Coding agents can author interpretable single-cell embedding models, here called blueprints, directly from the scientific literature, without ever training a computational model or being exposed to gene expression data. Such a blueprint is auditable by construction, but being purely prior-driven it captures only what the literature already describes and cannot, on its own, correct mismatches with a given dataset or surface structure the literature has not named.

Here, we let a coding agent refine the blueprint against data with structured feedback from a boosting autoencoder that performs structured dimensionality reduction with implicit feature selection. Its latent dimensions are interpretable, each linked to a small gene module, matching the gene-selected, named modules of the agent-authored blueprint. Trained on the data with the blueprint's axes as a prior, it learns additional axes that capture complementary structure the blueprint does not explain. Because the newly learned axes share the blueprint's gene-module vocabulary, contrasting them with the blueprint-anchored axes yields quantitative metrics and qualitative descriptors of cell-group gene programs and latent-space topology.

We first show that, across multiple datasets, agent-authored blueprints yield embeddings whose named axes faithfully discriminate the cell types they name and reach quality competitive with conventional, foundation-model, and program-informed baselines, while remaining batch-robust by construction. Building on this, we demonstrate the refinement on mouse cortex data, where the blueprint misses cell subtypes. The boosting autoencoder surfaces these as structure beyond the literature prior, and a coding agent folds them back into the blueprint as new axes that recover the missed subtypes.

Co-authors: Sonia Maria Krissmer, Marco Prinz, Harald Binder, Maximilian Frosch