single-cell

IRIS 是一种利用扰动筛选与迁移学习重建跨细胞类型细胞信号传导历史的新方法，发表于 Nature Methods。

该研究提出"加性分歧(AD)"指标，评估扰动预测模型在未见过的基因组合上的外推可靠性，发现专门的神经预测模型并未稳定优于简单加性基线，且偏离基线越大的预测误差越大，该信号可指导自动扰动筛选工作流。

The paper introduces Finsler Flow Matching (FFM), a framework that learns continuous stochastic dynamics from discrete Markov transition graphs using Finsler geometry to encode directional and diffusive structure, improving single-cell trajectory inference over existing generative approaches.

Arc Institute argues that virtual cell models can only predict how cell types respond to perturbations if they are trained on high-quality causal data across diverse contexts, so technical noise is not mistaken for real biology.

CellRFT introduces a reinforcement fine-tuning framework that uses biological evaluation as direct training feedback to improve single-cell perturbation models, addressing the mismatch between surrogate training losses and biological evaluation criteria.

Jingtian Zhou builds single-cell and computational frameworks to probe the 3D genome, contributing to major publications in Science, Cell, and Cancer Cell that advance understanding of cell states and diseases.

This paper introduces bioMoR, a biology-guided Mixture-of-Recursions framework for genomic learning that integrates structured biological knowledge into recursive Transformer architectures, improving efficiency and accuracy across omics benchmarks.

BioM-JEPA introduces a joint-embedding predictive architecture that learns single-cell representations by predicting graph-connected gene blocks instead of individual genes, showing improved efficiency and downstream performance in perturbation-response tasks.

This arXiv paper introduces CohortHijack, a robustness audit that removes non-target cells from single-cell query cohorts to test how annotation tools can be manipulated without altering the target cell's expression profile. It shows that structured removal and search strategies can change refined labels in popular pipelines while preserving the target, identifying query cohort composition as a vulnerability surface.

This paper studies a self-supervised task for generating single-cell gene expression vectors using an autoregressive transformer with a quantized VAE tokenizer. It reports scaling laws and a compute-optimal frontier for single-cell foundation models, with potential fine-tuning for perturbation prediction.

This paper introduces CoCoS, a contrastive pretraining framework that learns whole-cell representations from complementary transcriptomic views, addressing limitations of masked gene reconstruction in single-cell foundation models. Experiments on cell-type annotation and gene regulatory network inference show competitive transfer performance.

This paper analyzes the Virtual Cell Challenge benchmark for held-out CRISPRi perturbation prediction, finding that simple magnitude-based scalar features outperform deep MLP encoders, and that magnitude-only predictors transfer better across cell types.

GSK expands its $110M AI drug discovery partnership with Relation Therapeutics, emphasizing that clean, disease-specific biological data from lab experiments is more valuable than larger models. The move highlights an industry-wide pivot toward proprietary data as a competitive moat in AI-driven pharma.

Arc Institute announces the first whole-body map of DNA methylation and 3D genome organization at single-cell resolution across 16 human tissues, published in Science.

This paper introduces CoDiffGRN, a co-evolutionary discrete diffusion framework for gene regulatory network inference, along with a new benchmark BEELINE-KGC for inductive evaluation. It achieves state-of-the-art performance in novel regulatory discovery.

SpaCellAgent is a self-evolving LLM-based multi-agent framework that automates end-to-end spatiotemporal trajectory analysis from natural language queries, achieving over 40% improvement in analytical efficiency while maintaining expert-level performance.

PerturbCellRL introduces a reinforcement learning framework that post-trains a pretrained single-cell transcriptomic generator using cell-level verifiers as rewards, improving biological consistency of perturbation predictions beyond distributional matching.

This paper introduces PolyILR, a canonical orthonormal decomposition of the Aitchison tangent space that aligns with any tree topology, providing stable and interpretable features for compositional data such as microbiome and single-cell profiles.

The article introduces FreeBridge, a variational Schrödinger Bridge framework for modeling cellular transitions from high-content imaging data, learning stochastic transport constrained within a fixed cellular manifold to achieve competitive endpoint fidelity and reduce intermediate support violations.

CellBRIDGE is a new method that enhances optimal transport for scRNA-seq trajectory inference by incorporating ligand-receptor interaction costs to model cell-cell communication, improving alignment and enabling interpretable in silico perturbations.