Skip to content

FM to Virtual Cells — Supplementary

Long-form companion to the talk page. The talk page is the speaker's outline (~700 lines, slides + tables + decision trees). This page is the book of tables — what you read the night before the talk, what you look up mid-prep, what you cite from the talk's footnotes.

Eight load-bearing tables, each followed by the prose dossier that explains its rows. Appendix carryover at the end (Q&A scripts, datasets, ~110-reference reading list, timing cheat sheet, 2026 updates catalog).

Last updated: 2026-05-13. Field status reflects the May 2026 state.

§A. Model dossiers — the 11 anchor models

The master comparison table. Each row links to its detailed dossier below. The columns are the 5 questions every FM should answer; the rows are the 11 models the talk uses as exemplars. For a flat one-sentence-each glossary of every model named in the talk corpus — not just these 11 — see the model glossary.

# Model Family Params Compute disclosed? Public weights The one thing it's best at Where it breaks
A.1 scGPT single-cell 51M ❌ UNKNOWN (50× uncertainty band) ✅ MIT First sc-FM with gene + cell as tokens; defined the category Zero-shot perturbation prediction fails vs linear baseline
A.2 Geneformer V2 single-cell 104M / 316M ✅ DISCLOSED ($17k V2-104M) ✅ Apache-2.0 (HF) Domain-curated 104M beats general 316M at ⅓ compute — kills "scale wins" Same perturbation-prediction gap as scGPT
A.3 UCE single-cell 650M ❌ UNKNOWN ✅ MIT Cross-species (8 species via ESM2 protein-embedding bridge) Cross-species gain is positive but small; least transparent on cost
A.4 STATE single-cell (virtual cell) 600M SE + ST ❌ UNKNOWN (~$125k est.) ✅ open weights First production virtual cell at Tahoe-100M scale No online-update loop; you re-train to add perturbations
A.5 TranscriptFormer single-cell (cross-species generative) 100M+ ⚠️ partial ✅ open weights First generative cross-species sc-FM; 112M cells × 12 species × 1.53B years of evolution Cancer-specific evaluation not published; SOTA on healthy tissues only
A.6 Generative Virtual Cells single-cell (POC) tiny ✅ <$250 ✅ MIT The design pattern: joint planner + world model with validation gating A workshop POC, not a deployed model; scaling unproven
A.7 Virchow2 pathology 632M / 1.85B ✅ DISCLOSED (512× V100, ~$170k) ⚠️ CC-BY-NC Most transparent pathology FM; current SOTA on Campanella 2025 panel Lead over UNI2-h is ≤2 pts; ViT-H/14 + 200M tiles plateauing
A.8 UNI2-h pathology 681M ❌ UNKNOWN (~$75k est.) ⚠️ CC-BY-NC + HF-gated since Jan 2025 Full vertical stack: tile encoder → slide → vision-language → retrieval License-shift risk realized; opaque training cost
A.9 AlphaGenome genomic ~450M ⚠️ partial (TPU only) ⚠️ non-commercial (source Jan 2026) Single model wins 25/26 regulatory variant-effect benchmarks at 1-Mb context Can't do in-context learning; non-commercial license blocks industry
A.10 Evo2 genomic 7B / 40B ✅ DISCLOSED (2,048× H100, ~$5M) ✅ Apache-2.0 Only genomic FM with demonstrated ICL; 1M-token context $5M training cost concentrates capability at ~5 institutions
A.11 ESM-3 protein (multimodal) 1.4B / 7B / 98B ✅ DISCLOSED (1.07 × 10²⁴ FLOPs) ⚠️ 1.4B open, 7B/98B gated Cleanest compute disclosure in biology FMs; generated esmGFP de novo esmGFP is one anecdote; wet-lab validation at scale is rare

Two patterns the table makes visible: (1) Disclosure correlates with hardware sponsorship — every ✅ DISCLOSED model has an NVIDIA / DGX-Cloud / TPU partnership. (2) License posture is bifurcating — sc-FMs uniformly permissive (MIT/Apache); pathology uniformly restrictive (CC-BY-NC + gating); genomic split (Evo2 open, AlphaGenome closed).

A.1 scGPT (Cui et al. Nature Methods 2024)

The single-cell FM that defined the category.

  • Resources — ~51M params; 33M cells from CellxGene Census; GPU compute: UNKNOWN — Cui 2024 does not disclose; ESTIMATED ~10³–10⁵ A100-hours (a 50× uncertainty band); cost $2.6k–$250k [ESTIMATED]. Team: Bo Wang lab, U. Toronto + Vector Institute / Haotian Cui.
  • Framework — encoder-decoder transformer. Gene tokens + cell-state tokens. Masked-value-prediction pretraining objective.
  • Unique — first single-cell FM with both gene and cell as tokens, enabling cross-cell and cross-gene attention.
  • Gap exposed — scGPT's zero-shot perturbation predictions don't beat mean-of-training-perturbations baselines (Ahlmann-Eltze & Huber 2025). The architectural achievement did not translate to perturbation generalization. The 50× uncertainty band on training cost makes "scGPT is expensive to reproduce" claims unverifiable.

