CellScribe folds cell identity and state into one generative interface: 0.223 macro-F1 across 309 held-out types, ahead of a frozen-embedding probe

One generative model now returns both a cell type in the ontology hierarchy and a meta-program profile with activity scores, placing two axes that were previously handled separately behind a single interface. Earlier cross-tissue supervised annotators had fixed output heads confined to predefined types, while state inference relied on independent scoring such as ssGSEA or AUCell, or on NMF fitted retrospectively to a cohort, and so could not readily predict the state of a new cell. On the 309-type, 4.44M-cell study-holdout split of CELLxGENE, the full pipeline reaches 0.223 macro-F1 and 0.484 weighted-F1, against 0.194 and 0.366 for a supervised probe on frozen scFoundation embeddings; supervised fine-tuning alone already reaches 0.195 and 0.497.

The model uses each cell's expression profile as a modulation signal at every language-model layer rather than as an extra token inserted into the prompt. The usual route projects the embedding into one prompt token or, through a Perceiver-style resampler, into several tokens; AdaLN-Zero instead maps the 896-dimensional conditioning embedding to per-layer scale and shift through zero-initialised linear projections, recovering the pretrained sub-layer at initialisation. In a supervised-fine-tuning-stage comparison against AdaLN, a single linear token projection, a parameter-matched token projection and the Perceiver reduce typing macro-F1 by 43%, 20% and 11% respectively, and on the state task top-1 and hallmark accuracy fall by 50% to 80%.

Reinforcement-learning post-training scores generations by cell-ontology distance and meta-program agreement, pulling residual mistakes toward nearer ontology neighbours. Supervised fine-tuning fits each answer as a token sequence and does not distinguish confusing a closely related immune subtype from assigning the wrong lineage, nor missing a dominant hypoxia program from misranking it. After alignment, hierarchical F1 rises by 1.0, 1.3 and 1.6 points at exact, plus-or-minus 1 and plus-or-minus 2 tolerance, with the lift growing with tolerance; on the state task top-1 program accuracy, hallmark accuracy and NDCG@5 rise by 1.7, 2.1 and 5.7 points.

Span-conditioned in-context learning lets the model read several labelled cells in one prompt and name types entirely absent from supervised training. A closed-set classifier cannot emit a type outside its output head, and standard AdaLN binds one cell embedding to the whole token sequence, so it cannot tell several cells in the same prompt apart. On 50 types held out from all three training stages and 4,671 query cells, adding this post-training raises exact hierarchical F1 from 0.023 to 0.435, and to 0.695 over the full lineage against 0.069; removing the demonstration cells drops the exact value to 0.032.

The demonstrated setting is single-cell RNA sequencing: typing is evaluated on the 309-type, 4.44M-cell study-holdout split of CELLxGENE and state on 41 meta-programs and 257k cells from 3CA, while the unseen-type experiment uses 50 held-out types and 4,671 query cells. The intended reader works on single-cell annotation and tumour transcriptome analysis; coupling bulk RNA-seq deconvolution to CellScribe is ongoing work and will require prospective patient-level evaluation.

Probes built on pretrained scFoundation embeddings already perform strongly, showing that these encoders contain highly discriminative structure for cell typing, and state remains a more demanding structured prediction problem. Perturbation prediction is named as a natural next step to test whether this capacity extends beyond the current 41 NMF-derived meta-programs, and how AdaLN uses the cell embedding across decoder layers and decoding steps remains an open interpretability question.