Sources and Confidence

Provenance for every claim across these documents. Three tiers: verified by search, from the papers themselves, and unverified background knowledge. The last tier should be checked before citing.

Verified by web search (July 2026)

Ahlmann-Eltze, Huber & Anders — “Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines” Nature Methods, 2025. https://www.nature.com/articles/s41592-025-02772-6 Benchmarked 5 foundation models + 2 other deep models. Unseen combinatorial perturbations → don’t beat an additive model. Unseen genes → don’t beat predicting the training mean. Authors’ hypothesis: pretraining data is observational. Preprint: https://www.biorxiv.org/content/10.1101/2024.09.16.613342v5 · Code: https://zenodo.org/records/16092690

Kedzierska et al. — “Zero-shot evaluation reveals limitations of single-cell foundation models” Genome Biology, April 2025. https://genomebiology.biomedcentral.com/articles/10.1186/s13059-025-03574-x scGPT and Geneformer zero-shot are outperformed by HVG selection, scVI, and Harmony. Best batch-integration scores across all datasets came from plain HVG selection. Preprint: https://www.biorxiv.org/content/10.1101/2023.10.16.561085v2.full

UCE — Universal Cell Embeddings https://www.biorxiv.org/content/10.1101/2023.11.28.568918v1.full · https://www.nature.com/articles/s41586-026-10689-z 36M cells, 8 species (human, mouse, zebrafish, mouse lemur, crab-eating macaque, rhesus macaque, tropical clawed frog, pig). Genes tokenized via ESM2 protein embeddings (dim 5120). Confirmed: can represent any protein-coding gene from any species by sequence alone, including species absent from training. This is the claim Tier 1 leans on.

Tahoe-100M https://www.biorxiv.org/content/10.1101/2025.02.20.639398v1 · https://huggingface.co/datasets/tahoebio/Tahoe-100M 100M cells, ~1,100 drugs × 50 cancer cell lines, ~60,000 drug-cell combinations. Open-sourced Feb 2025 as the first contribution to Arc Institute’s Virtual Cell Atlas. Vevo Therapeutics + Parse Biosciences GigaLab + Ultima Genomics.

From the four PDFs’ own reference lists

• scGPT — Cui et al., Nature Methods 21:1470–1480, 2024. doi:10.1038/s41592-024-02201-0 — TransTissue [8]. 33M cells.
• CellFM — Zeng et al., Nature Communications 16:4679, 2025. doi:10.1038/s41467-025-59926-5 — TransTissue [49]. 100M human cells.
• CPA — Lotfollahi et al., Molecular Systems Biology 19:e11517, 2023. doi:10.15252/msb.202211517 — TransTissue [27].
• PerturbNet — Yu, Qian, Song & Welch, Molecular Systems Biology 21:960–982, 2025. doi:10.1038/s44320-025-00131-3 — TransTissue [47].
• PerturbAtlas — Zhang et al., NAR 53(D1):D1112–D1119, 2025. doi:10.1093/nar/gkae851 — TransTissue [51]. ⚠️ A database, not a generative method — miscited in TransTissue §7.
• Funk-SVD — Piatetsky, “Interview with Simon Funk,” SIGKDD Explorations 9(1), 2007.
• Sennrich et al. back-translation, ACL 2016 — cited in TransTissue as [36].
• Linformer (arXiv:2006.04768), Performer (arXiv:2009.14794) — TransTissue [45], [5].
• Open TG-GATEs — Igarashi et al., NAR 43(D1):D921–D927, 2015.
• S1500+ — Mav et al., PLoS One 13(2):e0191105, 2018.

From my training knowledge — NOT independently verified here

Flagged because these should be checked before citing — they come from background knowledge current to roughly mid-2025, not from a verified source.

• Geneformer — Theodoris et al., Nature 2023, ~30M cells, rank-value encoding. Confident on the mechanism, less on the exact cell count.
• scFoundation — ~50M cells, xTrimoGene asymmetric encoder-decoder. Cell count uncertain — verify.
• scBERT — gene2vec + Performer attention.
• GEARS — GNN over gene-ontology graph for unseen genetic perturbations.
• chemCPA — CPA + molecular structure encoder.
• Boiarsky et al. — logistic regression competitive with scGPT/Geneformer for cell-type annotation. Verify the citation.
• LINCS L1000 — ~1.3M signatures, ~978 landmark genes, ~20,000 compounds, ~80 cell lines. These are approximate. Verify all four numbers before using them in Tier 3.
• Rat↔human orthology ~80% clean 1:1, with cytochrome P450 families notably diverged. Directionally right; the exact figure depends on the mapping method and I would not cite 80% without a source.
• Perceiver / Perceiver IO as the closer architectural relative to the n×s bottleneck than Linformer/Performer — my read, not a claim from any paper.
• STATE (Arc Institute), CZI Virtual Cells Platform — exist as of my cutoff; current status unknown.
• Johnson et al. 2016 multilingual NMT with target-language tokens (<2es>), enabling zero-shot translation. High confidence, but worth citing from the original.

