Research Radar — 2026-05-18
Methods & AI
Computational
Do Larger Models Really Win in Drug Discovery? A Benchmark Assessment of Model Scaling in AI-Driven Molecular Property and Activity Prediction
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.04.29.721568
drug discovery model scaling benchmark foundation models molecular property prediction cheminformatics
Summary: Systematically benchmarks whether larger pretrained foundation models and LLMs genuinely outperform compact cheminformatics models and graph neural networks in drug discovery. Evaluates 78 model-task pairs across 26 endpoints spanning ADME, toxicity, and bioactivity, comparing molecular foundation models, LLMs, GNNs, and classical fingerprints. Reveals that model scaling does not uniformly translate to better performance — compact, task-specific models often match or exceed large pretrained models, especially on bioactivity endpoints, challenging the scale-centric narrative in AI-driven drug discovery.
Why it matters: The assumption that larger models automatically perform better has driven massive investment in molecular foundation models. This benchmark provides the first systematic evidence that this assumption does not hold across drug discovery tasks, with important implications for how computational resources should be allocated in pharmaceutical AI — smaller, well-tuned models may be the smarter choice for many real-world applications.
Why for Yiru: Model benchmarking and critical assessment of AI scaling are directly relevant to understanding when and how to apply computational methods in biomedical research. The finding that larger isn't always better mirrors broader debates about model scaling versus task-specific optimization.
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 DNA-guided RNA targeting genome editing gene regulation
Summary: Engineers a DNA-guided CRISPR–Cas12 system that directly targets cellular RNA rather than DNA. Unlike conventional CRISPR systems that use RNA guides, this approach uses short DNA oligonucleotides as guides, enabling Cas12 to recognize and cleave specific RNA transcripts in mammalian cells. Demonstrates efficient RNA knockdown with minimal off-target effects, expanding the CRISPR toolkit to a new guide chemistry with distinct properties from RNA-guided systems including enhanced stability and simplified synthesis.
Why it matters: DNA guides offer significant practical advantages over RNA guides — they are cheaper to synthesize, chemically more stable, and less immunogenic. A DNA-guided RNA-targeting CRISPR system represents a fundamentally new modality that could accelerate therapeutic RNA-targeting applications and enable RNA manipulation strategies not possible with current tools.
Why for Yiru: CRISPR technology evolution is foundational to functional genomics and potential therapeutic applications in cancer. A new guide chemistry expands the design space for programmable RNA targeting, relevant to understanding gene regulation in the TME.
Multiplex networks-based directed graph neural network for cancer driver gene identification
PLOS Computational Biology Published 2026-05-14 research article DOI: 10.1371/journal.pcbi.1014275
cancer driver genes graph neural network multiplex networks multi-omics network biology
Summary: Proposes a directed graph neural network (GNN) that integrates multiplex biological networks — including protein-protein interactions, gene regulatory networks, and pathway databases — to identify cancer driver genes. The directed GNN architecture captures directional regulatory relationships that undirected networks miss. Outperforms existing methods on benchmark datasets, and the multiplex integration reveals driver genes not detectable from any single network alone, including candidates validated in independent cancer cohorts.
Why it matters: Cancer driver gene identification remains a fundamental challenge in precision oncology. Most methods use a single interaction network, missing the multi-layered nature of gene regulation. A multiplex GNN that integrates directional regulatory information across network types represents a more realistic model of how genes collaborate in cancer.
Why for Yiru: Network-based approaches to understanding cancer biology connect to systems-level interests in the TME. GNN architectures that integrate multi-omics data are methodologically relevant to spatial and single-cell analysis pipelines.
CASPULE: A computational tool to study sticker spacer polymer condensates
PLOS Computational Biology Published 2026-05-14 research article DOI: 10.1371/journal.pcbi.1014282
biomolecular condensates phase separation computational biophysics polymer physics sticker-spacer model
Summary: Presents CASPULE, an efficient computational pipeline for simulating biomolecular condensates using the sticker-spacer polymer model. The tool enables systematic exploration of how sticker/spacer architecture — number, strength, and arrangement of interacting domains — governs condensate material properties, dynamics, and composition. Includes a user-friendly interface for parameterizing and running simulations without deep computational expertise, bridging the gap between condensate biology and polymer physics.
Why it matters: Biomolecular condensates are now recognized as fundamental organizing principles in cell biology, implicated in transcription, signaling, and disease. Accessible computational tools for condensate simulation lower the barrier to entry for experimental biologists seeking to understand how specific protein architectures drive phase behavior.
