How Orakl Oncology is using DINOv2 to accelerate cancer treatment discovery

Orakl Oncology, a spinoff from the renowned Gustave Roussy Institute in Europe, aims to accelerate cancer research and drug development by combining experimental, lab-based insights and machine learning. Their mission is to help researchers and developers identify effective therapies for cancer patients during clinical trials, streamlining the discovery process. To achieve this, Orakl Oncology conducts tests on lab-grown cancer cells, known as organoids, to simulate how drugs might perform on actual patients.

Thanks to the work of their academic collaborators from the Jaulin lab and CentraleSupelec, and as part of the RHU ORGANOMIC initiative, the team recognized that they needed quantitative solutions to interpret the vast amounts of imaging data that they were generating, so they sought out an efficient model that was fast and accurate. Meta’s DINOv2 stood out as the ideal choice, thanks to its ability to learn from vast image collections and empower high-performance computer vision models. While the researchers previously used models that were specialized for organoids, they found DINOv2 was more effective.

As an open source model, DINOv2 quickly made a major impact by saving the Orakl Oncology team time and increasing their efficiency, as they were able to quickly train DINOv2 on organoid images to more accurately predict patient responses in clinical settings based on lab data.

“The idea is to basically use images of these organoids to extract as much information as possible to make the prediction as accurate as possible,” says Gustave Ronteix, Co-Founder and CTO of Orakl Oncology. “So we’re moving from a qualitative description of images to something quantitative that you can then feed to your downstream models.”

Benefiting from the collective knowledge and contributions of the open source community, the team was able to quickly address early technical issues and overcome challenges that might have otherwise hindered their progress.

“DINOv2 outperformed other models with an accuracy improved by 26.8% compared to other techniques," says Leo Fillioux, a PhD student from the MICS lab, a joint collaboration between CentraleSupelec and the Université Paris-Saclay, and main contributor on the project.

The availability of DINOv2 has opened up new avenues for research by eliminating some of the more time-consuming aspects of the team’s work. For example, extracting relevant information from videos previously required labor-intensive analysis of individual frames or sequences. Now, with DINOv2, this data can be directly extracted from videos, allowing researchers to focus on downstream tasks.

Orakl Oncology quickly developed their platform, achieving in a short time what other biomedical technology organizations had taken years to build. “It shifts the focus away from the engineering and really enables us to go straight to the science, trying to find what is it that actually enables you to predict patient outcomes,” says Ronteix.