Agentic AI

Agentic AI is the computational backbone of AutopathX’s translational engine. By orchestrating multi-agent reasoning, multi-modal spatial omics, and therapeutic simulation across the tumor microenvironment (TME), our AI system moves beyond simple analysis — it thinks, tests, and iterates like a scientist. AutopathX Agentic AI dynamically models the cross-talk between key stromal and immune players, uncovering actionable patterns for precision immunotherapy.

From dendritic cells (DC) to macrophages, TAMs, CAFs, and TILs, our intelligent agents map and simulate immune dynamics across space, time, and treatment regimens. Powered by RAG-based systems, GPU compute, and expansive spatial datasets, Agentic AI doesn’t just follow — it leads translational discovery.

CRC TME-Guided Multi-Agent Framework: AIRE for Stratifying CAR T, TIL Synergies and Immunotherapy Resistance & Combinations

The effort showcases the conceptual foundation of the Agentic Immunotherapy Recommender Engine (AIRE)—a multi-agent framework designed to decode the colorectal cancer tumor microenvironment (CRC TME) and personalize therapy design. AIRE integrates spatial transcriptomics with domain-specific agents representing cellular subtypes such as dendritic cells (DCs), tumor-associated macrophages (TAMs), and cancer-associated fibroblasts (CAFs). Through collaborative reasoning among critic, synthesis, and manager agents, it analyzes key immune signaling motifs (CXCL10–CXCR3, CCL13–CCR2, ACTA2/TNC) and maps immune zonation defined by PD-L1 and IDO1 expression. A federated scoring rubric quantifies node-level relevance—including CXCR7 for DCs, ACTA2 for CAFs, and CCL13 for TAMs—by assessing spatial context and immune interaction strength, ultimately recommending actionable therapeutic strategies.

This agent-based federated graph approach transforms how immunotherapy candidates are prioritized, offering a scalable mechanism for precision decision-making across decentralized biomedical datasets. It enables dynamic integration of single-cell profiles, spatial histology, and clinical endpoints without compromising data ownership or interoperability. By harnessing the collective intelligence of specialized agents, researchers can simulate treatment synergies, identify resistance mechanisms, and stratify patients based on immune architecture—all essential for designing combination therapies that bridge CAR T, DC vaccines, TILs, and checkpoint blockade strategies. This framework marks a pivotal advancement in AI-guided immuno-oncology, paving the way for adaptive, context-aware precision interventions 

Adaptive Multi-Agent Architectures for Spatial Omics-Driven Immunotherapy in Colorectal Cancer Tumor Microenvironments

The framework presents an Adaptive Multi-Agent AI system engineered to transcend the inherent limitations of conventional multi-agent approaches, particularly in decoding the complexity of colorectal cancer (CRC) tumor microenvironments (TMEs). Unlike traditional agents that rely on rigid plan-execute cycles, AIME enables continuous strategy adjustment, domain-aware task decomposition, and real-time instantiation of specialized analytical agents. This dynamic adaptability is essential when navigating the heterogeneous, inflammatory landscapes of CRC TMEs, where evolving cell–cell interactions and spatial gradients present shifting diagnostic and therapeutic challenges.

The architecture integrates multimodal spatial omics platforms—Xenium spatial transcriptomics and CODEX spatial proteomics—through four sequential modules. The first module, Agent Specialization, aligns relevant spatial features (genes, proteins, tissue structures) to immunotherapeutic targets. Next, Dynamic Planning operationalizes therapeutic goals into context-aware subtasks such as niche classification and immune cell profiling. Actor Instantiation crafts tailored agents equipped with subtask-specific prompts and analytic tooling. Finally, Therapeutic Reasoning orchestrates collaborative inference among agents to generate transparent, traceable outputs—mapping immune niches, predicting treatment response, and identifying emergent spatial biomarkers across CRC samples.

