The State of the Art — Genomics & Single-Cell Foundation Models, July 2026

Every 2025–2026 claim here is sourced. Links at the bottom and inline where load-bearing. Anything drawn from older background knowledge rather than a checked source is flagged as such in 17_SOURCES.md.

The one-line summary, and it is not what you’d expect:

The field scaled to 100M+ cells, built genuinely impressive models, ran a competition with 1,200 teams — and the models still do not consistently beat naive baselines at perturbation prediction. The winning entries were deep learning fused with classical statistical features. Pure end-to-end learning has not solved this.

That is extraordinarily good news for you. Read on.

Part 1: How we got here (2021 → 2026)

1.1 The cell-as-sentence bet

Millions of public scRNA-seq cells + self-supervised objectives + transformers = the obvious play. The pitch was BERT’s, verbatim.

The design problem, as covered in 01_BACKGROUND.md §8.1: a cell is a set, not a sequence. No word order. So the real question every model answers differently is how do you encode the expression value? That’s the axis:

NLP framing: you are watching your own field’s 2018–2020 speed-run. scBERT is BERT. scGPT is GPT. UCE is XLM-R with a clever tokenizer. CellFM is “same thing, 3× data.” CellOS is JEPA.

Which means you know what comes next, roughly in order. That is not a small advantage. You’ve seen how this movie ends — including the part where someone checks the baselines.

1.2 The perturbation thread

Different question: given an unperturbed cell and an intervention, predict the perturbed cell. This is the one that matters for toxicogenomics, because a tox profile is a perturbation response.

• CPA (Lotfollahi, 2023) — TransTissue [27]. Additive disentangled latent: $$\mathbf{z}={\mathbf{z}}_{\text{basal}}+{\mathbf{z}}_{\text{drug}}(\text{compound, dose})+{\mathbf{z}}_{\text{cov}}$$ with an adversarial classifier forcing ${\mathbf{z}}_{\text{basal}}$ to carry no drug info. Then recombine at test time → compositional generalization.
• chemCPA — CPA + molecular structure encoder → unseen chemicals.
• GEARS — GNN over a gene-ontology graph → unseen genetic perturbations via the knowledge graph.
• PerturbNet (2025) — TransTissue [47]. Perturbation representation → latent → cell state via normalizing flows.
• PerturbAtlas (2025) — TransTissue [51]. ⚠️ A database of bulk RNA-seq perturbation datasets, not a generative method. TransTissue miscites it as one.

⭐ Look at CPA next to ToxCompl. $$\underset{\text{ToxCompl}}{\underbrace{b_i+b_j+P_{i,:}\cdot Q_{:,j}}}⟷\underset{\text{CPA}}{\underbrace{f({\mathbf{z}}_{\text{basal}}+{\mathbf{z}}_{\text{drug}}+{\mathbf{z}}_{\text{cov}})}}$$ Funk-SVD is the linear special case of CPA. this line of work has been building a linear, transductive CPA and calling it matrix completion. The two literatures are the same idea at different points on the nonlinearity axis, and nobody has said so in print. (16_MATH_NOTES.md §4 — and check the literature before claiming novelty.)

Part 2: The 2026 frontier

2.1 Tahoe-100M — the data unlock

Vevo Therapeutics + Arc Institute, Feb 2025.

• 100M cells, ~1,100–1,200 drugs × 50 cancer cell lines, ~60,000 drug-cell combinations
• ~50× larger than all previously public drug-perturbed single-cell data combined
• Open source, on HuggingFace, first contribution to Arc’s Virtual Cell Atlas

Single-cell perturbation data went from “scarce” to “abundant” in one release. Interventional, at scale.

2.2 Arc’s STATE — the flagship

Arc Institute, 2025–2026.

