Research Radar — 2026-05-04
Methods & AI
Computational
TxPert: using multiple knowledge graphs for prediction of transcriptomic perturbation effects
Nature Biotechnology Published 2026-05-01 research article DOI: 10.1038/s41587-026-03113-4
knowledge graph transcriptomic perturbation deep learning drug discovery
Summary: TxPert integrates multiple knowledge graphs with deep learning to predict transcriptomic responses to genetic and chemical perturbations, enabling systematic exploration of perturbation space beyond experimental coverage.
Why it matters: Predicting perturbation effects is a central challenge in functional genomics and drug discovery. Knowledge graph-based approaches that leverage structured biological relationships offer a complementary paradigm to purely sequence-based models.
Why for Yiru: Directly relevant to computational biology and AI-driven drug discovery — knowledge graph methods can be combined with single-cell and spatial data for perturbation modeling in immune contexts.
Tracing the rise of biomedical foundation models
Nature Biotechnology Published 2026-04-30 review DOI: 10.1038/s41587-026-03135-y
foundation model biomedical AI single-cell protein language model review
Summary: A comprehensive review tracking the emergence and evolution of foundation models across biomedical domains, from single-cell transcriptomics and protein structure to clinical applications, assessing capabilities, limitations, and future directions.
Why it matters: Biomedical foundation models are proliferating rapidly but fragmented across domains — this review provides a needed synthesis of the landscape, benchmarking progress and identifying gaps.
Why for Yiru: Essential reading for understanding where the field stands — directly relevant to decisions about which foundation model approaches to adopt for computational immunology and spatial omics.
DeepSeMS: revealing the hidden biosynthetic potential of the global ocean microbiome with a large language model
Nature Computational Science Published 2026-04-30 research article DOI: 10.1038/s43588-026-00983-1
large language model microbiome biosynthetic gene cluster natural product discovery
Summary: DeepSeMS applies a large language model to the global ocean metagenomic dataset to uncover previously hidden biosynthetic gene clusters, dramatically expanding the catalog of natural product biosynthetic potential.
Why it matters: Demonstrates how LLMs can be applied to metagenomic data at planetary scale to accelerate natural product discovery — a paradigm with implications beyond ocean microbiomes.
Why for Yiru: LLM-based mining of large biological sequence datasets is methodologically relevant to immune repertoire analysis and metagenomic approaches in tumor microenvironment studies.
Pretraining a foundation model for small-molecule natural products
Nature Machine Intelligence Published 2026-04-29 research article DOI: 10.1038/s42256-026-01226-8
foundation model natural products small molecule cheminformatics deep learning
Summary: A pretrained foundation model for small-molecule natural products that captures chemical space and bioactivity relationships, enabling downstream tasks such as property prediction and virtual screening.
Why it matters: Extends the foundation model paradigm to natural product chemistry — an area where structural diversity and limited data pose unique challenges for machine learning.
Why for Yiru: Small-molecule foundation models could eventually interface with immune receptor prediction and drug-target interaction modeling in computational immunology.
A multiobjective AI model for LNP engineering enhances tissue-selective mRNA delivery
Nature Biotechnology Published 2026-04-28 research article DOI: 10.1038/s41587-026-03109-0
AI-guided design lipid nanoparticle mRNA delivery multiobjective optimization drug delivery
Summary: A multiobjective AI model that optimizes lipid nanoparticle (LNP) formulations for tissue-selective mRNA delivery, enabling rational design of LNPs with improved specificity for target organs.
Why it matters: Tissue-selective delivery remains a major bottleneck for mRNA therapeutics — AI-driven LNP engineering could expand the therapeutic reach of mRNA beyond the liver.
Why for Yiru: AI-guided molecular design is a methodology with potential crossover to designing delivery vehicles for CAR-T engineering and in vivo cell reprogramming applications.
Biomedical discoveries
Biomedicine
Activated T cell extracellular vesicle DNA transfer enhances antigen presentation and anti-tumor immunity
Cancer Cell Published 2026-04-30 research article DOI: 10.1016/j.ccell.2026.03.023
T cell extracellular vesicle antigen presentation cancer immunotherapy immune evasion
Summary: Activated T cell-derived extracellular vesicles (ATEVs) transfer genomic DNA enriched in antigen processing and presentation genes into dendritic and tumor cells via granzyme B-mediated nuclear delivery, enhancing or restoring antigen presentation and overcoming immune evasion.