This adaptive framework offers significant promise for translational oncology, providing a modular and scalable system capable of navigating the biological complexity inherent to immunotherapeutic decision-making. Its fluid architecture supports iterative reasoning over evolving spatial data, enabling high-resolution biomarker discovery and therapeutic prioritization. As immunotherapy strategies continue to diversify, particularly in the context of spatially resolved inflammation and cellular crosstalk, AIME stands as a powerful computational ally—driving precision diagnostics and augmenting therapeutic design within CRC TMEs.

Orchestrating CRC TME Intelligence: Multi-Agent RAG Meets Federated Knowledge Graphs for Immunotherapy Precision

The effort Illustrated here a forward-thinking architecture in which Multi-Agent RAG systems replace traditional Retrieval-Augmented Generation workflows to tackle the complexity of spatial omics-driven CRC TME analysis. Unlike conventional RAG, which relies on a single LLM for retrieval and response synthesis, Multi-Agent RAG distributes tasks among specialized agents—enhancing interpretability, modularity, and precision. This framework enables agents to retrieve and reason across multi-modal spatial data, such as CODEX imaging, single-cell maps, and immunogenomic signatures. A Federated Knowledge Graph (FKG) built collaboratively by these agents allows integration without compromising patient privacy, harmonizing ontologies like GO, CL, and UMLS to reflect dynamic relationships between genes, cell types, and spatial niches.

Within this orchestration, various agents have domain-specific roles. The Manager Agent delegates tasks like TAM–CAF niche identification or checkpoint overlay. The Planner Agent breaks down objectives into retrieval, synthesis, and validation stages. Specialized Retriever Agents source data from Visium HD, TIMEDB, Human Cell Atlas, and TISCH (Tumor Immune Single-cell Hub), each bringing contextual evidence for signaling motifs or cell–gene proximities. The Tool Creation Agent integrates spatial analysis pipelines like Squidpy and Seurat, while Tool Agents execute them for cell–type deconvolution or cytokine pattern extraction. A Synthesizer Agent builds reasoning chains (e.g. CXCL10 → IDO1 → PD-L1), and a Critic Agent ensures biological plausibility by mapping terms to ontologies like GO:0006955 and CL:0000905, flagging inconsistencies or hallucinations before KG updates.

The figure contrasts system prompts with context engineering—highlighting adaptive context yields superior results in multi-agent reasoning. System prompts are rigid, predefined, and ill-suited for dynamic TME inference. In contrast, context engineering introduces clear, modular task flows suited to agent orchestration. Prompt example (shown) initiates analysis of DC–TAM zones, identifying CXCL10–IDO1 signaling and overlaying LZTS2–PD-L1 checkpoints. Another summarizes how DCs and macrophages shape the immune landscape and immunotherapy targets within CRC TME. These context-aware prompts yield biologically coherent outputs that reflect spatial co-localization, molecular cross-talk, and agent consensus—enhancing reproducibility and interpretability.

Ultimately, this multi-agent, immuno-trained workflow unlocks new dimensions in TME diagnostics. Agents orchestrate across spatial omics layers, integrating cell–cell proximity with signaling axes. The Federated Knowledge Graph architecture captures nuanced immunosuppressive or inflamed zones, built from distributed contributions that respect semantic diversity and data privacy. This enables deeper understanding of the dendritic cell–macrophage interplay, TAM–CAF co-localization, and their impact on T cell exhaustion and therapeutic resistance. Biomarker discovery, patient stratification, and immunotherapy target selection become more robust—reflecting the complexity of cytokine signaling, cell plasticity, and therapeutic modulation across CRC microenvironments.

Immune-Centric Agentic Orchestration for Adaptive Colorectal Cancer Diagnostics

The figure articulates a sophisticated, AI-enabled framework for agentic diagnostics in colorectal cancer (CRC) immunotherapy, integrating spatial transcriptomic profiling with multimodal orchestration of specialized diagnostic agents. Built atop platforms like CODEX and Xenium, the architecture features a federated agentic reasoning core that harmonizes real-time decision-making across pathology, genomics, clinical trial coordination, tumor microenvironment (TME) stratification, and dendritic cell (DC) surveillance. Each agent operates within a modular, human-in-the-loop interface optimized for regulatory deployment, clinical transparency, and diagnostic agility.

