Xianyang Zhang

Subcellular spatial transcriptomics captures where RNA is in tissue, but a single location holds too few molecules of any one gene to estimate composition alone. Every current method fixes in advance where to borrow -- a smoothing scale, a cell outline or a factor model -- and the fixed choice shapes what is visible. Kintsugi removes the fixed choice and lets held-out molecules decide, gene by gene, how much to borrow from spatial neighbours and from other genes at the same location. On a lung section measured by both Xenium and Visium HD, the data-chosen allocation placed an epithelial programme where the Xenium molecules were, ahead of smoothing, cell segmentation and a factor model; the result replicated across tissues and against protein. Across a 45-core pulmonary fibrosis cohort, separating composition from captured amount shows that a fibroblastic focus is not a place with more RNA but a place with different RNA: 2.8-fold higher in activated-fibroblast composition while segmented nuclear density is at most 1.08-fold higher.

The rapid expansion of single-cell RNA sequencing (scRNA-seq) has made accurate cell type annotation a critical bottleneck for biological discovery. Existing computational methods are often limited by reference data dependency, while emerging single Large Language Model (LLM) approaches are susceptible to model-specific biases and provide insufficient uncertainty quantification. To address these limitations, we introduce mLLMCelltype, a framework that harnesses collective intelligence-the emergent problem-solving capacity arising when multiple independent agents interact through structured deliberation to produce solutions exceeding individual capabilities-of multiple LLMs through an iterative deliberation process. Across 49 diverse datasets, our framework achieves a mean accuracy of 77.2%, a 15.7-percentage-point improvement over the best-performing single-LLM baseline (61.5%). The consensus mechanism demonstrates high robustness to noisy input and generalizes to datasets released after the LLMs' training. By providing transparent reasoning chains and robust consensus-based confidence metrics, mLLMCelltype minimizes manual annotation effort and enables reliable interpretation of complex cellular landscapes. The framework is available as an open-source package and an accessible web server.

Methods that map genetic risk to cells identify disease-relevant tissues and cell types but cannot test whether genetic effects concentrate at molecular interfaces between cells. Here we introduce EdgeMap, which integrates spatial transcriptomics with GWAS summary statistics to partition trait heritability into cell-intrinsic and intercellular components and to resolve the intercellular signal into specific ligand–receptor channels. Across 17 traits and five human tissues, edge heritability is enriched in biologically coherent trait–tissue pair-ings (3.8-fold; P = 4.4 × 10 ⁻⁶ ) and replicates across independent tissue sections, GWAS cohorts, and cell-segmented Visium HD. Per-pair decomposi-tion identifies 67 trait-specific channels (FDR < 0.10 ) organized into convergent pathway families—neurexin–neuroligin synaptic signaling in bipolar disorder, vascular adhesion in cardiovascular traits, and lipoprotein-clearance pathways in liver. Most edge genes are absent from standard gene-level prioritization, sup-porting intercellular communication as a complementary dimension of genetic architecture.

Identifying spatially variable genes (SVGs) is the first analytical step in spatial transcriptomics, determining which genes and pathways are prioritized for downstream validation. Yet the restricted spatial models of current detection methods create systematic blind spots that can exclude biologically coherent programs from discovery. Here we present F lash S, which reformulates kernel-based spatial testing in the frequency domain to detect arbitrary multi-scale expression patterns while scaling to millions of cells. In human cardiac tissue, this broader detection capacity recovers a coherent PGC-1 α -regulated mitochondrial biogenesis program—40 of 49 pathway genes spatially associated with ventricular cardiomyocytes—that PreTSA, a leading parametric alternative, largely misses (1 of 49 genes), a finding replicated in an independent cohort. Across 50 benchmark datasets spanning 9 platforms, F lash S achieves state-of-the-art ranking accuracy (mean Kendall τ = 0.935) and completes on the Allen Brain MERFISH atlas (3.94 million cells) in 12.6 minutes with 21.5 GB memory.

Spatial transcriptomics analyses often require coordinating specialized Python and R methods. When analytical intent is translated through ad hoc scripts or unconstrained LLM generated code, workflows can be difficult to inspect, rerun, or compare. We present ChatSpatial, a schema enforced orchestration framework in which an LLM selects versioned tools and parameters rather than executable code. Built on the Model Context Protocol, ChatSpatial integrates 65 methods across 15 analytical categories into 20 typed tools with curated defaults. In two cancer case studies, ChatSpatial coordinated cross ecosystem workflows that recovered the analytical structure and central conclusions of published ovarian cancer and oral squamous cell carcinoma analyses. Across three LLMs, schema enforcement raised parameter validity from 20% to 98% and stabilized method selection. In end to end benchmarks against two code generation alternatives, it yielded higher execution success and output concordance.

We propose a unified framework for fair regression tasks formulated as risk minimization problems subject to a demographic parity constraint. Unlike many existing approaches that are limited to specific loss functions or rely on challenging non-convex optimization, our framework is applicable to a broad spectrum of regression tasks. Examples include linear regression with squared loss, binary classification with cross-entropy loss, quantile regression with pinball loss, and robust regression with Huber loss. We derive a novel characterization of the fair risk minimizer, which yields a computationally efficient estimation procedure for general loss functions. Theoretically, we establish the asymptotic consistency of the proposed estimator and derive its convergence rates under mild assumptions. We illustrate the method's versatility through detailed discussions of several common loss functions. Numerical results demonstrate that our approach effectively minimizes risk while satisfying fairness constraints across various regression settings.

Coarsening Visium HD resolution from 8 to 64 µ m can flip cell-type co-localization from negative to positive ( r = −0.12 → +0.80 ), yet many widely used compositional deconvolution workflows require coarsening or subsampling at million-bin scale. Here we introduce FlashDeconv, which combines leverage-score importance sampling with sparse spatial regularization to achieve competitive benchmark accuracy while processing 1.6 million bins in 153 seconds on commodity hardware. Systematic multi-resolution analysis of Visium HD mouse intestine reveals a tissue-specific resolution horizon (8–16 µ m)—the scale at which this sign inversion occurs—validated by Xenium ground truth. Below this horizon, FlashDeconv provides, to our knowledge, the first sequencing-based quantification of Tuft cell chemosensory niches (15.3-fold stem cell enrichment). In a 1.6-million-bin human colorectal cancer cohort, FlashDeconv uncovers neutrophil inflammatory microdomains co-localized with immunoregulatory dendritic cells (mRegDC) at the tumor–stroma interface—spatial niches largely missed by discrete-label summaries, with RCTD doublet mode labeling only 2.3% of hotspot bins as neutrophil singlets.