Research Radar — 2026-06-07
Methods & AI
Computational
Structural motif search across the protein universe with Folddisco
Nature Biotechnology Published 2026-06-05 research article DOI: 10.1038/s41587-026-03162-9
structural bioinformatics protein structure motif search Foldseek embedding structure database protein function computational biology
Summary: Introduces Folddisco, a computational method and tool that enables structural motif search across the entire protein universe — millions of experimentally determined and predicted protein structures — at interactive speeds. The ability to search for specific three-dimensional structural motifs across all known protein structures has been a holy grail of structural bioinformatics: it would allow researchers to discover proteins that share a functional site (e.g., a catalytic triad, a ligand-binding pocket, a protein-protein interaction interface) even when those proteins share no detectable sequence similarity. Previous approaches were either too slow to scale to the rapidly growing protein structure databases (now containing hundreds of millions of structures thanks to AlphaFold and related methods) or too insensitive to detect subtle structural similarities. Folddisco addresses both challenges by building on the Foldseek framework — which uses a structural alphabet to encode 3D protein structures as sequences over a 20-letter alphabet, enabling ultra-fast structural comparisons using sequence-based search algorithms. Folddisco extends this with a motif-centric search strategy: instead of comparing whole structures, it extracts local structural motifs (defined by the user as a set of residues with specified 3D coordinates), encodes them using the structural alphabet, and searches for similar motifs across the database using an embedding-based acceleration strategy. The method achieves order-of-magnitude speedups over previous motif search tools while maintaining sensitivity to detect even subtle structural similarities. Folddisco is demonstrated on several challenging use cases: identifying remote homologs of enzymatic active sites across divergent protein families, discovering unexpected structural similarities between viral and host proteins that may indicate molecular mimicry, and systematically cataloguing the structural diversity of known functional motifs across the tree of life.
Why it matters: The ability to search for structural motifs at proteome scale fundamentally changes how researchers can explore protein function. Current approaches to annotating protein function rely heavily on sequence similarity — if a new protein looks like a known enzyme, it is assumed to have similar function. But many functionally analogous proteins have diverged beyond sequence recognition while retaining similar 3D structures at their active sites. Folddisco enables discovery of these "remote structural homologs," potentially revealing new biology: moonlighting functions of known proteins, evolutionary relationships obscured by sequence divergence, and structural mechanisms conserved across distant branches of life. The tool is particularly timely given the explosion of predicted protein structures from AlphaFold, ESMFold, and related methods — these databases contain enormous untapped biological information that Folddisco can help mine.
Why for Yiru: Structural motif search has direct applications to TME biology. Many protein-protein interactions that shape the TME — checkpoint receptor-ligand pairs (PD-1/PD-L1, CTLA-4/B7), chemokine-receptor interactions, integrin-ECM binding — are mediated by specific structural motifs. Folddisco could be used to systematically search for TME-relevant structural motifs across the human proteome, potentially identifying novel proteins that engage these interfaces and modulate immune-tumour interactions. The viral-host molecular mimicry application is also relevant: oncolytic viruses and tumour-associated viral proteins may contain structural motifs that mimic host immune regulatory proteins, and Folddisco could help identify these. More broadly, as structural biology enters the predicted-structure era, computational tools for mining structure-function relationships at scale will be essential for translating the structural data deluge into biological insight — exactly the type of tools a computational TME lab needs to master.
Explicit dynamic cross-strand interactions for DNA sequence language modelling
Nature Machine Intelligence Published 2026-06-04 research article DOI: 10.1038/s42256-026-01249-1
DNA language model genomic sequence deep learning cross-strand double helix transformer regulatory genomics epigenomics
Summary: Introduces a DNA sequence language model that explicitly models cross-strand interactions — the base-pairing relationships between the Watson and Crick strands of the DNA double helix — rather than treating DNA as a single-stranded linear sequence. Most genomic language models (e.g., DNABERT, Enformer, Nucleotide Transformer) treat DNA as a linear sequence of A, C, G, T, ignoring the fundamental double-stranded nature of DNA. This simplification discards important biological information: the two strands are complementary, and many regulatory processes (transcription factor binding, DNA methylation, chromatin organization) depend on the double-stranded context. The authors introduce cross-strand attention mechanisms that allow the model to learn dependencies between complementary positions on opposite strands, effectively modeling the "base-pairing grammar" of DNA alongside the sequential grammar. The model, trained on human and mouse genomes, outperforms single-strand models on a range of genomic prediction tasks including transcription factor binding site prediction, chromatin accessibility prediction, and DNA methylation state inference. The cross-strand architecture also enables the model to naturally handle reverse-complement sequences and learn strand-symmetric representations — properties that are biologically motivated but challenging for single-strand models. Ablation studies confirm that the cross-strand attention is the primary driver of improved performance, not simply increased model capacity.
Why it matters: Genomic language models are rapidly becoming essential tools in regulatory genomics, enabling prediction of variant effects, annotation of non-coding regions, and prioritization of disease-associated variants. However, current models operate on a simplified representation of DNA that discards the double-stranded structure — a fundamental physical property of the molecule. This work demonstrates that explicitly modeling the double helix improves predictive performance, suggesting that future genomic models should incorporate this inductive bias. The cross-strand attention mechanism is architecturally simple and could be incorporated into other genomic model architectures. More broadly, this work highlights a general principle: incorporating known biophysical properties of biomolecules into ML architectures (rather than relying solely on data-driven learning) can improve performance and biological plausibility.
Why for Yiru: Genomic language models are directly applicable to TME research for predicting the regulatory effects of non-coding variants. Germline and somatic variants in regulatory regions can influence TME composition by altering the expression of cytokines, chemokines, checkpoint molecules, and other immune modulators in tumour or stromal cells. A DNA language model that better captures the double-stranded context of regulatory elements could improve variant effect prediction for TME-relevant loci identified through GWAS or QTL studies. The cross-strand architecture may be particularly valuable for predicting the effects of variants in palindromic transcription factor binding sites — which are common in immune regulatory regions — where the double-stranded context determines binding specificity. More broadly, the principle of incorporating molecular structural priors into deep learning models applies to many problems in computational TME biology, from predicting TCR-pMHC binding (where both chains contribute) to modeling chromatin loop anchors (where DNA geometry matters).
