AI Drug Discovery Dataset Cascade Adds Protein Signals to 100M Cells

What was announced

On October 29, Parse Biosciences and Nomic Bio announced Cascade, a joint project to build a 100-million-cell dataset of human immune responses that can be used to train AI models.1 Parse, which QIAGEN owns, specializes in single-cell sequencing. Nomic is a Montreal company that measures proteins at high throughput.13 The dataset is meant to give drug developers detail on immune biology at both the cell level and the protein level, so they can find and validate new drug targets faster.5

The main feature is the pairing of two kinds of measurement. Cascade will combine 100 million single-cell RNA-sequencing profiles with more than 11,000 secretome profiles, which record the proteins cells release. Together they will cover how different immune cell types respond to more than 150,000 cytokine and drug perturbations, given alone or in combination.5 Parse's GigaLab, a service lab that can profile billions of cells a year, will produce the RNA data. Nomic's Omni 1000 platform will measure more than 1,000 secreted proteins in every sample.8

The project already has backers. GSK, Pfizer, Sanofi, Graph Therapeutics and Turbine have signed on as foundational partners.3 Nomic says data generation has started, and the finished dataset will be sold under license.6

The gap it targets

Both companies frame the project around a shortage of data. Parse co-founder and CTO Charlie Roco said biological AI modeling has hit a data bottleneck, and that very little data exists on how human immune cells respond to cytokines and drugs, even though those responses are central to drug discovery.1 Nomic CEO Milad Dagher put it in terms of communication: models need to capture both how a cell's state changes and the protein signals cells use to talk to each other.2

The trade press reads this as a direct criticism of how "virtual cell" models have been trained so far. Tech Times argues that the big perturbation datasets used for these models show how cells change their gene expression but not which proteins they then release to signal nearby cells.4 Its explanation is biological. RNA sequencing can show that a cytokine gene has been switched on. But how much of that protein is actually secreted depends on later steps, including translation, vesicle trafficking and secretory kinetics, and none of those show up in messenger RNA.4

The evidence for the second measurement layer comes from Nomic's own earlier work. Tech Times reports that Nomic's Perturb-PBMC proof-of-concept covered nearly 10,000 wells, a 191-protein panel, six donors and 80 cytokine perturbations, and found protein effects that RNA-based methods missed entirely.4 If that result holds up, transcriptomic and secretome data are not interchangeable, and a model trained on only one is missing part of the immune response.

Why it is feasible now

Cost has been the main obstacle. Tech Times reports that Omni 1000 launched in May 2025 at about $50 per sample end to end. Its underlying nELISA method uses bead-based assays in which each bead tests for one protein independently. That avoids the cross-reactivity that has usually limited multiplexed ELISAs to about 50–100 targets.4 The outlet argues that at Cascade's scale, this price point is what makes the project affordable at all.4

Parse's side has a track record. Tech Times says the company's GigaLab produced the Tahoe-100M dataset in one month. It also has an ongoing project with Tahoe Therapeutics to generate 300 million single-cell profiles, and a January 2026 partnership with Graph Therapeutics to build an immune cell atlas.4 Graph is now a Cascade partner as well, so Parse looks less like a project collaborator and more like a standing supplier of sequencing capacity to the virtual-cell field.

Where the coverage diverges

Comparing the reports turns up two clear differences.

First, what has been disclosed. Tech Times says the announcement left the experimental design open, including which immune cells, which disease contexts and which perturbation library.4 Nomic's own release answers some of those questions. It says the dataset uses blood immune cells (PBMCs) from 12 healthy donors and covers at least 14 immune cell types responding to cytokines, chemokines and drugs.3 Snowrock, citing Nomic's specifications, adds that the donors are six women and six men, with timepoints at four, eight and 24 hours, and more than 300 treatment conditions per donor at each timepoint. Nomic itself calls the dataset a reference for healthy immune response, and treats disease contexts as a later extension rather than part of the first build.37 Tech Times guessed that Graph's involvement might mean patient-derived disease samples are included.4 The published specifications point the other way, at least for now.

Second, what the headline numbers count. Some reports describe the data as matched 1,000-protein readouts across all 150,000-plus conditions.4 Snowrock gives a more careful account. The 100 million figure counts single-cell RNA profiles, while Nomic lists more than 11,500 secretome profiles and 12.2 million protein measurements. Each protein profile describes a whole sample of cells, not an individual cell.

The numbers add up. Twelve donors times three timepoints times roughly 300 conditions gives about 10,800 samples, which is in line with the 11,000-plus secretome profiles. Multiplying that by 14 cell types gives about 150,000, close to the "perturbation measurements" figure in Nomic's release.3 That suggests the 150,000 figure counts combinations of cell type and condition, not 150,000 different treatments. This is our inference, not something the companies have stated. Nomic's own release uses the term "perturbation measurements," while the Parse-side wording says "conditions."35

The difference matters for buyers. The dataset offers a lot of detail within a controlled experiment, but it does not cover a huge range of treatments. Snowrock notes that a model built for a specific disease would still need evidence from that disease setting.

The access question

Cascade is a commercial consortium, not an open resource. Snowrock reports that Parse's access page sells the data under partnership licenses, with launch-partner licensing open until December 11 and general licensing after that. Fees, delivery dates and license terms have not been published. Tech Times contrasts this with open efforts such as the Tahoe and Arc initiatives. It warns that academic researchers who have relied on public datasets should not count on access soon.4

We think that is the most important point about Cascade's effect on the field. Large pharma partners getting first access to the first large paired RNA-and-protein immune reference could give their internal models an advantage that academic groups cannot easily match. Nomic does say its protein measurements are designed to stay comparable across studies, so other datasets could be built to work alongside Cascade.7 That leaves some room for outside work, even if the core data stays licensed.

Where this sits relative to CRISPR and protein design

Cascade is not a gene-editing project. The announced perturbations are cytokines, chemokines and drugs, not CRISPR knockouts or other genetic changes.35 Its link to the CRISPR world is indirect. Virtual-cell models are generally described as predicting responses to both genetic and chemical stimuli, and better training data for one input type may improve predictions for the other.4 Pooled CRISPR screens remain a separate data source that Cascade does not provide.

Cascade is also not a protein-design tool, but it is relevant to that work. Generative protein-design efforts aimed at immune targets, such as engineered cytokines or binders that change signaling, need some way to predict what happens downstream once a cell receives the signal. A dataset that links changes in gene expression to the proteins cells actually release over time is the kind of evidence needed to test those predictions. This is our assessment. Neither company has claimed a protein-design use.

The verdict

The press releases are promotional, but the underlying idea is sound. Training immune models only on transcriptomes has always left out the signaling between cells that drives inflammation and drug response, and Nomic's earlier data supports that view.4 Cascade's design, with matched donors, three timepoints and a 1,000-protein panel, looks well suited to testing whether adding protein data improves predictions.

The claims should still be read with care. Neither company has reported any measured improvement in drug development from Cascade itself. Tech Times calls the question of whether models trained on it predict better, especially for treatments outside the training set, an empirical one the field still has to answer.4 The planned evaluations, such as holding back some donors or treatments and checking predictions against them, are possible experiments, not results. The dataset starts with 12 healthy donors, which cannot represent the full range of human immune variation.9

In our view, Cascade is the strongest sign yet that protein data is becoming a standard part of training virtual-cell models. In its first version, though, it is a healthy-donor reference sold mainly to large drug companies, not an industry-wide answer to the data shortage it was built to address.