Research Radar — 2026-06-15

Summary: Presents a systematic evaluation of agentic AI systems — large language model-powered autonomous agents — for biological discovery tasks spanning both standard analytical pipelines and open-ended scientific reasoning. While LLM-based agents have demonstrated competence in executing structured bioinformatic workflows (differential expression analysis, pathway enrichment, variant calling), biological discovery rarely consists of deterministic pipeline execution alone. Heterogeneous datasets, noisy measurements, and the need for iterative hypothesis generation and multimodal evidence integration are the norm. This study benchmarks AI agents across a spectrum of biological discovery tasks in both autonomous (agent acts independently) and copilot (agent assists a human scientist) settings. Tasks range from analysing multi-omics cancer datasets and generating mechanistic hypotheses to designing follow-up experiments. The authors systematically evaluate agent performance across dimensions including hypothesis quality, reasoning depth, factual accuracy, and reproducibility. Key findings reveal that current agents excel at structured data analysis but struggle with open-ended hypothesis generation requiring domain-specific biological intuition — the copilot setting consistently outperforms fully autonomous operation. The study provides a framework for evaluating agentic AI in scientific contexts and identifies specific failure modes (confirmation bias, over-interpretation of noise, hallucinated gene functions) that must be addressed before autonomous AI scientists become viable.

Why it matters: The vision of AI scientists that autonomously conduct biological research — generating hypotheses, designing experiments, analysing data, and iterating — has captured the imagination of the field and attracted substantial investment. But systematic evaluation of what current AI agents can and cannot do in real biological discovery contexts has been lacking. This study provides the first rigorous benchmark of agentic AI for biology, revealing a significant gap between the promise and the reality. The finding that copilot settings consistently outperform fully autonomous operation is practically important: it suggests that the near-term value of AI agents in biology lies in augmenting human scientists rather than replacing them. The identification of specific failure modes — hallucinated gene functions, confirmation bias, over-interpretation — provides a concrete roadmap for improvement. This work is essential reading for anyone building or evaluating AI systems for biological research, as it establishes both the current capability ceiling and the benchmarks by which progress should be measured.

Why for Yiru: The TME research workflow spans exactly the kind of tasks evaluated in this benchmark: from structured data analysis (identifying differentially expressed genes, characterizing cell types, mapping spatial niches) to open-ended reasoning (generating hypotheses about immune escape mechanisms, designing validation experiments, integrating evidence from transcriptomics, proteomics, and imaging). The finding that AI agents perform well on structured workflows but poorly on hypothesis generation has direct implications for how these tools should be integrated into TME research — as copilots for data analysis tasks, with human oversight for interpretation and experimental design. The benchmark framework itself could be adapted to evaluate AI agents specifically for cancer biology tasks, potentially creating a TME-specific evaluation suite. The copilot paradigm also aligns with how computational biologists already use AI tools like coding assistants and literature search — this study provides evidence for when and how these tools add value versus when they introduce risk.

Summary: Introduces a diagonal integration method for spatial multi-omics that integrates omics data from different modalities measured on separate tissue sections — a common practical scenario where simultaneous multi-omics profiling on the same section is technically challenging or cost-prohibitive. Current spatial multi-omics technologies that co-profile multiple modalities on the same tissue section face significant barriers in experimental complexity, reagent compatibility, and cost. This has motivated diagonal integration approaches that computationally align data from separate tissue sections, but existing methods designed for single-cell data overlook spatial information, leading to unreliable cross-modality anchoring. The authors present a method that explicitly incorporates spatial coordinates and tissue architecture into the integration process, using spatial neighbourhood graphs to constrain cross-modality alignment. The method handles the common case where the paired sections are not pixel-perfectly aligned — it models tissue deformation and rotation between sections. Applied to spatial transcriptomics and spatial proteomics data from consecutive tissue sections, the method recovers known spatial gene-protein relationships that single-cell-based integration methods miss, and identifies spatially coherent multi-omic signatures of tissue microenvironments. Benchmarking against existing methods demonstrates superior performance in preserving spatial coherence while achieving accurate cross-modality mapping.

Why it matters: The gap between what spatial multi-omics technologies promise (simultaneous profiling of transcriptome, proteome, and metabolome on the same tissue section) and what is practically achievable in most labs is substantial. Diagonal integration — computationally bridging modalities measured on separate sections — is the pragmatic path forward for most researchers. The key insight of this work is that ignoring spatial information during integration is not just suboptimal but actively harmful: single-cell-based methods can create spurious cross-modality links that violate tissue architecture. By explicitly incorporating spatial constraints, this method produces integration results that are both more accurate and more biologically interpretable. This is important because many biological questions — how does gene expression relate to protein localization in tumour-immune interfaces, or how do metabolic gradients correlate with transcriptional states — require multi-modal spatial data that is often collected on separate sections.

