scParadise

An automatic annotator trained on healthy blood meets a B cell it was never shown and calls it a monocyte — confidently, with a high probability score, and with nothing in the output to say anything went wrong. The error lands where a study cares most: the population that is new, diseased, or rare. scAdam, the annotation half of scParadise, scores every prediction three ways — the sensitivity of the loss to perturbing a cell's features, the entropy of its predicted class distribution, and the distance from its embedding to the centroid of its assigned class. Otsu's method sets the thresholds, and a cell that crosses them comes back as unknown rather than labelled.

scParadise is a Python framework from Lomonosov Moscow State University, with collaborators at the Gamaleya Center and the Centre for Strategic Planning of Russia's FMBA. It bundles three tools: scAdam for multi-level annotation with unknown cell type identification, scEve for cross-tissue modality imputation, and scNoah for standardized benchmarking. Rather than one general checkpoint, it maintains a hub of more than 50 pretrained scAdam models and five scEve models spanning tissues from PBMC and bone marrow to retina and adipose tissue across nine species; a user applies an existing checkpoint, or trains one for a tissue the hub does not cover.

Against CellTypist, scGPT, TOSICA, Azimuth, Symphony, scANVI and Seurat on human PBMC, pancreas and oropharyngeal cancer data plus a mouse aging brain atlas, scAdam led on balanced accuracy, F1-score, geometric mean and index balanced accuracy — and needs no prior clustering or reference mapping, only a normalized expression matrix.

#Key Features

• Unknown cell type identification: Gradient-, entropy- and distance-based uncertainty scores combine by single-score, majority-vote or summed-threshold rules, letting a model decline a label instead of misassigning one.
• Multi-level hierarchical annotation: Level-specific heads predict major type, subtype and cell state together, each receiving the cell embedding plus the softmax probabilities of the levels above.
• Cross-tissue modality imputation: scEve regresses CITE-seq surface protein abundance from gene expression, so unimodal scRNA-seq gains protein-level resolution — including in tissues the model never saw in training.
• Automated class balancing: Rare types are oversampled by convex interpolation between real cells in principal component space, raising balanced accuracy by 2.35% and index balanced accuracy by 2.44%.
• Benchmarking module: scNoah scores annotation and imputation methods on imbalance-aware metrics per cell type, with confusion matrices and per-cell prediction status drawn on the embedding.

#Technical Details

scAdam splits genes into chunks (16 by default) to build a 256-dimensional embedding sequence, then passes it through a five-block transformer encoder with eight attention heads and a 512-wide feedforward layer. Two-layer MLP heads per annotation level are trained with an adaptive hierarchical focal loss whose per-level weights update every 10 epochs from an exponential moving average of level losses. Training begins with masked gene modelling — 15% of genes masked, mean squared error on the masked positions, 50 epochs — then supervised fitting under AdamW with cosine annealing, mixup and noise augmentation, and early stopping. Optuna drives hyperparameter search with Hyperband pruning and stratified cross-validation, and warm-start fine-tuning carries a trained checkpoint onto a new gene set. scEve reuses the encoder with four blocks and four chunks, swapping the hierarchical classifier for a two-layer MLP regressor onto the target modality.

Published scAdam checkpoints report accuracies from 0.881 to 0.999, and scEve checkpoints cover between 54 and 270 protein targets. Across 32 test datasets spanning six annotation tasks, the balanced scAdam model failed to recover a cell type in only three cases, and scEve beat Azimuth and Seurat label transfer on every error, correlation and variance-explained metric tested.

#Applications

The immediate use is annotating a new dataset against a matched tissue checkpoint without hand-curating markers, but the sharper cases are those where existing labels are suspect. Re-annotating the Tabula Muris Senis atlas showed that cells recorded as granulocytes are exclusively neutrophils, and that a population labelled macrophages spans several distinct types. A subcutaneous adipose tissue checkpoint applied to visceral adipose tissue declined to label mesothelial and mast cells it had never seen, and imputing surface proteins into its NK cells resolved three previously unreported NKT subsets — CD56dim CD3+, CD56dim CD3+ CD4+ and CD56dim CD3+ CD8+ — confirmed by flow cytometry.

#Impact

scParadise argues that the useful unit for single-cell annotation is a maintained registry of tissue-specific checkpoints with published per-model metrics rather than one large general model, and gives that registry what it needs: a way to decline an unrecognized cell, and a shared vocabulary for comparing methods on imbalance-sensitive metrics. Its limits are inherited from supervision: scEve can only impute modalities present in its training data, and annotation errors in a training atlas propagate to every dataset the resulting checkpoint touches. Inference is cheap, but training a new checkpoint wants a GPU.