StPedf: Cell trajectory inference of spatial transcriptomics via spatial proximity embedding and spatial density-adaptive fusion
PLOS Computational Biology Published 2026-06-05 research article DOI:
spatial transcriptomics trajectory inference cell differentiation spatial proximity embedding single-cell computational method tissue architecture
Summary: Introduces StPedf, a computational method for inferring cell differentiation and state transition trajectories directly from spatial transcriptomics data by integrating spatial proximity information with gene expression similarity through a spatial density-adaptive fusion strategy. Cell trajectory inference (also called pseudotime analysis or lineage tracing) is a cornerstone of single-cell genomics — it reconstructs the continuous developmental or disease progression paths that cells traverse from gene expression snapshots. However, most trajectory inference methods (Monocle, Slingshot, PAGA, RNA velocity) were designed for dissociated single-cell RNA-seq data and ignore spatial information — they cannot distinguish between two cells that are transcriptionally similar but located in different tissue compartments undergoing different biological processes. StPedf addresses this by jointly modeling transcriptional similarity and spatial proximity: cells that are both transcriptionally similar and physically adjacent receive higher trajectory connectivity weights than cells that are transcriptionally similar but spatially distant. The spatial density-adaptive fusion strategy adjusts the relative weight of spatial vs. transcriptional information based on local cell density — in dense tissue regions where many cells are in close proximity, spatial information is weighted more heavily because physical adjacency more reliably indicates lineage relationships; in sparse regions, transcriptional similarity dominates. StPedf is validated on several spatial transcriptomics datasets (MERFISH, Visium, Slide-seq) from developing tissues and tumour samples, demonstrating that incorporating spatial information reveals trajectories that are biologically more coherent and recovers known spatial gradients of differentiation that are missed by non-spatial methods.
Why it matters: The rapid adoption of spatial transcriptomics technologies (10x Visium, MERFISH, Xenium, CosMx, Stereo-seq) means that most new single-cell datasets now include spatial coordinates. However, computational methods have lagged behind — many of the most popular analysis tools were designed for non-spatial data and cannot leverage the spatial dimension. StPedf represents an important step toward spatially-aware trajectory inference, and its density-adaptive strategy is a principled approach to a fundamental challenge: spatial proximity is more informative in some tissue regions than others. As spatial transcriptomics becomes the default modality for tissue profiling, methods like StPedf that natively integrate spatial information will become essential for extracting biological insight from these complex datasets.
Why for Yiru: Cell trajectory inference in the spatial context is directly relevant to TME biology. Tumour tissue contains multiple cell types undergoing coordinated transitions: T cells differentiating from naive to effector to exhausted states along spatial gradients of antigen and cytokine exposure; macrophages polarizing from M1-like to M2-like states along oxygen and metabolite gradients; cancer cells undergoing epithelial-to-mesenchymal transition at the invasive front. StPedf could be applied to spatial transcriptomics data from tumour sections to reconstruct these trajectories in their native spatial context, revealing which spatial niches drive which cell state transitions. This information could guide the rational design of combination therapies — for example, identifying the specific spatial location where T cells begin to express exhaustion markers would inform where and when to administer checkpoint inhibitors. The density-adaptive fusion strategy is also relevant to tumour tissue, which is highly heterogeneous in cell density (dense tumour nests vs. sparse stromal regions).
A multiscale, Bayesian inference approach to augment mechanistic models of cell signaling with machine-learning predictions of binding affinity
PLOS Computational Biology Published 2026-06-05 research article DOI:
Bayesian inference multiscale modeling cell signaling machine learning binding affinity mechanistic model systems biology parameter estimation
Summary: Presents a multiscale Bayesian inference framework that integrates machine-learning predictions of protein-protein binding affinities with mechanistic ordinary differential equation (ODE) models of cell signaling pathways, enabling principled propagation of uncertainty from molecular-scale predictions to cellular-scale signaling predictions. A fundamental challenge in systems biology is that mechanistic signaling models — which describe how extracellular signals are transduced through networks of protein interactions to produce cellular responses — require quantitative parameters (binding affinities, catalytic rates, degradation rates) that are difficult to measure experimentally for every interaction in the network. Machine learning methods can now predict these parameters from sequence and structural features, but ML predictions come with uncertainty that is rarely propagated to downstream model predictions. The authors address this gap with a Bayesian multiscale framework: at the molecular scale, a machine learning model (trained on experimentally measured binding affinities) predicts the affinity for each protein-protein interaction in the signaling network, outputting not just a point estimate but a posterior distribution. At the cellular scale, these posterior distributions serve as priors for a Bayesian inference procedure that calibrates the ODE model to cellular-scale experimental data (e.g., phospho-protein time courses, dose-response curves). This two-level Bayesian approach naturally propagates molecular-scale uncertainty to cellular-scale predictions and identifies which molecular parameters are most critical for accurate cellular predictions — guiding future experimental efforts. The framework is demonstrated on the MAPK/ERK signaling pathway, showing that incorporating ML-predicted binding affinities with quantified uncertainty improves model calibration and predictive accuracy compared to using literature-derived point estimates.
Why it matters: The integration of machine learning with mechanistic modeling is a grand challenge in computational biology. Mechanistic models provide interpretable, causal understanding of biological systems but are limited by parameter uncertainty. Machine learning models excel at prediction from data but lack mechanistic interpretability. This Bayesian multiscale framework bridges the two paradigms in a principled way, enabling mechanistic models to leverage the predictive power of ML while maintaining interpretability and uncertainty quantification. The approach is general — it can be applied to any signaling pathway and any ML predictor of biochemical parameters — and becomes increasingly valuable as ML methods for predicting molecular properties (binding affinities, enzyme kinetics, protein stability) continue to improve.