Why it matters: Reveals a previously unrecognized T cell–tumor communication mechanism mediated by EV-DNA transfer — opens a new avenue for acellular immunotherapy that could convert immunologically cold tumors to hot.
Why for Yiru: Directly relevant to tumor immunology and immunotherapy — understanding how T cell-derived vesicles reprogram antigen presentation has implications for CAR-T and adoptive cell therapy.
Metabolic and transcriptional plasticity supports CD8+ T cell resilience and anti-tumor immunity under nutrient stress
Immunity Published 2026-04-29 research article DOI: 10.1016/j.immuni.2026.04.004
CD8+ T cell T cell metabolism integrated stress response anti-tumor immunity metabolic plasticity
Summary: Reveals how the integrated stress response enables CD8+ T cells to switch between alternative biosynthetic modes when facing nutrient-depleted tumor microenvironments, preserving effector function and preventing dysfunction through metabolic and transcriptional plasticity.
Why it matters: T cell metabolic fitness in the tumor microenvironment is a major determinant of immunotherapy success — this work identifies the stress response machinery as a key resilience mechanism.
Why for Yiru: Understanding how T cells adapt metabolically to hostile tumor environments is directly relevant to engineering more resilient CAR-T cells and improving adoptive immunotherapy.
Tissue tension fosters macrophage-driven lipid peroxidation-induced DNA damage
Cancer Cell Published 2026-04-30 research article DOI: 10.1016/j.ccell.2026.03.022
tumor microenvironment macrophage tissue tension DNA damage lipid peroxidation fibrosis
Summary: Demonstrates that stromal fibrosis-generated tissue tension recruits macrophages via STAT3-mediated chemokines. Under mechanical stress, macrophages undergo lipid peroxidation, producing genotoxic aldehydes that induce epithelial DNA damage and drive tumor progression.
Why it matters: Establishes a mechanistic link between tissue mechanics, macrophage biology, and mutagenesis — a tripartite axis that explains how fibrosis contributes to cancer risk beyond physical barriers.
Why for Yiru: Connects macrophage biology, tumor microenvironment mechanics, and DNA damage — directly relevant to understanding how the physical TME shapes immune cell function and cancer evolution.
An SPP1-SOCS1 pathway constrains interferon responses in tumor-associated macrophages and shapes an immunosuppressive tumor microenvironment
Immunity Published 2026-04-27 research article DOI: 10.1016/j.immuni.2026.04.001
tumor-associated macrophage SPP1 SOCS1 interferon response immunosuppression immune checkpoint blockade
Summary: Identifies SPP1+ tumor-associated macrophages (TAMs) as enriched across cancer types and associated with immunotherapy resistance. SPP1 interacts with TRIM21 to limit SOCS1 ubiquitination, dampening IFN-γ-STAT1-ISG signaling and maintaining an immunosuppressive TME.
Why it matters: Reveals a novel SPP1-SOCS1 signaling axis that controls interferon responsiveness in TAMs — SPP1 is emerging as a key macrophage checkpoint target beyond its known roles in migration and survival.
Why for Yiru: SPP1+ macrophage biology is directly relevant to spatial omics analyses of the tumor microenvironment — understanding this pathway could inform macrophage-targeted immunotherapy strategies.
Krüppel-like factor 2 programs early exhausted T cell states and restrains antiviral immunity
Immunity Published 2026-04-27 research article DOI: 10.1016/j.immuni.2026.03.029
T cell exhaustion KLF2 CD8+ T cell chronic infection transcription factor
Summary: Identifies KLF2 as a central and specific regulator of CX3CR1+ effector-like exhausted CD8+ T cells during chronic infection. Disrupting the KLF2-dependent program enhances viral control without overt immunopathology, revealing a targetable exhaustion regulator.
Why it matters: KLF2 joins the growing catalog of transcription factors that program distinct exhausted T cell states — targeting exhaustion regulators without causing autoimmunity is a key challenge for chronic infection and cancer.