A.2 Geneformer V2 (HuggingFace, Dec 2024 update of Theodoris et al. Nature 2023)

The model that inverted the scale-wins narrative — and the cleanest compute disclosure in single-cell FM space.

  • Resources — V2-104M: 104M params, 104M-cell base + 14M cancer cells continual (CLcancer). DISCLOSED via NVIDIA BioNeMo: 64× A100 80GB × 4d 8h = 6,656 A100-hours = ~$17k on-demand. V2-316M: 128× A100 × 3d 19h = 11,576 A100-hours. Team: Theodoris lab, Gladstone Institutes / UCSF.
  • Framework — encoder-only BERT. Rank-based tokenization (cells become ordered gene lists, ranked by relative expression). MLM objective.
  • Unique — rank-based tokenization sidesteps the normalization problem entirely.
  • Gap exposed — V2-104M_CLcancer matches the 316M general-domain model at one-third the compute cost. Scale does not win; domain-curated pretraining does. V2-104M is the only single-cell FM in this matrix where the full training cost is publicly costable.

A.3 UCE — Universal Cell Embedding (Rosen et al. bioRxiv 2023 → Nature Methods 2024)

The cross-species single-cell FM.

  • Resources — 650M params; 33 transformer layers; ~36M cells × 8 species (Integrated Mega-scale Atlas; ESM2-tokenized via protein-sequence embeddings); GPU compute: UNKNOWN; ESTIMATED ~7 × 10⁴ A100-hours upper, ~1,275 floor; cost $3k–$175k [ESTIMATED]. Team: Leskovec + Quake labs, Stanford.
  • Framework — encoder-only ViT-style transformer. Gene-set tokenization mapped across species via orthology (ESM2 protein embeddings as the cross-species bridge). Self-supervised contrastive objective.
  • Unique — species-agnostic via protein-language embeddings. First sc-FM that handles 8 species in one embedding space.
  • Gap exposed — Cross-species transfer is positive but small in magnitude. UCE is the largest published sc-FM at 650M params but also the least transparent on training cost.

A.4 STATE (Arc Institute, 2025)

The Tahoe-100M-native virtual cell.

  • Resources — Two interlocking modules: State Embedding (SE-600M) at 600M params trained on 167M observational human cells; State Transition (ST) at undisclosed params trained on 100M+ perturbed cells (Tahoe-100M + Parse + Replogle); GPU compute: UNKNOWN; cost UNKNOWN (order-of-magnitude ~$125k). Team: Arc Institute (Hsu + Goodarzi labs). Adduri et al. bioRxiv 2025.06.26.661135.
  • Framework — bidirectional transformer with self-attention over sets of cells. SE learns observational embedding; ST predicts perturbation transitions. Trained offline on frozen snapshots.
  • Unique — first production-grade virtual-cell model at Tahoe-100M scale.
  • Gap exposed — STATE is trained on the snapshot; no online-update or active-learning loop in the published version. Arc's disclosure pattern (data scale foregrounded, compute buried) is structural.

A.5 TranscriptFormer (Pearce et al. Science 2025, CZ Biohub)

The first generative cross-species sc-FM. The model that turns CZ Biohub from a substrate-only lab into a model-shipping lab.

  • Resources — 100M+ params; 112M cells × 12 species spanning 1.53B years of evolution. Open weights (Apache-2.0). Team: CZ Biohub Pearce et al. Science 2025 · GitHub · bioRxiv 2025.04.25.
  • Framework — generative transformer; joint generation over gene tokens with phylogenetic conditioning.
  • Unique — first generative cross-species sc-FM. SOTA on cross-species cell-type classification (incl. species separated by 685M years) and zero-shot disease-state ID in human cells.
  • Gap exposed — Cancer-specific evaluation is unpublished. The published validation is on developmental, immune, and neural tissues. Track 7 (§C.7) — cancer cross-species — is the open project this model unlocks. Breaks the §3.11 "substrate vs model split" observation that was true through mid-2025.

A.6 Generative Virtual Cells (Lewis & Zueco, ICLR 2026 Gen² Workshop)

The methods paper proposing closed-loop virtual cells. Position paper, not a frontier training run.

  • Resources — small MLP/transformer trained on a toy Perturb-seq simulator. <$250 compute. <100 GPU-hours total. Team: AIXC Research / Lewis & Zueco.
  • Framework — joint planner + world model trained with validation gating.
  • Unique — jointly updated under validation feedback, not offline-trained on a frozen snapshot. The contribution is the design pattern.
  • Gap exposed — Workshop POC, not deployed; validation feedback assumes cheap high-throughput experiments. The gap between $250 workshop POC and $5M Evo2-scale model is enormous.