Why for Yiru: Cell signaling models are central to understanding how the TME processes extracellular signals. The TME is a signaling-rich environment: cytokines, chemokines, growth factors, and checkpoint ligands engage receptors on tumour, immune, and stromal cells, triggering signaling cascades that determine cell fate decisions (proliferation, apoptosis, differentiation, migration). Mechanistic models of these signaling networks could predict how perturbations — therapeutic antibodies, small molecule inhibitors, cytokine therapies — will alter TME cell states. However, building such models requires quantitative parameters for hundreds of protein interactions, many of which are poorly characterized. The Bayesian multiscale framework provides a solution: use ML to predict missing parameters from sequence/structure, then calibrate the model to available TME data (phospho-proteomics from tumour samples, cytokine response data from in vitro assays). The uncertainty propagation is particularly valuable for TME applications, where therapeutic decisions based on model predictions require confidence estimates. The identification of which parameters most influence predictions could also guide experimental design — which binding affinities should be measured to maximally improve TME signaling models?
Heuristic multi-site optimization for protein sequence design using Masked Protein Language Models
PLOS Computational Biology Published 2026-06-05 research article DOI:
protein design masked language model sequence optimization heuristic search protein engineering deep learning fitness landscape computational biology
Summary: Introduces a heuristic multi-site optimization strategy for computational protein sequence design that leverages masked protein language models (MPLMs) — such as ESM-2 and ProtBERT — to propose mutations that simultaneously optimize multiple positions in a protein sequence while preserving structural integrity and function. Computational protein design — the task of proposing amino acid sequences that fold into a desired structure and perform a desired function — has traditionally relied on physics-based energy functions (Rosetta, FoldX) that are computationally expensive and limited in their ability to capture complex sequence-structure relationships. Masked protein language models, trained on hundreds of millions of natural protein sequences, have learned rich representations of the "grammar" of protein sequences — which amino acids are compatible at which positions given the structural and functional context. The authors leverage this capability for protein design: given a starting protein sequence and a set of positions to mutate, the MPLM predicts the probability distribution over all 20 amino acids at each masked position, conditioned on the rest of the sequence. The key innovation is a heuristic search strategy that efficiently explores the combinatorial space of multi-site mutations — naive enumeration of all combinations is impossible for more than a few positions, but the heuristic search uses the MPLM's per-position probability rankings and pairwise coupling terms to identify high-probability multi-site mutants without exhaustive enumeration. The method is demonstrated on several protein engineering tasks: improving thermostability while maintaining catalytic activity, designing enzyme variants with altered substrate specificity, and optimizing antibody complementarity-determining region (CDR) sequences for enhanced binding.
Why it matters: Protein language models have revolutionized protein structure prediction (AlphaFold) and representation learning, but their application to protein design — the forward problem of creating new proteins with desired properties — is still in its early stages. This work demonstrates that MPLMs, which have been trained primarily as representation learners, contain sufficient information about sequence-structure-function relationships to guide protein design, and that heuristic search strategies can efficiently navigate the combinatorial space of multi-site mutations. This approach is computationally efficient (no physics-based simulations required) and can be applied to any protein for which a sequence is available, making it accessible to experimental labs without extensive computational resources. As MPLMs continue to improve with larger training sets and better architectures, ML-guided protein design will become increasingly powerful.
Why for Yiru: Protein design has direct applications to TME-targeted therapeutics. Engineered cytokines with altered receptor selectivity could preferentially activate anti-tumour immune cells while avoiding systemic toxicity. Designed antibody variants with optimized CDR sequences could achieve higher affinity for TME-specific antigens or better tumour penetration. Soluble receptor decoys could be designed to sequester immunosuppressive cytokines in the TME. The MPLM-based design approach described here could be applied to any of these problems — for example, using ESM-2 to propose mutations in IL-2 that bias binding toward the IL-2Rβγ receptor on effector T cells rather than the IL-2Rαβγ receptor on regulatory T cells. The heuristic multi-site search is particularly relevant because many TME protein engineering problems require simultaneous optimization of multiple properties (binding affinity, specificity, stability, expression level), which necessarily involves mutations at multiple positions.
Biomedical discoveries
Biomedicine
Acquired genetic and cell-state changes in IDH-mutant glioma progression
Nature Published 2026-06-03 research article DOI: 10.1038/s41586-026-10612-6
glioma IDH mutation tumour evolution single-cell epigenomics tumour microenvironment longitudinal cancer genomics
Summary: Presents a comprehensive longitudinal multi-omics analysis of IDH-mutant gliomas — the most common primary brain tumours in adults — tracing the genetic, epigenetic, and microenvironmental changes that accompany tumour progression from low-grade to high-grade disease. IDH-mutant gliomas initially present as slow-growing, lower-grade tumours (grade 2-3) but inevitably progress to aggressive, treatment-resistant high-grade tumours (grade 4) over years. Understanding the mechanisms driving this progression has been challenging because it requires longitudinal sampling — comparing the same patient's tumour at early and late stages — which is rarely feasible in clinical settings. The authors assembled a unique cohort of paired initial and recurrent IDH-mutant glioma samples from over 200 patients and performed integrated single-cell RNA-seq, single-cell ATAC-seq (chromatin accessibility), and whole-genome sequencing on each sample. Three major findings emerge. First, tumour progression is accompanied by the acquisition of specific genetic alterations — including CDKN2A deletion, PDGFRA amplification, and MYC activation — but these genetic events alone are insufficient to explain the full transcriptional and phenotypic changes observed. Second, epigenetic remodeling — particularly loss of the glioma-CpG island methylator phenotype (G-CIMP) and gain of enhancer activity at cell-cycle and mesenchymal genes — is a parallel and partly independent driver of progression, with chromatin changes preceding and enabling subsequent genetic alterations. Third, and most notably, the tumour microenvironment undergoes coordinated changes during progression: immune infiltration shifts from a T-cell-inflamed state in lower-grade tumours to a macrophage-dominated, immunosuppressive state in high-grade tumours, and this TME remodeling correlates with worse outcomes independent of tumour-cell-intrinsic features. The study provides an interactive resource enabling exploration of the longitudinal genomic, epigenomic, and microenvironmental changes across the IDH-mutant glioma progression spectrum.