• Trained on 167M observational cells + >100M perturbational cells across 70 human cell contexts
• Two modules:
  • SE (State Embedding) — an organized map of cell states
  • ST (State Transition) — predicts how expression shifts under drugs/genetic changes
• Reported: 50% improvement in distinguishing perturbation effects; 2× accuracy identifying true differentially expressed genes vs prior models

The architecture is worth noting for its own sake: separating “where am I” from “how do I move” is the same decomposition as CPA’s basal/drug split, and the same as ToxCompl’s $b_i$ / $P\cdot Q$ split. Three literatures, one idea.

2.3 Evo 2 — the DNA side

Nature, March 2026.

• 9.3 trillion DNA base pairs, all domains of life
• 7B and 40B parameters, 1 million token context, single-nucleotide resolution
• Predicts functional impact of variants — including BRCA1 pathogenicity — without task-specific fine-tuning
• Autonomously learned exon-intron boundaries, transcription factor binding sites, protein structural elements

Different axis from everything above: sequence → function, not expression → expression. Relevant to you mainly as evidence that the “genomics foundation model” project does work when the task is well-posed and the data is enormous.

2.4 The rest of the 2026 crop

CellScientist (closed-loop refinement), PRiMeFlow (expression heterogeneity in perturbation response), SCALE (conditional atlas-level endpoint transport), Cell-JEPA, AlphaCell, Chreode, VCWorld, dnaHNet, Evo2HiC.

Also a wave of benchmarks: VCBench, scBenchmark (7 tasks, 24 datasets, 3.2M cells), CellBench-LS (low-supervision), scFME (in-silico perturbation).

The benchmark explosion is itself the tell. Fields build benchmarks when they stop trusting each other’s numbers.

Part 3: The evaluation crisis ⭐

This is the most important section in this document. Read it twice.

The critique literature has gone from “a couple of skeptical papers” to the field’s central problem. Five results, escalating:

3.1 Kedzierska et al. — zero-shot doesn’t work

Genome Biology, April 2025.

scGPT and Geneformer, evaluated zero-shot, are outperformed by selecting highly variable genes — a 2010 heuristic involving no learning at all. Also beaten by scVI and Harmony. Best batch-integration scores across all datasets came from plain HVG selection.

Their hypothesis: masked language modelling on cells may simply not produce useful cell embeddings.

3.2 Ahlmann-Eltze, Huber & Anders — linear baselines win

Nature Methods, 2025.

Five foundation models + two other deep models, benchmarked on perturbation prediction:

• unseen combinatorial perturbations → don’t beat an additive model
• unseen genes → don’t beat predicting the mean of the training perturbations

Their diagnosis, and it’s the deep one: the pretraining data is observational. You cannot learn what happens when you push a system by only watching it sit still.

3.3 Souza & Mehta — parameter-free wins ⭐ NEW

“Parameter-free representations outperform single-cell foundation models on downstream benchmarks”, Feb 2026.

Simple, interpretable pipelines — careful normalization plus linear methods — reach state-of-the-art or near-SOTA across the standard benchmarks. Including:

outperforming foundation models on out-of-distribution tasks involving novel cell types and organisms absent from the training data.

Read that again. The out-of-distribution generalization case — the thing foundation models are supposed to be for — is where linear methods beat them.

Their conclusion: “the biology of cell identity can be captured by simple linear representations of single-cell gene expression data.”

3.4 The Virtual Cell Challenge 2025 — the field checked, at scale ⭐⭐

Arc Institute, results at NeurIPS 2025.

• 5,000+ registered, 114 countries
• 1,200+ teams submitted, 300+ final submissions
• Winners: BioMap Research (1st), Altos Labs (Generalist Prize, flow-matching generative model)

The organizers’ own conclusion:

“perturbation prediction models are not yet consistently outperforming naive baselines across all metrics”

and

“The winning approaches combined deep learning with classical statistical features, suggesting pure end-to-end learning is yet to solve this problem.”

This is not a skeptic’s paper. This is the field’s own flagship competition, run by the institute that built STATE and Evo 2, reporting that after 1,200 teams tried, naive baselines are still not consistently beaten.

