SAGE: Agentic Framework for Interpretable and Clinically Translatable Computational Pathology Biomarker Discovery

TL;DR

SAGE addresses the core gap in computational pathology by unifying knowledge-grounded biomarker discovery with interpretability and clinical feasibility. The framework builds a pathology-aware knowledge graph, generates testable hypotheses via a coordinated agent ensemble, and validates them through an automated, validation-first pipeline that translates hypotheses into executable analyses on real cohorts. Key contributions include multi-path ontological reasoning, a novelty-evaluating debate framework, and a validation loop that ties image-derived features to molecular and clinical endpoints. Demonstrated on bladder cancer, SAGE identifies biologically interpretable biomarkers and shows end-to-end feasibility, closing the loop from literature grounding to empirical validation with scalable, auditable workflows.

Abstract

Despite significant progress in computational pathology, many AI models remain black-box and difficult to interpret, posing a major barrier to clinical adoption due to limited transparency and explainability. This has motivated continued interest in engineered image-based biomarkers, which offer greater interpretability but are often proposed based on anecdotal evidence or fragmented prior literature rather than systematic biological validation. We introduce SAGE (Structured Agentic system for hypothesis Generation and Evaluation), an agentic AI system designed to identify interpretable, engineered pathology biomarkers by grounding them in biological evidence. SAGE integrates literature-anchored reasoning with multimodal data analysis to correlate image-derived features with molecular biomarkers, such as gene expression, and clinically relevant outcomes. By coordinating specialized agents for biological contextualization and empirical hypothesis validation, SAGE prioritizes transparent, biologically supported biomarkers and advances the clinical translation of computational pathology.

Figures (7)

• Figure 1: Overview of the SAGE framework for computational pathology. The pipeline comprises three stages: (1) Knowledge graph construction and reasoning (bottom left), integrating pathology-aware biomedical entities to support structured biological reasoning; (2) Hypothesis generation and refinement (top), where coordinated agents produce interpretable hypotheses linking image-derived pathology features with molecular and clinical endpoints; and (3) Validation and summarization (bottom right), which executes statistical analyses and generates clinically interpretable summaries. Figure \ref{['fig:sage_tools']} in the supplementary material details the pathology-aware tool orchestration subsystem.
• Figure 2: Biomedical Knowledge Graph Visualization. Expert-curated biomedical knowledge graph used for hypothesis generation in computational pathology. Nodes represent biomedical entities (e.g., genes, diseases, pathways, phenotypes), and edges denote semantic or mechanistic relationships mined from literature and medical ontologies. Colors indicate entity types (see legend). A zoomed-in subgraph highlights bladder cancer--related entities and dense multi-scale connectivity enabling context-aware reasoning.
• Figure 3: Pathology-informed overall survival analysis for joint FABP5--TLS stratification in TCGA-BLCA. Patients with high FABP5 expression and scarce TLS-like lymphoid aggregates on whole-slide pathology images exhibit significantly worse overall survival compared to patients with low FABP5 expression and abundant TLS-like aggregates.
• Figure 4: Overview of the AI-Enabled Literature Processing and Knowledge Graph Construction Workflow. Visualization of a three-phase workflow for AI-enabled data collection and processing: Phase 1: Keyword Expansion using OpenAI, MESH, and Semantic Scholar APIs, Phase 2: Large Scale Literature Crawling from high-impact sources, and Phase 3: Automated Article Processing and Storage via text extraction prioritization with optimized compression, for preparing the dataset for knowledge graph construction.
• Figure 5: Overview of the Tool Orchestration Subsystem. MCP-based tool orchestration subsystem enabling unified access to clinical and computational tools across oncology and histopathology suites. The MCP Orchestrator mediates between a coding agent/researcher and distributed client–server tools by analyzing prompts, selecting appropriate capabilities, and executing them over shared patient data and whole-slide images.