Why it matters: This study provides the most comprehensive view to date of how IDH-mutant gliomas evolve during progression, with three implications that extend beyond glioma. First, the finding that epigenetic remodeling can precede and enable genetic alterations suggests a "epigenetic priming" model of cancer evolution that may apply to other tumour types — epigenetic changes create permissive chromatin states that make subsequent genetic alterations more likely or more consequential. Second, the coordinated TME remodeling during progression — with a shift from T-cell-inflamed to macrophage-dominated immunosuppression — identifies a therapeutic window: early-stage tumours may be responsive to checkpoint immunotherapy (due to T cell infiltration), while late-stage tumours may require macrophage-targeted therapies. Third, the parallel yet partly independent contributions of genetic, epigenetic, and TME changes to progression argue against purely genetics-centric models of tumour evolution and support a more holistic view that integrates all three axes.
Why for Yiru: This study is a masterclass in how to integrate multiple data modalities to understand tumour evolution — and it provides a template directly applicable to studying TME remodeling during progression in other cancers. The finding that TME immune composition shifts systematically during glioma progression is consistent with emerging evidence in other solid tumours, and the longitudinal design provides causal evidence (not just cross-sectional correlation) that TME changes accompany — and potentially contribute to — malignant progression. The computational frameworks used — integration of scRNA-seq, scATAC-seq, and WGS from the same samples, longitudinal differential expression, TME deconvolution — are directly transferable to TME studies in other tumour types. The finding that TME features independently predict outcomes suggests that TME-based biomarkers should be incorporated into prognostic models alongside tumour-cell-intrinsic features. The "epigenetic priming" concept is also relevant to TME biology — do epigenetic changes in stromal or immune cells within the TME create permissive states for subsequent malignant progression?
Human microglial transitions at the Aβ–tau inflection point associate with divergent pathways to dementia and resilience
Nature Medicine Published 2026-06-04 research article DOI: 10.1038/s41591-026-04393-8
microglia Alzheimer's disease amyloid-beta tau neurodegeneration single-nucleus RNA-seq resilience dementia
Summary: Uses single-nucleus RNA sequencing of post-mortem human brain tissue to characterize the transcriptional states of microglia — the brain's resident immune cells — at the critical Aβ–tau inflection point, revealing that microglial responses at this juncture diverge into pathways associated with either progression to Alzheimer's dementia or cognitive resilience despite pathology. Alzheimer's disease (AD) is defined neuropathologically by the accumulation of amyloid-β (Aβ) plaques and tau neurofibrillary tangles, but the relationship between pathology and clinical symptoms is imperfect: some individuals maintain normal cognition despite substantial AD pathology ("resilient" individuals), while others develop dementia with relatively modest pathology. Microglia have been implicated in both protective and pathogenic roles in AD — they can clear Aβ and provide trophic support to neurons, but they can also drive neuroinflammation and synaptic pruning — and understanding what tips the balance between these opposing functions has been a central question. The authors profiled microglia from over 100 human donors spanning the full spectrum from no pathology to advanced AD, with a particular focus on the Aβ–tau "inflection point" — the stage at which tau pathology begins to spread beyond the medial temporal lobe, which clinically marks the transition from preclinical to prodromal AD. At this inflection point, microglia in individuals who later developed dementia adopted an "activated response" state characterized by upregulation of APOE, TREM2, and complement pathway genes — a state associated with increased phagocytic activity but also with inflammatory cytokine production and synaptic pruning. In contrast, microglia from resilient individuals — those with comparable Aβ and tau pathology but preserved cognition — adopted a "homeostatic maintenance" state characterized by expression of trophic factors, lipid metabolism genes, and anti-inflammatory mediators. This divergence was not simply a consequence of different pathology levels — the two microglial states were observed in individuals matched for Aβ and tau burden.
Why it matters: This study provides the most granular view to date of human microglial states at the critical juncture where AD pathology transitions from clinically silent to symptomatic. The identification of two divergent microglial response programs — one associated with neurodegeneration, one with resilience — has immediate therapeutic implications: rather than globally suppressing or activating microglia (which risks losing protective functions), therapies could aim to shift the balance from the activated/neurotoxic state toward the homeostatic/resilience state. The specific genes and pathways identified in each state provide concrete therapeutic targets. More broadly, this study exemplifies the power of single-cell genomics in human tissue to disentangle the cellular basis of disease heterogeneity — a theme that extends to cancer, where similar questions about "why do some patients progress while others don't?" can be addressed with similar approaches.
Why for Yiru: Microglia are the brain's tissue-resident macrophages, and the principles governing their functional states — how they toggle between protective and pathogenic roles depending on microenvironmental context — are directly analogous to the macrophage polarization dynamics that shape the TME. In both contexts, the key question is: what microenvironmental signals (cytokines, metabolites, cell-cell contacts, pathological protein aggregates vs. tumour antigens) drive macrophages/microglia toward tissue-destructive vs. tissue-protective states? The single-nucleus approaches used here to profile human microglia in their native tissue context are methodologically similar to approaches used to profile tumour-associated macrophages in the TME. The finding that microglial state — not just pathology burden — predicts cognitive outcomes parallels evidence in cancer that TME immune composition predicts outcomes independent of tumour mutational burden. The specific pathways identified (APOE, TREM2, complement) are also relevant to TME macrophages, where TREM2+ lipid-associated macrophages have recently been identified as an immunosuppressive population.