3.5 “The Metric Picks the Winner” — the rankings aren’t real ⭐

arXiv 2606.12639, June 2026.

Drug-response prediction on held-out chemistry (THP-1 cells, DRUG-seq). The finding:

the model ranking inverts almost end to end depending on which metric you use. Under an inverse-variance proxy, a linear fingerprint baseline appears to win and the deep models look pointless. Under the contest’s true active-set metric, the deep models win and the proxy’s “winner” is the worst chemistry-aware predictor.

Described as “the first [such demonstration] on real held-out drug chemistry.”

⭐⭐ This is GenTox §2, empirically, on drug response, in 2026.

GenTox proved in 2024 that PCC is unreliable for transcriptomic profiles (Theorem 1: unrelated profiles correlate at $\rho \to 1$; their own baseline scored ${\rho }_{\text{col}}=0.992$ with ${\rho }_{\text{row}}=0.0084$). The field has now independently rediscovered that metric choice determines the winner.

GenTox §2 was early and is now vindicated by the field’s own crisis. That’s not a small thing — it’s a genuine claim to priority on the most important open problem in the area. It’s also, right now, an unpublished draft with Figure ?? and [?] citations and two empty sections.

If you do one thing this year, finishing GenTox might be it.

3.6 The meta-problem

From the benchmark papers: “when benchmarking is conducted by the same group that introduced new methods, evaluations tend to exaggerate the performance of the latest models.” And: custom benchmarks select datasets and metrics that flatter the authors’ model.

Which applies to everyone, including the four papers you’re reading, and including anything you build. The defence is baselines you didn’t choose and metrics you committed to in advance.

Part 4: What this means for you

4.1 The field is exactly where you can help

Line these up:

Every single one of these supports the Tier-0 agenda in 14_RESEARCH_AGENDA.md. The baselines aren’t pedantry; they’re the field’s central methodological problem, and the four papers don’t report them.

The honest answer to “why don’t you just fine-tune scGPT?” in July 2026 is no longer a shrug. It’s:

“Because the field’s own flagship competition ran 1,200 teams and concluded perturbation models don’t consistently beat naive baselines; because linear methods beat FMs even out-of-distribution; because scGPT’s value encoder takes binned absolute counts and our data is log fold-change, which is a type error. We tested it anyway. Here’s what we found.”

That is a strong position. Skepticism, properly cited and properly tested, beats enthusiasm.

4.2 The bad news: someone did a version of your Tier 3 idea

“Task-adapted biological foundation models uncover perturbation-centric representations”, bioRxiv, June 30 2026 — three weeks ago.

They fine-tuned scGPT on >3M LINCS L1000 perturbation profiles with a supervised objective predicting perturbation identity. Results:

• latent space became perturbation-centric — same chemical/genetic perturbation aligns across heterogeneous conditions
• substantially outperformed both raw expression profiles and the original pretrained scGPT
• recovered 85–100% of perturbations in top-100 nearest neighbours
• perturbation classification accuracy 10–19% → 25–49%
• spontaneously captured chemical similarity and MOA despite never being trained on them

So my “the right FM for toxicogenomics is a perturbation-response model, not a cell-state model” thesis has been partially validated by someone else, three weeks ago.

What this changes:

• ❌ You cannot claim “nobody has adapted FMs to perturbation data.” They have.
• ✅ The direction is validated — which de-risks it substantially.
• ✅ They did human cell lines, L1000, in vitro. Nobody has done in vivo, multi-organ, cross-species, fold-change toxicogenomics.

What remains genuinely open:

1. They fine-tuned on L1000 Level 3 (normalized expression). DrugMatrix is fold-change. The value-encoding mismatch is unaddressed.
2. Cell lines in dishes, not organs in a live animal. No systemic exposure, no inter-organ communication.
3. Human. Not rat.
4. No cross-tissue anything. They aligned perturbations; they didn’t translate between organs.

