Dana‑Farber Researchers Develop AI Assistant to Help Oncologists Navigate Precision Cancer Treatments

Researchers at Dana‑Farber Cancer Institute have created an AI-based oncologist’s assistant designed to help clinicians identify approved precision oncology therapies for individual patients. The tool, built using a retrieval-augmented large-language model (RAG‑LLM) combined with the Molecular Oncology Almanac (MOAlmanac), achieved up to 95% accuracy on synthetic queries and 93% on real-world queries submitted by practicing oncologists. The model is currently available online as a research tool and the team hopes it will eventually support faster and evidence-based treatment decisions in clinical settings.

The Challenge of Precision Oncology

Precision oncology now offers more than 100 therapies that target specific cancer-driving mutations. Keeping up with approvals for these therapies is difficult because new drugs emerge rapidly, and regulatory approvals vary by region. Clinicians report that tracking all the latest FDA decisions and guidelines requires significant time and effort. Without a streamlined method, oncologists face the risk of missing treatment options for patients who could benefit from targeted therapies.

Combining AI with Curated Knowledge

The Dana‑Farber team led by Helena Jun, PhD, and Eliezer Van Allen, MD, developed an AI assistant to bridge this gap. They combined a generative LLM similar to those used in ChatGPT with MOAlmanac, a database that links molecular alterations to potential therapies. This approach, known as retrieval-augmented generation (RAG), allows the AI to retrieve the most relevant, up-to-date information before generating a recommendation.

Jun explained that their goal was to create a model that can provide actionable guidance without producing unsupported, outdated or speculative suggestions. To achieve this, the researchers trained the model to understand precision oncology terminology and clinical decision-making. They experimented with different prompting strategies, retrieval settings, and model configurations to ensure that recommendations were grounded in evidence. By combining the AI model with MOAlmanac, the team reduced errors and improved the relevance of outputs compared with using an LLM alone.

According to the study published in Cancer Cell on January 15, 2026, the RAG‑LLM was tested using both synthetic queries and real-world queries from practicing oncologists. The model achieved 95% accuracy on synthetic test sets and 93% accuracy on real clinician queries, outperforming an LLM-only approach. The high level of accuracy shows that augmenting a general-purpose AI model with domain-specific knowledge significantly improves reliability in a clinical context.

Limitations, Safety Measures and Regulatory Considerations

Despite its strong performance, the model may still produce incorrect outputs, especially when no approved therapy exists for a specific biomarker. To prevent speculation, the team set parameters to limit creativity and ensure the model reports the absence of approved treatments rather than inventing alternatives. This approach helps maintain safety and reliability when providing guidance to clinicians.

The assistant allows users to select a region, such as the United States or the European Union, reflecting differences in regulatory approvals. This feature ensures that recommendations align with the region’s approved therapies. It also provides transparency for clinicians working in multinational or multi-site healthcare settings.

Implications for Oncology Practice

Dana‑Farber researchers emphasize that the AI assistant currently functions as a research tool rather than a standalone clinical decision system, while additional validation is required before incorporating it into routine oncology workflows.

The team plans to expand testing, gather feedback from more oncologists, and design clinical trials to measure whether using the AI assistant improves treatment decisions and patient outcomes. They are also exploring ways to combine this retrieval-augmented model with other AI techniques to enhance performance further.

Once integrated into clinical workflows, RAG‑LLM tools could reduce the cognitive burden on oncologists and help standardize access to the latest precision medicine data. Hospitals and vendors should consider incorporating curated knowledge bases, monitoring pipelines, and governance processes to maintain accuracy, safety, and regulatory compliance.

Author

Richard Chau