Why for Yiru: Condensate biology is an emerging dimension of cellular organization that likely operates in the TME — from transcriptional condensates in cancer cells to signaling condensates at immune synapses. Computational tools for condensate simulation connect to broader interests in spatial organization of biomolecules.
Bio-BLIP: A Multimodal Architecture for Transferable Reasoning in Genomic Variant Interpretation
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.12.724740
genomic variant interpretation multimodal AI Q-former transfer learning clinical genomics
Summary: Introduces Bio-BLIP, a multimodal Q-former architecture designed for transferable reasoning across genomic variant interpretation tasks. The model integrates heterogeneous evidence — DNA sequence context, gene annotations, protein functional domains, and prior literature — through a shared multimodal representation space. Unlike task-specific fine-tuned models, Bio-BLIP demonstrates zero-shot and few-shot transfer to new variant interpretation tasks, including pathogenicity prediction and functional effect classification, without retraining.
Why it matters: Clinical variant interpretation currently requires laborious manual integration of diverse evidence types by trained curators. A multimodal AI that can reason across heterogeneous genomic evidence and transfer to new tasks could dramatically accelerate variant interpretation workflows and improve consistency in clinical genomics.
Why for Yiru: Multimodal AI architectures that integrate diverse biological data types are methodologically relevant to integrating multi-omics data in cancer research. The transfer learning paradigm connects to broader interests in building flexible computational tools for biomedical discovery.
CatIF-RL: Activity-Oriented Enzyme Sequence Design by Steered Inverse Protein Folding
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.11.724288
enzyme design inverse protein folding reinforcement learning denoising diffusion protein engineering
Summary: Presents CatIF-RL, a framework combining graph-based denoising diffusion inverse folding with reinforcement learning to design enzyme variants with enhanced catalytic activity. The inverse folding model generates sequences compatible with a target backbone, and RL steers generation toward activity-optimized variants by incorporating catalytic preference signals. Demonstrates significant activity improvements over wild-type enzymes across multiple families, creating variants that standard inverse folding alone would not discover.
Why it matters: Most protein design methods optimize for structural compatibility rather than function. CatIF-RL bridges this gap by explicitly steering sequence generation toward enhanced catalytic activity, representing a practical path toward AI-designed industrial and therapeutic enzymes with tailored functional properties.
Why for Yiru: Protein design with explicit functional optimization connects to interests in designing biologics and understanding structure-function relationships in molecular recognition — relevant to both therapeutic development and basic TME biology.
Biomedical discoveries
Biomedicine
Carboplatin with or without nivolumab in metastatic triple-negative breast cancer: a randomized phase II trial
Nature Communications Published 2026-05-16 research article DOI: 10.1038/s41467-026-73085-1
triple-negative breast cancer immunotherapy nivolumab carboplatin randomized trial chemotherapy
Summary: Reports results of a randomized phase II clinical trial evaluating carboplatin with or without nivolumab (anti-PD-1) in patients with metastatic triple-negative breast cancer. The combination of platinum-based chemotherapy with checkpoint inhibition is compared against chemotherapy alone, assessing progression-free survival, overall response rates, and correlative biomarker analyses. Provides evidence for the benefit of adding immunotherapy to a chemotherapy backbone in the metastatic TNBC setting, with exploratory analyses identifying potential predictive biomarkers of response.
Why it matters: TNBC remains the breast cancer subtype with the poorest outcomes and limited treatment options. This trial directly addresses whether combining PD-1 blockade with platinum chemotherapy improves outcomes over chemotherapy alone in the metastatic setting — a clinically actionable question that could influence standard-of-care decisions.
Why for Yiru: TNBC immunotherapy trials and their biomarker correlates are directly relevant to Boss's research focus on the tumor microenvironment and cancer immunotherapy. Understanding which patients benefit from checkpoint inhibition in breast cancer informs rational combination strategies.
Genomic, Clinical, and Spatial Predictors of Durable Response to BRAF/MEK Inhibition in BRAF-Mutant Melanoma
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.09.721157
melanoma BRAF targeted therapy spatial biomarkers durable response tumor microenvironment
Summary: Analyzes pre-treatment tumor samples from 155 BRAF-mutant metastatic melanoma patients to identify genomic, clinical, and spatial predictors of durable benefit from BRAF/MEK inhibition. Integrates whole-exome sequencing, transcriptomics, and spatial profiling of the tumor microenvironment. Identifies that specific immune-spatial architectures — including T cell proximity to tumor cells and tertiary lymphoid structure signatures — predict extended response to targeted therapy, independent of mutational burden. Suggests that pre-existing immune engagement within the TME gates the durability of targeted therapy benefit.