Plasma signals of lung tumor promotion for molecular cancer prevention
Cell Published 2026-06-04 research article DOI: 10.1016/j.cell.2026.05.005
lung cancer plasma proteomics cancer prevention biomarker tumour promotion early detection molecular epidemiology proteomics
Summary: Identifies plasma proteomic signatures that distinguish the tumour promotion phase from the earlier initiation phase in lung carcinogenesis, providing blood-based biomarkers for molecular cancer prevention strategies. Cancer prevention — intervening before invasive cancer develops — is the most effective way to reduce cancer mortality, but it is limited by the inability to identify which individuals with pre-cancerous lesions will progress to invasive cancer and which will not. Lung cancer is particularly challenging: low-dose CT screening detects many indeterminate pulmonary nodules, but most are benign, leading to unnecessary invasive procedures and anxiety. The tumour promotion phase — the transition from initiated (mutated) cells to expanding pre-malignant clones — is considered the optimal window for preventive intervention, but there are no clinical biomarkers that specifically detect this phase. The authors address this gap using plasma proteomics in a unique cohort: longitudinal plasma samples from individuals enrolled in lung cancer screening programs, collected before and during the development of lung cancer, alongside matched controls who had benign nodules. Using SomaScan (aptamer-based) and Olink (antibody-based) proteomic platforms measuring ~7,000 proteins, they identify a plasma protein signature that distinguishes individuals in the tumour promotion phase from those with benign nodules and from healthy controls. The signature includes proteins involved in inflammation (IL-6, CRP), extracellular matrix remodeling (COL4A1, MMP9), and immune regulation (PD-L1, B7-H3) — reflecting the systemic response to early tumour-stromal interactions. Importantly, the signature is detectable before nodules become radiographically suspicious, providing a potential window for preventive intervention.
Why it matters: This study addresses one of the most pressing needs in cancer prevention: biomarkers that can distinguish dangerous pre-cancerous lesions from benign ones. Current lung cancer screening relies on CT imaging, which has high sensitivity but low specificity — the vast majority of detected nodules are benign. A blood-based test that could triage CT-detected nodules into high-risk (requiring immediate intervention) and low-risk (safe to monitor) categories would dramatically reduce unnecessary procedures and healthcare costs. Beyond screening, the identification of specific plasma proteins associated with tumour promotion provides mechanistic insight into the systemic response to early tumorigenesis and identifies potential targets for preventive therapies — for example, anti-inflammatory agents or immune modulators that could intercept the promotion phase. The multi-platform proteomics approach also establishes a methodology for systematic plasma biomarker discovery that could be applied to other cancer types.
Why for Yiru: The TME is not just a local phenomenon — tumours systemically remodel the host through secreted factors that enter the circulation. The plasma proteomic signals identified in this study reflect early TME-host interactions during lung tumour promotion, and the specific proteins involved (immune checkpoints, ECM remodeling enzymes, inflammatory cytokines) are all TME-relevant. The finding that TME-derived signals are detectable in blood before tumours become clinically apparent suggests that plasma proteomics could be used to monitor TME status non-invasively — a "liquid biopsy of the TME." Computationally, the integration of plasma proteomics with tissue-based TME profiling (e.g., from resected nodules or biopsies) could connect circulating signals to their tissue sources and reveal how the systemic TME signature evolves during tumorigenesis. The proteomic platforms used (SomaScan, Olink) generate high-dimensional data that require sophisticated computational analysis — differential expression, machine learning classification, pathway enrichment — all of which are transferable skills for TME computational biology.
Single-cell mapping of regulatory DNA-protein interactions
Cell Published 2026-06-04 research article DOI: 10.1016/j.cell.2026.05.014
single-cell transcription factor DNA-protein interaction regulatory genomics chromatin CUT&Tag multi-omics epigenomics
Summary: Introduces a single-cell method for mapping regulatory DNA-protein interactions genome-wide, enabling the simultaneous profiling of transcription factor binding and chromatin-associated protein landscapes in individual cells. Understanding how transcription factors (TFs) and other DNA-binding proteins regulate gene expression is fundamental to biology, but traditional methods for mapping DNA-protein interactions (ChIP-seq, CUT&RUN, CUT&Tag) require thousands to millions of cells and produce population-averaged profiles that obscure cell-to-cell heterogeneity. This is a critical limitation because TF binding is highly dynamic and context-dependent — the same TF may bind different genomic loci in different cell types, cell states, or even individual cells within a seemingly homogeneous population. The authors develop a single-cell adaptation of CUT&Tag (Cleavage Under Targets and Tagmentation) that uses combinatorial barcoding to profile TF binding or histone modifications in thousands of individual cells. The method achieves high sensitivity (detecting thousands of binding sites per cell) and can be multiplexed to profile multiple TFs or histone marks simultaneously. Applied to human immune cells and cancer cell lines, the method reveals substantial cell-to-cell heterogeneity in TF binding that is not explained by differences in chromatin accessibility alone — cells with similar chromatin landscapes can have markedly different TF occupancy patterns, suggesting that TF concentration, post-translational modification, and protein-protein interactions contribute to binding variability beyond what chromatin state predicts. In cancer cells, the method identifies rare subpopulations with distinct TF binding profiles that may represent drug-tolerant persister cells.
Why it matters: This method fills a major gap in the single-cell genomics toolkit. While we can now profile transcriptomes (scRNA-seq), chromatin accessibility (scATAC-seq), and histone modifications (scCUT&Tag) at single-cell resolution, direct profiling of TF binding at single-cell resolution has been largely out of reach. TFs are the "executive branch" of gene regulation — they integrate signaling inputs and directly control which genes are expressed — so being able to measure TF binding heterogeneity is essential for understanding how genetically identical cells adopt different fates and functions. The finding that TF binding heterogeneity exceeds chromatin accessibility heterogeneity is particularly important: it means that scATAC-seq alone cannot fully capture the regulatory heterogeneity of a cell population, and that direct TF profiling is needed for a complete picture.
