Research Radar — 2026-05-16
Methods & AI
Computational
Deep learning models for chemical perturbation prediction do not yet utilise drug molecular features
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.13.724458
perturbation prediction deep learning drug features benchmarking L1000
Summary: Retrains seven state-of-the-art deep learning models for L1000 chemical perturbation prediction from scratch with zeroed or shuffled drug molecular inputs. Under cold-drug evaluation, ablation causes negligible performance change and a drug-free baseline (MLP using only cell-line basal expression) matches all models. Demonstrates that current architectures do not actually utilize drug molecular features for generalization to unseen compounds.
Why it matters: This finding is a bombshell for the perturbation prediction field — if leading models don't use drug features, the impressive reported performance may reflect memorization of cell-type baselines rather than learning drug mechanisms. Demands fundamental rethinking of architecture design and evaluation protocols.
Why for Yiru: Directly challenges the validity of perturbation prediction models Boss might use for drug response and CRISPR screen analysis. The cold-drug generalization failure is precisely the scenario encountered in translational research.
Deep peptide recognition profiling decodes TCR specificity and enables disease-associated antigen discovery
Nature Biotechnology Published 2026-05-13 research article DOI: 10.1038/s41587-026-03128-x
T cell receptor deep learning antigen discovery immunology peptide-MHC
Summary: Presents a deep-learning method that predicts T cell receptor specificity and discovers disease-associated antigens by profiling peptide recognition at unprecedented scale. The framework decodes the molecular rules governing TCR-peptide-MHC interactions, enabling identification of novel antigens in autoimmune and infectious disease contexts from the Garcia lab at Stanford.
Why it matters: TCR specificity prediction is a holy grail problem in immunology with direct applications to cancer immunotherapy, autoimmune disease, and vaccine design. A scalable deep-learning solution could transform antigen discovery pipelines.
Why for Yiru: TCR-antigen recognition is central to tumor immunology. Understanding which antigens T cells recognize in the TME could guide personalized immunotherapy design and biomarker discovery.
Advancing conversational diagnostic AI with multimodal reasoning
Nature Medicine Published 2026-05-14 research article DOI: 10.1038/s41591-026-04371-0
diagnostic AI multimodal LLM clinical AI Google Research
Summary: Google Research's AMIE (Articulate Medical Intelligence Explorer) system advances to request, interpret, and reason about multimodal medical data including imaging, lab results, and clinical notes during diagnostic dialogues. The system demonstrates improved diagnostic accuracy and clinical reasoning compared to previous unimodal versions.
Why it matters: Multimodal diagnostic AI that can actively request and interpret diverse data types represents a qualitative leap beyond chatbot-style medical AI. This is the closest thing to an AI clinical colleague that exists today.
Why for Yiru: AI systems that integrate diverse data modalities for clinical reasoning share conceptual architecture with multi-omics integration challenges. The reasoning frameworks developed here may inform how we build AI assistants for biomedical research.
Widespread DNA off-targeting confounds RNA chromatin occupancy studies
Nature Biotechnology Published 2026-05-15 research article DOI: 10.1038/s41587-026-03130-3
lncRNA chromatin off-target methodology RNA-DNA interactions
Summary: Demonstrates that many long noncoding RNA-DNA binding peaks detected using common chromatin occupancy assays arise from widespread DNA off-targeting artifacts rather than genuine RNA-chromatin interactions. The finding has major implications for the lncRNA field, where thousands of reported chromatin interactions may be technical artifacts rather than biological regulatory events.
Why it matters: If a substantial fraction of reported lncRNA-chromatin interactions are artifacts, the functional annotation of the noncoding genome requires systematic re-evaluation. This is a field-correcting finding of the kind that reshapes research programs.
Why for Yiru: Methodological rigor in genomics directly affects the reliability of regulatory inferences in cancer and immune cells. Awareness of assay-specific artifacts is essential when interpreting epigenetic and regulatory datasets.
DNA-guided CRISPR–Cas12 for cellular RNA targeting
Nature Biotechnology Published 2026-05-15 research article DOI: 10.1038/s41587-026-03129-w
CRISPR Cas12 RNA targeting DNA guide gene editing
Summary: Demonstrates that Cas12 nucleases can use chemically modified DNA guides (ΨDNA) instead of RNA guides, which switches their targeting specificity from DNA to cellular RNA. This DNA-guided CRISPR system enables precise, programmable RNA targeting in living cells without the stability and delivery challenges of RNA guides.
