Insilico Medicine Accelerates Drug Discovery Using Amazon SageMaker

Founded in 2014, Insilico harnesses AI and deep learning techniques to identify promising therapeutic targets and design therapeutics with optimized properties. In traditional drug discovery, researchers must perform numerous iterations of testing and synthesizing molecules to achieve successful results. Insilico streamlines the initial stages by employing ML to design strong candidate molecules for further evaluation in wet labs. “Our technology solutions can consider far more parameters than a human could when designing a molecule manually,” says Daniil Polykovskiy, senior director of technology at Insilico. “For each step, we have automated processes or ML models that assess the feasibility of a candidate.” Currently, the company has thirty programs in development, including five in phase 1 clinical trials and one in phase 2.

Behind the scenes, Insilico’s rapid prototyping team develops and trains a diverse range of new models to continuously enhance its ML-based solutions. This team consistently explores new frontiers and collaborates with other Insilico teams to provide models that are tailored to specific research needs. To facilitate this process, the team requires access to GPU resources, which it previously obtained through a different cloud provider. However, the previous architecture relied on static virtual machines and Kubernetes clusters that necessitated manual scaling, leading to inefficiencies—sometimes hardware was underused, whereas, at other times, teams had to wait for resources to become available.

 “We needed a streamlined approach to build models without focusing on hardware or implementation details,” says Polykovskiy. “We also required a centralized environment where teams could access sufficient resources, collaborate, train and evaluate models, and share results seamlessly.” To achieve this, Insilico began migrating to Amazon SageMaker in 2024.

On Amazon SageMaker, Insilico reduced the time required to iterate and deploy a new ML model from 50 to 3 days­­, thus accelerating the process by over 16 times. The company’s previous ML pipeline, which was developed in-house, involved sequentially trying various models, training them with different parameters, and selecting the ideal one for deployment. Using Amazon SageMaker, Insilico can perform these steps in parallel, automatically scale up resources for model testing, and access cutting-edge GPUs and CPUs as needed. “The whole training pipeline is more efficient on Amazon SageMaker,” says Polykovskiy. “We can rapidly scale our experiments and select the backend and hardware that we need for each model.” The rapid prototyping team deploys updated models every 2 weeks rather than every few months, a roughly 83 percent time reduction. By accelerating the model-release cycle, Insilico can advance scientific discoveries at a rapid pace.

Insilico has also reduced the time and effort that are required of its DevOps team to maintain and scale infrastructure. Previously, requesting and deploying new resources through the DevOps team took around 1 week, but using Amazon SageMaker, teams can independently deploy and scale resources as needed. “The teams know their resource budget, so they know what they can train and when, and they don’t need to wait for resources to be allocated by another team,” says Polykovskiy.

The rapid prototyping team at Insilico uses the scalable resources in Amazon SageMaker to train new foundation models and test models alongside other Insilico teams. “Using Amazon SageMaker, we have very flexible computing resources and can collaborate with multiple teams to run tests in parallel,” says Polykovskiy. In one example, the team received a hypothesis to test and delivered results—a new model that was trained on a specific dataset—within 1 day. Insilico has migrated half of its technology teams to Amazon SageMaker as of July 2024 and continues to migrate the rest. “The teams see a lot of value in Amazon SageMaker,” says Polykovskiy. “Everyone feels the need to move forward with more flexible access to compute.”