Research Radar — 2026-06-30
Methods & AI
Computational
Accurate trajectory inference in time-series spatial transcriptomics with structurally-constrained optimal transport
Nature Communications Published 2026-06-29 research article DOI: 10.1038/s41467-026-74927-8
spatial transcriptomics AI deep learning
Summary: Time-series spatial transcriptomics (ST) data can show dynamics of biological structures in situ. Here, the authors present an algorithm for cellular trajectory inference from time-series ST that leverages knowledge of spatially-coherent biological structures that develop as a unit.
Why it matters: Introduces a principled framework for inferring cellular trajectories from time-series spatial transcriptomics, addressing a critical gap in spatial biology where temporal dynamics are underexplored.
Why for Yiru: Directly relevant to Yiru's spatial transcriptomics work — this method enables trajectory inference from longitudinal ST data, which could be applied to study tissue development and tumor progression in his projects.
G-LATO: Inference of Spatial Latent Ordering via Deep Gaussian Processes
bioRxiv (bioinformatics) Published 2026-06-23 preprint DOI: 10.64898/2026.06.23.734031
spatial transcriptomics deep learning AI
Summary: Spatial transcriptomics enables the study of cells within their native tissue context, yet identifying gradients of cellular development remains challenging. We introduce a deep Gaussian process model to address this gap. Our method recovers spatially smooth gradients explaining observed gene expression. We illustrate our method on healthy liver and glioblastoma data in reconstructing known spatial organisation and uncovering new pathological gradients, thus providing robust inference for spatial biology.
Why it matters: Offers a principled Bayesian approach to recovering spatial gradients in ST data, overcoming limitations of existing methods that lack uncertainty quantification and spatial smoothness constraints.
Why for Yiru: Highly relevant to Yiru's spatial transcriptomics analyses — deep Gaussian processes provide a probabilistic framework for identifying developmental gradients in tissues, directly applicable to his tumor microenvironment studies.
Impact of molecular multimodality on neural network models for prediction tasks related to drug discovery
Nature Communications Published 2026-06-29 research article DOI: 10.1038/s41467-026-74487-x
deep learning AI multi-omics
Summary: Here, the authors examine whether combining multiple molecular data types improves drug discovery models, finding multimodal approaches boost predictions with effective fusion, and even simple late-fusion methods can reach state-of-the-art performance.
Why it matters: Provides systematic evidence that multimodal molecular data fusion — even with simple architectures — improves drug discovery predictions, with practical implications for computational drug development pipelines.
Why for Yiru: Relevant to Yiru's interest in deep learning and multi-omics integration — the finding that simple late-fusion can match complex architectures is practically useful for his own multimodal modeling efforts.
Experiment-guided AlphaFold3 resolves measurement-consistent protein ensembles
Nature Biotechnology Published 2026-06-29 research article DOI: 10.1038/s41587-026-03166-5
AI deep learning foundation model
Summary: A generative model using AlphaFold3 as prior infers protein conformational ensembles consistent with measured experimental data.
Why it matters: Demonstrates how generative AI priors (AlphaFold3) can be fused with experimental data to resolve conformational ensembles, bridging computational prediction and experimental measurement in structural biology.
Why for Yiru: Illustrates a paradigm for integrating foundation models with experimental data that could inspire similar approaches in Yiru's spatial and single-cell analyses.
Metrics for Distinguishing Biological and Interventional Change in AI Models
bioRxiv (bioinformatics) Published 2026-06-24 preprint DOI: 10.64898/2026.06.24.733252
AI deep learning computational biology
Summary: Statistical and machine-learning models of longitudinal biological data evaluate change by comparing each new observation against the trajectory implied by prior observations, assuming the process generating that trajectory is stable. Here the authors introduce two subject-level metrics that quantify the geometric signature an interventional change leaves in the data substrate, enabling distinction between biological and interventional change.
Why it matters: Provides a formal statistical framework for distinguishing natural biological variation from intervention-induced changes in longitudinal models, addressing a fundamental confound in perturbation analysis.
Why for Yiru: Conceptually relevant to Yiru's work with perturbation screens and longitudinal studies — these metrics could improve how treatment effects are distinguished from biological variability in his experiments.
Can a Tissue-derived Progression Signature Accurately Predict Colorectal Cancer Stage Transitions in Blood?
bioRxiv (bioinformatics) Published 2026-06-23 preprint DOI: 10.64898/2026.06.23.734006
cancer biomarker AI machine learning
Summary: Colorectal cancer (CRC) is challenging to track because its molecular changes are very complex as the disease progresses. In this study, the authors develop a machine learning framework integrating monotonic progression and StepMiner approach. A balanced 74-gene signature is used for machine-learning classification with Random Forest. External validation shows strong performance in tissue-based datasets but poor performance in plasma and blood-based datasets, highlighting biological differences between transcriptomic profiles.
