biomarker extraction

This study addresses the challenge of effectively integrating heterogeneous proteomic data—specifically, whole-sample mass spectrometry (MS) and multiplexed protein array profiles—in pancreatic cancer research. To overcome the limitations of conventional approaches that naively concatenate multi-source features, the authors propose a novel model fusion framework that explicitly models and leverages the heterogeneity between data sources through a tailored integration strategy. This approach synergistically exploits the complementary strengths of each modality rather than treating them as homogeneous inputs. Experimental results demonstrate that the proposed method significantly outperforms both single-modality models and standard fusion baselines in pancreatic cancer classification, yielding substantial improvements in diagnostic accuracy. The work thus offers a principled and effective paradigm for integrating heterogeneous multi-omics data in biomedical applications.

Current cell state discovery relies on dimensionality reduction, visualization, and manual clustering interpretation; however, intra-cluster heterogeneity frequently compromises biomarker identification accuracy, resulting in high trial-and-error costs and poor interpretability. To address this, we propose a novel framework integrating Mixture-of-Experts (MoE) modeling with interactive visual analytics: the MoE model automatically learns nonlinear associations between cell subpopulations and gene biomarkers without imposing rigid clustering assumptions; concurrently, the visual interface enables biologists to iteratively formulate, test, and refine state hypotheses while incorporating domain knowledge to guide model optimization. Case studies on real single-cell datasets demonstrate that our approach significantly improves biomarker detection accuracy and biological interpretability, successfully aiding the discovery of novel cell states and reducing analytical uncertainty by 42% (per expert assessment) compared to conventional methods.

In clinical proteogenomics, converting raw multi-omics data into reliable, novel biological hypotheses remains a major challenge due to the lack of automated, interpretable frameworks. Method: We propose PROTEUS—the first fully automated hypothesis generation framework that uniformly models the scientific discovery process as an evolvable, interpretable research process graph. It integrates large language models, modular workflow simulation, graph neural network–based representation learning, and an automatic open-scoring mechanism to enable end-to-end analysis of heterogeneous high-throughput data. Contribution/Results: PROTEUS unifies exploratory analysis, statistical testing, and iterative hypothesis generation within a single graph structure, supporting open-science–driven autonomous discovery. Evaluated on 10 public clinical multi-omics datasets, it generated 360 hypotheses; external validation and automated assessment demonstrated significant improvement in the reliability–novelty trade-off. This advances general-purpose AI toward domain-specialized scientific discovery systems.

This study addresses the challenge of improving molecular subtype characterization and clinical outcome prediction in breast cancer by jointly modeling protein sequence semantics and quantitative expression levels. We propose a novel integrative framework that fuses protein sequence embeddings—generated by ProtGPT2—with transcriptomic or proteomic expression data to construct biologically interpretable, discriminative multi-omics features. The method combines ensemble K-means clustering, XGBoost classification, PPI network analysis, and feature importance ranking to enable fine-grained subtype stratification and mechanistic insight extraction. It identifies key protein modules—including KMT2C, CLASP2, and MYO1B—that coordinately regulate hormonal signaling, cytoskeletal remodeling, and drug resistance pathways. In survival prediction and biomarker status classification tasks, our approach achieves F1-scores of 0.88 and 0.87, respectively—significantly outperforming conventional expression-only baselines.

Molecular profiling for cancer diagnosis and therapy selection typically relies on costly, invasive genomic assays. This study addresses the need for non-invasive, cost-effective alternatives using routine hematoxylin and eosin (H&E)-stained whole-slide images (WSIs). Method: We develop a multitask AI system built upon Virchow2—a foundation model pretrained on 3 million WSIs—and introduce pathological representation disentanglement coupled with clinical annotation alignment to enable pan-cancer molecular biomarker prediction from H&E slides alone. Contribution/Results: Our model simultaneously predicts 80 molecular biomarkers across diverse cancer types (mean AU-ROC = 0.89), encompassing alterations in 505 genes, activity of five core signaling pathways, DNA repair deficiency, tumor mutational burden (TMB), microsatellite instability (MSI), and chromosomal instability (CIN). Validated on 38,984 patients and 47,960 H&E slides, it identifies histological correlates for 40 biomarkers and links 58 to clinically actionable therapeutic targets—advancing digital pathology–driven precision oncology and companion diagnostics.

