Single-cell reference mapping, integration and perturbation. Explore models, applicable tests and evaluation requirements.
These are candidates, not a ranked list. An applicable test may need a trained head, a scoring adapter, a specialist comparison, or access to a hosted service. No result is implied by appearing here.
Public checkpoints; specify exact version before evaluation.
Official source ↗Paper-reported results →Cell embeddings need train-donor reference labels and a classifier.
In-silico perturbation representation needs a predeclared response decoder.
Public code and downloadable checkpoints; use the unfine-tuned whole-human model for a new task.
Official source ↗Paper-reported results →Use frozen whole-human cell embeddings with the same reference labels as baselines.
Fine-tuning or integration procedure must be fitted only on allowed donors.
Public code; model weights have separate terms that must be checked.
Official source ↗Paper-reported results →Cell embeddings plus a train-donor head; confirm model-weight terms.
Public software; train a task-specific model on the permitted split.
Official source ↗Paper-reported results →Train a latent reference on permitted donors, then classify with the same reference labels.
Public code; task-specific training data required.
Official source ↗Paper-reported results →Task-trained perturbation predictor; not a pretrained general foundation model.
Each candidate test needs a defined dataset, split, metric and run protocol before it can become a benchmark.
Map unseen donors to a labelled cell-type reference.
Test whether cell identity is retained across donors and batches.
Predict expression changes after unseen perturbations.
Model sources and access descriptions checked 2026-09-16. Confirm licenses, terms and checkpoint revisions before any run.