OncoAgent: AI Decision System and Agent Workflow for Glioblastoma

Glioblastoma is the most aggressive type of brain tumor, and its molecular subtypes are closely related to treatment selection and prognosis. Traditional subtype classification relies on expert interpretation of genomic data, which is time-consuming and subjective. OncoAgent is a production-grade AI medical decision system that combines deep learning classification, RAG evidence retrieval, and agent workflow to build an end-to-end solution, providing a new technical paradigm for precision oncology.

Glioblastoma is the most aggressive type of brain tumor, and its molecular subtypes are closely related to treatment selection and prognosis. Traditional subtype classification relies on expert interpretation of genomic data, which is time-consuming and subjective, creating an urgent need for more efficient and objective AI solutions.

System Architecture

Adopts a modular design, with core components including PyTorch deep learning classifier, RAG evidence retrieval pipeline, agent workflow engine, FastAPI service layer, and Docker containerization deployment solution.

Deep Learning Model

Trained on TCGA bulk RNA sequencing data to capture non-linear correlations between genes; classification accuracy and generalization ability are improved through cross-validation and hyperparameter optimization.

Agent Workflow

The main agent decomposes tasks, while sub-agents perform subtasks such as data preprocessing, model inference, and evidence retrieval, supporting flexible orchestration and error handling.

Experiment Tracking

Integrates MLflow to record hyperparameters, metrics, and model versions, meeting collaboration, iteration, and compliance requirements.

To address the interpretability issue of classification results, OncoAgent integrates a RAG pipeline. When outputting subtype predictions, it automatically retrieves relevant research literature, clinical trial data, and diagnosis/treatment guidelines, providing credible knowledge sources for decision-making and enhancing clinicians' trust.

Production Deployment

FastAPI builds high-performance inference services; Docker enables environment isolation, supporting horizontal scaling and operation monitoring functions.

Data Security

Uses data desensitization, encrypted transmission, fine-grained access control, and audit logs; containerization ensures tenant data isolation.

Clinical Integration

Supports standardized data interfaces and medical information exchange standards, enabling seamless integration with hospital HIS/LIS systems and outputting structured diagnostic reports.

OncoAgent demonstrates the paradigm of integrating multiple AI technologies into a clinical decision system. With the accumulation of multi-omics data and model improvements, such systems will play a greater role in tumor diagnosis, treatment recommendation, and other links, ultimately improving patient treatment outcomes.