Research Radar — 2026-05-01

Summary: Comprehensive benchmarking of Xenium spatial transcriptomics across over 40 breast and lung tumor sections, systematically dissecting technical noise including transcript spillover. The authors introduce SPLIT (Spatial Purification of Layered Intracellular Transcripts), a method that resolves mixed transcriptomic signals to improve background correction and cell-type resolution, revealing T-cell exhaustion signatures associated with malignant cell colocalization that would otherwise remain obscured.

Why it matters: Xenium is one of the most widely adopted spatial transcriptomics platforms, yet its technical limitations have been poorly characterized. This is the most comprehensive benchmarking study to date and SPLIT provides a practical solution for signal refinement that will improve the quality of spatial analyses across the field.

Why for Yiru: Directly relevant to spatial transcriptomics methodology. The T-cell exhaustion colocalization finding demonstrates how improved signal processing can uncover clinically meaningful immune-tumor spatial relationships. The benchmarking framework and SPLIT method are immediately applicable to TME spatial analysis pipelines.

Summary: A comprehensive survey of biological foundation models that maps the rapid evolution of this field, cataloging current architectures, training strategies, and applications from single-cell omics to protein design, while providing an outlook for future developments including multi-modal integration and clinical translation.

Why it matters: Foundation models are reshaping computational biology at an unprecedented pace. A systematic survey that traces where the field has been and where it is heading is essential reading for anyone building or applying these models in biomedical research.

Why for Yiru: Core to Boss's research interests in biomedical AI and representation learning. The survey touches on single-cell foundation models, multi-omics integration, and the pathway toward clinically deployable AI — all directly aligned with ongoing work.

Summary: TEA-GCN (Two-tier Ensemble Aggregation Gene Co-expression Network) leverages unsupervised transcriptomic dataset partitioning and multi-metric co-expression scoring to construct robust gene networks from over 450,000 public RNA-seq samples across 12 species, outperforming state-of-the-art methods in gene function prediction and regulatory network inference with enhanced cross-species conservation.

Why it matters: Public transcriptomic data is massively underutilized for network inference due to batch effects and sample heterogeneity. TEA-GCN provides a scalable, explainable framework that makes this resource accessible for functional genomics across species.

Why for Yiru: Gene regulatory network inference is foundational for understanding TME cell states and interactions. The ensemble approach and NLP-based explainability could inform how we construct and interpret cell-type-specific networks from spatial and single-cell data.

Summary: A method combining DNA large language models with graph domain adaptation for automated cell type annotation of single-cell ATAC-seq data, addressing the challenge of transferring annotations across different chromatin accessibility datasets and experimental conditions.

Why it matters: Cell type annotation remains a major bottleneck in single-cell epigenomics. Leveraging DNA language models — which capture genomic sequence grammar — combined with graph-based domain adaptation is a novel paradigm that could generalize across diverse single-cell modalities.

Why for Yiru: The DNA LLM + graph domain adaptation architecture is directly relevant to Boss's interests in foundation models and representation learning for single-cell biology. The domain adaptation framework could be extended to spatial transcriptomics and cross-modality integration.

Summary: Systematic evaluation of amino acid similarity-aware kmer representations for T-cell receptor repertoire classification, assessing how biochemical properties of amino acids can improve the representation of immune receptor sequences for downstream machine learning tasks.

Why it matters: TCR repertoire analysis is a key computational immunology challenge with implications for cancer immunotherapy monitoring and biomarker discovery. Better sequence representations directly enable more accurate repertoire-based diagnostics.

Why for Yiru: Directly relevant to computational immunology interests. TCR repertoire analysis connects to tumor immunology and immunotherapy response prediction. The representation learning angle ties into broader interests in biomedical AI.