Xihong Lin

The multiomic Variant-to-Gene (mV2G; https://mv2g.hbliulab.org) is a comprehensive atlas that integrates diverse functional genomic evidence to prioritize tissue-specific variant-to-gene (V2G) associations. While genome-wide association studies (GWAS) have identified millions of associations between genetic variants and diseases, translating these findings into biological mechanisms remains challenging because >90% of variants reside in noncoding regions. Existing V2G resources provide complementary regulatory evidence but are fragmented and often lack tissue-specific interpretation. To address this challenge, we constructed the mV2G atlas by integrating 24 types of functional genomic evidence across 50 human tissues, including molecular quantitative trait loci, enhancer–gene predictions, 3D chromatin interactions, and experimental validation. The atlas contains 188 634 118 evidence-supported V2G pairs involving 13 618 039 variants and 69 521 genes. To prioritize high-confidence V2G pairs, we developed a CRISPR-informed, evidence-weighted model. Using this model, we prioritized 3 348 174 high-confidence V2G pairs, 88% of which were tissue-specific. The mV2G atlas provides searchable variant- and gene-centered interfaces, an interactive browser for visualizing variants, target genes, cis-regulatory elements, and chromatin states, as well as downloadable datasets. By integrating complementary regulatory evidence into a unified framework, mV2G provides an accessible resource for interpreting the functional and phenotypic impact of genomic variation in relevant tissues for human diseases.

Autonomous agents, from virtual laboratories to biological data-analysis agents and co-scientists, have begun to turn scientific goals into hypotheses. Whether they can go further and convert real, noisy and complex biological data into executable model hypotheses and architectures that encode testable assumptions remains unclear. Predicting how cells respond to genetic and chemical perturbations (i.e., modeling virtual cells) is a natural test of this ability: it is central to decoding disease mechanisms and therapeutic discovery, yet it demands cross-disciplinary expertise rarely concentrated in a single research group. Here, we introduce CellForge, an organized team of autonomous agents for single-cell perturbation modeling. Given a dataset and a task, it surveys the literature, parses the data, identifies relevant baselines, and formulates a model hypothesis through critic-mediated discussion among agents specialized in data modeling, single-cell biology, and deep learning. It then writes the training code, runs the experiments, and returns a trained predictor including architecture, code, and validation results. The architecture itself is the hypothesis: CellForge produces it end-to-end, with no human specifying the design, loss, or training protocol. Across six perturbation datasets spanning transcriptomic, epigenomic, and proteomic contexts, CellForge yields models that are competitive with hand-crafted state-of-the-art results on established tasks and extends the same workflow to paired multiomic readouts. Blinded human evaluators rate CellForge’s research plans as more scientifically sound and innovative than those from single LLMs or deep-research systems. In summary, we show autonomous agents can complete the full hypothesis-to-validation cycle in single-cell biology, with an AI effectively building another AI, in a complex setting that currently requires a team of human experts.

The multiomic Variant-to-Gene (mV2G, https://mv2g.hbliulab.org) is a comprehensive atlas that integrates diverse functional genomic evidence to prioritize tissue-specific variant-to-gene (V2G) associations. While genome-wide association studies (GWAS) have identified millions of associations between genetic variants and diseases, translating these findings into biological mechanisms remains challenging because >90% of variants reside in noncoding regions. Existing V2G resources provide complementary regulatory evidence but are fragmented and often lack tissue-specific interpretation. To address this challenge, we constructed the mV2G atlas by integrating 24 types of functional genomic evidence across 50 human tissues, including molecular quantitative trait loci, enhancer-gene predictions, three-dimensional chromatin interactions, and experimental validation. The atlas contains 188,634,118 evidence-supported V2G pairs involving 13,618,039 variants and 69,521 genes. We further developed a unified tissue-specific V2G prioritization framework and prioritized 1,530,420 high-confidence functional V2G pairs involving 1,131,316 unique variants, with 87% exhibiting tissue-specificity. The mV2G atlas provides searchable variant- and gene-centered interfaces, an interactive browser for visualizing variants, target genes, cis-regulatory elements, and chromatin states, as well as downloadable datasets. By integrating complementary regulatory evidence into a unified framework, mV2G provides an accessible resource for interpreting the functional and phenotypic impact of genomic variation in relevant tissues for human diseases.

The UK Biobank Pharma Proteomics Project (UKB-PPP) generated plasma proteomic data for 54,219 participants. However, proteomic measurements remain unavailable for approximately 90% of the UK Biobank participants. This substantial missingness significantly limits the power for proteome-wide genetic discovery, and downstream genome-wide association studies (GWAS) using traditional imputation methods can yield spurious associations if prediction models are misspecified. We propose the Synthetic Phenotype Assisted Linear Mixed models (Syn-PALM), a robust and computationally scalable synthetic data-assisted framework for GWAS in the presence of partially observed measurements. Syn-PALM jointly analyzes partially observed proteomic measurements and complete synthetic data predicted for all samples using machine learning models, while accounting for cryptic relatedness and population structure using mixed models. Syn-PALM is robust to prediction model misspecification and increases statistical power as prediction accuracy improves. Through simulations, ablation analyses, and GWAS analysis of the UKB-PPP data, Syn-PALM improves power over existing methods and boosts proteome-wide genetic discovery.

