Recursion OS: The Heart of
Our Operations
Interested in our public data sets and models? Visit RxRx.ai.
Getting
Started
Step 1
Step 2
Step 3
Step 4
Step 5
Step 6
Step 7
Step 8
Step 9
Step 10
Step 11
Step 12
Step 13
The Journey of an Experiment
Generating one of the largest relatable data sets in pharma is only possible with Recursion’s automated high throughput labs. We’ve generated ~36 PB of proprietary data across phenomics, transcriptomics, proteomics, ADME, and InVivomics. Click through to follow a single experiment from the start of the process to the end.
Our labs can do up to 2.2 million samples per week.
The Journey of an Experiment
Central to our mission is the Recursion Operating System (OS), a platform powered by one of the world’s largest proprietary biological and chemical datasets. Instead of looking narrowly at a handful of diseases with existing therapeutic hypotheses, we build Maps of Biology and Chemistry that broaden our search and allow us to explore unknown areas of disease biology.
Our labs can do up to 2.2 million samples per week.
Tissue Culture
First, cells are cultured in large batches in our Tissue Culture labs. To focus on scalability, we’ve developed innovative methods for growing, freezing, thawing, and experimenting with these cells in large quantities. Our first maps are in HUVEC and NGN2 neurons; we’ve produced hundreds of billions of each.
We’re probably one of the largest, if not the largest, producers of HUVEC cells in the world. We can create over 100 billion cells per year for our high throughput experiments.
Tissue Culture
First, cells are cultured in large batches in our Tissue Culture labs. To focus on scalability, we’ve developed innovative methods for growing, freezing, thawing, and experimenting with these cells in large quantities. Our first maps are in HUVEC and NGN2 neurons; we’ve produced hundreds of billions of each.
We’re probably one of the largest, if not the largest, producers of HUVEC cells in the world. We can create over 100 billion cells per year for our high throughput experiments.
CRISPR at scale
In order to run experiments, we need to intervene in the cell (i.e. perturb it) in order to mimic diseases or to test treatments. The primary way we model diseases on our platform is by knocking out a gene’s function with CRISPR-Cas9 editing. We systematically combine the cells with the programmed CRISPR guide set for each gene, one at a time. We also introduce slight variations in the guides for each gene and we do many instances or replicates of each combination so that we create a more robust experimental signal for each gene.
CRISPR At Scale
In order to run experiments, we need to intervene in the cell (i.e. perturb it) in order to mimic diseases or to test treatments. The primary way we model diseases on our platform is by knocking out a gene’s function with CRISPR-Cas9 editing. We systematically combine the cells with the programmed CRISPR guide set for each gene, one at a time. We also introduce slight variations in the guides for each gene and we do many instances or replicates of each combination so that we create a more robust experimental signal for each gene.
Cell Seeding
Cells are seeded into experiment plates. These are essentially grids of miniature test tubes, with each plate containing 1536 miniature test tubes called ‘wells’. Each well will ultimately contain a unique experimental condition with some unique combination of cells and a reagent or condition that we are testing in that well.
Cell Seeding
Cells are seeded into experiment plates. These are essentially grids of miniature test tubes, with each plate containing 1536 miniature test tubes called ‘wells’. Each well will ultimately contain a unique experimental condition with some unique combination of cells and a reagent or condition that we are testing in that well.
Compound management
This automated process is repeated for thousands of plates. In different runs, our platform screens whole genome CRISPR knockout, or we profile millions of compounds in the same manner to compare phenotypes and transcriptomes.
All of the reagents for each experiment are stored in an automated storage system that directly integrates with the liquid transfer and plate automation work cells.
Automated Compound Management
This automated process is repeated for thousands of plates. In different runs, our platform screens whole genome CRISPR knockout, or we profile millions of compounds in the same manner to compare phenotypes and transcriptomes.
All of the reagents for each experiment are stored in an automated storage system that directly integrates with the liquid transfer and plate automation work cells.
Incubation
The plates move through various stages of the assay while the cells take on the effects of the perturbation, getting washed, new cell food or ‘media’ is added, the plates are incubated. This all takes a few days.
Incubation
The plates move through various stages of the assay while the cells take on the effects of the perturbation, getting washed, new cell food or ‘media’ is added, the plates are incubated. This all takes a few days.
Cell Imaging
Then high content microscopes take pictures of each well. This can be done at different points along the assay, giving us a longitudinal signal for how the cells change over time as the reagent coexists in the cell environment. We use Brightfield imaging which gives us the most information about the experiment over time.
Our phenomics images come off of our platform at a rate of millions per week.
Cell Imaging
Then high content microscopes take pictures of each well. This can be done at different points along the assay, giving us a longitudinal signal for how the cells change over time as the reagent coexists in the cell environment. We use Brightfield imaging which gives us the most information about the experiment over time.
Our phenomics images come off of our platform at a rate of millions per week.
