Sonrai accelerates cancer research using Amazon SageMaker AI

From the beginning, Sonrai has been all in on AWS. Sonrai benefits from the comprehensive tool set offered by AWS, taking opportunities to customize as needed. “If we were trying to build a lot of these features ourselves, we simply couldn’t,” says Matt Lee, director of AI and medical imaging at Sonrai.

The company needed a cloud solution that was secure, scalable, and robust enough to handle large datasets with confidential patient information. Its experiments often need to run in parallel and can be computationally intensive. To help its customers efficiently analyze data and make precision medicine discoveries, Sonrai used Amazon SageMaker for MLOps, which delivers high-performance production machine learning (ML) models quickly at scale, for its Sonrai Discovery solution. For example, using Amazon SageMaker AI with MLflow, which offers a managed, serverless MLflow capability, Sonrai helps its customers track models and data seamlessly. “We can seamlessly track multiple parallel experiments and serve up the really important results from each of those experiments using Amazon SageMaker AI with MLflow,” says Lee.

Because Sonrai’s customers submit results to regulatory bodies, all data analysis needs to be secure and traceable. Sonrai uploads data securely using Amazon Simple Storage Service (Amazon S3), which offers object storage built to retrieve any amount of data from anywhere. Using Amazon SageMaker AI with MLflow, Sonrai can track both its models and the data getting fed into the model, seamlessly capturing any data that changes during the data cleaning stage. (See figure 1 below.) “We’re really confident that everything we do is reproducible, not just by ourselves but by a third party if needed,” says Lee.

Sonrai keeps costs low using the performant, cost-effective AWS infrastructure for AI. During initial research, the company uses cost-efficient instance types. When it’s time to run an experiment, Sonrai scales up using a compute instance that returns results faster. Sonrai can also run and track experiments in parallel using MLflow to increase efficiency. “The scalability of AWS is really useful to a company like ours,” says Lee. “We wouldn’t be able to afford servers on premises that are as powerful as the ones AWS offers because they would be idle most of the time.”

Sonrai’s infrastructure is built entirely on AWS, which means the company doesn’t need a physical presence in a region to meet regulations like the General Data Protection Regulation. Sonrai can process data where it’s stored on AWS to serve its global customer base. “We can spin up an instance of our solution in a region of the world in a matter of hours,” says Lee. “We simply couldn’t do that with an on-premises infrastructure.”

Sonrai’s bioinformaticians, ML engineers, and AI specialists also save time by working together to generate results. Using Amazon SageMaker Studio, a single web-based interface for complete AI model development, Sonrai can make sure everyone is on the same page and results are replicable. “We’ve greatly reduced iteration time using AWS,” says Lee. “It makes collaboration seamless when we’re all working off the same data and writing reports to the same locations.”

Sonrai estimates that collaborating using Amazon SageMaker AI doubles team productivity because employees don’t need to wait for results. In a project with one customer working in early cancer detection, Sonrai processed 300 patient samples with nearly 9,000 features, looking for a biomarker that would predict a particular cancer. Using AWS, the team accelerated model deployment for cancer research, completing the final report within 6 months and developing a model containing key features to identify at-risk patients.

Sonrai is continuing to expand its use of Amazon SageMaker AI to other use cases, including foundation model usage. The company has built a pipeline to use pathology foundation models to aid in prognostic and classification tasks. Using Amazon SageMaker AI, Sonrai estimates that it will be able to process hundreds of images—gigabytes in size—in less than 1 hour. “If we tried to do that on premises, we wouldn’t have the scale to process the data in parallel,” says Lee. “Using AWS services, we have much quicker time to insight with large data.”