Why it matters: While most melanoma patients respond to BRAF/MEK inhibitors, the durability of response varies dramatically and biomarkers to identify long-term responders are lacking. This study provides the first integrated spatial-genomic predictor of durable targeted therapy benefit, which could inform treatment sequencing decisions between targeted therapy and immunotherapy.
Why for Yiru: Spatial biomarkers of therapy response and tumor-immune architecture are directly aligned with Boss's research interests. The integration of spatial profiling with genomic and clinical predictors exemplifies the multimodal approach needed to understand TME determinants of therapy outcome.
Unraveling lncRNA diversity at a single cell resolution and in a spatial context across different cancer types
Nature Methods Published 2026-05-14 research article DOI: 10.1038/s41592-026-03071-4
lncRNA single-cell spatial transcriptomics cancer noncoding RNA tumor heterogeneity
Summary: Develops a computational and experimental framework to profile long non-coding RNA (lncRNA) diversity at single-cell resolution and in spatial context across multiple cancer types. Integrates single-cell RNA-seq with spatial transcriptomics to map lncRNA expression patterns to specific tumor regions, cell types, and cellular neighborhoods. Reveals cancer type-specific and shared lncRNA programs, identifies lncRNAs that mark specific tumor cell states, and demonstrates that spatial lncRNA patterns add prognostic information beyond coding gene expression alone.
Why it matters: The noncoding genome — particularly lncRNAs — remains a largely unexplored dimension of cancer biology, especially at single-cell and spatial resolution. This study provides a foundational resource and methodology for incorporating lncRNAs into spatial cancer atlases, potentially revealing new biomarkers and therapeutic targets invisible to coding-gene-focused analyses.
Why for Yiru: Single-cell and spatial profiling of the noncoding transcriptome directly extends the toolkit for TME characterization. LncRNAs that mark specific tumor-immune cell states or neighborhoods could provide new axes for understanding TME heterogeneity.
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
genomic medicine population genomics Chinese population rare disease precision medicine Hong Kong
Summary: Reports findings from over 20,000 participants in the Hong Kong Genome Project (HKGP), spanning a rare disease cohort with suspected genetic disorders and a population cohort undergoing genomic screening. Provides a foundational resource for precision medicine in the Chinese population, identifying population-specific variant frequencies, novel disease-gene associations, and actionable pharmacogenomic variants. Demonstrates the clinical utility of population-scale genomics for both rare disease diagnosis and preventive genomic screening.
Why it matters: Most genomic medicine resources are built on European-ancestry populations, creating disparities in variant interpretation for Asian populations. The HKGP fills a critical gap by establishing a large-scale Chinese reference dataset, directly enabling more accurate genetic diagnosis and personalized medicine for billions of people.
Why for Yiru: Population genomics and precision medicine are relevant to understanding the genetic architecture of complex diseases including cancer susceptibility. The Chinese-specific reference data is particularly relevant given the demographics of the research community.
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 reasoning clinical AI large language models medical diagnosis
Summary: Develops a conversational diagnostic AI system with multimodal reasoning capabilities that can engage in clinical dialogue, interpret diverse data types including imaging, lab results, and patient histories, and generate differential diagnoses. Evaluated against physician performance on standardized clinical cases, the system demonstrates diagnostic accuracy approaching or matching specialists across multiple specialties. The multimodal architecture integrates vision, language, and structured clinical data through a unified reasoning framework.
Why it matters: Diagnostic errors affect millions of patients annually. A conversational AI that can reason across multimodal clinical data — rather than just process text or images in isolation — represents a step toward AI systems that mirror how clinicians actually diagnose: by integrating diverse information streams through dialogue and reasoning.
Why for Yiru: AI in clinical medicine and multimodal reasoning are broadly relevant to computational approaches in healthcare. The integration of diverse data types for decision support parallels challenges in integrating multi-omics and spatial data for biological discovery.