Why it matters: Highlights a critical translational gap between tissue-derived signatures and liquid biopsy performance, with important implications for developing clinically deployable cancer progression biomarkers.
Why for Yiru: Relevant to Yiru's interest in cancer biomarkers — the discordance between tissue and blood signatures underscores the importance of matching training data to clinical application context, a lesson for his own biomarker work.
Biomedical discoveries
Biomedicine
Droplet single-cell CRISPR screens identify regulators of T cell-mediated target-cell killing
bioRxiv (immunology) Published 2026-06-23 preprint DOI: 10.64898/2026.06.23.734054
single-cell CRISPR screen T cell cancer immunotherapy
Summary: Cytotoxic CD8 T cells kill target cells through brief cell-cell encounters, but pooled genetic screens cannot readily link perturbations in individual T cells to the fate of the target cells they engage. The authors developed droplet single-cell CRISPR screening to pair individual primary human CD8 T cells with cancer cells, measure rapid target-cell death, and recover sgRNAs from phenotype-defined droplets. The platform recovered regulators of TCR signaling, synapse formation, granule exocytosis and cytotoxic differentiation, identifying negative regulators of killing including PTEN, RASA2, FOXO1, AFAP1L2 and mTORC1 components. Transient pharmacologic mTORC1 inhibition reproduced a rapid-killing state and improved antitumor efficacy.
Why it matters: Establishes a transformative platform for functional genomics in T cells at single-cell resolution, directly linking genetic perturbations to cytotoxic function and identifying mTORC1 as a druggable checkpoint for enhancing T cell killing.
Why for Yiru: Highly relevant to Yiru's work in CAR-T and cancer immunotherapy — the identification of mTORC1 as a negative regulator of T cell killing presents a directly actionable strategy for enhancing CAR-T cell potency through transient pharmacological inhibition.
PD-L1 deletion or blockade regulate macrophage antigen presentation and checkpoint molecule surface levels
bioRxiv (immunology) Published 2026-06-23 preprint DOI: 10.64898/2026.06.23.734016
macrophage tumor microenvironment immune immunotherapy cancer
Summary: Macrophages in the tumor microenvironment upregulate PD-L1 expression, suppressing T cells through PD-1 ligation. However, the macrophage-intrinsic role of PD-L1 is less clear. Using genetic deletion and antibody blockade approaches, the authors show that neither dramatically alters macrophage polarization markers or phagocytic capacity. Both conditions consistently reduced surface CD80 and had disparate effects on MHC-I, MHC-II, CD86, PD-L2, and PD-1 expression. This reveals that PD-L1 intrinsically shapes the macrophage antigen presentation landscape and checkpoint receptor profile in a manner distinct from its canonical T cell-suppressive role.
Why it matters: Reveals a previously underappreciated cell-intrinsic role for PD-L1 in macrophages beyond T cell suppression, with direct implications for how checkpoint blockade therapies reprogram the tumor immune microenvironment.
Why for Yiru: Directly relevant to Yiru's interest in tumor microenvironment and macrophage biology — understanding PD-L1's macrophage-intrinsic functions could inform how checkpoint inhibitors reprogram the TME beyond T cell reactivation.
Inflammatory immune modulators of AML lung infiltration and respiratory failure
Nature Immunology Published 2026-06-29 research article DOI: 10.1038/s41590-026-02582-8
immune cancer T cell tumor microenvironment
Summary: Paraskevopoulou et al. examine lung damage that occurs in patients diagnosed with acute myeloid leukemia (AML). In mouse models, AML infiltration leads to inflammatory remodeling of the lung niche, resulting in extensive damage and fibrosis in a process driven by galectin-9 and IL-33 axis signaling.
Why it matters: Uncovers a specific inflammatory mechanism (galectin-9/IL-33 axis) driving AML-induced respiratory failure, identifying potential therapeutic targets for a devastating complication of leukemia.
Why for Yiru: Relevant to Yiru's interest in cancer immunology and tumor microenvironment — the galectin-9/IL-33 axis represents an immune checkpoint mechanism with potential relevance beyond AML to solid tumor contexts.
Spatio-DARLIN enables robust and efficient in situ lineage tracing in mice at single-cell resolution
Nature Methods Published 2026-06-29 research article DOI: 10.1038/s41592-026-03151-5
single-cell spatial transcriptomics cell atlas
Summary: Spatio-DARLIN combines the DARLIN mouse model with sequencing-based spatial transcriptomics to enable high-resolution spatial lineage tracing in mice.
Why it matters: Provides a powerful tool for linking clonal history with spatial context in mammalian tissues, opening new frontiers in developmental biology, cancer evolution, and tissue regeneration research.
Why for Yiru: Highly relevant to Yiru's spatial transcriptomics and single-cell work — Spatio-DARLIN could be applied to trace tumor clonal evolution in spatial context, directly complementing his tumor microenvironment research.
