Allcyte

Challenge & Solution

In the start-up phase, the company was able to rely on institutional high-performance computing (HPC) resources. As analysis needs grew and Allcyte evolved into an independent company, image analysis was outsourced to Google Cloud Platform (GCP). Since 2017, Allcyte has analyzed millions of images of patient samples to determine the likely clinical efficacy of a large number of cancer treatments simultaneously -- prior to the launch of costly and risky clinical trials.

About Allcyte

Allcyte is a biotech company that combines advanced microscopy with Artificial Intelligence to advance cancer research in order to support new drug development. Allcyte's innovative Pharmacoscopy® platform enables the assessment of cancer drug efficacy at the single cell level. In collaboration with CLOUDPILOTS and Google, Allcyte implemented a flexible GPU cluster to implement the next generation of AI-powered high-throughput microscopy and image analysis

Allcyte was founded as a spin-off of the Research Center for Molecular Medicine (CeMM) of the Austrian Academy of Sciences. In the start-up phase, the company was able to draw on institutional high-performance computing (HPC) resources. As the need for analysis increased and Allcyte evolved into an independent company, image analysis was outsourced to Google Cloud Platform (GCP). Since 2017, Allcyte has analyzed millions of images of patient samples to determine the likely clinical efficacy of a large number of cancer treatments simultaneously -- before the start of costly and risky clinical trials. Supported by increasingly realistic, patient-centric models, Pharmacoscopy® has proven to be an effective tool in the early stages of the drug development process as well as precision medicine. 

In 2019, Allcyte's computational team was ready for the next challenge: how can Pharmacoscopy be applied beyond blood cancers to multi-faceted cancer indications? Disseminated tumor cells (DTCs) can occur during the course of multiple cancer indications, including pancreatic, ovarian, and lung cancer. However, while blood cancer cells are homogeneous and can therefore be analyzed using proven classical image analysis methods, DTC are highly morphologically mixed. This requires the precise recognition of complex objects in a high-throughput manner; a task for which modern region-based Convolutional Neural Networks (R-CNNs) are ideally suited. Developing these AI solutions for Pharmacoscopy® required rapid prototyping and deployment of a novel GPU-based analysis pipeline in the Cloud.

Allcyte's new DTC analysis pipeline was not suitable for traditional HPC environments due to the intermittently, demanding workload: An on-premises solution would have required maintaining excess capacity at all times. The Cloud-based Kubernetes solution allows data to be accessed directly from its location in the Cloud, while the machines used can scale flexibly in highly parallel operations without network bottlenecks. Automatic cluster scaling allows dozens of powerful GPU computers to be accessed during peak demands and reduced to a single machine during standby.

In addition to the performance requirements of the analysis, Allcyte has high demands on traceability and reproducibility. At the same time, continuously improved versions of AI models should be traceable. To keep track, Allcyte relies on integrated monitoring by Stackdriver. In addition, Cloud Build facilitates the development, integration, and deployment cycle so that a new version of the software on Github can be brought into a deployable form within minutes. As a result, every single analysis becomes traceable and is consistently reproducible down to the specific version of the software environment.