T cells compete via reverse MHC class I signaling at the synapse with dendritic cells to secure Golgi recruitment for activation
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.13.724773
T cell activation immunological synapse MHC class I reverse signaling dendritic cell IL-12
Summary: Reveals a previously unrecognized mechanism of T cell competition at the immunological synapse: T cells engage MHC class I on dendritic cells via reverse signaling to trigger polarized IL-12 secretion from DCs. Using live-cell imaging and functional assays, shows that T cells compete for this reverse signaling — the T cell that successfully engages DC MHC class I secures Golgi recruitment and IL-12 delivery to its own synapse, gaining an activation advantage over competing T cells. This competitive mechanism may explain how high-affinity T cell clones outcompete lower-affinity ones during immune responses.
Why it matters: T cell competition for DC-derived signals shapes the quality and clonal composition of immune responses, yet the molecular mechanisms remain poorly understood. This reverse MHC class I signaling mechanism adds a new dimension to immunological synapse biology and may explain why some T cell clones dominate responses while others fail to activate.
Why for Yiru: T cell activation mechanisms and immunological synapse biology are fundamental to understanding antitumor immunity. Competitive dynamics at the DC-T cell interface could influence which tumor-reactive T cell clones expand during immunotherapy, with implications for vaccine and cell therapy design.
Cross-disciplinary watchlist
Other Fields
Microtubule dynamics control the direction of cardiomyocyte growth
Science Published 2026-05-14 research article DOI: 10.1126/science.adz1970
cardiomyocyte microtubule cell growth cytoskeleton mRNA localization heart
Summary: Discovers that microtubule dynamics act as a molecular toggle controlling whether adult cardiomyocytes grow in length or width — two fundamentally different modes of cardiac growth with distinct functional consequences. Increasing microtubule stability drives cellular widening by redirecting mRNA export and translation along the cell width, while microtubule destabilization promotes elongation. This directional growth control is mediated through spatially coordinated mRNA transport and localized translation at specific subcellular domains, revealing how cytoskeletal dynamics encode spatial information for cell shape determination.
Why it matters: Pathological cardiac hypertrophy (widening) versus physiological growth (elongation) have opposite effects on heart function. Understanding the microtubule-based toggle that controls growth direction opens the door to therapeutic strategies that promote beneficial cardiac remodeling while suppressing pathological hypertrophy — a major goal in cardiovascular medicine.
Why for Yiru: The concept of cytoskeletal dynamics encoding spatial information for cell fate decisions is broadly relevant to understanding how cells integrate mechanical and biochemical signals. Similar microtubule-based spatial coding mechanisms may operate in immune cell polarization and migration within the TME.
Simultaneous single-cell profiling of the transcriptome and proteome
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.14.724921
single-cell multi-omics transcriptomics proteomics method development multi-modal profiling
Summary: Describes a workflow enabling simultaneous profiling of the transcriptome and proteome from the same single cell. Individual cells are isolated by automated dispensing into minimal MS-compatible lysis volume, followed by sequential mRNA capture and protein supernatant recovery for independent downstream processing — RNA-seq for the transcriptome and mass spectrometry-based proteomics for the protein fraction. Demonstrates correlated and anti-correlated mRNA-protein relationships across individual cells that would be invisible to either modality alone, revealing post-transcriptional regulation that shapes functional cell states.
Why it matters: mRNA levels explain only ~40% of protein abundance variation, yet most single-cell studies rely on transcriptomics alone. True single-cell multi-omics that captures both mRNA and protein from the same cell provides a more complete picture of cell state and reveals the extent of post-transcriptional regulation — a critical missing dimension in single-cell biology.
Why for Yiru: Multi-modal single-cell profiling is directly relevant to comprehensive TME characterization. Understanding mRNA-protein discordance in tumor and immune cells could reveal previously hidden functional states not captured by transcriptomics alone.
Exploring the relationship between vascular remodelling and tumour growth using agent-based modelling
PLOS Computational Biology Published 2026-05-15 research article DOI: 10.1371/journal.pcbi.1012967
agent-based model vascular remodelling tumor growth tumor microenvironment radiotherapy oxygen
Summary: Develops a multiscale agent-based model (ABM) investigating how mechanical interactions between proliferating tumor cells and surrounding vasculature affect oxygen supply, tumor growth dynamics, and radiotherapy response. The model extends existing tumor spheroid models by incorporating vessel deformation due to mechanical forces between vessel walls and tumor cells. Reveals that tumor-induced vascular compression creates spatially heterogeneous hypoxia that profoundly shapes both growth patterns and treatment response, with implications for optimizing radiotherapy fractionation schedules.
