Research Radar — 2026-06-02

Summary: Presents the Conditional Monge Gap framework, a conditional optimal transport method that models the distribution shift between unperturbed and perturbed single-cell populations, enabling generalizable perturbation outcome prediction. A central challenge in single-cell perturbation modeling is that experimental data cover only a tiny fraction of possible perturbations — most gene knockouts, drug treatments, and genetic combinations have never been experimentally profiled. The authors address this by formulating perturbation response as a conditional optimal transport problem: given a control cell population and a perturbation condition, predict how the full transcriptomic distribution shifts. The Monge Gap formulation learns a transport map parameterized by the perturbation condition, enabling the model to generalize to perturbation conditions not seen during training. This is critical because the combinatorial space of multi-gene perturbations is astronomically large, and experimental characterization of even a modest fraction is infeasible. The framework is evaluated on single-cell perturbation datasets including drug response and CRISPR knockout screens, demonstrating superior generalization to held-out conditions compared to existing conditional generative models. The conditional formulation naturally handles compositional perturbations — multiple drugs or gene knockouts applied simultaneously — by learning to compose transport maps from individual perturbation embeddings.

Why it matters: Predicting perturbation responses computationally could dramatically accelerate functional genomics and drug discovery by prioritizing experiments and identifying synergistic combinations in silico. The conditional optimal transport formulation is mathematically principled and addresses a key weakness of existing methods: overfitting to seen perturbation conditions. The ability to generalize to unseen single and combinatorial perturbations makes this framework immediately useful for designing CRISPR screens, predicting drug synergy, and identifying genetic interactions. The Monge Gap approach also naturally produces interpretable transport maps that reveal which genes and pathways are most affected by a perturbation.

Why for Yiru: Perturbation modeling in the TME context could predict how specific genetic or pharmacological interventions reshape the tumour-immune ecosystem — for example, predicting which checkpoint combinations would most effectively reprogramme T cell exhaustion or macrophage polarization. The conditional transport framework could model how perturbations shift immune cell states along differentiation trajectories, identifying interventions that push cells toward anti-tumour phenotypes. More broadly, this approach could be applied to model how tumour mutations (natural perturbations) alter the cellular composition and communication networks of the TME.

Summary: Introduces SPAC-seq (Spatial Perturbation And Capture sequencing), a technology that combines high-throughput CRISPR perturbation libraries with spatial whole-transcriptome readout, and TARDIS, a companion statistical toolkit for spatial perturbation analysis. CRISPR screens have revolutionized functional genomics by enabling systematic gene knockout or activation followed by phenotypic readout, but traditional screens operate on dissociated cells and lose all spatial context — they cannot reveal how a gene perturbation affects tissue organization, cell localization, or intercellular communication. SPAC-seq addresses this by performing CRISPR perturbations in intact tissue contexts and capturing both the perturbation identity and the spatial transcriptome of each cell. TARDIS then provides statistical methods for analyzing these data, including differential expression testing that accounts for spatial autocorrelation, identification of spatially variable perturbation effects, and inference of how perturbations alter ligand-receptor interactions between neighboring cells. Together, SPAC-seq and TARDIS enable discoveries in spatial functional genomics including how specific genes control cell localization within tissue niches, how perturbations reshape microenvironmental organization, and which signaling pathways mediate perturbation effects across cell types.

Why it matters: This represents a convergence of two of the most powerful tools in modern genomics — CRISPR screening and spatial transcriptomics — into a single platform. The key advance is moving from "what does this gene do to a cell?" to "what does this gene do to a tissue?" This is transformative for understanding multicellular processes like tumour-immune interactions, developmental patterning, and tissue regeneration, where a gene function cannot be understood in isolated cells. The TARDIS statistical framework also addresses the non-trivial analytical challenges of spatial perturbation data, including the need to distinguish direct perturbation effects from secondary effects propagated through cell-cell communication.

Why for Yiru: SPAC-seq is directly applicable to studying TME organization — one could systematically perturb immune checkpoint genes, chemokine receptors, or metabolic enzymes in tumour-immune co-cultures or organoids and map how each perturbation reshapes the spatial organization of the TME. This would reveal which genes control immune cell infiltration, exclusion, and niche formation — questions central to understanding immunotherapy response and resistance. The ligand-receptor inference module of TARDIS is particularly valuable for identifying the molecular mediators of spatial reorganization following perturbations. Computational methods for analyzing SPAC-seq data — especially for integrating perturbation effects across multiple spatial scales — represent a rich area for methodological development.

Summary: Introduces SIGnature, a method that combines explainable artificial intelligence (XAI) with single-cell RNA foundation models to score gene importance for cell-type and state classification, enabling scalable cross-dataset analyses. Single-cell foundation models — large transformer models pre-trained on tens of millions of single-cell transcriptomes — have demonstrated impressive zero-shot capabilities for cell-type annotation, batch integration, and gene program discovery. However, these models function largely as black boxes: it is difficult to understand which genes drive their predictions for a given cell type or state. SIGnature addresses this by applying post-hoc interpretability methods (specifically, gradient-based attribution) to RNA foundation models, producing gene-level importance scores that quantify each gene contribution to the model classification decision. These scores can be computed for any cell type or state of interest and are shown to be robust across foundation model architectures. The authors demonstrate that SIGnature-derived gene importance scores recover known lineage markers, identify novel cell-state regulators, and enable comparative analysis across datasets — for example, identifying which genes distinguish exhausted T cells across different tumour types or immunotherapy contexts.

