Six datasets went in. Six architectures came out. None of them were chosen from a menu — each was specified by the Method Design agents during debate, then implemented by the code-generation agent and trained end to end.
Descriptions are from the paper, arXiv:2508.02276; the complete generated training scripts are in its supplementary material.
These are outputs, not products. CellForge does not ship weights, and re-running it on the same dataset will produce a different architecture. What is stable is the process and the quality band, not the artifact. Treat each card below as a worked example of what the system produces.
| Model | Dataset | Modality | Core idea |
|---|---|---|---|
| CPA-X | Adamson | scRNA-seq | Disentangled perturbation embedding + adversarial covariate removal |
| scGen-X | Norman | scRNA-seq | Latent arithmetic over gene pairs with a non-additive interaction term |
| ChemCellFlow | Srivatsan | scRNA-seq | Sinkhorn conditional OT + 6-layer normalizing flow |
| CPA-Traj | Schiebinger | scRNA-seq | Trajectory-aware VAE, time as a continuous covariate |
| totalGAT | Papalexi | CITE-seq | Graph attention + RNA↔protein cross-attention, dual heads |
| ChromDDPM | Liscovitch-Brauer | scATAC-seq | Denoising diffusion over 200k-peak accessibility |
All six land in the 10–30M parameter range and train in 4–8 hours on a single GPU.
Adamson 2016 · CRISPRi in K562 · scRNA-seq · 111k cells / 33k genes
An extension of the compositional perturbation autoencoder. Cells are encoded into a latent space that is explicitly factorised: perturbation effect on one axis, cell-state and technical covariates on another. An adversarial objective on the covariate branch pushes batch and cell-context information out of the perturbation embedding, so composing a perturbation onto an unseen cell context stays meaningful.
Result: PCC 0.9883 — the best of every method compared on this dataset.
Why the agents converged here: Adamson is single-gene CRISPRi with dense coverage per perturbation. The hard part is not the perturbation model, it is not letting cell-state variation leak into it. The design spends its capacity on disentanglement.
Norman 2019 · combinatorial CRISPRa · scRNA-seq · 84k cells / 17k genes
scGen predicts a perturbation as a vector offset in latent space. That works for single perturbations and breaks for combinations, because the effect of activating genes A and B together is frequently not the sum of activating each alone.
scGen-X keeps the latent-arithmetic backbone and adds an explicit interaction term for gene pairs, so the model can represent synergy and antagonism rather than being forced into additivity.
Result: MSE_DE 0.1736.
This is the clearest case in the paper of the agents diagnosing a specific failure mode of a known method and patching exactly that.
Srivatsan 2020 · small-molecule perturbation · scRNA-seq · 81k cells / 18k genes
Two pieces:
- Sinkhorn conditional optimal transport — drug response is treated as transporting the unperturbed cell distribution to the perturbed one, conditioned on compound identity and dose. Entropic (Sinkhorn) regularisation makes the OT problem tractable at single-cell scale.
- A 6-layer normalizing flow — an invertible density model on top, so the output is a full conditional distribution rather than a point estimate.
Drug response is dose-dependent and heterogeneous across cells; a conditional distribution is the honest output type, and the flow provides it.
Schiebinger 2019 · cytokine stimulation time course · scRNA-seq · 65k cells / 17k genes
A trajectory-aware VAE. The key decision: time is conditioned on as a continuous covariate, not a discrete class label. Treating timepoints as categories throws away the ordering and the intervals, and cannot interpolate to an unmeasured time. Continuous conditioning keeps both.
Result: DEG recall 0.535 — one of the weaker numbers in the paper, and reported as such. Time-resolved response prediction is not solved.
The trajectory-aware encoder has no counterpart in the seed literature corpus. It was not retrieved. It came out of the debate, and it is one of the two pieces of evidence in the paper that the Ideation stage does more than recombine what it read.
Papalexi 2021 · CRISPR · CITE-seq · 171k cells / 18k genes + 200 proteins
The only genuinely multi-modal task: predict both the transcriptome and 200 surface proteins from the same perturbation.
- Graph attention over a gene interaction network, so the perturbation effect propagates along known biology rather than through a dense layer.
- Cross-attention between the RNA and protein branches, letting each condition on the other — protein abundance is not a deterministic function of its transcript, and the model is allowed to learn that relationship.
- Separate decoder heads per modality, because RNA counts and protein counts have different noise models and different dynamic ranges. One shared head would force a bad compromise.
Result: protein recall 0.420. Surface-protein prediction is the hardest readout in the benchmark.
Liscovitch-Brauer 2021 · CRISPR · scATAC-seq · 58k cells / 200k peaks
A denoising diffusion probabilistic model over the chromatin accessibility profile, conditioned on the perturbation. Evaluated with ROC-AUC and PR-AUC rather than regression metrics, because accessibility is near-binary and extremely sparse.
The 200k-peak feature space is an order of magnitude wider than the transcriptomic tasks. Iterative denoising handles that better than a single-shot decoder, which tends to collapse toward the marginal accessibility profile.
Like the CPA-Traj encoder, the diffusion denoiser has no counterpart in the seed corpus. Diffusion for perturbation-conditioned accessibility was not retrieved; it was proposed.
- Sample size is one. Each card is a single run on a single dataset. They are existence proofs that the system produces coherent, competitive designs — not evidence that these are the best designs for these datasets.
- The names are the agents' own. "CPA-X" and "scGen-X" are what the debate called them, and they signal honest lineage: these are extensions of published methods, not clean-sheet inventions. The two clean-sheet components are called out explicitly above.
- Competitive is not always winning. The paper reports competitive-or-better across the board with a clear win on Adamson. It does not claim state of the art on all six.
Full numbers: RESULTS.md. How the designs are produced: ARCHITECTURE.md.