This is what “keeping up with the literature” costs and buys. Three weeks ago, one of my four research directions got half-scooped. It’s not dead — it’s narrowed and de-risked. That’s usually what happens, and it’s why you check before you commit six months.

4.3 The toxicogenomics gap is real and still open

I searched specifically for a toxicogenomics foundation model. There isn’t one. The search returned: DILI prediction from ML, Att-RethinkNet (multilabel on TG-GATEs), ToxAssay (an R package), ADMET/toxicity reviews. No foundation model.

What exists nearby:

• Rat → human hepatocyte translation (PLOS One 2020): CNN + bottleneck DNN translate drug-induced expression rat→human, explicitly “circumventing the current reliance on orthologs”, beating classical ML. This is TransTissueFormer’s architecture and problem, one species-axis over — and it’s not cited in any of the four papers.
• Transfer learning rat in vitro → human in vivo (PLOS One 2023).

Both are worth reading closely. They are the closest prior work to TransTissueFormer in existence — same architecture family, same problem, one axis over — and neither is cited in any of the four papers. Whether the cross-species and cross-tissue lines should be unified is an open question.

4.4 DrugMatrix’s moat, stated precisely

Read the bottom three rows.

Ahlmann-Eltze says FMs fail on perturbation because their data is observational. DrugMatrix is entirely interventional. The Virtual Cell Challenge says pure end-to-end doesn’t work and hybrids win. this program builds hybrids — matrix factorization plus attention plus learned embeddings — which is exactly the winning recipe.

The group is small, under-resourced relative to Arc, and has data nobody else has for the question they’re asking. That’s a real position.

Part 5: What I’d verify next

Things I did not find and would want to know:

1. Did anyone answer Ahlmann-Eltze? Is there any model that now reliably beats linear baselines on unseen perturbations?
2. Virtual Cell Challenge 2026 — is there a second edition? Did the baseline gap close?
3. STATE’s actual numbers vs baselines — the “50% improvement” is Arc’s own reporting. Independent benchmark?
4. Is there a bulk-transcriptomics FM? All of this is single-cell. Conspicuous absence.
5. Cross-species perturbation FMs. UCE handles cross-species embedding; anyone done cross-species perturbation response?
6. The GeniE extrapolation question. How much of BioSpyderWT is model output? Affects everything downstream.

I can run any of these.

Sources

Verified today (July 2026):

• Virtual Cell Challenge 2025 Wrap-Up — Arc Institute — 1,200 teams; baselines not consistently beaten; hybrids won
• Virtual Cell Challenge: Toward a Turing test for the virtual cell — Cell
• NeurIPS 2025: Altos Labs Wins Generalist Prize — GEN
• Arc Institute’s first virtual cell model: STATE — 167M observational + 100M perturbational cells, 70 contexts
• Parameter-free representations outperform single-cell foundation models — arXiv 2602.16696 — Souza & Mehta, Feb 2026
• The Metric Picks the Winner — arXiv 2606.12639 — June 2026; ranking inverts with metric
• Task-adapted biological foundation models uncover perturbation-centric representations — bioRxiv — June 30 2026; scGPT fine-tuned on 3M+ LINCS profiles
• Genome modelling and design across all domains of life with Evo 2 — Nature — March 2026
• Tahoe-100M — bioRxiv
• Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines — Nature Methods
• Zero-shot evaluation reveals limitations of single-cell foundation models — Genome Biology
• Use of deep learning to translate drug-induced gene expression rat → human primary hepatocytes — PLOS One
• Transfer learning to predict drug-induced human in vivo expression from rat data — PLOS One
• VCBench — bioRxiv
• CellOS — bioRxiv
• awesome-foundation-model-single-cell-papers — GitHub — keep an eye on this

From training knowledge, NOT verified — check before citing: Geneformer cell counts, scFoundation scale, scBERT details, GEARS, chemCPA, LINCS L1000 exact figures (~1.3M / 978 / ~20k / ~80), rat-human orthology ~80%. See 17_SOURCES.md.