Transforming variant curation using Amazon Bedrock with Exact Sciences

Exact Sciences is committed to improving the lives of patients with rare diseases through prevention, early detection, and personalized treatments. To aid treatment recommendations, the company’s genetic experts use variant curation, a process by which they analyze the detected variants to determine if they are causing disease in the patient. This involves analyzing complex literature for supporting evidence, a time-consuming task that can impact test turnaround time.

With more than 100,000 cases per year, the company needed a solution that could intelligently analyze published literature to expedite the research process. “Time can be crucial for children in critical care settings,” says Benjamin Bluhm, director of AI research and data innovation at Exact Sciences. “We need to deliver high-quality answers as fast as possible.”

 

In exploring ways to consolidate information from research papers, clinical studies, inheritance patterns, and patient data, Exact Sciences became interested in the potential of generative AI. The company was already using secure services on AWS for its bioinformatics pipelines and data storage. So, when Exact Sciences decided to move forward with developing its first research and development generative AI solution, it turned to AWS.

Exact Sciences engaged the AWS Generative AI Innovation Center and its team of dedicated scientists, strategists, and AWS Partners. Working together, the AWS and Exact Sciences teams prioritized internal use cases by reviewing the company’s resource allocation at each step of the variant curation and phenome abstraction processes. They discovered that extracting insights from documents would greatly shorten the overall time to reporting. Moreover, the use case was a great fit for state-of-the-art generative AI approaches, so both teams were confident that the proof of concept would go into production.

The Generative AI Innovation Center worked closely with Exact Sciences’ scientists to develop a proof of concept that would meet the company’s quality and HIPAA requirements. The Generative AI Innovation Center also connected the company with subject matter experts who helped architect its solution with healthcare best practices top of mind.

To meet data security and compliance requirements, Exact Sciences built its solution on Amazon Bedrock, which is HIPAA-eligible and provides access to a wide range of large language models (LLMs). During the engagement, Exact Sciences tested several LLMs to assess their accuracy and capabilities. “We needed a model that was both predictable and easy to work with,” says Bluhm. When Anthropic released the latest version of its state-of-the-art model, Claude, Exact Sciences knew that Anthropic’s Claude in Amazon Bedrock would be ideal. “Claude outperformed all the other models in our evaluation,” says Bluhm.

Working together, the teams built a proof of concept for the Variant Curation Accelerator in just 6 weeks.

Instead of a geneticist beginning a time-consuming research process that could span days, the Variant Curation Accelerator automates this process. The solution looks for identified genes or variants and searches for relevant papers, which are run through Amazon Textract and analyzed using the LLMs in Amazon Bedrock. From there, the results are integrated into a user interface via API for human review. “Using Amazon Bedrock gave us a secure platform for multiple models, making it possible for sensitive information to be securely handled,” says Bluhm.

Exact Sciences spent approximately 1 year getting their tools production-ready. To aid integration into workflows, the company gave its geneticists the ability to interact with the solution while in development. Based on user feedback, Exact Sciences built guardrails that help uphold accuracy. For example, Exact Sciences implemented checks of whether phenotypic information gathered from the patient match those for a given gene in the literature and provided citations. “We display the source PDF and highlight the actual text that informed the model,” says Bluhm. “That has really resonated with our medical curation team and built trust in the solution.”

Since going live with the Variant Curation Accelerator, the company has seen at least a 30 percent reduction in research times,¹ helping get critical insights to clinicians and patients faster. Additionally, Exact Sciences has boosted its team’s case review capacity. “By freeing up some of the variant curation time, our genetic experts can spend more time focusing on the complicated cases—ones that require deeper analysis,” says Bluhm.

¹ PreventionGenetics. Data on file. August 2025. Marshfield, WI.

With its increased research capacity, Exact Sciences can now automatically look deep into functional studies from a magnitude of research and tie them to actual patients—a new capability for the company. “When those studies cite a 2-year-old with a rare set of conditions, and your patient is a 2-year-old with those conditions, that precision of clinical relevance really supports patient-centric diagnostics,” says Bluhm. “We can now do that at scale.”

In addition to being able to distill content from published research, Exact Sciences also developed solutions that aid discovery of patient conditions from unstructured clinical notes. Clinical notes are also run through Amazon Textract—a machine learning service that automatically extracts text, handwriting, and data from scanned documents—creating markdown text for the LLMs in Amazon Bedrock to interact with. “By streamlining the review process through highlighted clinical note text, we’re empowering more experienced geneticists and medical experts to focus on phenotype abstractions,” says Bluhm.

Looking ahead, Exact Sciences will continue to innovate with foundation models in Amazon Bedrock. The company’s goal is to provide accurate diagnostics to providers and their patients using the Variant Curation Accelerator’s framework to extract insights from medical literature. The company will increase the scope of its generative AI usage and plans to add agentic AI. “On AWS, we don’t have to worry about model access, security, and scalability,” says Bluhm. “We get to focus on the quality of the solution and the patient impact instead.”