A.7 Virchow2 (Paige + MSK, Zimmermann et al. arXiv 2408.00738)

The current pathology FM SOTA — and the most transparent pathology FM on hardware.

  • Resources — 632M-param ViT-H/14; 3.1M WSIs / ~225,000 patients / ~2B tiles / ~200 tissue types. DISCLOSED: 512× NVIDIA V100 32GB, batch 4,096, AdamW. ESTIMATED ~1.4 × 10⁵ V100-hours = ~$170k. Virchow2G (1.85B) on same fleet. Team: Paige + MSK / Eric Zimmermann (V2), Eugene Vorontsov (V1).
  • Framework — DINOv2 self-supervised pretraining on mixed-magnification 224×224 H&E (V2 adds IHC) tiles; attention-pooling for slide-level aggregation.
  • Unique — dual transparency: hardware count + patient-level provenance.
  • Gap exposed — Lead over UNI2-h is small (~1–2 points). The pathology FM curve appears to be plateauing around ViT-H/14 + 200M tiles.

A.8 UNI2-h (Mahmood lab, Jan 14 2025 update of Chen et al. Nature Medicine 2024)

The Mahmood-stack flagship — and the opaque counterpoint to Virchow2.

  • Resources — 681M-param ViT-H/14; 200M+ tiles across 350K+ WSIs (H&E + IHC). GPU compute: UNKNOWN ("extensive A100 hours" in HF card). ESTIMATED ~3 × 10⁴ A100-hours = ~$75k. HF gated since Jan 2025. Team: Mahmood Lab, Harvard/MGB.
  • Framework — DINOv2 on histology tiles. Companion models: TITAN (slide-level), PathChat / PathChat-2 (vision-language), SlideSeek (retrieval).
  • Unique — full vertical stack. PathChat DX = first generative-AI pathology tool with FDA Breakthrough Designation (Jan 2025).
  • Gap exposed — Two compounding gaps: opaque training cost (can't argue scaling vs curation) + license-shift risk realized via Jan 2025 HF gating.

A.9 AlphaGenome (DeepMind, Avsec et al. Nature 2025)

The current genomic FM SOTA on variant effects.

  • Resources — ~450M params [VENDOR-ASSERTED]; ENCODE + GTEx + FANTOM5 + 4DN across 5,930 human + 1,128 mouse tracks; 1-Mb input windows. DISCLOSED hardware: 8× TPU v3 per replica with sequence parallelism. Total TPU-hours UNKNOWN. ESTIMATED ~$200k ($46k/fold × 4-fold ensemble + distillation). Source code released Jan 2026; non-commercial license. Team: Google DeepMind / Žiga Avsec.
  • Framework — U-Net + transformer bottleneck. 1-Mb input window. Multi-track output prediction simultaneously.
  • Unique — single-model multi-track prediction at 1-Mb context. Replaces Enformer + Borzoi + Splam etc.
  • Gap exposed — Cannot do in-context learning. The non-commercial license is the binding industry constraint, not compute.

A.10 Evo2 (Arc Institute + NVIDIA + Stanford + UCB + UCSF, Brixi et al. Nature 2026)

The generative + ICL genomic FM. Second-most-expensive biology FM publicly trained.

  • Resources — 7B and 40B params; OpenGenome2: 8.8T tokens (~9.3B bp) from 100,000 species. DISCLOSED: 2,048× NVIDIA H100 on DGX Cloud (AWS), several months. ESTIMATED ~2 × 10⁶ H100-hours = ~$5M. Discloses FLOPs ratio: "~150× more compute than AlphaFold, ~2× the FLOPs of ESM-3" — the cleanest cross-model compute comparison in 2025-2026 biology. Open weights (Apache-2.0). Team: Arc + NVIDIA + Stanford + Berkeley + UCSF / Garyk Brixi.
  • Framework — StripedHyena 2 (hybrid state-space + attention). Byte-level (single-nucleotide) tokenization. Causal LM objective. 1M-token context.
  • Unique — only genomic FM with demonstrated in-context learning.
  • Gap exposed — Prompt-engineering is ad-hoc; no systematic prompt-template benchmark for biological FMs. The 2,048-H100 fleet is only available to ~5 institutions worldwide.

A.11 ESM-3 (EvolutionaryScale, Hayes et al. Science 2025)

The 98B-param multimodal protein FM. Sets the floor for "frontier biology FM compute" at 10²⁴ FLOPs.