Why it matters: Expands the CRISPR toolkit to RNA targeting using DNA guides, which are cheaper to synthesize, more stable, and easier to deliver than RNA guides. Opens new possibilities for transient RNA manipulation without permanent genetic changes.
Why for Yiru: Programmable RNA targeting has applications in studying post-transcriptional regulation, modulating immune gene expression, and potentially targeting oncogenic transcripts in cancer cells.
TwinSAR: An Adaptive Kernel-based Algorithm with logit-transformed Z-score Filtering for Chemical Twin Detection in Large-scale Virtual Screening
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.12.724687
virtual screening cheminformatics chemical similarity drug discovery
Summary: Introduces TwinSAR, a statistically principled algorithm for identifying chemical twins based on stoichiometric proximity rather than topological fingerprints. Uses adaptive kernel methods with per-block calibration and logit-transformed Z-score filtering. Validated in a BCL-2 virtual screening pipeline, reducing a 327K-compound library to 390 candidates.
Why it matters: Chemical similarity by elemental composition is orthogonal to standard fingerprint-based methods and can retrieve scaffold hops that Tanimoto similarity misses. The statistical rigor and scalability to million-compound libraries make it practically deployable.
Why for Yiru: Virtual screening methodology connects to interests in computational drug discovery and chemical biology approaches for target identification and validation.
Biomedical discoveries
Biomedicine
Immune-remodeling mRNAs expressing IRF8 or NIK generate durable antitumor immunity in multiple cancer models
Nature Biotechnology Published 2026-05-13 research article DOI: 10.1038/s41587-026-03115-2
mRNA therapy cancer immunotherapy IRF8 NIK lipid nanoparticles TME
Summary: Demonstrates that lipid nanoparticle-delivered mRNAs encoding immune transcription factors IRF8 or NIK reprogram the tumor microenvironment and generate durable antitumor immunity across multiple cancer models. The immune-remodeling approach, from the Langer/Anderson labs at MIT, represents a new class of mRNA cancer immunotherapy distinct from neoantigen vaccines or CAR-T.
Why it matters: mRNA-based immune remodeling via transcription factor delivery is a conceptually novel immunotherapy strategy. Unlike vaccines that target specific antigens, this approach reprograms the entire TME toward immune activation — potentially effective across heterogeneous tumors.
Why for Yiru: TME reprogramming and mRNA-based immunotherapy are at the intersection of Boss's interests in tumor immunology and translational cancer research. Understanding how transcription factor delivery reshapes immune cell states could inform spatial TME analysis.
LVV SMRTcap reveals extensive proviral variation in lentiviral vector-transduced CAR T cells
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.13.724601
CAR T cells lentiviral vector long-read sequencing quality control cell therapy
Summary: Adapts a long-read capture approach (SMRTcap) to simultaneously map integration sites and assess proviral integrity at single-molecule resolution in CAR T cells. Reveals that 40% of integrated vectors in research-grade CAR T cells carry recurrent deletions removing promoters or CAR cassettes. Deletions are present in viral stocks and detected — at lower frequency — in clinical CAR T products pre- and post-infusion. Also finds widespread G-to-A hypermutation consistent with APOBEC editing introducing premature stop codons.
Why it matters: Current CAR T QC assays (VCN, integration site analysis) are blind to proviral deletions and hypermutation that could compromise therapeutic efficacy or safety. This method reveals a hidden dimension of CAR T product heterogeneity with direct clinical implications.
Why for Yiru: CAR T cell therapy quality and the genomic integrity of therapeutic vectors are directly relevant to cancer immunotherapy research. The long-read sequencing methodology also has applications in spatial and single-cell genomics.
