Research Radar — 2026-06-29

Summary: Introduces a weight-only geometric audit framework that scores genes by embedding geometry to identify representational outliers in single-cell foundation models (Geneformer, scGPT, scFoundation). Shared outliers are enriched for loss-of-function intolerance and disease association, and geometric anomaly predicts perturbation sensitivity (ρ=0.725).

Why it matters: Provides the first systematic audit of embedding geometry in single-cell foundation models, revealing that tokenization strategy shapes representational quality and that geometric anomalies predict real biological fragility.

Why for Yiru: Directly relevant as these evaluation methods could be applied to audit any single-cell foundation model Yiru works with, ensuring reliable gene representations for downstream analysis.

Summary: Introduces PertOmni, a CLIP-style multimodal representation learning framework that aligns transcriptomic perturbation signatures with text embeddings of gene/compound descriptions and cell painting images. Uses masked contrastive objectives for within-cell-type discrimination across small molecule and CRISPRi perturbation datasets.

Why it matters: Enables unified representation learning across perturbation modalities (small molecules, CRISPR, cell painting), addressing the key challenge of generalizing across perturbation types and datasets.

Why for Yiru: Yiru works with perturbation screens (CRISPR, drug) — this framework could directly improve how perturbation effects are modeled and predicted in his research.

Summary: CONCISE is a statistical method for spatially constrained co-expression and ligand-receptor interaction inference that jointly models spatial autocorrelation, molecular count variation, measurement errors, and spatial proximity. Shows most existing methods produce inflated false-positive rates while CONCISE achieves well-calibrated inference.

Why it matters: Addresses a critical methodological gap in spatial transcriptomics — spurious co-expression due to spatial autocorrelation — which has been a known confound in ligand-receptor analysis.

Why for Yiru: Essential for Yiru's spatial transcriptomics work — provides more reliable cell-cell communication inference than existing tools.

Summary: SPEAK is an LLM-based method for identifying spatial domains from SRT data using prior knowledge from both LLMs and human experts. Constructs spatial context prompts for each spot based on cell types and marker genes, enabling zero-shot inference, expert-guided fine-tuning, and prototype updating. Validated on STARmap, Visium, MERFISH, and Xenium data.

Why it matters: Pioneers the use of LLMs for spatial domain identification, combining biological priors with flexible expert-guided refinement across multiple platforms.

Why for Yiru: Directly applicable to Yiru's spatial transcriptomics projects — offers a more interpretable and adaptable approach to domain identification than current methods.

Summary: Develops a visually interpretable deep learning model predicting T-cell-enriched immune scores from H&E whole-slide images using multimodal spatial supervision from CD8/PAX8/ERG IHC. The H&E DL Immune score correlated with T-effector RNA scores across clinical trial cohorts (ρ=0.71) and predicted ICI response.

Why it matters: Demonstrates that H&E morphology alone can predict T-effector status and ICI response when properly constrained by spatial supervision, addressing a critical unmet need for clinically deployable predictive biomarkers.

Why for Yiru: Relevant for Yiru's interest in computational pathology and immunotherapy biomarkers — shows how multimodal spatial supervision improves histological immune profiling.

Summary: Proposes an ambiguity-aware multi-stage framework for spatial cell-type annotation combining hybrid spatial feature clustering with constrained LLM inference. Reduces cluster-level ambiguity from 16.1% to 2.27% on Xenium cholangiocarcinoma data; preserves mixed clusters rather than forcing labels.

Why it matters: Addresses the underappreciated problem of overconfident cell-type assignments in spatial transcriptomics by explicitly modeling and flagging ambiguous annotations.

Why for Yiru: Useful for Yiru's spatial analysis pipeline — especially important when annotating heterogeneous tumor microenvironments where ambiguous cell states are biologically meaningful.