Initiating the workflow, the pathology agent undertakes high-resolution spatial segmentation of CRC tissue, extracting regions of interest (ROIs) by identifying lymphocyte-enriched niches through quantitative immune density mapping and morphometric stratification. These annotated ROIs serve as foundational inputs for downstream reasoning agents, informing therapy eligibility, immune phenotyping, and trial readiness. Through real-time tissue parsing, the pathology agent ensures deep visibility into immune infiltration patterns that inform both prognostic stratification and therapeutic targeting.

The DC agent builds on this substrate by deconvoluting heterogeneous dendritic cell populations within the CRC TME. It classifies DC subtypes—cDC1, cDC2, pDC, migratory, and LAMP3+ variants—using marker-driven overlays and proteomic features, reconstructing their functional topology across antigen presentation, cytokine trafficking, and chemokine signaling. Spatial mapping of these subtypes enables the modeling of DC–TAM cross-talk networks, where mutual signaling cascades influence immunotherapy outcomes and TME remodeling. The agent’s extract further highlights DC subtype localization, tumor-type specificity, and role adaptation under inflammatory or immunosuppressive pressure.

Agentic coordination expands in the adaptive cell therapy layer, where orchestrated integration of signaling data is executed across CAR-T, CAR-NK, NK cells, TILs, and DCs. Multi-agent processing pipelines ingest spatial and single-cell omics, reconstructing immune pathway hierarchies and mapping co-expression networks relevant to cell therapy design. Specialized agents invoke logic frameworks to stratify therapy zones and extract actionable markers, adapting to tissue heterogeneity, immune activation thresholds, and cellular proximity scores. Proprietary tools facilitate seamless parsing of spatial proteomes and transcriptomes, transforming heterogeneous molecular inputs into therapy-eligible diagnostic outputs.

This unified agentic architecture represents a translational leap in CRC immunotherapy by facilitating precision diagnostics that are adaptive, scalable, and biologically contextualized. Its capacity to parse multimodal spatial data, coordinate reasoning across agents, and generate interpretable outputs aligns with clinical imperatives around personalized medicine, trial optimization, and biomarker-driven treatment selection. By embedding explainable AI principles and modular task allocation, the system advances the diagnostic interface from static annotation to dynamic reasoning, offering a deployable blueprint for next-generation oncologic decision support.

Agentic AI-Driven Zonation Mapping of Hepatic Injury in DILI and MASH

Depicted herein is a federated multi-agent AI architecture designed to resolve overlapping hepatic injury phenotypes in Drug-Induced Liver Injury (DILI) and Metabolic dysfunction-associated steatohepatitis (MASH). This integrative framework leverages spatial transcriptomics, histopathology, and single-cell RNA sequencing to enable autonomous agents to classify fibrosis, steatosis, necrosis, and ballooning degeneration across hepatic zonation. Agentic orchestration across modalities facilitates high-resolution mapping of injury trajectories, particularly within midlobular zone 2—a nexus of regenerative and fibrotic signaling.

DILI, though statistically rare, poses high regulatory risk due to its association with post-market drug withdrawals. MASH, conversely, exerts a substantial economic and clinical burden, progressing toward cirrhosis and hepatocellular carcinoma. The convergence of these pathologies necessitates precision diagnostics capable of disentangling shared and distinct molecular signatures.

Spatial omics platforms such as 10x Genomics Visium and NanoString CosMx enable zonation-resolved profiling, revealing differential expression of markers including GLUL, FGG, and SAA1. Deep learning applied to histopathology images identifies MASH-specific features—ballooning, bridging fibrosis, steatosis, and necrosis—with high fidelity, forming a curated training corpus for large language models deployed on MCP servers to support scalable, explainable diagnostics.

