Research Radar — 2026-07-02

Summary: This study generates a genome-scale CRISPRi perturbation atlas in human induced pluripotent stem cells, systematically mapping transcriptional regulatory networks across the genome. By profiling the effects of thousands of knockdowns on gene expression, the resource reveals key regulators of pluripotency, differentiation, and cell identity. The atlas establishes a foundational reference for understanding gene regulation in stem cells and beyond.

Why it matters: Provides the first comprehensive transcriptional regulatory map of human pluripotent stem cells at genome scale, setting a new benchmark for functional genomics and perturbation-based gene regulatory network inference.

Why for Yiru: Directly relevant to my interest in CRISPR screens and perturbation biology — this atlas provides a reference framework for designing and interpreting perturbation screens in immune cells and cancer models.

Summary: Kang et al. present BulkFormer, a foundation model pretrained on large-scale human bulk RNA-seq profiles that captures gene-gene interactions and expression context. The model improves performance across diverse bulk transcriptome tasks, and the study demonstrates that pretraining data modality is a key determinant of downstream foundation model performance.

Why it matters: Challenges the prevailing single-cell-centric paradigm for transcriptomic foundation models by showing that bulk RNA-seq pretraining can outperform single-cell pretraining for many important downstream tasks.

Why for Yiru: Directly applicable to my research on multi-omics integration and biomarker discovery in cancer immunotherapy — BulkFormer's approach to modeling gene interactions could enhance my analysis of bulk transcriptomic data from tumor microenvironments.

Summary: General-purpose tabular foundation models (TabICL, TabPFN) outperform specialized single-cell foundation models (scGPT, scLAMBDA, PRESAGE, STACK, Prophet) at perturbation prediction across cell-level, pseudobulk, and whole-embryo settings. This demonstrates that in-context learning on tabular data provides a strong and scalable alternative to domain-specific architectures for biological prediction.

Why it matters: Challenges the assumption that domain-specific foundation models are necessary for biological tasks, with significant implications for how the field designs and deploys AI models for perturbation biology.

Why for Yiru: Highly relevant to my work on perturbation prediction and single-cell analysis — tabular in-context learning could simplify my modeling pipeline for predicting immune cell responses to genetic perturbations.

Summary: MintCNA integrates statistical modeling with deep learning to detect copy number alterations from paired single-cell multi-omics data. Its attention-guided convolutional autoencoder denoises the signal while multivariate change-point detection identifies genome-wide chromosomal breakpoints, outperforming existing single-omics CNA callers.

Why it matters: Fills a critical gap in single-cell multi-omics analysis by providing a unified CNA caller that leverages information across data modalities, improving accuracy for cancer genomics applications.

Why for Yiru: Relevant to my work on multi-omics and cancer genomics — MintCNA's integrative approach for copy number profiling is useful for characterizing genomic instability in tumor immune microenvironments.

Summary: ClairS applies deep learning to accurately detect somatic small variants from long-read sequencing data of tumor-normal pairs. It addresses the unique challenges of long-read data, including higher error rates and complex alignment patterns, achieving high accuracy for cancer genomics applications.

Why it matters: As long-read sequencing becomes increasingly adopted in cancer genomics, accurate somatic variant calling from this data type is essential for precision oncology and translational research.

Why for Yiru: Somatic variant calling is foundational to cancer genomics — this long-read approach may improve mutation detection sensitivity in my tumor microenvironment studies.

Summary: Penumbria is a general-purpose 3D cell segmentation framework using U-Net with xLSTM bottleneck blocks, a Global Zernike Phase Layer for optical aberration correction, and a Scaled Geocaps Layer for multi-scale feature routing. It achieves state-of-the-art accuracy across diverse cell morphologies and imaging conditions, outperforming Cellpose-SAM and StarDist-3D.

Why it matters: Advances the state of the art in 3D cell segmentation, a critical bottleneck for high-content imaging and spatial biology at single-cell resolution.

Why for Yiru: While not my primary focus, advances in cell segmentation are relevant to spatial omics and imaging-based tissue analysis of the tumor microenvironment.