Research Radar — 2026-06-04

Summary: Introduces MetaSTAARlite, a summary-statistics-based pipeline that enables powerful, functionally informed rare-variant meta-analysis across biobank-scale whole-genome sequencing (WGS) cohorts. As WGS becomes the standard for large-scale population and disease studies — with biobanks such as UK Biobank, All of Us, and FinnGen generating WGS data for hundreds of thousands of participants — the computational challenge shifts from single-cohort analysis to meta-analysis across multiple cohorts. Individual-level data sharing is often restricted by privacy regulations and data use agreements, making summary-statistics-based meta-analysis the only practical approach. However, existing rare-variant meta-analysis tools are either not scalable to WGS-scale data (hundreds of millions of variants) or lack the functional annotation integration that gives methods like STAAR their power. MetaSTAARlite addresses both gaps: it operates on per-variant summary statistics (score statistics and covariance matrices) that can be shared without individual-level data, incorporates functional annotation weights across multiple annotation categories (epigenetic, evolutionary conservation, protein structure), and scales linearly with sample size through efficient matrix operations. The method supports both single-variant and aggregate rare-variant tests, gene-centric and sliding-window analyses, and conditional analysis to identify independent signals. Applied to biobank-scale WGS data, MetaSTAARlite substantially increases discovery power for rare-variant associations while maintaining well-calibrated type I error rates.

Why it matters: Rare variants (minor allele frequency < 1%) collectively explain a substantial portion of missing heritability for complex diseases, but individual rare variants have effect sizes too small to reach genome-wide significance without extremely large sample sizes. MetaSTAARlite provides the computational infrastructure to combine rare-variant evidence across all available WGS cohorts, potentially unlocking discoveries that no single cohort could achieve alone. The summary-statistics-based design also respects data privacy constraints, making it compatible with the federated analysis models increasingly mandated by biobanks and funding agencies. As WGS replaces genotyping arrays in population genomics, tools like MetaSTAARlite will become essential for maximizing the scientific return on these massive investments.

Why for Yiru: Rare-variant analysis is increasingly relevant to cancer genomics — germline rare variants may influence TME composition, immunotherapy response, and cancer susceptibility in ways that common variants cannot capture. MetaSTAARlite could be applied to combine rare-variant evidence across cancer cohorts from multiple biobanks, potentially identifying rare variants that affect immune-related phenotypes (cytokine levels, immune cell counts, checkpoint expression). The functional annotation weighting framework is also conceptually transferable — one could develop TME-specific annotation weights that prioritize variants affecting genes expressed in relevant immune or stromal cell types. More broadly, the computational approach to federated rare-variant analysis exemplifies the type of privacy-preserving, scalable methods needed for multi-center cancer genomics consortia.

Summary: Presents the largest single-cell cis-expression quantitative trait locus (eQTL) study to date, mapping genetic regulation of gene expression across 2.2 million single cells from intestinal biopsies and peripheral blood of 421 individuals, including 125 with inflammatory bowel disease (IBD). Most genetic variants associated with complex diseases lie in non-coding regions, making it difficult to identify which genes they regulate and in which cell types. Traditional eQTL studies use bulk tissue RNA-seq, which averages gene expression across all cell types and misses cell-type-specific regulatory effects. This study addresses this limitation by performing single-cell eQTL mapping in intestinal tissue — directly in the disease-relevant organ — across multiple intestinal cell types including epithelial cells, stromal cells, and diverse immune populations. Three key findings emerge: first, cell-type-level eQTLs are systematically more distal to transcription start sites and enriched in enhancers compared to tissue-level eQTLs, consistent with enhancers being the primary mediators of cell-type-specific gene regulation. Second, enhancer-enriched, cell-type-resolved eQTLs co-localize with IBD GWAS loci at substantially higher rates than tissue-level eQTLs — meaning that cell-type resolution directly improves the ability to connect disease-associated variants to effector genes and cell types. Third, many IBD-associated variants regulate different genes in different cell types, highlighting the inadequacy of the "one variant, one gene, one cell type" model. The authors provide an interactive resource enabling querying of eQTL results across cell types and genetic variants.