Why for Yiru: TF binding heterogeneity is directly relevant to TME biology. The functional diversity of immune cells in the TME — why one CD8+ T cell becomes a highly cytotoxic effector while its neighbor becomes exhausted — is ultimately driven by differences in TF activity (e.g., T-bet, Eomes, TOX, TCF7). Single-cell TF binding maps in TME-infiltrating immune cells could reveal the regulatory logic underlying these divergent cell fates and identify the specific genomic loci where fate-determining TFs bind. In cancer cells, TF binding heterogeneity could explain drug-tolerant persister cell emergence — rare cells with distinct TF occupancy at stress-response and drug-efflux genes may survive initial therapy and seed relapse. Computationally, single-cell TF binding data require new analysis methods — peak calling at single-cell resolution, integration with scRNA-seq and scATAC-seq, and inference of TF cooperation and competition networks — presenting rich opportunities for computational TME research.
Replaying germinal center evolution on a quantified affinity landscape
Cell Published 2026-06-05 research article DOI: 10.1016/j.cell.2026.05.013
germinal center affinity maturation B cell antibody evolution immune repertoire single-cell immunology
Summary: Reconstructs the evolutionary dynamics of antibody affinity maturation in germinal centers (GCs) by combining high-throughput measurement of antibody-antigen binding affinities with single-cell lineage tracing, then uses these quantitative affinity landscapes to "replay" GC evolution in silico and identify the selective forces that shape antibody repertoires. Germinal centers are specialized microanatomical structures in lymph nodes where B cells undergo iterative rounds of mutation (somatic hypermutation) and selection to produce high-affinity antibodies — the basis of effective humoral immunity and the goal of most vaccines. Despite decades of study, fundamental questions remain: how strong is the selective advantage conferred by a given improvement in binding affinity? Do GCs optimize for maximal affinity or for other properties such as breadth or stability? And why do some GCs produce highly potent antibodies while others yield mediocre responses? The authors address these questions by developing a system to measure the binding affinity of thousands of individual B cell receptors (BCRs) from single GC B cells against their cognate antigen, then linking these affinity measurements to the cells' clonal lineage trees (reconstructed from BCR sequence data). This produces a quantitative "affinity landscape" — a mapping of how affinity changed along each branch of the evolutionary tree. By modeling the GC as an evolutionary process on this landscape, the authors can replay GC evolution under different selective regimes and determine which regimes produce outcomes matching the observed data. Key findings include: selection for affinity is strong but not absolute — B cells with modest affinity improvements can survive and continue to mutate; GCs balance affinity optimization with maintenance of clonal diversity; and the most successful GCs are those that achieve an optimal balance between exploration (generating diverse mutants) and exploitation (selectively expanding the best mutants).
Why it matters: This study provides the most quantitative view of GC affinity maturation to date, with implications for vaccine design. Current vaccines are designed empirically — immunize and measure the antibody response — but a predictive understanding of how GC dynamics produce high-affinity antibodies could enable rational vaccine design: which antigens, adjuvants, and boosting schedules will optimally drive GCs toward broadly neutralizing antibodies? The finding that GCs balance affinity optimization with diversity maintenance challenges the simplistic view that "stronger selection is always better" and suggests that vaccine strategies that maintain clonal diversity may ultimately produce better responses than those that aggressively select for the highest-affinity clones. The in silico GC replay framework provides a general tool for understanding evolutionary dynamics in other contexts where quantitative fitness landscapes can be measured.
Why for Yiru: Germinal center biology connects to the TME in several ways. Tertiary lymphoid structures (TLS) — ectopic lymph-node-like structures that form in tumours — contain GC-like B cell follicles, and the presence of TLS is associated with improved responses to immunotherapy. Understanding the evolutionary dynamics of affinity maturation in tumour-associated TLS could reveal whether these structures produce anti-tumour antibodies and how to therapeutically enhance this process. More broadly, the evolutionary framework developed here — measuring fitness landscapes from single-cell data and replaying evolution in silico — is directly applicable to understanding clonal evolution in the TME. Tumour cells, like B cells, undergo mutation and selection, and the quantitative fitness landscape concept could be applied to understand how specific mutations confer selective advantages in the TME context. The balance between exploration (generating diversity through mutation) and exploitation (selectively expanding fit clones) is a fundamental tension in both GC and tumour evolution.
Cross-disciplinary watchlist
Other Fields
Multiplexed, precise genome engineering in monocots with twin prime editing systems
Nature Biotechnology Published 2026-06-05 research article DOI: 10.1038/s41587-026-03174-5
prime editing genome engineering monocot plant biotechnology CRISPR multiplexed crop improvement gene editing
Summary: Extends prime editing — a "search-and-replace" genome editing technology — to enable simultaneous, precise editing at multiple genomic loci in monocot plant species including rice and wheat, overcoming technical barriers that previously limited prime editing to single-locus edits with low efficiency in these economically critical crops. Prime editing, developed in 2019, uses a Cas9 nickase fused to a reverse transcriptase, guided by a prime editing guide RNA (pegRNA) that specifies both the target site and the desired edit. It can introduce all types of point mutations, small insertions, and small deletions without requiring double-strand breaks or donor DNA templates, making it more precise and versatile than traditional CRISPR-Cas9 editing. However, prime editing efficiency has been low in plants, particularly monocots (grasses including rice, wheat, maize, which are the world's most important food crops), and multiplexed editing — making edits at multiple loci simultaneously — has been essentially impossible. The authors address both challenges. For efficiency, they develop "twin prime editing" — using two pegRNAs targeting opposite DNA strands at the same locus, with each pegRNA encoding a complementary edit — which substantially increases editing efficiency by providing two independent opportunities for the edit to be incorporated. For multiplexing, they develop a polycistronic tRNA-processing system that expresses multiple pegRNAs from a single transcript, which are then processed into individual functional pegRNAs by endogenous tRNA-processing enzymes. Combining these innovations, they demonstrate simultaneous precise editing at up to four loci in rice and wheat, targeting agronomically important genes involved in yield, disease resistance, and grain quality.
