in evolutionary biology, agriculture, and natural product discovery.
Why for Yiru: Foundation models that generalize across biological contexts are a major trend in computational biology. The ability to transfer knowledge across species has conceptual parallels to cross-tissue and cross-condition transfer learning relevant to TME research.
Hong Kong Genome Project: A Population Resource for Genomic Medicine
Nature Medicine Published 2026-05-15 research article DOI: 10.1038/s41591-026-04410-w
genomic medicine population genetics Chinese population reference genome variant catalog precision medicine
Summary: Reports the initial findings and data resource from the Hong Kong Genome Project, a large-scale population genomics initiative characterizing genetic variation in the Southern Chinese population. The project catalogs population-specific variants, establishes allele frequency references, and identifies medically actionable variants with different frequencies compared to European reference datasets, highlighting the importance of population-matched genomic resources for precision medicine in Asian populations.
Why it matters: Most genomic reference datasets are heavily biased toward European populations, limiting the accuracy and equity of genomic medicine for non-European populations. The Hong Kong Genome Project helps close this representation gap and provides critical infrastructure for precision medicine in Chinese and broader Asian populations.
Why for Yiru: Population-specific genomic variation is relevant to understanding cancer susceptibility, drug metabolism, and immunotherapy response across different patient populations. The Hong Kong resource may contain variants relevant to cancer predisposition and treatment outcomes in Chinese patients.
Molecular Subtyping of COPD Using Variational Autoencoders on Multi-Omics Data
Nature Communications Published 2026-05-19 research article DOI: 10.1038/s41467-026-72989-2
COPD VAE molecular subtyping multi-omics deep learning disease heterogeneity
Summary: Applies variational autoencoders (VAEs) to multi-omics data from COPD patients to identify molecular subtypes beyond conventional clinical classifications. The VAE-based approach discovers latent molecular subtypes with distinct transcriptomic, proteomic, and metabolomic profiles that correlate with disease progression trajectories and treatment responses. This molecular taxonomy reveals heterogeneity within clinically defined COPD that has implications for personalized treatment strategies.
Why it matters: COPD is a heterogeneous disease where clinical classification fails to capture the molecular diversity driving different disease trajectories. Deep learning-based molecular subtyping could enable personalized treatment approaches in a disease affecting hundreds of millions worldwide.
Why for Yiru: Variational autoencoders for molecular subtyping are methodologically relevant to cancer classification and TME stratification. The approach of using deep generative models to discover latent disease subtypes from multi-omics data is directly applicable to tumour heterogeneity analysis.
Oncolytic Measles Virus in a Mesothelioma-on-Chip Model Reveals TME Remodeling
bioRxiv Published 2026-05-14 preprint DOI: 10.1101/2026.05.12.724508
oncolytic virus organ-on-chip mesothelioma TME virotherapy 3D model
Summary: Studies oncolytic measles virus therapy in a mesothelioma-on-chip microfluidic model that recapitulates key features of the tumour microenvironment including 3D architecture, stromal components, and immune cell interactions. The organ-on-chip platform reveals how oncolytic measles virus remodels the TME — altering cytokine gradients, immune cell recruitment, and tumour-stroma interactions — in ways not observable in conventional 2D cultures. Identifies TME remodeling mechanisms that contribute to both therapeutic efficacy and resistance.
Why it matters: Oncolytic virotherapy is an emerging cancer treatment modality, but understanding how viruses interact with the complex 3D TME has been limited by inadequate model systems. Organ-on-chip models bridge the gap between simple cell culture and animal models, enabling mechanistic studies of virus-TME interactions.
Why for Yiru: Organ-on-chip models of the TME are valuable platforms for studying TME biology under controlled conditions. The demonstration of oncolytic virus-mediated TME remodeling connects to interests in how therapeutic interventions reshape immune and stromal compartments in tumours.