Research Radar — 2026-06-21

Summary: Introduces Morpho-FM, a weakly supervised framework that predicts spatial gene expression from routine H&E whole-slide images by conditioning a pretrained single-cell transcriptomic foundation-model prior on local histological neighbourhoods. Routine H&E histology captures tissue architecture at clinical scale but lacks a direct molecular readout of the transcriptional programmes that organise tumour epithelium, stroma, vasculature, and immune compartments. Spatial transcriptomics provides this context, yet cost, workflow complexity, and sparse sampling limit routine use. Most existing histology-to-expression models are trained de novo on small paired cohorts and remain weakly constrained when extrapolating from sparse measurements to dense, tissue-wide molecular maps. Morpho-FM addresses this by using a lightweight morphology-to-transcriptome adapter that maps cached whole-slide histology features into a transcriptomic decoder, enabling prediction at measured locations, dense full-section reconstruction, and re-aggregation to the original measurement support. Across harmonized prostate cancer benchmarks, Morpho-FM achieved the strongest overall performance among five representative methods (mean per-gene Pearson correlations of 0.286-0.298). Controlled ablation analyses identified pretrained transcriptomic initialization as a reproducible source of performance gain. Beyond accuracy benchmarks, Morpho-FM recovered ERBB2-enriched tumour compartments, boundary-associated molecular gradients, and annotation-aligned tissue domains across breast cancer datasets. The framework demonstrates that transcriptomic foundation-model priors serve as an effective constraint for morphology-conditioned molecular decoding.

Why it matters: The ability to infer spatial gene expression from routine H&E slides would dramatically expand the reach of spatial transcriptomics. H&E staining is performed on virtually every tumour biopsy worldwide, yet the molecular information encoded in tissue morphology remains inaccessible without expensive spatial transcriptomics assays. Morpho-FM demonstrates that foundation models pretrained on massive single-cell transcriptomic datasets can serve as powerful priors for this inference task — effectively allowing the model to "hallucinate" plausible molecular profiles that are consistent with both the observed histology and the biological constraints learned from millions of cells. If validated prospectively, this approach could enable retrospective spatial transcriptomic analysis of archival tissue collections and provide molecular context for routine pathology workflows at minimal additional cost.

Why for Yiru: This work is directly aligned with TME spatial omics research. The ability to predict spatial gene expression from H&E images opens the door to large-scale, low-cost spatial analysis of tumour-immune interactions — precisely the kind of tool needed to study how immune infiltration patterns relate to clinical outcomes across thousands of archival samples. The foundation-model approach is particularly relevant: by conditioning on pretrained transcriptomic representations, Morpho-FM can potentially capture immune cell type signatures and cell-cell interaction programmes that are encoded in tissue morphology. The framework could be extended to predict spatial distributions of immune checkpoint expression, cytokine gradients, or T cell receptor repertoire features from H&E alone. Methodologically, the use of weak supervision and transfer learning from foundation models is a design pattern worth adopting for other TME prediction tasks where paired training data are scarce.

Summary: Presents smDeepFLUOR, a deep learning framework for classifying single-molecule fluorescence signals with high accuracy and throughput. Single-molecule fluorescence microscopy has become an indispensable tool for studying biomolecular dynamics, interactions, and conformations at the nanoscale. However, analysing the resulting data — distinguishing genuine single-molecule signals from background noise, classifying different molecular states, and quantifying transition kinetics — remains a labour-intensive bottleneck requiring expert manual curation. smDeepFLUOR addresses this by training deep neural networks on large datasets of labelled single-molecule fluorescence traces, enabling automated, high-throughput classification of fluorescence time series into distinct molecular states. The method achieves near-human-level accuracy while processing thousands of traces in seconds — a task that would take days of manual analysis. The framework is validated across multiple fluorophore types and experimental conditions, demonstrating robustness to variations in signal-to-noise ratio and photobleaching kinetics.

Why it matters: Single-molecule biophysics experiments generate vast amounts of data that are currently bottlenecked by manual analysis. Automating this pipeline with deep learning not only accelerates discovery but also reduces analyst bias — different human experts may classify the same trace differently, whereas a trained model provides consistent, reproducible classifications. The ability to process thousands of traces rapidly also enables statistical analyses (e.g., transition rate distributions, state occupancy histograms) that would be impractical with manual curation. As single-molecule techniques scale to multiplexed and high-throughput formats, automated analysis tools like smDeepFLUOR become essential infrastructure.

Why for Yiru: Single-molecule fluorescence techniques are increasingly applied to study immune receptor signalling — for example, tracking individual T cell receptor–pMHC binding events, monitoring conformational changes in checkpoint receptors, or observing cytokine receptor dimerization dynamics. Automated classification of these signals could accelerate TME-relevant biophysical studies by enabling higher-throughput analysis of receptor-ligand interactions. More broadly, the deep learning approach to time-series classification demonstrated here could be adapted to other single-cell trajectory data types — such as classifying cellular response dynamics from live-cell imaging of T cell activation or tumour cell drug responses.

Summary: Introduces SMARTPocket, an atomic-level geometric deep learning framework for predicting RNA-small molecule binding pockets directly from three-dimensional RNA structure. RNAs are important therapeutic targets, yet identifying ligandable small-molecule binding pockets remains a major barrier to RNA-targeted drug discovery. SMARTPocket represents RNA as full-atom point clouds and uses transfer learning from more than 110,000 protein binding interface structures to overcome the limited number of experimentally elucidated RNA-ligand complexes. Across four established single-chain benchmarks and three broader curated benchmarks, SMARTPocket consistently outperforms existing RNA pocket predictors and general biomolecular modelling approaches. The model generalizes to apo RNA structures when conformational changes are modest, identifies cryptic ligandable pockets, and recapitulates experimentally validated binding sites in the SARS-CoV-2 frameshifting element and an RNA aptamer evolved to bind small molecules. SMARTPocket-guided docking further improves near-native RNA-ligand pose recovery and computational efficiency compared with blind docking.