No evidence of immunosurveillance in mutation-hotspot-driven clonal hematopoiesis
Nature Genetics Published 2026-06-29 research article DOI: 10.1038/s41588-026-02594-y
immune cancer T cell
Summary: Analysis of the UK Biobank cohort demonstrates that there is no significant MHC-I- or MHC-II-driven negative selection against expansion of mutant clones in clonal hematopoiesis.
Why it matters: Challenges the prevailing hypothesis that immune surveillance constrains clonal hematopoiesis, with important implications for understanding early cancer evolution and immunoprevention strategies.
Why for Yiru: Relevant to Yiru's interest in cancer immunology — the finding that MHC-driven immunosurveillance does not limit clonal hematopoiesis suggests alternative selective pressures operate at the earliest stages of malignant transformation.
Stroke-induced lipocalin-2-expressing red pulp macrophages reprogram peripheral immunity
bioRxiv (immunology) Published 2026-06-23 preprint DOI: 10.64898/2026.06.23.733904
macrophage immune T cell
Summary: Acute ischemic stroke induces profound systemic immune alterations. The authors identify lipocalin-2 (LCN-2) as a rapidly induced regulator of stroke-associated immunosuppression. LCN-2 is strongly upregulated in splenic red pulp macrophages within 24h and 7 days post-stroke. LCN-2-expressing RPMs form immunological synapses with CD3+ T cells. Recombinant LCN-2 reprogrammed T cells and monocytes toward tolerogenic phenotypes. Human spleens likewise displayed LCN-2-expressing CD68+ RPMs.
Why it matters: Identifies a novel mechanism of systemic immunosuppression after stroke mediated by splenic red pulp macrophages and lipocalin-2, revealing a tissue-resident macrophage-T cell cross-talk axis with therapeutic implications.
Why for Yiru: Relevant to Yiru's interest in macrophage biology — the LCN-2- expressing RPM-T cell axis represents a new paradigm for how tissue-resident macrophages reprogram systemic immunity, with potential parallels in the tumor microenvironment.
Cross-disciplinary watchlist
Other Fields
Retargeted serine integrases for one-step, precise integration of large DNA sequences in human cells
Nature Biotechnology Published 2026-06-29 research article DOI: 10.1038/s41587-026-03186-1
CRISPR gene editing
Summary: Serine integrases are retargeted to streamline genomic integrations of DNA.
Why it matters: Simplifies large DNA integration in human cells, offering a powerful tool for gene therapy and synthetic biology that could accelerate cell engineering for research and therapeutic applications.
Why for Yiru: Relevant to Yiru's CAR-T interests — more efficient large DNA integration could streamline CAR construct delivery and knock-in engineering for next-generation cell therapies.
Causally measuring aging and rejuvenation through transcriptomic damage
bioRxiv (bioinformatics) Published 2026-06-26 preprint DOI: 10.64898/2026.06.26.734659
AI biomarker aging transcriptomics
Summary: Aging is caused by the progressive accumulation of damage. The authors present a computational framework to quantify damage from standard RNA-sequencing data, capturing four classes of aberrant transcript structures. They constructed a transcriptomic damage-based aging (tDamAge) clock using machine learning models that predicts age and detects transcriptomic shifts under pro-aging and anti-aging conditions. Interventions such as caloric restriction, rapamycin, and methionine restriction lowered tDamAge.
Why it matters: Introduces a causal, transcriptomic damage-based aging clock that directly measures molecular damage rather than correlating with chronological age, enabling quantification of rejuvenation interventions.
Why for Yiru: Relevant to Yiru's interest in AI and transcriptomics — the framework's ability to detect transcriptomic shifts from interventions using standard RNA-seq data could be adapted to measure treatment effects in cancer and immune contexts.
EnzyKAN: Protein Language Model Embeddings and Kolmogorov-Arnold Network Variants for Enzyme Commission Classification
bioRxiv (bioinformatics) Published 2026-06-23 preprint DOI: 10.64898/2026.06.23.734004
AI deep learning foundation model
Summary: A reproducible investigation of Kolmogorov-Arnold Network (KAN) variants for enzyme classification using protein language model embeddings. KANs use learnable edge functions rather than fixed ones. Utilising ESM-2 650M embeddings, MLP achieved macro F1=0.750 while SineKAN achieved macro F1=0.716 for seven-class EC classification. The work demonstrates that KANs are competitive but do not yet exceed conventional baselines for protein sequence tasks.
Why it matters: Provides a rigorous benchmark of KAN architectures on a biological sequence task, showing that despite their theoretical appeal, KANs do not yet surpass well-tuned MLPs for protein function prediction.
Why for Yiru: Useful context for Yiru's work with deep learning in biology — the finding that KANs match but don't exceed MLPs is an important methodological sanity check when choosing architectures for biological sequence and expression modeling.