Why it matters: Precise, multiplexed genome editing in crop plants is a transformative capability for agriculture. Most agronomic traits (yield, drought tolerance, disease resistance) are polygenic — controlled by multiple genes — so achieving significant improvements requires editing multiple loci simultaneously. Traditional breeding takes years to decades to combine favorable alleles at multiple loci; CRISPR-based gene knockout is faster but cannot introduce the precise nucleotide changes often needed (e.g., mimicking natural beneficial alleles, introducing gain-of-function mutations). Twin prime editing provides the precision of prime editing with the efficiency needed for practical crop improvement. The demonstration in rice and wheat — staple crops that feed billions — is particularly significant. This technology could accelerate the development of climate-resilient, high-yielding crop varieties at a time when food security is threatened by climate change.
Why for Yiru: While plant genome editing may seem distant from TME research, the technological innovations — particularly the twin pegRNA strategy and polycistronic pegRNA expression — are generalizable to mammalian systems. Multiplexed prime editing in human cells faces similar efficiency challenges to those addressed here, and the twin prime editing approach could be adapted to improve editing efficiency in T cells, macrophages, or other immune cells for TME applications. Multiplexed editing is particularly relevant to engineering cell therapies — an ideal CAR-T cell might require simultaneous edits at multiple loci (TRAC knockout to eliminate endogenous TCR, PD-1 knockout to prevent exhaustion, insertion of a CAR transgene, and a safety switch). The computational challenge of designing multiple pegRNAs that work together efficiently is non-trivial and presents opportunities for optimization algorithms. More broadly, keeping abreast of advances in genome editing is essential because these tools increasingly enable the precise genetic modifications needed to study and therapeutically manipulate the TME.
Xenophagocytosis blockade enhances interspecies chimerism
Cell Published 2026-06-05 research article DOI: 10.1016/j.cell.2026.05.016
xenotransplantation chimerism phagocytosis stem cell immunology regenerative medicine cross-species transplantation
Summary: Identifies xenophagocytosis — the engulfment and elimination of foreign (xenogeneic) cells by host phagocytes — as a critical barrier to interspecies chimerism, and demonstrates that blocking this process substantially enhances the contribution of donor cells to chimeric tissues. Interspecies chimerism — generating animals that contain cells from two different species — is a powerful approach with two major applications: (1) growing human organs in animal hosts for transplantation, addressing the severe shortage of donor organs; and (2) creating more accurate animal models of human disease by "humanizing" specific tissues. However, generating stable interspecies chimeras has proven extremely difficult — donor cells (e.g., human pluripotent stem cells injected into animal embryos) typically contribute poorly to host tissues and are rapidly eliminated. The mechanisms underlying this rejection have been poorly understood. The authors discover that host macrophages and other phagocytic cells recognize xenogeneic cells through a "don't-eat-me" signal mismatch — the donor cells lack the species-matched CD47 and other surface proteins that normally protect cells from phagocytosis. By engineering donor cells to express host-matched CD47, or by pharmacologically blocking the phagocytic receptors (SIRPα) on host macrophages, xenophagocytosis is suppressed and donor cell contribution to chimeric tissues is substantially enhanced across multiple organ systems. This finding is demonstrated across several interspecies combinations including human-into-pig and human-into-mouse.
Why it matters: This study identifies a specific, druggable mechanism underlying the long-standing barrier to interspecies chimerism. The potential to grow human organs in animal hosts — so-called "organ farming" — could solve the organ shortage crisis, but this vision has been stalled by the inability to achieve high levels of human cell chimerism in host animals. The identification of CD47-SIRPα mismatch as a key mechanism — and the demonstration that it can be overcome — represents a major step forward. More broadly, the findings highlight the importance of innate immune barriers (phagocytosis, not just adaptive immunity) in xenotransplantation and chimerism, which have historically been underappreciated compared to T-cell and antibody-mediated rejection. The approach of engineering donor cells to express host-matched immunomodulatory proteins could be generalized to other innate immune checkpoints.
Why for Yiru: The CD47-SIRPα axis is well-established as a key immune checkpoint in the TME — tumour cells upregulate CD47 to avoid being eaten by tumour-associated macrophages, and anti-CD47 therapies are in clinical development. This study extends the importance of this axis to interspecies cell recognition, revealing that species-matching of CD47 is critical for immune evasion. The connection suggests that tumour cells' upregulation of CD47 may represent a co-option of a fundamental self-recognition mechanism that normally operates at the species level. For TME research, this highlights CD47-SIRPα as an even more central immune checkpoint than previously appreciated. The methods for engineering cells to express specific immunomodulatory proteins are directly transferable to engineering TME-targeted cell therapies. More broadly, the humanized animal models enabled by enhanced chimerism could eventually include humanized immune systems in mice, providing better models for studying human TME biology and testing immunotherapies.
