Single-cell AI's real edge is transfer, not routine annotation
A benchmark of single-cell foundation models against simple baselines, plus new perturbation, annotation, and segmentation tools, shows where the big models earn their keep and where they do not.
- In a preprint benchmark, foundation models outperformed simple baselines in cross-species label transfer, while within-species annotation performance was comparable to PCA. Self-reported m:…5089
- In a preprint benchmark, scFoundation achieved up to 75% directional accuracy in zero-shot in silico perturbations in K562 cells, and Geneformer V2 networks recovered patient-level tumor differential expression with 2.9 fold enrichment compared with 2.0 for co-expression. Self-reported m:…5089
- In a preprint, SPEAR outperformed existing methods in recovering differential gene expression and distributional structure across genetic, pharmacological, and cytokine perturbation benchmarks. Self-reported m:…5465
- In a preprint, scPCIP outperformed marker-based and existing plant deep-learning methods for cross-species annotation, and leaf bundle-sheath cells were found to carry a root endodermis program in 8 of 12 studies. Self-reported m:…0224
- In a preprint, Tracel improved cell-type marker assignment and reduced contamination from neighboring cells in spatial transcriptomics, working across imaging-based and sequencing-based platforms. Self-reported m:…8610
Where the big models win
Single-cell foundation models have arrived with large claims, and a preprint benchmark published November 4 tests them against simple baselines instead of against each other. The result is a split decision. Foundation models outperformed simple baselines in cross-species label transfer, and scFoundation achieved up to 75% directional accuracy in zero-shot in silico perturbations in K562 cells. Geneformer V2 networks recovered patient-level tumor differential expression with 2.9 fold enrichment, compared with 2.0 for co-expression. Where the task involves moving knowledge across species or predicting an unseen perturbation, the big models help. m:…5089
Where the task stays at home, they do not. Within-species annotation performance of foundation models was comparable to PCA, a decades-old linear method. And gene-specific knockout effects in organoids accounted for only 12-35% of observed effects, with the remainder explained by expression location. That is a useful corrective: much of what looks like a gene-specific effect may be where a gene is expressed, not what it does. m:…5089
Perturbation modeling gets a better map
SPEAR, another preprint, attacks the perturbation problem directly. It uses spectral positional encoding derived from the graph Laplacian to ground gene positions in co-expression network geometry, and adaptive repositioning lets gene attention mechanisms reorganize based on perturbation-conditioned hidden states without explicit supervision. The authors report it outperformed existing methods in recovering differential gene expression and distributional structure across genetic, pharmacological, and cytokine benchmarks, and that it captures latent positional relationships mapping to known transcription factor-target interactions. m:…5465
Annotation across species, in plants
The cross-species finding gets independent support from plants. scPCIP annotates plant cell types using conserved gene programs rather than species-specific markers, and the authors report it outperformed marker-based and existing plant deep-learning methods for cross-species annotation. One result illustrates why: leaf bundle-sheath cells carry a root endodermis program in 8 of 12 studies. Cell identity, in this view, is a reusable program, not a species-specific label. m:…0224
Cleaner maps of tissue
Tracel addresses the measurement layer underneath all of this. Spatial transcriptomics assigns transcripts to cells, and transcripts that diffuse or spill across boundaries contaminate the result. Tracel improved cell-type marker assignment and reduced contamination from neighboring cells, and it is platform-agnostic, functioning across both imaging-based platforms such as Xenium and CosMx and sequencing-based Visium HD. It also preserves within-cell-type variation, leading to more accurate differential expression analysis across spatial locations. m:…8610
The same honesty about scope applies to the new tools. SPEAR's reported advantage is in recovering differential expression and distributional structure across perturbation benchmarks, and it captures latent positional relationships that map to known transcription factor-target interactions, which gives biologists a checkable mechanism rather than a black-box score. Tracel's advantage is similarly concrete: cleaner cell assignments on data from platforms researchers already use, from Xenium and CosMx to Visium HD. m:…5465 m:…8610
Read together, these preprints describe a field grading itself honestly. Foundation models are not a general upgrade over simple methods; they are a transfer tool. The practical question for a team choosing a method is narrower than the marketing suggests: are you annotating cells in a well-studied species, or are you moving knowledge somewhere it has never been? The answer decides whether the big model is worth it. m:…5089
What is still unresolved
- All four are preprints; the benchmark results have not passed peer review.
- The benchmark's baselines and datasets are the authors' choices; other baselines or datasets could change the split between transfer wins and within-species parity.
- Zero-shot perturbation accuracy of up to 75% is reported for K562 cells; generalization to other cell types is not shown in the cited claims.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.