Why it matters: This study represents the state of the art in connecting non-coding genetic variation to disease mechanisms at cellular resolution. The finding that cell-type-level eQTLs dramatically improve GWAS co-localization — and that enhancers are the key regulatory elements driving this improvement — has immediate implications for how post-GWAS functional studies should be designed. The observation that single variants can regulate different genes in different cell types challenges the prevailing paradigm of variant-to-gene (V2G) mapping and suggests that the regulatory consequences of disease-associated variants are more complex and context-dependent than appreciated. For IBD specifically, identifying the specific intestinal cell types through which genetic risk operates could guide cell-type-targeted therapeutic development.

Why for Yiru: The single-cell eQTL framework developed here is directly applicable to the TME. One could perform single-cell eQTL mapping in tumour and adjacent normal tissue to identify genetic variants that regulate gene expression specifically in TME cell types — tumour cells, T cells, macrophages, fibroblasts. These TME-specific eQTLs could reveal why certain germline variants influence cancer risk or immunotherapy response — for example, a variant that alters chemokine expression specifically in tumour-associated macrophages. The computational methods for cell-type-resolved eQTL mapping and GWAS co-localization provide a template for similar analyses in cancer. The finding that enhancers are the primary drivers of cell-type-specific regulation also connects to Yiru interest in the epigenetic basis of TME heterogeneity.

Summary: Presents PepAnno, a comprehensive web server for multi-functional peptide annotation powered by a structure-aware, multi-view geometric deep learning framework. Peptides are gaining prominence as therapeutic candidates — they occupy a unique niche between small molecules and biologics, combining the target specificity of antibodies with the tissue penetration of small molecules. However, computational tools for predicting peptide bioactivities have lagged behind those for small molecules and proteins, with many existing tools relying on sequence-only representations that ignore 3D structural information. PepAnno addresses this gap with a dual-stream architecture: a Transformer branch processes pre-trained sequence embeddings to capture evolutionary and biochemical information, while a GATv2 (graph attention network) branch processes predicted 3D structural graphs to capture geometric features such as secondary structure elements, surface properties, and folding patterns. A cross-modal attention mechanism fuses these two representations, enabling the model to learn which features — sequence-based or structure-based — are most predictive for each bioactivity task. The model supports simultaneous prediction across seven key bioactivities (antimicrobial, anticancer, antiviral, antifungal, antibiofilm, hemolytic, and cell-penetrating), consistently matching or outperforming task-specific methods. The web server also provides automated physicochemical profiling, 3D structure visualization, and access to integrated peptide databases.

Why it matters: Therapeutic peptides represent a rapidly growing drug class — over 80 peptide drugs are currently approved, with hundreds in clinical development for indications ranging from diabetes to cancer. Computational tools that can rapidly screen peptide libraries for desired bioactivities while avoiding undesirable properties (toxicity, hemolysis) could dramatically accelerate peptide drug discovery. The structure-aware design of PepAnno is particularly important because peptide function depends critically on 3D conformation — an alpha-helical antimicrobial peptide and a disordered peptide with the same sequence composition can have completely different activities. The web server format also democratizes access, making sophisticated deep learning tools available to experimental researchers without computational expertise.

Why for Yiru: Peptide-based therapeutics are an emerging modality for TME-targeted therapy. Therapeutic peptides could be designed to block immunosuppressive checkpoint interactions (PD-L1/PD-1), disrupt chemokine-receptor axes that recruit immunosuppressive cells, or deliver cytotoxic payloads selectively to tumour cells expressing specific surface markers. PepAnno could be used to screen candidate peptide libraries for TME-relevant properties — for example, identifying peptides predicted to penetrate tumour cells while avoiding hemolysis. More broadly, the dual-stream architecture combining sequence and structure representations provides a template for designing multimodal deep learning models for other biological molecules relevant to the TME, such as cytokines and chemokines.

Summary: Introduces MOGT (Multi-omics Graph Transformer Network), a semi-supervised graph neural network that models biological networks derived from multi-omics data to predict disease-associated genes and identify drug repurposing candidates. Genome-wide association studies have identified thousands of genetic loci associated with neuropsychiatric and neurodegenerative disorders, but translating these statistical associations into mechanistic understanding and therapeutic opportunities remains a major challenge. MOGT addresses this by integrating multiple types of biological networks — protein-protein interactions, gene co-expression, and functional annotation similarities — into a heterogeneous graph, then using a graph transformer architecture with prototype-based attention to learn which network neighborhoods are most predictive of disease gene status. The prototype augmentation is the key innovation: rather than learning gene representations solely from local network neighborhoods, MOGT learns a set of prototype vectors that represent archetypal "disease gene network patterns," and each gene representation is augmented by its similarity to these prototypes. Applied to Parkinson disease (PD), MOGT outperforms existing methods in disease gene prediction and identifies high-risk genes that are then used to query the Connectivity Map (CMAP) database for drugs that reverse the disease-associated gene expression signature. Among 10 predicted candidates, the drug UK-356618 was experimentally validated in a primary neuron model, reversing PD-associated gene expression abnormalities and improving cellular phenotypes.