Profilin 1 maintains cell cycle fidelity to prevent unscheduled genome doubling and polyploidy in cancer
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.12.724607
whole-genome doubling polyploidy cell cycle cancer osteosarcoma PFN1
Summary: Identifies Profilin 1 (PFN1) loss as a driver of whole-genome doubling (WGD) in p53-proficient cells. Using FUCCI live-cell imaging and single-cell genomics, shows PFN1-deficient cells bypass mitosis and undergo endoreplication. Genome-doubled cells retain proliferative capacity and undergo aberrant mitoses, amplifying genomic instability. PFN1 loss also promotes MDM2-mediated p53 checkpoint evasion and chemoresistance, with orthotopic xenografts showing enhanced metastasis.
Why it matters: WGD is a major route to aneuploidy and therapy resistance in cancer, but how p53-proficient cells tolerate genome doubling is poorly understood. PFN1 emerges as a novel safeguard whose loss enables a specific route to polyploidy — potentially a therapeutic vulnerability.
Why for Yiru: Genome instability mechanisms and therapy resistance in cancer are fundamental to understanding tumor evolution and treatment failure. The link between cytoskeletal protein loss and genome doubling connects cell biology to genomic outcomes.
Rac2 Hyperactivity Drives Neutrophil Degranulation, Myeloperoxidase Deficiency, and Lymphopenia
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.12.723629
neutrophils Rac2 immunodeficiency MPO T cell lymphopenia innate immunity
Summary: Using a Rac2+/E62K gain-of-function mouse model, shows that hyperactive Rac2 primes neutrophils for premature primary granule degranulation, depleting myeloperoxidase needed for microbial killing. Neutrophils are hyperactivated yet functionally compromised. In spleen, degranulating neutrophils accumulate and selectively clear T cells via phosphatidylserine exposure, providing a mechanism for the T cell lymphopenia observed in RAC2E62K patients.
Why it matters: Reveals the paradoxical mechanism by which a gain-of-function mutation causes immunodeficiency: hyperactivation triggers granule depletion that leaves neutrophils unable to kill, while simultaneously driving T cell clearance. The Rac2-neutrophil-T cell axis is a novel immunoregulatory circuit.
Why for Yiru: Neutrophil biology in immune dysfunction and the unexpected connection between innate cell hyperactivity and adaptive lymphopenia offers insights into immune cell cross-talk relevant to tumor-associated neutrophil function.
Comprehensive Analysis Reveals Adaptive DNA Repair and Replication Stress Networks in Genomically Unstable Breast Cancer
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.13.724208
breast cancer genomic instability DNA repair replication stress synthetic lethality
Summary: Multi-cohort analysis of breast cancer datasets reveals that genomically unstable tumors upregulate compensatory DNA repair and replication stress tolerance programs rather than achieving true repair restoration. Identifies pathway-specific co-occurrence and mutual exclusivity patterns among DDR genes and major drivers, nominating context-specific synthetic lethal targets for BRCA-mutant and other high-FGA breast cancers.
Why it matters: Reframes genomic instability as a state of adaptive repair network rewiring rather than simple repair deficiency. The context-specific synthetic lethal opportunities identified could guide precision oncology beyond PARP inhibitors.
Why for Yiru: DNA repair network adaptation in cancer and synthetic lethality approaches are directly relevant to understanding therapeutic vulnerabilities in genomically unstable tumors including those in the TME.
Characterization of tumor interactions with the immune system in an autochthonous mouse model of glioblastoma
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.13.724869
glioblastoma tumor microenvironment immune checkpoint palasading necrosis TGF-beta
Summary: Characterizes immune responses in an autochthonous mouse glioblastoma model using spectral flow cytometry and IHC. Large tumors generate PD1+ CD4 T cell populations at leptomeningeal/perivascular border sites, explaining pharmacodynamic responses in neoadjuvant anti-PD1 trials. However, palisading necroses avidly recruit myeloid cells that produce copious TGF-β, creating an immunosuppressive barrier that may explain why responses are insufficient for clinical benefit.
Why it matters: Provides a mechanistic explanation for the mixed results of checkpoint inhibitor trials in glioblastoma: immune cells are recruited but immediately suppressed by necrosis-driven TGF-β. Suggests combination strategies targeting both checkpoints and TGF-β signaling.
Why for Yiru: The spatial organization of immune suppression in brain tumors — with border-site T cell activation and necrosis-driven myeloid suppression — is a spatial TME biology question directly relevant to Boss's research interests.