DILI drugs stratification via dose-dependent clustering reveals hepatotoxic signatures across agents such as acetaminophen, rifampicin, and imatinib. Single-cell RNA sequencing from control, MASH, and DILI samples elucidates endothelial and immune cell dynamics, highlighting shared midlobular LSEC markers—STAB2, OIT3, DNASE1L3—informing agent training and orchestration for nuanced interpretation of injury states and therapeutic windows.

Collectively, this agentic and generative AI framework redefines computational hepatology by integrating spatial, histological, and transcriptomic intelligence. It advances mechanistic insight, early detection, and personalized intervention in DILI and MASH, while establishing a foundation for auditable, multi-agent systems aligned with clinical, regulatory, and translational imperatives.

Agentic GenAI Framework for Multimodal Stratification of MASH Liver Pathology on HIPAA-Compliant BioMCP Infrastructure

Recent advances in agentic artificial intelligence have enabled the development of modular, multimodal systems capable of autonomous reasoning across complex biomedical domains. In the context of liver pathology, a multi-agent orchestration framework has been constructed to identify and interpret histological features associated with MASH (Metabolic dysfunction-associated steatohepatitis). Specialized agents trained on high-resolution histopathological inputs demonstrate proficiency in detecting fibrosis, ballooning degeneration, steatosis, and bridging fibrosis, with bounding-box confidence metrics guiding spatial localization. These agents operate within a privacy-preserving architecture that integrates spatial transcriptomics, single-cell RNA sequencing, and genomics, thereby facilitating cross-modal inference while maintaining clinical interpretability.

Feature extraction is achieved through generative AI algorithms that segment histological inputs into discrete pathological entities, which are subsequently clustered using hierarchical machine learning pipelines. These pipelines stratify image tiles into biologically relevant categories—such as cirrhosis, early fibrosis, and normal parenchyma—enabling robust classification and annotation. The agentic framework is underpinned by a cognitive loop encompassing perception, planning, reasoning, and execution, supported by modular toolchains and memory architectures. This design allows agents to iteratively refine their outputs in response to multimodal inputs and contextual feedback, thereby enhancing diagnostic precision and scalability across heterogeneous datasets.

To ensure regulatory compliance and clinical applicability, the orchestration layer is deployed on BioMCP, a HIPAA-compliant cloud infrastructure optimized for PHI-sensitive workflows. BioMCP supports literature synthesis, variant interpretation, and clinical trial matching, positioning it as a translational bridge between computational pathology and precision medicine. The identification of MASH liver injury features as diagnostic biomarkers holds significant promise for immunotherapeutic stratification, early disease detection, and longitudinal monitoring. By aligning agentic AI with regulatory frameworks and clinical imperatives, this architecture exemplifies a scalable paradigm for responsible AI deployment in biomedical research and diagnostics.

Agentic AI-Driven Responsible LLM Evaluation Benchmarks for MASH Liver Phenotyping

The proposed architecture and implementation of autonomous multi‐agent systems is pivotal for the in-silico recapitulation of complex hepatic pathophysiology, exemplified by metabolic dysfunction–associated steatohepatitis (MASH). Within this framework, discrete agents dedicated to fibrosis progression, hepatocellular ballooning, steatosis, and inflammatory signaling operate in concert to simulate the multifactorial progression across fibrosis stages. By integrating multimodal endpoints—including histopathological imaging, scoring, and spatial transcriptomic profiles—this architecture enables high‐resolution mapping of disease trajectories and facilitates systematic interrogation of intercellular crosstalk and microenvironmental influences.

The establishment of rigorous benchmark suites is instrumental for the iterative refinement of biomedical large language models (LLMs). Domain‐specific tasks such as factual accuracy validation, reasoning over curated knowledge graphs, and multimodal data fusion delineate clear performance objectives. Additionally, privacy‐preserving summarization of protected health information and clinical question–answering benchmarks enforce operational safety and compliance. End‐to‐end evaluation pipelines quantify model competence in knowledge retrieval, analytic plan execution, and concordance with gold‐standard annotations, thereby driving incremental enhancements in precision and interpretability.

