Recursion released a preprint on applying deep-learning-driven analysis of cellular morphology to develop a scalable “phenomics” platform. The preprint demonstrates the capabilities of Recursion’s platform to model complex immune biology and screen for new therapeutics.
Over the following months the resulting disease, subsequently named COVID-19, spread across the rest of the world and was declared a pandemic by the World Health Organization on March 11th, 2020. The COVID-19 pandemic has impacted millions worldwide and has the potential to cause a worldwide recession. At the current moment, there are no available vaccines.
Recursion, a digital biology company industrializing drug discovery, conducted several experiments in April 2020 to investigate therapeutic potential for COVID-19 from a library of FDA-approved drugs, EMA-approved drugs or compounds in late-stage clinical trials for modulation of the effect of SARS-CoV-2 on human cells. The resulting experiments were then compiled into the RxRx19a dataset, which is composed of 305,520 images and corresponding deep learning embeddings at nearly 450 gigabytes of data. RxRx19a provides the largest publicly available set of human cellular morphological data to researchers around the world who are working to make advances in the fight against the COVID-19 pandemic.
The cells were then fixed, stained and imaged at 96 hours post-infection. African green monkey kidney epithelial cells (Vero) were also infected as a control condition. The HRCE and Vero cells both demonstrated robust phenotypes compared to the mock and irradiated controls. HRCEs were selected for further high-throughput screening due to their disease relevance and robust disease-specific phenotype.
Chemical Suppressor screens were conducted by treating HRCE cells in six half-log doses with six replicates per dose for each compound approximately two hours after cell seeding (concentrations tested may vary for certain reference compounds studied). At 24 hours post-seeding, cells were infected with SARS-CoV-2 and incubated for 96 hours until fixation, staining and imaging. Recursion then evaluated 1,672 compounds in HRCE and referenced compounds in both HRCE and Vero using fluorescent microscopy images of five channels that illuminate different organelles of the cell. Images were processed using Recursion’s proprietary deep learning neural network to generate high-dimensional featurizations of each image for the identification of distinct phenotypic profiles.
RxRx19a is the first morphological dataset that demonstrates the rescue of morphological effects of COVID-19. Through RxRx19a, researchers in the scientific community will have access to both the images and the corresponding deep learning embeddings to analyze or apply to their own experimentation. The embeddings are 1024-dimensional vectors with one vector for each image and come from Recursion’s internal model trained on additional cell types and perturbation modalities. We provide these embeddings to more easily enable researchers without significant compute resources to still explore and uncover insights from this data. Scientific researchers can use the data to further demonstrate how high-content imaging can be used for compound efficacy screening. Results and conclusions drawn from the in vitro experiments and targeted hypothesis-driven research will contribute to the growing body of scientific data in the fight against COVID-19.
The RxRx19a dataset is highly similar in nature to RxRx1a, a dataset previously released by Recursion in June 2019, although there are some key differences. For ease of comparison and understanding, we provide the following table highlighting the primary differences:
This work is licensed under the Creative Commons Attribution 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/ or send a letter to Creative Commons, PO Box 1866, Mountain View, CA 94042, USA.
Note, that this license applies only to the RxRx19a dataset, not RxRx1.
A CSV containing the experiment design, e.g. what cell type and treatment are in each well. The schema is provided in the README.
A large CSV file containing all of the deep learning embeddings for each image.
1,527,600 8-bit PNG 1024x1024 images. These images are downsampled from the original 2048x2048 16-bit versions. The directory structure is explained in the README.