Cross-disciplinary watchlist
Other Fields
A generative artificial intelligence approach for peptide antibiotic optimization
Nature Machine Intelligence Published 2026-05-13 research article DOI: 10.1038/s42256-026-01237-5
generative AI antibiotic design antimicrobial peptides drug resistance de la Fuente lab
Summary: Presents ApexGO, a generative AI approach that redesigns peptide antibiotics to better kill drug-resistant bacteria. The de la Fuente-Nunez lab validates AI-designed candidates in laboratory tests and mouse infection models, with redesigned peptides matching or outperforming standard antibiotics against multidrug-resistant pathogens.
Why it matters: Antimicrobial resistance is a top-10 global health threat. Generative AI that can optimize existing peptide antibiotics — rather than screening from scratch — accelerates the path from computation to clinic. The in vivo validation in mouse infection models raises the bar for AI-driven antibiotic discovery.
Why for Yiru: AI-driven molecular design with experimental validation in disease models exemplifies the computational-to-translational pipeline that Boss's research aspires to. The generative optimization framework has conceptual parallels to designing perturbations for cellular reprogramming.
Population-scale genomic medicine with the Hong Kong Genome Project
Nature Medicine Published 2026-05-15 research article DOI: 10.1038/s41591-026-04410-w
population genomics Chinese population genomic medicine Hong Kong clinical genetics
Summary: Reports genomic analyses from the Hong Kong Genome Project covering over 20,000 participants, providing clinically relevant variant information for the Chinese population. Offers implementation insights for genomic medicine initiatives in East Asian populations, which remain severely underrepresented in global genomic databases.
Why it matters: East Asian populations are dramatically underrepresented in genomic medicine — most clinical variant databases are Eurocentric. A population-scale resource for Chinese genomics directly addresses this disparity and enables more accurate genetic diagnosis and risk prediction.
Why for Yiru: Population-specific genomic resources are essential for understanding cancer predisposition, pharmacogenomics, and immune-related genetic variation in the populations Boss's research may ultimately serve.
Hydrogen Peroxide induces resistance to DNA damage in a localization and p53 dependent manner
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.13.724825
ROS p53 hormesis chemoresistance DNA damage p21
Summary: Uses chemo-genetic H2O2 production to study ROS-dependent adaptive responses with subcellular precision. Shows nucleosomal H2O2 production provides p53/p21-dependent resistance to subsequent high-dose H2O2 and chemotherapy-induced DNA damage, while mitochondrial H2O2 does not. Brief p53 stabilization protects p53-wildtype cells from chemotherapy, suggesting a hormesis-based strategy to selectively shield healthy tissue during cancer treatment.
Why it matters: The concept of using controlled p53 activation to protect normal tissue during chemotherapy — while p53-mutant tumors remain vulnerable — is a creative therapeutic strategy. The subcellular localization dependency of the hormetic response reveals unexpected spatial logic in stress signaling.
Why for Yiru: Oxidative stress biology and p53 signaling are fundamental to cancer biology. The concept of selectively protecting normal tissue during therapy through hormesis is an elegant therapeutic concept that connects basic biology to clinical application.
Pan-Viral Conformational Landscapes of Frameshifting Elements Reveal Length-Dependent Plasticity and Antisense-Driven Structural Reprogramming
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.13.724818
RNA structure frameshifting virus antisense oligonucleotides molecular dynamics
Summary: Comparative structural analysis of programmed ribosomal frameshifting elements from JEV, WNV, HCV, and HIV using integrative computational modeling. Reveals conserved core architecture with length-dependent conformational plasticity. Demonstrates antisense oligonucleotide binding can reprogram FSE architectures, disrupting native structural motifs — positioning ASO-mediated structural perturbation as a strategy for modulating viral gene expression.
Why it matters: Frameshifting elements are essential for many RNA viruses and represent underexplored antiviral targets. The demonstration that ASOs can reprogram RNA structures rather than simply block them opens a new dimension of RNA-targeted therapeutics.
Why for Yiru: RNA structure-function relationships and computational modeling of biomolecular dynamics are broadly relevant to understanding gene regulation, RNA-based therapeutics, and the structural biology of noncoding elements.