Brightness demixing for simultaneous multi-target imaging in 3D single-molecule localization microscopy
Nature Methods Published 2026-06-05 research article DOI: 10.1038/s41592-026-03118-6
super-resolution microscopy single-molecule localization multi-target imaging SMLM optics fluorescence imaging biophysics
Summary: Introduces a computational method called brightness demixing that enables simultaneous imaging of multiple distinct molecular targets in three-dimensional single-molecule localization microscopy (SMLM) by using the intrinsic brightness (photon count) of individual fluorophore blinking events to distinguish different target species, eliminating the need for multiple spectrally distinct fluorophores or sequential imaging rounds. SMLM techniques (PALM, STORM, DNA-PAINT) achieve ~10-20 nm resolution by localizing individual fluorescent molecules with high precision, but they have a fundamental multiplexing limitation: to image multiple protein species, each must be labeled with a spectrally distinct fluorophore, but the number of usable spectral channels is limited to 3-4 by the available fluorescent dyes and optical filters. Sequential imaging (staining, imaging, stripping, re-staining) can extend multiplexing but multiplies imaging time and risks tissue damage. Brightness demixing circumvents these limitations by exploiting a previously underutilized dimension of the SMLM signal: the number of photons detected per localization event. Different fluorophores, or the same fluorophore in different local environments, produce localization events with different characteristic brightnesses. By fitting a mixture model to the brightness distribution of all localization events in a field of view, the authors can computationally assign each event to a specific target species with high accuracy. This allows 6-8 targets to be imaged simultaneously using just 1-2 spectral channels, effectively doubling or tripling the multiplexing capacity of SMLM. The method is demonstrated on cellular structures including the nuclear pore complex, focal adhesions, and the endocytic machinery, revealing spatial relationships between protein components that were previously inaccessible in a single experiment.
Why it matters: This computational approach elegantly expands the multiplexing capacity of SMLM without requiring any new hardware or fluorophore chemistry — it works with standard SMLM setups and commercially available dyes. This democratizes high-plex super-resolution imaging, making it accessible to labs that cannot invest in specialized multi-channel SMLM systems. The brightness dimension has been underutilized in SMLM analysis, and this work demonstrates that it contains substantial information that can be exploited computationally. The concept of "demixing" mixed signals using per-event statistics could inspire similar approaches in other single-molecule techniques (single-molecule FRET, nanopore sequencing, single-particle tracking) where each event carries multiple measurable parameters.
Why for Yiru: Super-resolution imaging of the TME is a powerful approach for understanding the nanoscale spatial organization of immune synapses, checkpoint receptor clustering, and ECM architecture. Current SMLM studies of the TME are limited to 2-4 targets per experiment, which is insufficient to capture the complexity of multi-protein complexes like the immunological synapse (which involves dozens of proteins organized in concentric supramolecular activation clusters). Brightness demixing could enable, for example, simultaneous imaging of 6-8 proteins at the T cell-tumour cell interface — capturing TCR, CD3, CD8, PD-1, PD-L1, and signaling adaptors in a single experiment — to reveal how checkpoint blockade reorganizes the synapse architecture. The computational method is directly applicable to SMLM data that TME researchers already collect. More broadly, the principle of extracting additional information dimensions from existing data through computational analysis (rather than new hardware) is a recurring theme in quantitative biology that computational TME researchers should embrace.
Cell differentiation can underpin the reproducibility of morphogenesis
PLOS Computational Biology Published 2026-06-04 research article DOI:
morphogenesis cell differentiation computational modeling reproducibility developmental biology pattern formation robustness tissue development
Summary: Uses computational modeling to demonstrate that cell differentiation — the process by which cells adopt specialized fates during development — can actively enhance the reproducibility and robustness of tissue morphogenesis, rather than being merely a downstream consequence of pattern formation. Morphogenesis is the biological process that generates the shape and structure of tissues and organs, and a central question in developmental biology is how this process achieves its remarkable reproducibility — embryos from the same species develop nearly identical anatomical structures despite environmental variation, stochastic gene expression, and imprecise cellular behaviors. The dominant paradigm holds that morphogenesis is orchestrated by morphogen gradients — diffusible signaling molecules that form spatial concentration gradients and instruct cells to adopt specific fates at specific positions — with cell differentiation being a passive readout of this positional information. The authors challenge this view with a computational model that couples cell differentiation dynamics to tissue mechanics: as cells differentiate, they change their mechanical properties (adhesion, stiffness, contractility), and these changes feed back to influence tissue shape and patterning. The model shows that this feedback can actively correct errors in morphogen interpretation — if a cell misreads its position and begins to differentiate incorrectly, the resulting mechanical mismatch with its neighbors generates forces that either correct the cell's fate or physically relocate it to a position matching its adopted fate. This differentiation-mechanics feedback makes morphogenesis more robust to noise in both morphogen signaling and mechanical perturbations, providing an explanation for the robustness of development that purely morphogen-based models cannot account for.
Why it matters: This study proposes a paradigm shift in how we think about the relationship between cell differentiation and morphogenesis. Instead of differentiation being downstream of pattern formation, it may be an active participant in ensuring that pattern formation is reproducible. The concept of differentiation-mechanics feedback has implications beyond embryonic development: it may explain how tissues maintain homeostasis in adults (cells that differentiate inappropriately are mechanically eliminated), how wound healing is coordinated (differentiating cells at the wound edge generate forces that guide closure), and how tumours disrupt tissue architecture (cancer cells' aberrant differentiation states generate mechanical conflicts with normal tissue). The computational framework is general and can be adapted to model morphogenesis in any tissue where cell differentiation and mechanics are coupled.
Why for Yiru: The concept of differentiation-mechanics feedback is directly relevant to TME biology. Tumour tissue is characterized by aberrant differentiation states — cancer cells dedifferentiate or transdifferentiate, immune cells adopt exhaustion or suppressive states, fibroblasts activate to a contractile phenotype — and these differentiation changes are accompanied by mechanical changes (altered adhesion, stiffness, contractility) that reshape the TME. The computational framework developed here could be adapted to model how TME mechanical remodeling feeds back to influence cell states — for example, how increasing matrix stiffness drives fibroblast activation, which further increases stiffness, creating a feed-forward loop that promotes tumour progression. The robustness concepts are also relevant: TMEs are highly heterogeneous, yet tumours reliably progress — understanding the feedback mechanisms that make tumour development robust to perturbations could reveal vulnerabilities. More broadly, the coupling of differentiation and mechanics is a largely unexplored dimension of TME biology that this computational framework helps to formalize.
Generated at 2026-06-07 10:00 +0800 by DeepSeek-V4-Pro via Clawdie for Yiru’s Research Radar