  • Resources — 98B params largest (also 1.4B / 7B); 2.78B proteins, 771B unique tokens; 1.07 × 10²⁴ FLOPs DISCLOSED. ESTIMATED ~9.9 × 10⁵ H100-hours; cost ~$2.5M–$8M. Team: EvolutionaryScale / Thomas Hayes. 1.4B open under Cambrian NC; 7B/98B Forge-API-gated.
  • Framework — multimodal generative transformer with 7 token tracks: sequence + 3D coordinates + structure tokens (4,096-codebook VQ-VAE) + SS8 + SASA + function keywords + InterPro annotations.
  • Unique — multimodal generation + cleanest compute disclosure. Generated esmGFP de novo (36% identical to avGFP / equivalent to ~500M years of evolution).
  • Gap exposed — esmGFP is one anecdote. The 1.4B open variant is what most academic reproductions use — the 98B weights are gated.

§B. Lane dossiers — the 9 small-lab application lanes

How a small lab can ship FM-area work without a $5M GPU budget. Lanes 8–9 are net-new in 2026 (opened by xVERSE / TranscriptFormer / VCHarness).

Lane Typical $ Compute Time Risk 2-3 exemplars Top venues
B.1 Embeddings as features $0–$500 inference-only, 1× consumer GPU weeks low AACR #5470 UNI2-h+MIL HER2; AACR #1441 UNI2-h+CLAM prostate; AACR #2758 UNI2-h+ABMIL rhabdo Cancer Discovery, JCO CCI, npj Precision Oncology, CSHL BDS 2026
B.2 PEFT / LoRA / adapters $500–$5k 1–8× A100 for days 2–4 mo medium sc-FM PertAdapter (ICLR 2026 LMRL); scGPT-spatial LoRA; GenePT/scELMo ICLR / NeurIPS LMRL, Genome Biology, ISMB 2026
B.3 Domain-specific small FM $10k–$50k 8–64× A100 ~1 week 6–12 mo medium-high Geneformer V2-104M_CLcancer ($20k); Nicheformer ($25k); scTab + CellPLM Nat Methods, Genome Biology, Cell Systems, CSHL BDS 2026
B.4 Negative results / replication $0–$2k inference + linear regression 3–6 mo low scientifically Ahlmann-Eltze; Kedzierska; Wenkel; Wu Nat Methods; Liu Adv Sci; parameter-free; CellBench-LS; Han et al.; PertEval-scFM; Csendes scPerturBench; cellular-dynamics zero-shot — 11 papers in 2025-2026 Nat Methods, Genome Biology, Adv Sci, RECOMB 2026
B.5 Benchmark / dataset curation $0–$5k + curator time de minimis compute 6–18 mo low Replogle Perturb-seq; Open Problems v2; HEST-1k; Tahoe-100M NeurIPS D&B, Scientific Data, Nat Methods Resource, Single Cell Genomics 2026
B.6 FM-wrapper tools / pipelines $0–$5k dev box + inference for tests 6–12 mo low (adoption is the risk) signifinder (Bioconductor); mia / miaTime / miaViz; SpatialData; CytoVerse (browser-native) GBCC / EuroBioC, Bioinformatics, JOSS, F1000Research
B.7 FM-aided wet-lab / clinical $5k–$50k FM-side mostly inference 12–24 mo low (FM side) AACR #4163 KRONOS; AACR #1442 Virchow2+VIDCellTyper; AACR #1444 H-optimus+ABMIL; Janowczyk Nat Med 2025 (UNI EGFR, 197 patients, 43% testing avoidance — the first deployment-grade clinical paper) Clin Cancer Res, JCO CCI, npj Precision Oncology, Nat Med (deployment), AACR Annual
B.8 Data-augmentation engine (new 2026) $0–$2k inference + small classifier 3–6 mo medium (synthetic-data bias control) xVERSE (rare cell types resolved with 4 cells); TranscriptFormer (cross-species synthesis); scDiffusion; CellPLM Nat Methods Resource, Genome Biology, NeurIPS LMRL workshop
B.9 FM-aided experimental design / active learning (new 2026) $1k–$10k FM-side inference per iteration 12–18 mo medium (wet-lab partner; baseline matters) CRADLE-VAE (Bunne); scPRINT; VCHarness (BioMap, autonomous virtual-cell builder); rBio (reasoning interface) Nat Methods, Cell Systems, Mol Syst Biol, NeurIPS active-learning workshops

Decision tree: see the talk page §6.11. Combinations: see talk page §6.13 — six load-bearing two-lane patterns. The single most-cited posters at AACR 2026 combine two lanes, not one.

B.1 Lane 1 — FM embeddings as features ($0–$500)

A 1B-param model trained on someone else's $200k cluster, used as a frozen encoder, is the cheapest 2026 way to publish a clinically-meaningful classifier — and the field has stopped pretending the wrapper architecture is the contribution.

