From personal GPUs to production-grade protein prediction on AWS

A Healthcare & Life Sciences business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas.

For import-based models such as ESMFold, a researcher uploads a protein sequence to an Amazon S3 bucket. That upload triggers an AWS Lambda function which invokes a SageMaker asynchronous inference endpoint. Because the endpoint is asynchronous, the request enters an internal queue and the caller is told where to find the result rather than waiting on the line.

By the end of the engagement, the client has a scalable AWS environment that removes the compute ceiling of local hardware and supports continued development of its therapeutic platform. Acceptance is defined against a working SageMaker environment, validated permissions and a demonstrated end-to-end inference workflow, with ownership transferred to the client's own AWS account.

• Compute unlocked — On-demand GPU provisioning on SageMaker replaces constrained personal hardware, so sequence size and iteration speed are no longer capped by a single machine.
• Scalable inference — An asynchronous, auto-scaling endpoint processes large protein sequences and notifies the team on completion, keeping expensive GPUs busy only when there is work to do.
• Cost-aware storage — Data is managed across S3, EBS, EFS and FSx for Lustre, matching each workload to the right storage tier.
• Owned and governed — The platform runs in the client's own AWS account with proper roles and permissions, so the team controls both cost and data.
• Handover — Runbooks, Loom walk-throughs and a knowledge-transfer workshop leave the client's team able to operate and extend the environment.

With a production-ready SageMaker foundation in place, the client can scale its protein prediction work in step with its science, ready to take on larger models and new therapeutic targets.