Positioning agentic AI as a horizontal knowledge substrate underscores its transformative potential across clinical, research, pharmaceutical, commercial, and consumer‐facing digital verticals. In each domain, agentic workflows automate complex processes—from real‐time data ingestion and hypothesis generation to decision‐support synthesis—while preserving auditability and regulatory traceability. Evaluation dimensions encompassing robustness, transparency, ethical alignment, and domain‐specific efficacy enable systematic comparison of agentic approaches in applications ranging from genomics discovery to therapeutic development and clinical decision support.

The convergence of multi‐agent architectures with stringent LLM evaluation protocols charts a course toward transparent, reasoning‐driven, and hallucination‐resistant AI in biomedicine. By enforcing quantifiable benchmarks and continuous performance monitoring at each agentic node, this methodology cultivates stakeholder confidence and regulatory alignment. Ultimately, such a framework accelerates translational research, optimizes clinical workflows, and fosters responsible deployment of AI‐driven innovations in healthcare.

A Modular Framework for Agentic AI in Histopathology: From Foundation Models to Autonomous Liver Disease Diagnostics

The figure presents a modular framework for integrating foundation models and agentic AI into computational pathology, with a focus on liver histopathology. While architectures such as ViT-B/16 and DINOv2 serve as high-performance vision encoders trained via self-supervised learning (SSL) on gigapixel whole-slide images (WSIs), they are best understood as backbone models. In contrast, pathology-specific foundation models—such as GigaPath and CONCH—are pretrained across diverse histological and textual corpora, enabling generalization across diagnostic tasks. These models capture fine-grained morphological features and contextual patterns, supporting downstream applications in disease classification, biomarker discovery, and treatment response prediction. Their scalability and multimodal adaptability position them as core components in translational research and high-throughput clinical workflows.

In the context of metabolic dysfunction-associated steatotic liver disease (MASLD) and its progressive phenotype, MASH (Metabolic dysfunction-associated steatohepatitis), foundation models are leveraged to classify histological features such as steatosis, ballooning, inflammation, and fibrosis. By integrating multimodal data—including WSIs and clinical metadata—these models facilitate phenotype stratification and support precision diagnostics. The use of multimodal AI agents enhances interpretability and diagnostic accuracy.

Emerging capabilities of agentic AI in pathology include autonomous slide interpretation, multimodal integration across histology, radiology, and genomics, and agentic reasoning for planning and decision-making. However, limitations persist in current large language models (LLMs) and vector embeddings, including static representations, sparse multimodal fusion, limited temporal reasoning, and computational bottlenecks in real-time environments. These constraints hinder the deployment of fully autonomous agents in clinical workflows.

To address these gaps, the transition from foundation models to agentic AI requires modular architectures that separate perception, reasoning, and action. Fine-tuning with domain-specific instructions, real-time optimization, and federated learning are essential to enable scalable, privacy-preserving deployment. Explainability mechanisms such as saliency maps and natural language rationales are critical for clinical trust and adoption.

Overall, the integration of foundation models and agentic AI holds transformative potential for MASH liver pathology and related hepatic disorders. By enabling scalable, interpretable, and autonomous diagnostic workflows, these technologies can accelerate biomarker discovery, improve patient stratification, and support therapeutic decision-making in hepatology and beyond.

Agentic AI for Histopathology: Learning Steatotic Features in MASH and DILI

Metabolic dysfunction–associated steatohepatitis (MASH) presents a diagnostic challenge because its hallmark features—steatosis, hepatocyte ballooning, and inflammation—often overlap with injury patterns seen in drug‑induced liver injury (DILI). This convergence complicates clinical interpretation and underscores the need for intelligent systems that can disentangle subtle histological signatures.