Anchor exemplar — AACR 2026 poster #5470 (academic, not Mahmood/Paige): UNI2-h tile embeddings → attention-MIL → HER2 prediction in breast cancer. AUC 0.715, tied with full UNI2-h, but with lower runtime and lower scanner cost — the publishable finding is parity at a fraction of deployment cost. Compute: <$500, inference-only.

Other AACR 2026 examples: #1441 (UNI2-h + CLAM for prostate-cancer risk stratification) · #2758 (UNI2-h + ABMIL for PAX3/7::FOXO1 fusion detection in rhabdomyosarcoma — first published H&E predictor of a pathognomonic fusion in this rare paediatric disease).

B.2 Lane 2 — PEFT / LoRA / adapters ($500–$5k)

A <1% parameter adapter is the highest-leverage technical contribution a small lab can make — it inherits the base model's $250k of pretraining and slots into a public benchmark.

  • Exemplar A — sc-FM Perturbation Adapter (Maleki et al., ICLR 2026 LMRL): drug-conditional adapter, <1% trainable params, frozen scGPT/Geneformer/scFoundation backbone. Win: zero-shot generalisation to unseen drugs and cell lines where CPA / GEARS / scGen cannot extrapolate.
  • Exemplar B — scGPT-spatial / LoRA-fine-tuned scGPT: LoRA rank-8 adapters on scGPT's attention layers for Visium decomposition. ~$1.5k on a single 8× A100 node over 24–48h.
  • Exemplar C — GenePT / scELMo (Yale, Stanford): frozen LLM embeddings of gene names + a logistic-regression head. ~$200 compute. Matches or beats scGPT zero-shot on cell-type annotation.

B.3 Lane 3 — Domain-specific small FM ($10k–$50k)

Continual-pretrain a smaller general model on your domain corpus. 2025 evidence shows this consistently matches the 3× larger general model on in-domain tasks.

  • Exemplar A — Geneformer V2-104M_CLcancer (Dec 2024): $16,640 disclosed for V2-104M base + ~<$3k for the 14M-cancer-cell continual step ≈ $20k all-in. Matches V2-316M at ⅓ parameters and ⅓ compute. The cleanest published evidence that domain curation beats parameter count.
  • Exemplar B — Nicheformer (Theis, bioRxiv 2024): 80M-param transformer pretrained on ~110M dissociated + spatial cells with a niche-aware objective. ~$25k estimated.
  • Exemplar C — scTab + CellPLM: both beat scGPT and Geneformer on cell-typing benchmarks at 10–20× less training cost.

B.4 Lane 4 — Negative results / replication / critique ($0–$2k)

The single most-published lane of 2025–2026. 11 critique papers are now in the canon (see §E for the catalog). To differentiate a 2026 Lane-4 paper, target an uncovered axis: (1) donor-split benchmarks; (2) cross-tissue transfer; (3) time-resolved perturbation; (4) cancer-cell-line → primary-tumor transfer; (5) rare-cohort robustness.

Anchor exemplar — Ahlmann-Eltze, Gerard, Huber (EMBL Heidelberg) — Nature Methods 2025: pure inference on six published FMs + a one-line mean-of-training-perturbations linear baseline. <$2k compute. Retired the entire sc-FM perturbation-prediction leaderboard.

B.5 Lane 5 — Benchmark curation / dataset contribution ($0–$5k + time)

Held-out splits, donor-split benchmarks, and modality-specific evaluation suites are the most under-priced contributions in the FM era — your dataset becomes infrastructure every subsequent model has to cite.

Exemplars: Replogle et al. 2022 Cell (Weissman lab — genome-wide Perturb-seq) · Open Problems for Single Cell (Luecken, Lance, Dann et al. — Nat Biotechnol 2024) · HEST-1k (Jaume et al., NeurIPS D&B 2024). Tahoe-100M (Vevo + Arc, Feb 2025) is the contribution pattern — open-source as benchmark infrastructure.

B.6 Lane 6 — FM-wrapper tools / pipelines ($0–$5k)

Scverse and Bioconductor still want maintainers. A well-designed wrapper that lets non-Python users feed their data to scGPT or UNI will out-cite the underlying FM in clinical-application papers.

Exemplars: signifinder (Calura lab, U Padua) · mia / miaTime / miaViz (Borman, Turku) · SpatialData (Marconato, Palla et al., Theis + Stegle labs) · CytoVerse (bioRxiv 2026.01 — browser-native scFM via ONNX).

B.7 Lane 7 — FM-aided wet-lab / clinical application ($5k–$50k)

Once an FM is publicly available, using it as instrumentation inside a clinical or wet-lab study is a clinical paper that happens to use a frozen FM.

