Skip to content

scFoundation

Family: sc-fm Modality: Single-cell RNA-seq (continuous-valued, ~20K-gene vocabulary, human) Released: 2024 (Nature Methods) License: Apache-2.0 (code); separate model-weight license Code/checkpoint: github.com/biomap-research/scFoundation Also surfaced in: agentic-ai, single-cell-spatial-omics, virtual-cells

scFoundation (xTrimoscFoundationα) is a 100M-parameter single-cell foundation model from BioMap / Tencent AI Lab and Tsinghua University, built on the xTrimoGene asymmetric encoder–decoder transformer. It is the first single-cell FM at the 100M scale that jointly models ~20,000 genes in a continuous-valued (non-binned) representation, and it introduces a Read-Depth-Aware (RDA) pretraining task that lets the model project low-quality cells onto the manifold of high-depth cells. Pretrained on 50M+ human single-cell transcriptomes, it claims state-of-the-art transfer to perturbation prediction, drug-response prediction, and read-depth enhancement.

What it does

scFoundation produces cell- and gene-context embeddings from raw single-cell expression vectors and is designed to be plugged into downstream architectures — GEARS for perturbation, DeepCDR / SCAD for drug response — as a frozen or lightly fine-tuned feature extractor. The RDA pretraining task means the model has been explicitly trained to map a downsampled (low-read-depth) cell to the same embedding as its high-depth counterpart, which is useful when shallow scRNA-seq is all that's available.

How it works

Architecture. scFoundation uses the xTrimoGene backbone: an asymmetric encoder–decoder transformer where the encoder is a vanilla-transformer block that processes only nonzero, nonmasked genes (a small, information-dense subset), and the decoder is a Performer (linear-attention) block that operates over all ~20K genes (Hao et al., Nature Methods 2024; xTrimoGene preprint). This asymmetry keeps the heavy compute on the sparse input while still producing a full-length output. An embedding module converts continuous gene-expression scalars into learnable high-dimensional vectors so the model can ingest raw counts rather than binned tokens (Hao et al., Nature Methods 2024).

Tokenization. Genes are addressed by a fixed 19,264-gene index (OS_scRNA_gene_index.19264.tsv) — a continuous-valued vocabulary where each cell is a (gene, scalar-expression) pair list, not a ranked or binned sequence (biomap-research/scFoundation README).

Pretraining corpus + objective. Pretraining uses over 50 million human single-cell transcriptomic profiles (Hao et al., Nature Methods 2024). The novel objective is Read-Depth-Aware (RDA) pretraining: each training pair consists of a masked expression vector plus two total-count indicators T and S (target depth and source depth). The model is asked to recover the masked expression at the target depth T given an input downsampled to depth S — so RDA simultaneously performs (a) traditional masked-token reconstruction and (b) a denoising / depth-enhancement task that maps shallow profiles to deep ones (Hao et al., Nature Methods 2024; HyperAI commentary).

Scale. The authors trained scFoundation at three sizes — 3M, 10M, and 100M parameters — and report that masked-expression-prediction quality scales monotonically with parameter count (Hao et al., Nature Methods 2024). The flagship 100M-parameter checkpoint covers ~20,000 genes. Pretraining wall-clock GPU-hours are not stated in the manuscript (unverified — paper does not report total training cost).

Fine-tuning recipe. scFoundation is positioned as a plug-in embedding service for task-specific models. For perturbation, the recipe is GEARS-with-scFoundation-embeddings (the scFoundation encoder produces 19,264 × 512 gene-context embeddings that are fed into GEARS as the gene-relation graph). For drug response, embeddings are inserted into DeepCDR (bulk RNA-seq → drug IC50) or SCAD (single-cell drug sensitivity). The repository ships scripts for each of these stacks (biomap-research/scFoundation README).

Downstream tasks evaluated. Read-depth enhancement, drug response prediction (DeepCDR, SCAD), perturbation prediction (GEARS), gene module inference, cell mapping / integration, and cell-type annotation (Hao et al., Nature Methods 2024; biomap-research/scFoundation README).

Headline benchmarks

  • Perturbation prediction (GEARS + scFoundation). scFoundation embeddings reportedly improve GEARS Pearson-Delta on the Adamson/Norman/Replogle benchmarks vs. the GEARS baseline alone (Hao et al., Nature Methods 2024). Third-party re-evaluation, however, finds that scFoundation reaches only 0.552 / 0.459 / 0.269 / 0.471 Pearson-Delta on Adamson / Norman / Replogle-K562 / Replogle-RPE1 respectively — worse than a Train-Mean baseline (0.711 / 0.557 / 0.373 / 0.628) and worse than scGPT on most splits (0.641 / 0.554 / 0.327 / 0.596) (Ahlmann-Eltze et al., Nature Methods 2025).
  • Cell-type annotation. In a head-to-head benchmark, the top three single-cell FMs by combined usability + accuracy are scGPT, CellPLM, and Geneformer — with scFoundation reported as competitive but not dominant on intra-dataset annotation (Liu et al., Advanced Science 2025).
  • Scale law. Within the scFoundation family, the 100M variant outperforms 10M and 3M variants on masked-expression prediction monotonically (Hao et al., Nature Methods 2024).

Why it matters for AACR / virtual cells

scFoundation is the most-cited single-cell FM at the 100M-parameter scale that ships drug-response stacks (DeepCDR, SCAD) out of the box, which is directly aligned with the AACR audience's interest in translating cell-state models into IC50 or RECIST predictors. The RDA pretraining task is also a practical fit for clinical scRNA-seq, which is often shallower than research-grade cell atlases — RDA embeddings should, in principle, be more robust to PBMC samples generated under hospital-throughput conditions. Both features explain why AACR 2026 posters pair scFoundation with scGPT when building patient-response classifiers.

Known limitations / open questions

  • Perturbation gains contested. Independent benchmarks find that scFoundation underperforms a Train-Mean baseline on standard Perturb-seq splits (Ahlmann-Eltze et al., Nature Methods 2025).
  • Pretraining cost not reported. The Nature Methods paper does not state pretraining GPU-hours or hardware (unverified — paper does not report).
  • Human-only. Unlike UCE, scFoundation is trained on human transcriptomes only; mouse / PDX work requires separate handling (Hao et al., Nature Methods 2024).
  • Model-weight license is separate from code. The Apache-2.0 license covers code; the model checkpoint is distributed under its own "Model License" agreement, which users must accept before commercial use (biomap-research/scFoundation README).
  • No tumor-pretrained variant. Unlike scGPT (which ships a 5.7M-cell pan-cancer checkpoint), scFoundation's released weights are a single pan-tissue model — there is no cancer-domain-tuned variant to benchmark against (unverified — repository lists only the 100M pan-tissue checkpoint at the time of writing).

Sources