Why for Yiru: T cell exhaustion programs are directly relevant to understanding CAR-T cell dysfunction in solid tumors and identifying strategies to maintain effector function in immunosuppressive environments.
Genetic variation reveals a homeotic long noncoding RNA that modulates human hematopoietic stem cells
Cell Published 2026-05-01 research article DOI: 10.1016/j.cell.2026.04.014
hematopoietic stem cell lncRNA HOXA9 genetic variation blood cancer
Summary: A blood cancer-protective inherited variant in the homeotic lncRNA HOTSCRAMBL alters HOXA9 splicing, fine-tuning hematopoietic stem cell self-renewal while constraining HOXA-driven blood cancers.
Why it matters: Identifies a naturally occurring human genetic variant that protects against blood cancer through lncRNA-mediated control of stem cell programs — a powerful example of human genetics informing stem cell biology.
Why for Yiru: Hematopoietic stem cell biology underpins many immunotherapy approaches including CAR-T manufacturing and bone marrow transplantation — understanding endogenous regulatory mechanisms informs engineering strategies.
Cross-disciplinary watchlist
Other Fields
A multimodal large language model for materials science
Nature Machine Intelligence Published 2026-04-24 research article DOI: 10.1038/s42256-026-01214-y
large language model multimodal AI materials science scientific discovery
Summary: A multimodal large language model that integrates text, crystal structures, and property data for materials science, enabling composition-to-property prediction, synthesis planning, and knowledge retrieval across diverse materials domains.
Why it matters: Demonstrates how multimodal LLMs can serve as unified scientific assistants in domains with heterogeneous data types — a paradigm applicable to many scientific fields.
Why for Yiru: The multimodal integration approach used in materials science (text + structure + properties) mirrors challenges in biomedical AI where diverse data types (imaging, sequencing, clinical) must be unified.
NOEM: efficient and scalable finite element method enabled by reusable neural operators
Nature Computational Science Published 2026-04-28 research article DOI: 10.1038/s43588-026-00974-2
neural operator finite element method physics simulation computational science scientific machine learning
Summary: NOEM introduces reusable neural operators that dramatically accelerate finite element method simulations by learning solution operators that generalize across geometries and boundary conditions, achieving orders-of-magnitude speedups.
Why it matters: Neural operator approaches that generalize across problem instances represent a major advance for physics-informed machine learning — moving beyond single-problem training to reusable simulation surrogates.
Why for Yiru: Neural operators that accelerate physical simulations have conceptual parallels to models that accelerate biological simulations — the reusable architecture concept could inspire approaches for biological network modeling.
From embodied intelligence to physical AI
Nature Machine Intelligence Published 2026-04-24 perspective DOI: 10.1038/s42256-026-01239-3
physical AI embodied intelligence robotics AI systems
Summary: A perspective on the transition from embodied intelligence research to physical AI systems that can interact with and manipulate the real world, discussing the challenges of grounding AI in physical constraints and real-time sensory feedback.
Why it matters: The embodied-to-physical AI transition represents a frontier where AI must contend with real-world physics — the principles developed here for robustness and real-time adaptation have cross-domain relevance.
Why for Yiru: The concept of AI systems that must operate under physical constraints and real-time feedback has parallels to biomedical AI systems that must work within biological constraints and patient-specific variability.
Adopting a human developmental visual diet yields robust and shape-based AI vision
Nature Machine Intelligence Published 2026-04-24 research article DOI: 10.1038/s42256-026-01228-6
computer vision developmental AI shape recognition training curriculum robustness
Summary: Training AI vision systems on a curriculum that mimics the visual experience of human infants — starting with simple shapes and gradually increasing complexity — yields more robust and shape-based visual representations that generalize better than standard training paradigms.
Why it matters: Challenges the dominant paradigm of training on massive uncurated datasets by showing that developmentally inspired curricula produce more human-like and robust visual systems — curriculum design matters as much as model architecture.
Why for Yiru: The principle that training data curriculum design can dramatically improve model robustness is transferable to biomedical AI — thoughtful data ordering may improve foundation models for biology.