The 2026 frozen-encoder menu is now bigger than 4 options. Pathology: UNI / UNI2-h, Virchow / Virchow2 / Virchow2G, CONCH, CHIEF, PathChat / PathChat-DX, Phikon / Phikon-v2 / H-optimus-0, Prov-GigaPath, Hibou, PLUTO, TITAN, mSTAR, GPFM, PathOrchestra, PathPT. Single-cell: scGPT, Geneformer V2, scFoundation, UCE, CellPLM, CancerFoundation, scMulan, LangCell, Nicheformer (spatial), OmniCell (spatial), TranscriptFormer (cross-species generative).

Beyond AACR — the first published real-world clinical deployment: Janowczyk et al. 2025 Nature Medicine — fine-tuned UNI identifies EGFR mutations from H&E in a 197-patient clinical trial, potentially preventing further genetic testing in 43% of cases.

Lane 7 + Track 5 (UQ) is the regulatory-grade combination. The honest 2026 Lane-7 paper includes a post-hoc UQ wrapper (deep ensemble, MC dropout, Laplace, conformal). FDA-pathway requires it.

B.8 Lane 8 — FM as generative data-augmentation engine ($0–$2k) — new 2026

Before April 2026, FMs were encoders (Lane 1) or adapters' substrate (Lane 2). xVERSE and TranscriptFormer changed the menu: a generative cell FM can synthesize labeled training data for downstream tasks.

Anchor — xVERSE (Jiang & Xie, April 2026): synthesizes virtual cells indistinguishable from real (AUROC ≈ 0.5) and resolves rare cell types with as few as 4 cells.

A 12-month project: pick a rare-cancer scRNA-seq cohort (<500 cells published). Synthesize a 10–50× expanded labeled corpus with xVERSE or TranscriptFormer. Train downstream classifier on (real + synthetic) vs (real only). Report (1) classifier-AUC lift, (2) calibration on held-out real test set, (3) qualitative validation that synthetic cells preserve known cancer signatures.

Caveat: synthetic cells can launder the FM's training-distribution biases into your downstream classifier. The honest write-up requires a held-out real test set and a calibration plot.

B.9 Lane 9 — FM-aided experimental design / active learning ($1k–$10k FM-side) — new 2026

The closed-loop version of Lane 7. Lane 7 uses a frozen FM as instrumentation; Lane 9 uses the FM to prioritize which experiment to run next. This is the pattern JPM-2026 buyers (Recursion, Insitro, Latent Labs, Vevo) actually pay for. Bunne 2025 Cell names it as a canonical virtual-cell use case.

Companions: VCHarness (BioMap, bioRxiv 2026.04) — autonomous AI system that constructs perturbation-response models AND prioritizes next experiments. rBio (CZ Biohub) — natural-language interface to active-learning queries.

A 12-month project: partner with a wet-lab running CRISPRi or drug-perturb screens. Implement FM-guided selection on 200 candidate KOs in a single cancer cell line. Two arms: (1) FM-guided 50 perturbations, (2) random or literature-prior–weighted 50. Compare phenotype enrichment + cost.

Caveat: random selection isn't enough as baseline; use StringDB-prioritized as the honest comparator.

§C. Track dossiers — the 9 small-lab innovation tracks

How a small lab can contribute methods work on FMs (not just with them). Tracks 7–9 are net-new in 2026.

Track Open problem Compute Exemplar Paired track
C.1 Mechanistic interpretability What do scGPT/Geneformer/UNI actually learn? <$2k Simon & Zou 2026 SAE atlas; Adams 2025 PNAS protein-FM SAEs; Hibou-LP 2024 pathology SAE T6
C.2 New pretraining objectives that target causality Next-gene-prediction optimizes correlation — that's why FMs lose to linear baselines $5–20k Lopez/Hsu 2025 Nat Methods causal repr; CINEMA-OT; STATE pre-training T6, T9
C.3 Compositional generalization benchmarks + theory Does A+B generalize when model saw only A and B? <$2k Hetzel 2024 NeurIPS LMRL; Norman 2019 substrate; Boiarsky 2023 baseline-warning T8, T9
C.4 Architectures with biology-specific inductive biases BERT clones ignore pathway / network / lineage priors $5–15k xVERSE (transcriptomics-native, +17.9% vs LM-derived); MAP (knowledge-driven); HyenaDNA, Caduceus, Enformer T2
C.5 UQ / OOD detection for clinical-grade FMs Every FDA path needs calibrated uncertainty; FMs don't have it $1–5k van Amersfoort 2024 deep DUQ; Angelopoulos 2023 conformal; no published clinical-grade pathology-FM UQ as of May 2026 B7
C.6 Causal evaluation frameworks What's the correct test for causality, post-Ahlmann-Eltze? <$3k Wenkel latent-additive baseline; Lopez/Hsu causal repr C2, C9
C.7 Cross-species transfer / phylogenetic priors (new 2026) Does cross-species pretraining help or hurt for cancer biology? <$2k TranscriptFormer (open); UCE 8-species ablations; no cancer-specific benchmark published §5.2 gap
C.8 Synergistic-info evaluation of multimodal FMs (new 2026) Which fusion strategies buy cross-modal information vs being redundant? <$2k Microsoft Beyond Alignment / SIS metric (Feb 2026); apply across multimodal FM zoo (scGPT-spatial, Nicheformer, OmniCell, CLM-X, CELLama, SpatialFusion, mSTAR, TITAN) C3, C9
C.9 Causal transportability benchmarks (new 2026) What's the right test for cross-context generalization, post-VCsNC? <$3k Virtual Cells Need Context (bioRxiv 2026.02) — names the framework; multi-context benchmark unwritten. Pearl + Bareinboim transportability formula + Replogle/Norman as substrate C6, C8

