Use Case: 

Artera leverages Union in its mission to personalize cancer therapy

Artera uses Union to power its AI-enabled predictive and prognostic cancer tests. Artera runs inference predictions on the Union platform by interpreting patient data and biopsy slides to help clinicians determine a patient’s risk level and whether the patient will benefit from a certain therapy.

About Artera

Following cancer diagnosis, patients and oncologists work together to decide on the course of therapy to most likely have the best outcome with the fewest side effects for the patient. Often, this choice is ambiguous and difficult to make.

Artera is a leading precision medicine company developing AI tests that personalize decision-making for cancer therapy. The company offers AI-driven tests, such as the ArteraAI Prostate Test, that assess a patient’s overall long-term risk of outcomes such as distant metastasis and prostate cancer-specific mortality and predict a patient’s likelihood of benefiting from a particular therapy. To accomplish this, Artera’s tests use a multimodal solution interpreting both the pathology slides and a patient’s clinical data. The ArteraAI Prostate Test is the first test that can predict therapeutic benefits for patients with localized prostate cancer, enabling physicians to make treatment decisions with more confidence. The team recently achieved a key milestone with Medicare for the test, which is now the first and only predictive test for therapy personalization recognized in the National Comprehensive Cancer Network® (NCCN®) Guidelines for prostate cancer.

Outgrowing a homegrown solution

Interpreting gigapixel-sized images (100,000 x 100,000) is no easy feat. Artera’s products are not powered by a single model but instead are a composite of many models in a larger inference pipeline.

Artera’s need for a workflow orchestration platform arose when the number of its production inference pipelines grew. Initially, a pipeline used a single container for packaging, distributing, and executing. Over time, this approach was not feasible as the number of pipelines increased. A single container could not meet the different resource requirements, manage dependencies well, or address the scalability needs of managing and deploying some pipelines without affecting others. Maintaining, updating, and debugging multiple pipelines in a single container led to challenges in managing environments between teams, iterating on new ideas, and maintaining production workloads. Moreover, as the AI inference container grew in size, it led to an increase in workflow cold start times, node uptime, and compute costs.

Artera decided to evaluate orchestration solutions. The team started with a comprehensive list of capabilities to evaluate. The key considerations were as follows:

  • Artera’s model development was already happening in Python, so a Pythonic solution was important 
  • Its infrastructure was broadly reliant on Kubernetes and potential solutions needed to be Kubernetes native
  • Its requirements for reproducible pipelines needed to span both production and critical aspects of development
  • The solution would ideally require as few engineering resources as possible to free up the Artera team to focus on building more models to support cancer therapy decision-making.

Finding a purpose-built solution

Artera narrowed in on Union as a potential solution with a short-list of other Pythonic- and Kubernetes-based orchestration solutions. Union’s Pythonic solution would help developers quickly onboard to the platform while its native Kubernetes support would fit well with Artera’s existing infrastructure solution. Union quickly rose to the top with superior capabilities that could help Artera’s development velocity, including faster developer onboarding and iteration cadence. Additionally, the Union team’s proven expertise in building and maintaining Flyte infrastructure and supporting users was compelling for Artera to use Union’s fully managed orchestration solution.

A partner that scales with us

Union’s type engine helped Artera get more reliable predictions, by minimizing incompatible data types leading to runtime errors or unexpected behavior. Reda Oulbacha, Machine Learning Developer at Artera, remarked, “Union’s type engine is something that many of its competitors don’t do as well.” He added, “It’s important for us to have these strong guarantees about what are the inputs and what are the outputs.” This capability helped Reda’s team with reliability, rapid iteration, and control costs.

The shared common components in its pipelines also made Union’s composability capability a key advantage. Reda remarked, “The strong composability with Flyte allows us to compose tasks and workflows flexibly, and that unlocks a lot of reuse components for us.” The composability aspect also helps Artera with reproducibility from reusing the same components. This increases iteration speed and control costs.

Union’s benefit extended to managing expensive compute resources efficiently. The granular, task-level control of resources enabled Artera to use graphics processing units (GPUs) for necessary components only. Combined with interruptible instances, the Artera team could balance performance and costs of compute resources. A 50% reduction in container image pull times helps use expensive GPUs resources efficiently. Additionally, the Artera team no longer had to specify and communicate GPU instance requirements to their platform team, as Reda noted, “With Union, we can do that directly in the code.” This eliminated the need to engage their platform team and cut down additional cycle time of code review, deployment, etc.

“With Union, we can co-locate pod specifications and compute resource requirements directly with our tasks and workflows. It reduces at least by 2x the time it takes to make changes to the compute resources for our workflows.”
— Reda Oulbacha, ML Developer

Artera also benefits from Union’s parallelization capability in feature extraction, both for efficient resource utilization and 50% reduction in processing time.

“And so I think the team support has been very responsive and has been amazing and very open to feedback. The Union team involved us in the development of features for the user story. And that was great. They also guide us and support our platform team on infra guidance and recommendations. Working with Union is like extending our team.”
— Reda Oulbacha, ML Developer

The Artera team found Union’s support particularly compelling, both for proactive feature development design partnerships and reactive support for platform and user issues. Hari Ananthakrishnan, Principal Engineer, commented about Union’s willingness to engage, evaluate, and implement a new cloud service important to their networking requirements. 

“We have had some asks around multi-region, and it's one thing that other providers have fallen short on. We’re already needing to scale across multiple geographic regions and multiple availability zones. Having a partner who can scale with us is mission critical”
— Nathan Silberman, VP of Machine Learning and Engineering

By leveraging Union, Artera has been able to operate with a leaner team, allowing them to focus on other tasks more directly related to their business. Union and Artera plan to deepen their engagement. Union’s simplified operations with global availability will enable Artera to develop in one location and deploy to multiple locations as it grows in the United States and expands to other locations.