Why for Yiru: TME research relies heavily on spatial multi-omics to understand how different cell types organize and communicate within tumours. In practice, most labs cannot perform simultaneous spatial transcriptomics and proteomics on the same section — they profile consecutive sections with different modalities. A robust diagonal integration method that preserves spatial coherence would be immediately useful for integrating the growing body of publicly available spatial data from different modalities. For example, one could integrate spatial transcriptomics data showing T cell exhaustion gradients with spatial proteomics data showing checkpoint ligand expression patterns — even if these were measured on different sections. The tissue deformation modelling is particularly relevant for clinical specimens where sections can warp during processing. This method could enable more comprehensive computational reconstructions of the TME by combining data from multiple published studies that used different spatial modalities on similar tumour types.

Summary: Develops an integrated antigen discovery platform combining a deep learning prediction model, pooled mammalian cell screening, and peptide-HLA-E binding validation to systematically identify cancer-associated HLA-E peptides. HLA-E is an attractive immunotherapy target because it is minimally polymorphic (unlike classical HLA-A/B/C, which vary enormously across individuals) and broadly expressed across human populations and cancer types — a single HLA-E-targeted therapy could potentially treat patients regardless of their HLA genotype. However, systematic discovery of cancer-associated HLA-E peptides has been constrained by sparse training data and the technical difficulty of HLA-E immunopeptidomics. The authors address this with MHC Attention, a deep learning model that predicts HLA-E peptide binding using an attention-based architecture trained on a combination of existing HLA-E binding data and transfer learning from the more abundant classical HLA class I data. The model is integrated with a pooled mammalian cell screening platform that experimentally validates predicted peptides at scale, and peptide-HLA-E binding assays confirm candidates. The platform identified multiple cancer-associated HLA-E peptides, including peptides derived from known tumour antigens and previously uncharacterized cancer-specific transcripts. The identified peptides are validated for their ability to elicit T cell responses, demonstrating the pipeline's potential for identifying bona fide immunotherapy targets.

Why it matters: The major limitation of T cell-based cancer immunotherapies targeting peptide-HLA complexes is HLA polymorphism: a therapy targeting a peptide presented by HLA-A*02:01 only works for the ~40% of patients who carry that allele. HLA-E, by contrast, is essentially monomorphic — nearly everyone has the same HLA-E molecule. This makes HLA-E an extraordinarily attractive target for "off-the-shelf" T cell therapies that could treat any patient regardless of HLA type. The technical barrier has been identifying which peptides HLA-E presents on cancer cells, because HLA-E immunopeptidomics is far more difficult than classical HLA class I. This platform — combining deep learning with high-throughput experimental validation — opens the door to systematic HLA-E antigen discovery. If cancer-specific HLA-E peptides can be reliably identified, they could form the basis of TCR-based therapies, bispecific molecules, or vaccines with unprecedented population coverage.

Why for Yiru: HLA-E is expressed in many tumour types and plays complex roles in immune regulation — it can both present antigens to T cells and inhibit NK cell killing through interaction with NKG2A. Understanding the landscape of HLA-E-presented peptides in the TME is directly relevant to tumour immunology. The deep learning model developed here could be applied to TME-specific transcriptomic data to predict which HLA-E peptides are presented in different tumour contexts. The transfer learning approach — leveraging abundant classical HLA data to improve predictions for the data-sparse HLA-E — is methodologically instructive and could be applied to other non-classical MHC molecules (HLA-G, MR1, CD1). For computational immunology, this work demonstrates how combining deep learning with experimental validation can overcome data scarcity in niche but important immunological problems.

Summary: Investigates how whole-genome duplications during early vertebrate evolution contributed to the expansion of brain cell-type diversity by comparing single-cell transcriptomes from five chordates spanning the vertebrate lineage: human, mouse, lizard, lamprey, and amphioxus. The complex brains of vertebrates contain more cell types than those of their closest invertebrate relatives, and two rounds of whole-genome duplication (WGD) occurred during early vertebrate evolution. However, it has been unclear whether the duplicated genes (ohnologues) directly facilitated cell-type innovation. Using single-cell transcriptomic atlases, the authors mapped the evolutionary origins of brain cell types and tracked how ohnologues contributed to cell-type-specific gene expression programmes. They report that many cell-type families defined by conserved core transcription factors in vertebrates do not show simple one-to-one orthology with invertebrate cell types — instead, WGD provided the raw genetic material for subfunctionalization and neofunctionalization that enabled new cell-type identities. Specific ohnologue pairs show complementary expression patterns across related cell types, consistent with subfunctionalization driving cell-type diversification. The study provides a genome-scale framework connecting deep evolutionary events to the cellular complexity of the modern vertebrate brain.

Why it matters: The origin of vertebrate brain complexity is one of the great questions in evolutionary biology. While whole-genome duplication has long been hypothesized as a driver of vertebrate innovation, directly linking ancient WGD events to specific cell-type innovations has been challenging. This study bridges evolutionary genomics and single-cell biology to provide the most direct evidence to date that WGD-derived ohnologues contributed to the diversification of brain cell types. The conceptual framework — using single-cell transcriptomics to trace how duplicated genes acquired new cell-type-specific functions — is broadly applicable beyond neuroscience to any organ system. It also provides a template for understanding how gene duplication events in other lineages (including cancer, where whole-genome doubling is common) may drive cellular diversification.