Why it matters: Tumor-induced vascular compression is a recognized but poorly quantified phenomenon in solid tumors. An ABM that mechanistically links tumor growth mechanics to vascular function and treatment response provides a framework for understanding spatial heterogeneity in therapy outcomes and could inform personalized treatment scheduling.
Why for Yiru: Agent-based modeling of tumor-vascular interactions directly addresses spatial TME biology — specifically how physical forces shape the metabolic and therapeutic landscape of tumors. The radiotherapy response component adds clinical translation relevance.
VesiclePy: A machine learning vesicle analysis toolbox for volume electron microscopy
PLOS Computational Biology Published 2026-05-14 research article DOI: 10.1371/journal.pcbi.1013499
volume electron microscopy vesicle analysis machine learning neuroscience image analysis toolbox
Summary: Presents VesiclePy, a machine learning-based analysis toolbox for detecting, segmenting, and quantifying synaptic vesicles and other small vesicular structures in volume electron microscopy (vEM) data. The tool combines deep learning-based detection with morphological classification to enable whole-neuron vesicle mapping at nanoscale resolution. Validated across multiple brain regions and sample preparation protocols, VesiclePy automates what previously required months of manual annotation.
Why it matters: Volume EM provides unprecedented views of cellular ultrastructure, but the data deluge far outstrips manual analysis capacity. Automated vesicle analysis tools unlock the ability to map complete vesicle distributions across entire neurons, enabling systematic studies of synaptic organization and neurotransmitter systems at scale.
Why for Yiru: Advanced imaging analysis and ML-based tool development are broadly relevant to spatial biology. The challenges of automating ultrastructure analysis in vEM parallel those in spatial transcriptomics and multiplexed imaging — making the ML approaches transferable across imaging modalities.
Membrane-anchored influenza neuraminidase vaccine drives human-like broadly protective B cell responses
bioRxiv Published 2026-05-15 preprint DOI: 10.1101/2026.05.13.724804
influenza vaccine neuraminidase broadly neutralizing antibodies B cell response universal flu vaccine germinal center
Summary: Develops a membrane-anchored, folding-domain-free neuraminidase (mNA) immunogen that elicits superior germinal center B cell and broadly protective antibody responses compared to soluble NA. In non-human primates, mNA immunization induces cross-reactive memory B cell responses targeting conserved NA epitopes, expanding clones with the DR motif that mediates broad protection across influenza subtypes. The membrane-anchored format better mimics native NA presentation on virions, improving the quality and breadth of the B cell response.
Why it matters: Current influenza vaccines primarily target the highly variable hemagglutinin head domain, requiring annual reformulation. Neuraminidase is more conserved and a promising target for universal flu vaccines. A membrane-anchored NA immunogen that elicits broadly protective responses in primates represents meaningful progress toward a vaccine that protects against diverse influenza strains.
Why for Yiru: Vaccine design principles — particularly how antigen format shapes B cell response quality — are relevant to understanding how to elicit effective antitumor immune responses. The germinal center biology and broadly neutralizing antibody concepts have parallels in designing cancer vaccines that target conserved tumor antigens.
Sleep chart of biological ageing clocks in middle and late life
Nature Published 2026-05-13 research article DOI: 10.1038/s41586-026-10524-5
sleep biological ageing epigenetic clock aging longitudinal epidemiology
Summary: Maps the relationship between sleep patterns and multiple biological ageing clocks using longitudinal data from large population cohorts in middle and late life. Identifies specific sleep traits — including sleep duration, timing, regularity, and quality — that are independently associated with accelerated or decelerated epigenetic ageing. Reveals a U-shaped relationship where both short and long sleep duration associate with faster biological ageing, and demonstrates that changes in sleep patterns over time track with changes in ageing clock measurements.
Why it matters: Sleep is a modifiable health behavior, yet its relationship to the fundamental biology of ageing has been unclear. Establishing that specific sleep traits are independently associated with epigenetic ageing clocks provides mechanistic insight into how sleep influences healthspan and suggests that sleep optimization could be a practical intervention to slow biological ageing.
Why for Yiru: Biological ageing and its modifiable determinants are relevant to understanding organismal and cellular aging processes that influence cancer risk, immune function, and treatment tolerance. The epigenetic clock methodology also connects to broader interests in molecular biomarkers of physiological state.