Why it matters: The prototype-augmented graph learning approach is methodologically innovative — it addresses a fundamental limitation of graph neural networks, which tend to learn myopic representations biased toward local network structure. By augmenting each gene representation with global prototype information, MOGT learns richer representations that capture higher-order patterns in biological networks. The experimental validation of a predicted drug — going from GWAS hits through computational prediction to wet-lab validation — demonstrates a complete translational pipeline for drug repurposing that could be applied to many other diseases. The framework is general and could incorporate additional data modalities and network types.

Why for Yiru: The drug repurposing pipeline demonstrated here — disease gene prediction from multi-omics networks followed by CMAP-based drug identification — is directly applicable to TME research. One could apply MOGT to identify genes associated with immunotherapy resistance using GWAS or transcriptome-wide association study data from cancer cohorts, then query CMAP for drugs that reverse the resistance-associated gene expression signature, identifying candidates for combination immunotherapy. The prototype-augmented graph learning approach could also be adapted to TME-specific networks — for example, a cell-cell communication graph where prototypes might represent archetypal "immune-hot" or "immune-cold" TME architectures. The successful experimental validation provides confidence in computational drug repurposing pipelines for TME applications.

Summary: Presents a systematic comparison of four simulation-based inference (SBI) methods — Approximate Bayesian Computation (ABC), Neural Posterior Estimation (NPE), NPE with temporal embeddings, and Preconditioned Neural Posterior Estimation (PNPE) — applied to epidemic models of increasing complexity with explicit attention to structural and practical identifiability. Epidemic models are essential for understanding disease transmission, forecasting outbreak trajectories, and evaluating intervention strategies, but calibrating these models to data is challenging because their likelihood functions are often analytically intractable. SBI methods circumvent the likelihood by learning the relationship between model parameters and simulated data, enabling Bayesian inference for models where traditional methods fail. The authors evaluate these four methods across epidemic models ranging from simple SIR (susceptible-infected-recovered) to complex age-structured and spatially explicit models, under varying observational noise and fixed simulation budgets. Key findings include: neural methods (NPE, PNPE) generally produce more faithful posterior distributions than ABC under constrained simulation budgets; PNPE achieves strong performance but requires substantially more computational resources and may need reconditioning for each new observation; temporal embeddings improve inference for models with complex epidemic dynamics by capturing sequential dependencies; and ABC remains the most computationally efficient option, though it produces more conservative (wider) posterior estimates. The study also highlights the critical importance of identifiability assessment — many epidemic model parameters cannot be uniquely estimated from available data, regardless of the inference method used.

Why it matters: Simulation-based inference is rapidly gaining adoption across computational biology — from ecology and epidemiology to systems biology and single-cell genomics — because it enables statistically rigorous parameter inference for models where the likelihood is intractable. However, practitioners face a bewildering choice of methods with limited guidance on which to use. This systematic benchmark provides that guidance, with actionable recommendations based on model complexity, simulation budget, and available computational resources. The explicit treatment of identifiability is particularly valuable — it serves as a reminder that no inference method can extract information that is not present in the data, and that identifiability analysis should be a mandatory step in any SBI workflow.

Why for Yiru: Simulation-based inference methods are directly applicable to mechanistic models of the TME. Mathematical models of tumour-immune interactions — including ODE-based models of cytokine signaling, agent-based models of cell migration and interaction, and stochastic models of immune repertoire dynamics — typically have intractable likelihoods but can be simulated forward. SBI could be used to calibrate these models to spatial transcriptomics, multiplexed imaging, or longitudinal tumour growth data, enabling inference of parameters such as immune cell infiltration rates, tumour cell killing rates, and cytokine diffusion coefficients. The benchmarking results directly inform which SBI method to choose for TME models of varying complexity. The identifiability analysis framework is also essential for TME models, which often have many parameters relative to available data — identifying which parameters can and cannot be estimated from a given experimental design is critical for model-based experimental planning.