Why it matters: RNA-targeted drug discovery is an emerging frontier that has been held back by the difficulty of identifying druggable pockets in RNA structures. Unlike proteins, which have well-characterized binding pockets with defined chemical properties, RNA structures are more dynamic, more charged, and have fewer examples of successful small-molecule targeting. SMARTPocket addresses this data scarcity through transfer learning from the much larger corpus of protein-ligand complexes, effectively teaching the model what a "bindable" pocket looks like in atomic detail and then adapting that knowledge to RNA. The ability to identify cryptic pockets — binding sites that are not visible in static structures but emerge through conformational dynamics — is particularly valuable, as many therapeutically relevant RNA structures (riboswitches, splicing regulatory elements, viral RNA elements) sample multiple conformations.

Why for Yiru: RNA biology is increasingly recognized as important in the TME. Non-coding RNAs, RNA-binding proteins, and RNA modifications regulate immune cell function, tumour cell plasticity, and therapy resistance. SMARTPocket could be applied to identify druggable pockets in TME-relevant RNAs — for example, structured elements in mRNAs encoding immune checkpoints, cytokines, or chemokine receptors. The geometric deep learning approach, which operates directly on 3D atomic coordinates, is also methodologically interesting as a general framework for structure-based prediction tasks that could extend to protein-RNA interactions, RNA modifications, or aptamer design for TME-targeted delivery.

Summary: Presents Tox21mer, a 43.5-million-parameter transformer that encodes each Tox21 concentration-response curve together with assay metadata into a 768-dimensional representation. The U.S. Tox21 collaboration has generated a large reference library of high-throughput concentration-response assays, but extracting transferable knowledge from these data has been challenging. Tox21mer was pretrained on approximately 2.5 million curves from 102 assay protocols and 6,727 compounds using masked-response reconstruction as the primary objective, with low-weight auxiliary supervision on assay outcome and AC50. The learned representation supported a macro-F1 of 0.985 for three-class outcome prediction (agonist, antagonist, inactive) and an R² of 0.87 for log₁₀(AC50). A masked-only pretraining variant retained near-baseline probe performance, indicating that the representation is learned largely from the self-supervised objective rather than from auxiliary labels. The embeddings formed coherent groupings by curve-class category. The model can support extrapolation to untested compounds through integration with chemical features or distillation into chemistry-only student models.

Why it matters: Foundation models are transforming many areas of biology, but their application to toxicology and drug safety has lagged behind fields like protein structure prediction and genomics. Tox21mer demonstrates that self-supervised pretraining on large-scale assay data can produce representations that capture essential pharmacological properties — whether a compound is an agonist or antagonist, and at what concentration it is active. The self-supervised approach is particularly noteworthy: it suggests that the "grammar" of concentration-response curves can be learned without explicit labels, potentially enabling transfer to new assays or compounds where labelled data are scarce. This is directly relevant to drug repurposing and safety screening applications.

Why for Yiru: Drug repurposing and computational pharmacology are increasingly relevant to TME research as immunotherapy combinations and targeted therapies proliferate. Tox21mer's self-supervised representations could be used to screen existing drugs for potential TME-modulating activity — for example, identifying compounds that affect immune cell viability or cytokine production at specific concentrations. The ability to predict concentration-response curves for untested compounds could also guide dose selection for in vitro TME perturbation experiments. Methodologically, the transformer architecture applied to structured assay data (rather than sequences or images) is an interesting paradigm that could be adapted to other high-throughput screening data types relevant to TME research.

Summary: Takes a network approach to understand DNA methylation clocks by building a co-methylation network from 12 public blood datasets. Biological age predicts health and lifespan better than chronological age, and DNA methylation-based "clocks" are leading molecular proxies. However, different established clocks share a vanishingly small number of CpG sites, many of which show weak associations with age, and clocks often do not transfer across array platforms. By building a co-methylation network of the sites showing the strongest age correlation and pruning weak links, the authors find a small number of large modules of covarying CpGs surrounded by many small modules and singleton sites. These modules are biologically interpretable, associated with CpG island contexts, and enriched for distinct Gene Ontology functions. Mapping five established clocks (Horvath, Hannum, AltumAge, Skin & Blood, Han) onto this network reveals that they select some CpGs from the same module, suggesting they are more similar than they appear. A simple clock retaining one CpG per module matches the performance of established clocks, while a clock built from module-level principal components outperforms all five established clocks in three validation cohorts and is transferable across array platforms.

Why it matters: DNA methylation clocks are widely used as biomarkers of biological aging, but their biological basis has remained opaque — different clocks use different CpGs, and it has been unclear whether they capture the same underlying biology or different aspects of aging. This network analysis reveals that despite apparent differences, established clocks sample from shared co-methylation modules, suggesting a common underlying structure. The demonstration that a simple module-based clock outperforms existing clocks while being transferable across platforms is practically important: it suggests a more principled approach to clock construction that should generalize better across populations and technologies.

Why for Yiru: Epigenetic aging is relevant to cancer immunology because TME cells — both tumour and immune — can exhibit accelerated or decoupled epigenetic aging patterns that affect function. Tumour- infiltrating T cells, for example, can acquire exhaustion-associated methylation changes that resemble aging. The network-based approach to methylation analysis could be applied to TME methylation data to identify co-methylation modules associated with immune infiltration, checkpoint expression, or therapy response. More broadly, the demonstration that network analysis can reconcile apparently discordant biomarkers into a coherent framework is a methodological lesson applicable to other TME biomarker development efforts.