This study addresses how spatial biologists can guide and validate complex tissue data analysis tasks executed by AI agents. Building upon the Claude Science agent, the authors employ contextual inquiry, formative pilots, and observational experiments to propose four key design directions: execution control, familiar views, source information transparency, and cross-environment accessible verification. The work reveals the epistemic mechanisms through which scientists rely on visual evidence to evaluate AI-generated results. Furthermore, it constructs a comprehensive empirical model of the analytical workflow encompassing both interactive control and verification. Ultimately, this research contributes a systematic design framework for human-AI collaborative scientific discovery, offering actionable insights into integrating intelligent agents within rigorous biological research practices.

This work addresses the critical barriers to deploying clinical-grade AI biomarker models in computational pathology—namely, the absence of standardized intermediate representations, provenance tracking, and reproducible evaluation frameworks. To overcome these challenges, we establish the first shared benchmark framework for computational pathology, leveraging The Cancer Genome Atlas (TCGA) cohort to provide structured intermediate representations, predefined data splits, trained models, and evaluation metrics, with cross-validation on an independent Memorial Sloan Kettering Cancer Center (MSKCC) cohort. The framework integrates pathology foundation models (PFMs) to extract features from H&E whole-slide images, combined with multiple instance learning, quality control metadata, spatial coordinate mapping, and OncoKB annotations. Across 33 tumor–biomarker tasks, a high-performing subset of eight tasks achieved mean AUROCs of 0.831 on TCGA and 0.801 on MSKCC, demonstrating cross-institutional stability and establishing a reproducible, comparable foundation for AI-driven biomarker development.

This study addresses the lack of existing multimodal integration methods that translate omics data into testable morphological hypotheses to guide evidence retrieval from histopathology images. The authors propose a closed-loop framework that, for the first time, maps DNA methylation and miRNA features into morphological intent vectors, retrieves relevant pathological regions from structured text via TF-IDF, and employs a cosine similarity gating mechanism to trigger a vision-language model for deterministic refinement. This approach establishes a lexically auditable retrieval-and-verification pipeline, reducing reliance on implicit semantic matching in embedding spaces. Evaluated on the TCGA-BRCA dataset, the method achieves new state-of-the-art performance across multiple clinical tasks—including ER, PR, and HER2 status prediction, molecular subtyping, and risk stratification—significantly outperforming current multimodal fusion and vision-language model baselines.

Current evaluations of bioinformatics agents overemphasize answer correctness while neglecting workflow auditability and scientific credibility. This work proposes a Function–Evidence–Validation (FEV) tri-dimensional evaluation framework centered on inspectable workflow trajectories, shifting the primary focus to workflow correctness for the first time. Through systematic literature review, trajectory analysis, and cross-domain benchmark mapping, the study comprehensively analyzes 109 agent systems and 28 evaluation resources across subfields including genomics, single-cell and spatial omics, and protein science. The findings reveal that while agents perform adequately in planning and execution, they exhibit significant deficiencies in reproducibility, traceability, external validation, and prospective experimental design. This research provides both theoretical grounding and practical guidance for developing transparent, auditable next-generation bioinformatics agents.

Clinical pathway modeling traditionally relies on manual design and struggles to adapt to disease variants and comorbidities. To address this, we propose a two-stage process mining framework: (1) automatic discovery of process models from electronic health records (using Synthea-simulated SARS-CoV-2 data) via algorithms such as Heuristic Miner; and (2) dynamic expansion of the clinical pathway knowledge base through conformance checking, enabling subtype- and comorbidity-aware fine-grained modeling. Our key contribution lies in tightly coupling process mining with iterative, feedback-driven knowledge base updates—balancing real-world practice diversity with model interpretability. Experimental evaluation demonstrates that our method achieves 95.62% AUC in pathway identification and 67.11% arc simplicity, significantly outperforming static modeling approaches.