Leading performance on cell-type annotation, batch integration and perturbation prediction is reported for the best-established model in this group, and strong few-shot results on network biology and disease gene prioritisation for the other.
All of these are representation models: they produce embeddings, and the analysis you care about is a head trained on top. That is why reported accuracy is task-dependent throughout and why local evaluation is quick — the heads train in minutes.
Every model on this page runs as part of a managed AI pipeline in our GPU clusters, with a dedicated private cluster in our cloud or an on-premise installation where medical governance requires it. Output is decision support for a qualified professional to review, not a diagnosis.
Models in this group take single-cell gene-expression matrices or profiles. Each table gives three hardware tiers — Minimum, the smallest configuration on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is what one server of that tier processes per hour. Use these figures for initial sizing only. Before production we benchmark your own data to confirm accuracy, latency, throughput and cost.
scGPT
Vendor: University of Toronto (Bo Wang lab)
What it does: leads on cell-type annotation, batch integration and perturbation prediction — the broadest single base for single-cell work, and heads train on top of it in minutes.
Leading performance on cell-type annotation, batch integration and perturbation prediction (independent, Nature Methods 2024).
Requirement
Minimum
Medium
High
GPU type
RTX 3090
L40S 48 GB
A100 80 GB
VRAM
24 GB
48 GB
80 GB
vCPUs
8
16
32
RAM
32 GB
64 GB
128 GB
Server
1× RTX 3090 24 GB
1× L40S 48 GB
1× A100 SXM 80 GB
Rate (cells/hour)
≈ 30,000
≈ 105,000
≈ 270,000
Geneformer
Vendor: Broad Institute / MIT (Ellinor lab), Theodoris Lab
What it does: is strong few-shot on network biology and disease gene prioritisation, and supports in-silico perturbation before an experiment is run.
Strong few-shot performance on network biology and disease gene prioritisation (independent, Nature 2023).
Requirement
Minimum
Medium
High
GPU type
RTX 3090
L40S 48 GB
A100 80 GB
VRAM
24 GB
48 GB
80 GB
vCPUs
8
16
32
RAM
32 GB
64 GB
128 GB
Server
1× RTX 3090 24 GB
1× L40S 48 GB
1× A100 SXM 80 GB
Rate (cells/hour)
≈ 30,000
≈ 105,000
≈ 270,000
scFoundation
Vendor: BioMap
What it does: is a large single-cell foundation model for annotation and prediction across downstream benchmarks.
Task-dependent across single-cell downstream benchmarks.
Requirement
Minimum
Medium
High
GPU type
L40S 48 GB
A100 80 GB
H100 80 GB ×2
VRAM
48 GB
80 GB
160 GB
vCPUs
16
32
64
RAM
64 GB
128 GB
256 GB
Server
1× L40S 48 GB
1× A100 SXM 80 GB
2× H100 SXM 80 GB
Rate (cells/hour)
≈ 10,000
≈ 35,000
≈ 90,000
scBERT
Vendor: scBERT authors
What it does: encodes expression profiles for cell-type annotation at a small footprint, cheap enough for a whole atlas.
Task-dependent across single-cell annotation tasks.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
12 GB
24 GB
80 GB
vCPUs
6
12
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (cells/hour)
≈ 90,000
≈ 315,000
≈ 810,000
Universal Cell Embeddings (UCE)
Vendor: UCE authors
What it does: produces cell embeddings that transfer across datasets, which is what makes combining data from several sources workable.
Task-dependent across cell-type annotation and cross-dataset transfer.
Requirement
Minimum
Medium
High
GPU type
RTX 3090
L40S 48 GB
A100 80 GB
VRAM
24 GB
48 GB
80 GB
vCPUs
8
16
32
RAM
32 GB
64 GB
128 GB
Server
1× RTX 3090 24 GB
1× L40S 48 GB
1× A100 SXM 80 GB
Rate (cells/hour)
≈ 30,000
≈ 105,000
≈ 270,000
Choosing between them
Pick on the task mix and the size of your atlas. Where you need perturbation prediction, the models differ meaningfully; where you need annotation, several are close and the choice is practical. Sizing below is per model rather than per dataset, since memory scales with genes retained.
Accuracy figures above are those the producers and independent evaluations report, on their own test sets. They are a shortlist tool, not a prediction of what you will see. At the start of a project we run a short proof of concept on a sample of your own data, which replaces them with real figures — so the cost and the schedule for the full engagement are known before anything is committed.
Send a representative sample, your expected volume and your latency target for single-cell analysis. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster — with the cost per unit of work known before you commit.