In sync with the JP Morgan Health Care Conference, Recursion Pharmaceuticals has unveiled Phenom-Beta, a deep learning model designed to transform cell microscopy images into meaningful biological representations. “Phenom-Beta allows you to take images of human cells from a microscope or other sources and turn them into mathematical representations of biology,” said Chris Gibson, CEO of Recursion, in an NVIDIA-hosted call with journalists.
Hosted on the NVIDIA BioNeMo platform, Phenom-Beta is the company’s first in a potential series of foundation models for external use. The name “Phenom-Beta” references both “phenomenal” and “phenomics.” The scientist Steven A. Garan coined “phenomics” in 1996 to refer to the study of phenotypes, the observable characteristics or traits of an organism influenced by genetic predisposition and environmental factors.
Recursion bets the new frontier in phenomics could echo the rise of genomics
Chris Gibson, Ph.D., CEO of Recursion Pharmaceuticals
“We think [phenomics] will become as big and exciting as genomics,” Gibson said. Referring to the platform as “a new layer of biology” in line with domains such as proteomics and metabolomics, Gibson said. In the past, tapping phenomics required building hundreds of millions of experiments to gather the requisite data to train models. “And now that Recursion has spent over a billion dollars building these tools, we want to be able to share some of what we’ve built and the foundation model we built with the rest of the industry,” Gibson said.
Recursion created Phenom-Beta to flexibly process cellular microscopy images into general-purpose embeddings at scales ranging from small projects to billions of images. The platform promises to shed light on a spectrum of cellular responses, particularly focusing on how cells react to a variety of chemical or genetic perturbations.
Trained on Recursion’s proprietary dataset comprising more than 2 million images, Phenom-Beta can process cellular images to reveal subtle phenotypic changes, often imperceptible to the human eye. This model was trained using the RxRx3 dataset, a publicly available dataset that contains 6-channel fluorescent microscopy images and associated deep learning embeddings. Despite being trained on a specific imaging assay based on Cell Painting, a technique that uses multiple fluorescent dyes to stain different components of the cell, Phenom-Beta is a channel-agnostic model. It therefore can be used on a variable number of channels and any channel order.
Recursion has several compounds in phase 1 and 2 studies.
With Phenom-Beta phenomics release, Recursion shares foundation, keeps house in order
While Recursion is opening up access to Phenom-Beta, it doesn’t mean that Recursion is giving away all of its secrets. “We still have access to over 50 petabytes of proprietary biological data that we’re using for our own internal programs at Recursion with five programs in the clinic now and then also in our partnerships with companies like Genentech and Bayer” Gibson said.
By publicly releasing Phenom-Beta, Recursion aims to “accelerate the sharing of other models” across the industry. Gibson believes that openly sharing their foundational model built on years of proprietary work will “move all of us forward faster.” Hosted on NVIDIA’s BioNeMo platform and DGX Cloud infrastructure, Phenom-Beta makes Recursion’s advanced approach more accessible to other organizations. As Gibson noted, this allows “academic groups perhaps — at some point in the future industry groups” to analyze smaller image datasets and still extract meaningful biological insights.
Filed Under: Cell & gene therapy, Data science, Drug Discovery and Development, Genomics/Proteomics, machine learning and AI