In this work, a histopathology agent is trained through deep learning to recognize steatotic features in MASH liver tissue. By learning layered representations—from detecting fat droplets displacing hepatocyte nuclei to quantifying their zonal distribution—the agent builds a robust understanding of disease morphology. This layered learning enables it to distinguish canonical MASH pathology from confounding DILI patterns.

Extending beyond image analysis, the framework integrates foundation large language models to orchestrate outputs into an autonomous, agentic AI system. This allows the agent not only to detect and classify steatosis but also to contextualize findings within broader clinical and pharmacological data. The result is a scalable, explainable platform that advances precision diagnostics at the intersection of MASH and DILI.

Decoding Tumor Microenvironment in Liver Cancer (HCC) through Agentic AI: Integrating Histopathology and Spatial Omics for Immunotherapy Design

Understanding the tumor microenvironment (TME) in hepatocellular carcinoma (HCC) is critical for advancing immunotherapy and precision oncology. Cellular cross-talk within the TME—particularly between malignant hepatocytes, immune infiltrates, and stromal components—drives resistance, immune exclusion, and therapeutic response. Key signaling pathways such as Wnt/β-catenin, TGF-β, JAK/STAT, and VEGF orchestrate these interactions, shaping the immunological landscape and influencing disease progression. Agentic AI offers a transformative approach by integrating multimodal data—histopathology, spatial transcriptomics, and single-cell RNA sequencing—to model these complex dynamics with high resolution and interpretability.

This integrative framework leverages federated knowledge graphs to unify spatial and molecular features across tissue compartments, enabling the elucidation of immune evasion mechanisms and fibrosis-linked progression. Annotated histological slides and fibrosis staging provide morphological context, while spatial overlays and transcriptomic signatures reveal immune exclusion zones and pathway activation states. Agentic AI agents interact with foundation models to interrogate biological networks, stratify patients by immune phenotype, and generate actionable insights for targeted immunotherapy. This approach not only enhances mechanistic understanding of HCC but also establishes a scalable blueprint for AI-driven discovery in complex tumor ecosystems.

Causal Inference-Driven Multi-Agent Reasoning in CRC Tumor Microenvironment: Decoding Immune Dynamics for Precision Immunotherapy

Causal inference offers a principled framework for uncovering mechanistic relationships within the colorectal cancer tumor microenvironment (CRC TME), especially when applied to multimodal spatial omics data. By moving beyond correlation, it enables the identification of directional signaling motifs and cellular interactions—such as dendritic cell (DC) recruitment via CXCL10–CXCR3, or immunosuppressive feedback through TAM–CAF–T cell circuits. In spatially resolved datasets like CODEX and Xenium, causal modeling helps disentangle overlapping immune niches, stratify inflamed versus excluded zones, and simulate the impact of therapeutic interventions. This is particularly critical in CRC, where the dynamic interplay between DCs, CAFs, and TAMs governs immune infiltration, checkpoint expression, and resistance mechanisms. Causal inference thus serves as a computational lens to decode the functional architecture of the TME and guide precision immunotherapy design.

To achieve biologically coherent causal graphs, the integration of context engineering prompts and domain expertise is essential. Context-aware prompts allow agents to dynamically tailor their reasoning to spatial features, immune phenotypes, and therapeutic goals—unlike static system prompts that constrain adaptability. Domain-specific agents, such as Ontology Mapper and Critic, embed biomedical priors from ontologies like MONDO and GO, ensuring that inferred relationships reflect validated biology. This orchestration across specialized agents enables modular decomposition of complex tasks, such as mapping DC–TAM signaling or CAF-mediated exclusion zones and supports counterfactual reasoning for intervention modeling. In this framework, causal inference becomes not just a statistical tool but a biologically grounded strategy for understanding disease progression, predicting therapeutic response, and designing context-sensitive interventions across heterogeneous CRC microenvironments.