Analysis that is mine, not from any paper

Now VERIFIED in code (code/experiments/)

1. The linearity of the Funk-SVD-augmented task. ✅ CONFIRMED — exp02_rank_test.py. Ridge on augmented pairs scores PCC 1.0000 / MAE 0.0000; the analytic construction $W=[P_{KI}1]A^{+}$ with no fitting also scores 1.0000; $q_c$ recovers at correlation 1.000000. It’s a property of Funk-SVD’s algebra, so it should transfer to real data — but run it there before saying so.

   ⚠️ The rank claim was wrong. The original form asserted $\text{rank}(W)\approx r+1=301$. It’s $\approx 2$, flat across $r\in \{\mathrm{5..80}\}$. $r+1$ is an upper bound; Funk-SVD underuses its capacity and spends it on whatever variance dominates. The corrected claim is stronger (the task can be near rank-one) — but the $r+1$ form is wrong and shouldn’t be repeated.

2. Row-wise PCC is undefined for the mean predictor. ✅ PROVEN + VERIFIED — 16_MATH_NOTES.md §3.5, confirmed in exp01. Constant rows ⟹ ${\text{Var}}_j({\hat{G}}_{i,:})=0$ ⟹ no denominator. This is what makes row-wise PCC the metric that exposes it.

3. The degenerate ${\alpha }_{j}y$ adversary scores ${\rho }_{\text{col}}=1.0000$ exactly. ✅ VERIFIED — toxsim/metrics.py. GenTox §2.3’s argument made concrete.

4. Table 3 decomposes exactly into study panels. ✅ VERIFIED — toxsim/fit_panels.py. NNLS: residual 0.0, mean abs error 0.0, all 24 structural zeros reproduced, 24 of 47 candidate panels used. Not in any of the four papers.

   ⚠️ Not unique — 36 equations, 47 candidates, underdetermined. It’s a valid decomposition, not the study design. The original study metadata would settle it.

5. Missingness is MNAR, not MAR. 🟡 Strongly suggested by (4), not proven. TransTissue §5 explicitly assumes MAR. The structural zeros (BR-LI = 0 where independent sampling predicts ~40) are hard to explain otherwise. But “study design drove missingness” → “missingness correlates with the values” still needs its own argument.

6b. No gene-gene attention in TransTissueFormer. ✅ CONFIRMED against Figure 1. The figure draws (B,n) → [n×s] → (B,s) → [1×r] → (B,s,r) → L stacked layers. The transformer operates on 512 pooled slots; genes are gone before layer 1. All gene-gene interaction is linear, in the bottleneck.

6c. The first attention layer is inert. ✅ VERIFIED — code/demo_transtissueformer.py. Reading the 1×r projection as shared (as Figure 1 labels it), all $s$ tokens are collinear: token matrix rank 1, attention score matrix rank 1, attention output rank 1. A rank-1 score matrix means every query attends identically up to a scalar.

⚠️ NOT fatal, and this correction matters. The MLP’s nonlinearity breaks the symmetry — rank recovers to full by block 3. The cost is waste, not correctness: one dead attention layer, fixable with an 8,192-parameter slot embedding (0.18% of the model). “The architecture is broken” would be wrong.

⚠️ Figure 1 draws one trapezoid per slot but labels the op once. If the 1×r is per-slot, this is already handled. The paper alone cannot settle it — check the code.

Still UNVERIFIED — check before repeating

6. The Tier 0e protocol concern. Withholding only G[t,b] leaves the test column’s other tissues in the factorization, whereas deployment has only the source. Follows from Algorithm 1 as written; the implementation may differ from the pseudocode. Stated as a question, not a finding.

7. Funk-SVD as the linear special case of CPA. b_i + b_j + P·Q vs z_basal + z_drug + z_cov. My reading. I have not searched whether anyone has published this — check before claiming novelty.

8. The multi-task failure is a missing target-token. 🟡 The formal argument (conditional-mean-under-squared-loss, 16_MATH_NOTES.md §4.2) is solid. Whether adding $e_{\tau }$ actually fixes it at n=425 is untested. Stated as an open question, not a prediction.

9. “The right FM for toxicogenomics is a perturbation-response model, not a cell-state model.” 🔴 A thesis, not a finding.

10. The NLP analogies throughout. Interpretive framing, not literature. The tissue↔language and back-translation↔matrix-completion mappings are tight; others are merely suggestive. An analogy is a way to see a problem, not evidence about it.

A note on (1)

The rank argument was originally stated as “rank ≈ 301.” Testing confirmed the substance (the map is exactly affine) and refuted the specifics (the rank is ~2, because $r+1$ is only an upper bound). That is the normal outcome of checking a derivation against an implementation — and the reason the agenda puts “run the baselines” before “draw conclusions.”