Heritability enrichment analysis using data from genome-wide association studies is often used to understand the functional basis of genetic architecture. Stratified linkage disequilibrium score regression (S-LDSC) is a widely used method-of-moments estimator for heritability enrichment, but S-LDSC has low statistical power compared with likelihood-based approaches. We introduce graphREML, a precise and powerful likelihood-based heritability partitioning and enrichment analysis method. It utilizes summary statistics from genome-wide association studies and sparse linkage disequilibrium graphical models, which make likelihood calculations tractable. We validate our method using extensive simulations and in analyses of a wide range of real traits. On average across traits, graphREML produces enrichment estimates that are concordant with S-LDSC, indicating that both methods are unbiased; however, graphREML identifies 2.5 times more significant trait-annotation enrichments, demonstrating greater power compared with the moment-based S-LDSC approach. Furthermore, graphREML flexibly models the relationship between the annotations of an SNP and its heritability, producing well-calibrated estimates of per-SNP heritability.

Epstein–Barr virus (EBV) infects more than 95% of adults worldwide but is associated with endemic nasopharyngeal carcinoma (NPC) specifically in southern China^1,2,3,4. Here, through a stepwise host–EBV genome interaction analysis, we identify a genetic interaction between HLA-A*11:01 and the high-risk EBV variant 85841G as a key determinant of NPC risk. Individuals carrying a susceptible HLA-A background (HLA-A*11:01⁻ or HLA-A*02:07⁺) and infected with the high-risk 85841G EBV form a dual-risk subgroup with substantially elevated, interaction-driven NPC risk, far exceeding the effects of host or virus alone. This dual-risk subgroup comprises 20.5% of the population and accounts for approximately 47% of NPC cases. We show that EBV 85841G encodes an EBNA3B peptide that binds to HLA-A*11:01 and elicits specific T cell responses capable of lysing EBV⁺ B cells transformed by 85841G-carrying strains, and is associated with reduced salivary viral load and lower NPC risk among A*11:01 carriers. Evolutionary analysis reveals that 85841G arose via ancient recombination events between northern and southern EBV and subsequently underwent clonal expansion in southern China, leading to co-enrichment of interacting host and viral risk factors that, in turn, contribute to NPC endemicity. These findings reveal a markedly stratified, interaction-driven risk architecture in NPC and highlight opportunities for precision prevention.

Rare genetic variants are increasingly recognized as important contributors to human trait architecture, with noncoding variants accounting for a substantial portion of the heritability. These variants tend to be less polygenic and more biologically specific than common variants, remaining understudied across large biobanks. Here we analyzed whole genome sequencing (WGS) data from up to 490,549 UK Biobank participants to assess the effects of rare coding and noncoding variants across 1,342 phenotypes, including 944 diseases, 76 clinical biomarkers, and 322 metabolomics traits. We developed and applied STAARpipelinePheWAS , a scalable framework for WGS phenome-wide rare variant association analysis, identifying 49,121 genome-wide significant gene-trait pairs. Our study presents a comprehensive map of noncoding rare variant associations in both disease and biomarker domains. Many associations were undetected in prior exome- or array-based studies and were enriched in drug targets and biologically coherent pathways. All results are publicly accessible through an interactive portal ( https://www.staarphewas.org/ ), offering a foundational resource for rare variant discovery, functional interpretation, and translational genomics.

Whole genome sequencing (WGS) studies play a pivotal role in studying the genetic underpinnings of human diseases and traits. High quality and reproducible variant calling is the cornerstone for the success of downstream analyses, including WGS association studies and polygenic risk prediction. This paper compares the data quality, performance, and concordance of two widely used WGS variant callers, the Genome Analysis Toolkit (GATK) and Variant Tool set that discovers short variants (VT), using 60 532 multi-ancestry whole genomes sequenced by the Centers for Common Disease Genomics (CCDGs) of the NHGRI Genome Sequencing Program. Our findings show that both QCed GATK and VT pipelines yield highly consistent and reliable called Single Nucleotide Variants (SNVs) in large-scale WGS studies, supporting their agreements in joint variants calling. However, the two pipelines exhibit greater discrepancies in calling insertions and deletions (INDELs).

Whole genome sequencing (WGS) studies in large biobanks provide an unprecedented opportunity to study the rare-variant (RV) effects on the natural history of human diseases by analyzing censored time-to-event (TTE) phenotypes, such as age at disease diagnosis, disease progression, and lifespan. Unlike existing methods developed for continuous and categorical phenotypes, rare-variant association tests (RVATs) for TTE phenotypes in large biobanks face several major challenges, including heavy censoring, cryptic relatedness, and population structure. We introduce GATE-STAAR (Genetic Analysis of Time-to-Event phenotypes via the variant-Set Test for Association using Annotation infoRmation), a powerful and computationally efficient frailty model framework for RVATs of TTE phenotypes in large biobanks. GATE-STAAR accounts for high censoring rates, cryptic relatedness, and population structure in large biobanks, while incorporating multifaceted variant functional annotations to improve power and result interpretability. We propose a rare-variant saddlepoint approximation method to effectively address heavy censoring in WGS TTE analysis. We demonstrate through extensive simulations that GATE-STAAR is powerful while maintaining proper control of type I error rates. We apply GATE-STAAR to analyze the WGS data of approximately 400,000 UK Biobank participants of white British ancestry across a variety of TTE phenotypes, and validate the findings using participants of European ancestry from the All of Us Research Program. These analyses uncover RV associations with age at diagnosis of a range of diseases.