C.1 Track 1 — Mechanistic interpretability

The first wave broke in 2025–2026. Published SAE work uses general-domain sc-FMs (scGPT, Geneformer on CELLxGENE). No one has run SAEs on a cancer-curated sc-FM (Geneformer V2-104M_CLcancer, CancerFoundation) to ask: do cancer-domain features differ, and do they predict therapy response or resistance?

A 12-month project (revised): train SAEs on Geneformer V2-104M_CLcancer's residual stream, compare feature atlas vs V2-104M_general. Validate by knockout in held-out perturbation data. Compute: <$2k. Now it's a cancer mechanism paper, not a first-SAE paper.

C.2 Track 2 — New pretraining objectives that target causality

Next-gene-prediction (scGPT) and masked-gene-modeling (Geneformer) optimize correlation. That's structurally why both lose to linear baselines on perturbation prediction.

A 12-month project: design a counterfactual pretraining objective — for each cell, predict gene expression under a held-out perturbation using invariant risk minimization (Arjovsky 2019) or contrastive perturbation losses. Pretrain a 20M-param model on Tahoe-100M; show it beats scGPT-33M on Norman + Replogle held-out splits. Compute: $5–20k.

C.3 Track 3 — Compositional generalization benchmarks + theory

When a model sees perturbations A and B separately, does it generalize to A+B? Norman 2019 has 287 combinatorial splits — but no published sc-FM has been formally evaluated for compositional generalization with theoretical guarantees.

A 12-month project: build a held-out compositional-perturbation benchmark from Norman + Replogle. Score every sc-FM. Then prove a theoretical lower bound on what an additive baseline achieves. Find regimes where FMs should beat additivity — and show whether they do. Compute: <$2k.

C.4 Track 4 — Architectures with biology-specific inductive biases

All current biology FMs are domain-agnostic transformer clones. But biology has rich exploitable structure: gene-gene interaction networks, pathway hierarchies, cell-lineage trees, regulatory grammar, evolutionary conservation.

A 12-month project: build a graph-attention transformer that takes a gene-gene interaction graph (StringDB / Reactome / OmniPath) as a fixed prior on attention weights. Pretrain on a 1M-cell subset of CELLxGENE. Hypothesis: pathway-aware attention beats general attention at this scale. Compute: $5–15k.

Real exemplars (2026 wave): xVERSE (transcriptomics-native, beats LM-derived sc-FMs by 17.9% on representation), MAP (knowledge-driven, zero-shot prediction for unprofiled drugs), TxPert (Wenkel et al., Nat Biotech 2026) — uses multiple knowledge graphs as the inductive bias for transcriptomic-perturbation prediction. TxPert is the reckoning answering itself: Wenkel co-authored the 2025 latent-additive critique, and TxPert is the methodological response from inside the critique camp. The "all sc-FMs are scGPT-shaped" frame is out of date.

C.5 Track 5 — UQ / OOD detection for clinical-grade FMs

Every FDA-deployable model needs calibrated uncertainty and OOD detection. None of the current biology FMs provide either. PathChat DX got Breakthrough Designation in spite of this.

A 12-month project: implement four UQ methods on top of frozen UNI2-h embeddings — (1) deep ensembles, (2) MC dropout, (3) Laplace approximation, (4) conformal prediction. Evaluate calibration + selective-prediction accuracy + OOD detection on TCGA → CPTAC distribution shift. Compute: $1–5k.

Real exemplars: van Amersfoort et al. 2024 (deep deterministic UQ); Angelopoulos & Bates 2023 conformal prediction tutorial. No published clinical-grade UQ work for pathology FMs as of May 2026.

C.6 Track 6 — Causal evaluation frameworks (post-Ahlmann-Eltze)

Wenkel proposed latent-additive as the new baseline — but that's still a correlational baseline. What's the right test for whether a model recovered causal structure?