Why for Yiru: The evolutionary framework developed here — tracking how gene duplication enables new cell-type identities — has parallels in cancer biology. Whole-genome doubling is a common event in tumour evolution, occurring in ~30% of human cancers, and is associated with both worse prognosis and altered immune recognition. The analytical approaches used to trace ohnologue contributions to cell-type diversification could be adapted to study how genome doubling in cancer cells creates new transcriptional states, metabolic programmes, or immune evasion phenotypes. The comparative single-cell transcriptomics framework also provides methodological lessons for cross-species TME comparisons — understanding which aspects of tumour-immune interactions are conserved across species is essential for translating preclinical immunotherapy findings.

Summary: Introduces TopoMIL, a framework that extracts the representative topological structure of cell and tissue distributions in microscopic images and integrates these topological features into multiple instance learning (MIL) for diagnostic classification. In computational pathology, MIL has become the dominant paradigm for analysing whole-slide images, where a patient sample is represented as a bag of image patches (instances) and only the sample-level label is available for training. However, existing MIL frameworks largely overlook the spatial organization and topological structure of cells within tissue — features that pathologists routinely use for diagnosis (gland architecture, immune infiltration patterns, tumour-stroma interfaces). TopoMIL addresses this by computing persistent homology features from cell-graph representations of each image patch, capturing topological properties such as connected components, cycles, and voids at multiple spatial scales. These topological features are integrated into the MIL attention mechanism, allowing the model to attend to patches based on both their visual content and their role in the tissue's topological structure. The framework is evaluated on multiple diagnostic tasks across cancer types, demonstrating consistent improvement over standard MIL approaches, particularly for tasks where tissue architecture is diagnostically informative.

Why it matters: Computational pathology has made remarkable progress with MIL-based approaches, but the field has largely treated tissue images as collections of independent patches, ignoring the spatial relationships that are central to histopathological diagnosis. TopoMIL's integration of topological features — which capture multi-scale spatial organization in a mathematically principled way — represents an important conceptual advance. Persistent homology can capture features that are difficult to encode in standard convolutional neural networks: the connectivity of tumour cell clusters, the structure of immune infiltrates, the fragmentation of glandular architecture. The consistent improvement across multiple diagnostic tasks suggests that topology provides complementary information to visual appearance, and the framework is general enough to be integrated into any MIL pipeline. This work is part of a broader trend of bringing topological data analysis into biomedical machine learning.

Why for Yiru: The TME is fundamentally a spatial structure — the organization of tumour cells, immune cells, stroma, and vasculature determines immune access, metabolic gradients, and treatment response. Topological features that capture tissue architecture at multiple scales are directly relevant to TME analysis. For example, the topological structure of immune infiltrates (dispersed vs. aggregated, peripheral vs. intra-tumoural) is a known prognostic factor but is difficult to quantify with standard image analysis. TopoMIL could be applied to histological images from TME studies to extract topological features that predict immunotherapy response or identify spatial patterns associated with immune exclusion. The integration of topology with attention-based MIL also provides a template for building spatially aware models of the TME that go beyond cell-type proportions to capture tissue organization.

Summary: Explores the intersection of genomic language models and DNA sequence compression, benchmarking different tokenization strategies and producing an information-content map of the human genome through the lens of language model compressibility. Genomic language models — transformer models trained on DNA sequences to learn the statistical structure of genomes — are increasingly used for variant effect prediction, regulatory element identification, and sequence generation. A fundamental design choice is how to tokenize DNA: single nucleotides (A/C/G/T), k-mers of fixed length, or learned byte-pair encoding tokens. Each choice affects both model performance and the interpretability of the learned representations. The authors systematically benchmark tokenization strategies across compression efficiency (how well the model captures the statistical structure of the genome) and downstream task performance. A key contribution is an information-content map that visualizes which genomic regions are highly compressible (repetitive, simple sequence) versus information-rich (complex regulatory regions, coding exons) — providing a new lens on genome organization. The information-theoretic framework also reveals that different genomic features leave distinct signatures in language model representations, suggesting new approaches for genome annotation.

Why it matters: Genomic language models are becoming foundational tools in computational biology, but the field lacks systematic understanding of how tokenization choices affect what models learn. This study provides that understanding through the lens of compression — a principled information-theoretic framework. The finding that different tokenization strategies capture different aspects of genomic information has practical implications for model design: the best tokenizer for variant effect prediction may differ from the best tokenizer for regulatory element discovery. The information-content map is a conceptually elegant contribution that reframes genome annotation as a compression problem — regions that are hard to compress contain more information and are likely functionally important. This perspective could inform everything from variant prioritization to synthetic genome design.

Why for Yiru: Genomic language models are increasingly applied to cancer genomics for tasks like identifying driver mutations, predicting variant pathogenicity, and characterizing mutational processes. Understanding how these models represent genomic information is directly relevant to using them effectively for TME genomics. The information-content framework could be applied to tumour genomes to identify regions of unusually high or low genomic complexity — potentially revealing structural variants, mutational hotspots, or regions under selective pressure. The compression-based perspective also connects to the broader question of how much information is encoded in tumour genomes and how much of that information is accessible to language models — a question with implications for using these models in clinical settings.