Trekseq
After imaging, plates are transferred to our next high throughput platform: Transcriptomics, which give us an understanding of gene expression. The plates are processed again with barcodes that bind to each mRNA, giving it a unique identifier. The transcriptomics barcodes can then be ‘read’ by an instrument called a sequencer, resulting in a large data file that is representative of the well’s unique transcriptome
We are one of the largest transcriptomics data producers in the world.
Trekseq
After imaging, plates are transferred to our next high throughput platform: Transcriptomics, which give us an understanding of gene expression. The plates are processed again with barcodes that bind to each mRNA, giving it a unique identifier. The transcriptomics barcodes can then be ‘read’ by an instrument called a sequencer, resulting in a large data file that is representative of the well’s unique transcriptome
We are one of the largest transcriptomics data producers in the world.
Deep learning & AI
All data from transcriptomics and phenomics are analyzed through a series of mapping transformations where they are embedded by our AI models into a mathematical space, allowing us to calculate metrics about and between each perturbation.
Deep Learning & AI
All data from transcriptomics and phenomics are analyzed through a series of mapping transformations where they are embedded by our AI models into a mathematical space, allowing us to calculate metrics about and between each perturbation.
Industrialized Workflows
Together these data are used to trigger actions in our industrialized workflows. They build our maps of biology and in those maps we discover relationships which we can further test in our labs.
Industrialized Workflows
Central to our mission is the Recursion Operating System (OS), a platform powered by one of the world’s largest proprietary biological and chemical datasets. Instead of looking narrowly at a handful of diseases with existing therapeutic hypotheses, we build Maps of Biology and Chemistry that broaden our search and allow us to explore unknown areas of disease biology.
Design
From the hits identified, synthesis-aware generative AI workflows are used to design optimized drug candidates to meet target candidate profiles and remove likely liabilities. Models are used to score and rank potential molecules, with active learning used to select panels of the most informative compounds to make.
Design
From the hits identified, synthesis-aware generative AI workflows are used to design optimized drug candidates to meet target candidate profiles and remove likely liabilities. Models are used to score and rank potential molecules, with active learning used to select panels of the most informative compounds to make.
Make
After hits are identified by our industrialized workflows, we select only the very best insights to pursue as potential programs. The molecule is optimized, tested for safety, and developed further into a potential medicine on our translation platform.
Make
After hits are identified by our industrialized workflows, we select only the very best insights to pursue as potential programs. The molecule is optimized, tested for safety, and developed further into a potential medicine on our translation platform.
Test
Next, compounds are tested using our automated biology assays. Molecules automatically pass through panels of increasingly more complex assays to confirm activity against their target and in a range of in vitro and cell-based assays. Multiple other parameters covering ADME, PK, and toxicity are also assessed.
Automating these processes reduces manual handling of experiments, decreasing the time and cost of the majority of biological assays by >75%.
Test
Next, compounds are tested using our automated biology assays. Molecules automatically pass through panels of increasingly more complex assays to confirm activity against their target and in a range of in vitro and cell-based assays. Multiple other parameters covering ADME, PK, and toxicity are also assessed.
Automating these processes reduces manual handling of experiments, decreasing the time and cost of the majority of biological assays by >75%.
Learn
Experimental values obtained from the testing of molecules are automatically fed back into our models to improve them. Iterative design cycles evolve molecules towards our goal, as we learn our way through the project. A final molecule is optimized and designed further into a potential medicine on our translation platform.
Learn
Experimental values obtained from the testing of molecules are automatically fed back into our models to improve them. Iterative design cycles evolve molecules towards our goal, as we learn our way through the project. A final molecule is optimized and designed further into a potential medicine on our translation platform.
Meet LOWE, Recursion’s LLM agent for drug discovery
LOWE is the next evolution of the Recursion OS. LOWE supports drug discovery programs by orchestrating complex workflows. These workflows chain together a variety of steps and tools, from finding significant relationships within Recursion’s Maps of Biology and Chemistry to generating novel compounds and scheduling them for synthesis and experimentation.
In June 2024, we announced that Bayer will be the first beta-user of LOWE.
Precision chemistry design
Using generative AI, predictive models, and experimentation, we design, synthesize, and test novel molecules that are optimized for efficacy, selectivity, safety, and bioavailability.
LOWE - an LLM agent for drug discovery
LOWE is an LLM agent that enables scientists and technologists to directly query the RecursionOS, identifying novel targets, generating novel compounds, and even scheduling compounds for synthesis and experimentation.
In June 2024, we announced that Bayer will be the first beta-user of LOWE.
MASSIVE COMPUTING POWER
Modeling biology and chemistry using one of the largest relatable data sets in pharma is made possible by Recursion’s Top 100 Super Computer. Click through to see what it took to build.
.png){kind=small}
Decoding biology to radically improve lives.®
Decoding biology to radically improve lives.®