Why it matters: As foundation models become the default tools for single-cell analysis, methods for interpreting their predictions become essential — both for scientific discovery and for clinical trust. SIGnature transforms foundation models from annotation tools into discovery engines: instead of just labeling cells, the model now tells you which genes define each cell state. This is particularly valuable in disease contexts where the goal is not just to classify cells but to identify the molecular drivers of pathogenic cell states. The cross-dataset scalability is also important — SIGnature can identify conserved gene programs that define, for example, immunotherapy-responsive T cell states across multiple patient cohorts and cancer types.

Why for Yiru: SIGnature could be applied to TME single-cell atlases to systematically identify the genes that define functionally important TME cell states — exhausted T cells, immunosuppressive macrophages, antigen-presenting DCs, inflammatory CAFs. Comparing SIGnature scores across treatment-naive vs. post-immunotherapy tumours could reveal which genes are most important for therapy-induced cell state transitions. Foundation model interpretability is also relevant to Yiru computational methods work — understanding what single-cell foundation models know about the TME could inform the design of TME-specific foundation models or fine-tuning strategies.

Summary: Presents Spatial-DC, a graph-enhanced transfer learning method that computationally profiles single-cell-type-resolved signatures from spatial proteomics data. Spatial proteomics technologies (such as imaging mass cytometry, MIBI, and CODEX) can measure 40-100 protein markers in tissue sections with single-cell resolution, but the resulting data have two key limitations: each cell proteomic profile is sparse (only a subset of the proteome is measured), and cell-type annotation requires reference data that may not be available for the specific tissue context. Spatial-DC addresses both challenges through graph-enhanced transfer learning. First, it constructs a spatial neighborhood graph where edges connect physically adjacent cells, enabling the model to learn from both a cell own marker expression and the expression patterns of its neighbors. Second, it uses transfer learning to project cells from the spatial proteomics space into a reference single-cell transcriptomics space, enabling cell-type annotation and the imputation of unmeasured protein expression. The method generates refined cell-type distribution maps and reconstructs cell-type-resolved proteomes, enabling high-resolution downstream analysis at both single-cell-type and spatial-context levels, including identification of spatially organized immune niches and cell-cell interaction hotspots.

Why it matters: Spatial proteomics is increasingly used in clinical and translational research because proteins are the direct targets of most therapeutics and provide functional information that transcriptomics cannot. However, the analytical toolchain for spatial proteomics has lagged behind that for spatial transcriptomics. Spatial-DC bridges this gap by bringing concepts from single-cell transcriptomics (reference mapping, imputation, cell-type annotation) into the spatial proteomics domain, while respecting the unique properties of protein data. The graph-enhanced approach is particularly well-suited to tissue analysis because it explicitly models the spatial dependencies that define tissue architecture.

Why for Yiru: The TME is defined by spatially organized protein-level interactions — checkpoint ligands binding to receptors, cytokines diffusing through tissue, and extracellular matrix remodeling. Spatial-DC could enhance the resolution of TME analysis from spatial proteomics data by imputing unmeasured immune checkpoint and cytokine proteins, annotating rare TME cell types, and identifying spatial neighborhoods where specific protein-level interactions are enriched. The method also provides a framework for integrating spatial proteomics with scRNA-seq reference atlases, which could connect TME protein architecture to transcriptomically defined cell states.

Summary: Presents a systematic benchmarking study comparing 43 computational methods for identifying topologically associating domains (TADs) from Hi-C and related chromatin conformation capture data, along with TADShop, a software tool that integrates the best-performing methods. TADs are megabase-scale chromatin interaction domains that partition the genome into regulatory neighborhoods — genes and their enhancers are typically confined within the same TAD, and disruption of TAD boundaries can lead to enhancer hijacking and oncogene activation in cancer. Despite the biological importance of TADs, there are over 40 published methods for calling TADs from Hi-C data, with no consensus on which method performs best or under what conditions. This study benchmarks all major TAD callers across multiple Hi-C resolutions, sequencing depths, and biological contexts using both simulated data with ground truth and experimental data. The authors find substantial variability in TAD calls across methods and identify a subset of top-performing methods that are robust across conditions. TADShop provides a unified interface for running recommended methods and comparing their outputs.

Why it matters: The proliferation of TAD-calling methods without systematic benchmarking has created confusion in the 3D genomics field, with different studies reaching different conclusions partly because of methodological choices. This study provides the community with an authoritative comparison that will guide method selection for years to come. The finding that method performance varies dramatically across conditions is important — it means that one-size-fits-all recommendations are insufficient, and users should validate TAD calls with multiple methods. This is also relevant to cancer genomics because TAD disruption is a recurrent mechanism of oncogene activation.

Why for Yiru: 3D genome organization is increasingly appreciated as a determinant of gene regulation in the TME — enhancer hijacking through TAD disruption can activate oncogenes or immune evasion genes, and 3D chromatin architecture influences how tumour cells respond to epigenetic therapies. Computational methods for integrating 3D genome data with single-cell transcriptomics and spatial data could reveal how TAD-level chromatin changes contribute to TME heterogeneity. The benchmarking framework in TADShop could also serve as a template for systematic comparisons of other genomics methods relevant to TME analysis.