GSK and Pfizer join a 100 million-cell AI data project

Parse Biosciences and Nomic Bio announced plans on October 29 to produce Cascade, a dataset intended to help drug researchers train AI models to predict immune responses. The project will combine measurements of gene activity in 100 million individual cells with measurements of proteins released by groups of cells into surrounding fluid. Parse’s announcement

The companies named GSK, Pfizer, Sanofi, Graph Therapeutics and Turbine as foundational partners. Parse reported partner commitments; neither announcement disclosed payments or individual contract terms. Nomic said data generation is underway. Researchers are being offered a data partnership while the companies build the resource. Partner commitments Production update

Cascade gives those prospective buyers a specific research proposition: compare what happens inside an immune cell with the signals it sends to other cells after treatment. That pairing could help scientists choose experiments and test computer predictions about drug responses. The announcements describe intended uses, without reporting a demonstrated improvement in drug development from Cascade itself. Project description

Gene activity and protein signals answer different questions

RNA sequencing measures molecules a cell produces as it uses genetic instructions. Messenger RNA carries instructions that cells use to make proteins. Reading those molecules helps scientists identify which genes are active and how activity changes between cells or conditions. The National Human Genome Research Institute describes the transcriptome as the collection of RNA sequences in a cell. Single-cell measurements preserve differences that an average across many cells can obscure. RNA sequencing Transcriptomes

Protein measurements add information about what cells produce and release. Gene activity alone cannot fully describe how proteins are made, broken down or secreted. A methods paper by Nomic and academic researchers explains why direct protein measurements are needed alongside RNA data. Protein measurement research

Parse will produce the RNA measurements in its GigaLab service laboratory. Nomic will use its Omni 1000 platform to measure more than 1,000 secreted proteins from matched samples. Matching the records by donor, treatment and time should let researchers examine how changes in cellular gene activity relate to later protein signals. Production and matching

The RNA records describe individual cells; the protein records measure fluid collected from groups of cells in matched samples. Nomic’s earlier blood-cell dataset calls this fluid the supernatant. Protein concentration means an amount measured in a given volume of fluid. It describes the group’s output, without assigning each measured protein to one sequenced cell. One protein profile can therefore match RNA records from many individual cells. Sample-level protein measurements

Nomic’s protein measurement method has a published experimental history. Its November 2025 Nature Methods paper reported measuring 191 proteins across 7,392 blood-cell samples, yielding about 1.4 million measurements. Researchers compared responses across six donors using 80 different cytokines, signaling proteins used by immune cells, alongside other stimuli. They measured protein concentrations after 24 hours and compared responses with unstimulated samples from the same donors. The study involved the platform’s developers, and its disclosures identify Nomic employment or ownership interests among several authors. Published method and disclosures

The cell count measures depth within a defined experiment

Nomic’s Cascade specifications list 12 healthy donors, six female and six male, with measurements at four, eight and 24 hours. The planned design covers at least 14 immune cell types and more than 300 treatment conditions per donor at each timepoint. Those conditions include immune-signaling molecules and drugs, used individually or in combinations. Study specifications

The 100 million figure counts individual-cell RNA profiles within that design. Nomic separately lists more than 11,500 secretome profiles and 12.2 million protein measurements. A secretome is the collection of proteins released by cells. Each protein profile therefore describes a sample containing cells, rather than every sequenced cell receiving its own protein measurement. What Cascade counts

That distinction affects what a buyer can study. Many measured cells can reveal variation among cell types within the sampled donors. The reference begins with healthy donors. Nomic describes disease samples and additional donors as ways to extend Cascade, rather than populations already included in the announced reference. A model intended for a particular disease would need evidence relevant to that setting. Reference population and extensions

Licensed access joins an existing supply of open data

Parse’s access page offers Cascade through a data partnership license, with launch-partner licensing open through December 11 and general-partner licensing afterward. Drugmakers and other researchers are prospective licensees; Parse and Nomic supply the measurements. Public materials do not disclose fees, delivery dates, license duration or how payments would be divided between the suppliers. The access page directs interested organizations to a contact form. Cascade access

Commercial data will sit alongside resources researchers can already obtain openly. Tahoe-100M’s maintained dataset card describes an atlas of gene activity across 50 cancer cell lines exposed to 1,100 small-molecule treatments. It provides data and metadata downloads under a CC0 dedication. Tahoe studies cancer cells; Cascade adds secreted-protein measurements and timed immune responses. Tahoe-100M dataset

Another data-generation effort illustrates who can finance an open resource. In January, Tahoe Therapeutics, Arc Institute and Biohub announced a joint program for more than 120 million single-cell data points across 225,000 drug-patient interactions. Each organization committed multiple millions of dollars, according to Arc. The plan provides for initial sharing among the three organizations and eventual public release, rather than an immediate public download. Joint funding and release plan

The suppliers have published measurement totals and a licensing timetable, but the reviewed access pages do not specify delivery file formats, experimental controls included with a release or the schedule for receiving its RNA and protein records. Published access information

Predicting a new response remains a separate test

Arc’s 2026 Virtual Cell Challenge tests a problem relevant to such purchases: whether models can predict responses in cellular settings they have never seen treated. Entrants must predict the effects of reducing particular genes’ activity in six cell lines, using measurements of their untreated state. Arc holds back experimental responses to judge the predictions. That task tests transfer to unfamiliar biology rather than reproduction of a model’s training examples. Challenge design

Nomic proposes a comparable evaluation opportunity within Cascade: researchers could leave some donors or treatments out of model training, then test predictions against those withheld measurements. Following protein responses over time would also let them check whether an early cellular signal predicts a later response. Those are potential experiments enabled by the design; neither supplier reports their outcome yet. Proposed model evaluations

December 11 is the next published access milestone. Before committing, prospective partners still need prices, delivery dates and a clear account of which measurements arrive with each release. Published access timetable