A 12-month project: build a causal-recovery benchmark with three test sets: (a) MR-validated genetic effects, (b) validated drug MOA on cell lines (LINCS L1000), (c) ENCODE TF-target relationships. Score every sc-FM. Define a "causal F1" metric. Publish as a public leaderboard. Compute: <$3k.

C.7 Track 7 — Cross-species transfer / phylogenetic priors — new 2026

TranscriptFormer pretrained on 12 species × 1.53B years of evolution and beat human-only sc-FMs on cross-species cell-type classification. But: does cross-species pretraining extend to cancer biology, where mouse models, organoids, and PDX systems are the daily substrate?

A 12-month project: build a cancer cross-species transfer benchmark — paired mouse-tumor scRNA-seq + matched human tumor (TCGA-style or patient-matched PDX cohorts). Score TranscriptFormer vs human-only sc-FMs on (1) mouse-to-human cell-state mapping, (2) drug-response transfer from GEMM studies, (3) immune-microenvironment cross-species alignment. Compute: <$2k.

No published cancer cross-species sc-FM benchmark as of May 2026.

C.8 Track 8 — Synergistic-information evaluation of multimodal FMs — new 2026

Microsoft's Beyond Alignment paper (Feb 2026) introduced the Synergistic Information Score (SIS) — a metric for whether a multimodal FM extracts new information from cross-modal interactions vs just repeating unimodal signal. No one has applied SIS across the multimodal cell FM zoo yet.

A 12-month project: implement SIS on ≥5 published multimodal sc / spatial / pathology FMs on a common task set. Identify which fusion strategies actually buy synergy and which are redundant-aligned-fusion in disguise. Compute: <$2k.

C.9 Track 9 — Causal transportability benchmarks — new 2026

Virtual Cells Need Context, Not Just Scale (bioRxiv 2026.02) frames the sc-FM-doesn't-generalize problem as a causal transportability problem in Pearl's sense. The paper names the framework but does not give the field a transportability benchmark suite.

A 12-month project: define a transportability benchmark with explicit cross-context splits — (a) train on cell-line X, test on cell-line Y under same drug; (b) train on healthy donors, test on cancer patients; (c) train on bulk-screened perturbations, test on combinatorial; (d) train on one timepoint, test on another. Score every published sc-FM using a transportability-explicit metric. Compute: <$3k.

The first paper that operationalizes this owns the post-reckoning evaluation citation for the next 3–5 years.

§D. Compute & resources

Detailed compute / cost / team / data table with arithmetic and DISCLOSED / ESTIMATED / UNKNOWN tags lives in _resources-matrix.md. Key rows summarized below.

Model GPU type × count GPU-hours $ (on-demand) Disclosure tag Source
Geneformer V2-104M 64× A100 80GB 6,656 ~$17k ✅ DISCLOSED NVIDIA BioNeMo recipe
Geneformer V2-316M 128× A100 11,576 ~$30k ✅ DISCLOSED NVIDIA BioNeMo recipe
Virchow2 512× V100 32GB ~140,000 ~$170k ✅ DISCLOSED Zimmermann 2024 arXiv
Evo2 2,048× H100 ~2,000,000 ~$5M ✅ DISCLOSED Brixi 2026 Nature + NVIDIA
ESM-3 (98B) (10²⁴ FLOPs disclosed) ~990,000 H100-eq $2.5–8M ✅ DISCLOSED Hayes 2025 Science
AlphaGenome 8× TPU v3 per replica × 4-fold ensemble + distill UNKNOWN ~$200k est. ⚠️ partial Avsec 2025 Nature
scGPT UNKNOWN 10³–10⁵ A100-h (50× band) $2.6k–$250k est. ❌ UNKNOWN Cui 2024 Nat Methods
UNI2-h UNKNOWN ("extensive A100") ~30,000 est. ~$75k est. ❌ UNKNOWN Chen 2024 HF card
UCE UNKNOWN 1,275–70,000 A100-h $3k–$175k est. ❌ UNKNOWN Rosen 2024 Nat Methods
STATE UNKNOWN unknown ~$125k est. ❌ UNKNOWN Adduri 2025 bioRxiv
TranscriptFormer UNKNOWN unknown unknown ⚠️ partial Pearce 2025 Science
Generative VC POC 1× consumer GPU <100 <$250 ✅ DISCLOSED Lewis & Zueco ICLR 2026

Methodology: FLOPs ≈ 6 × parameters × tokens (Chinchilla heuristic, Epoch AI 2024). Hours → cost at on-demand: A100 80GB ~$2.5/hr (DGX cloud), V100 32GB ~$1.20/hr, H100 ~$2.5–4/hr, TPU v3 ~$2/hr (vendor pricing).

The patterns: 1. Disclosure correlates with sponsorship: every DISCLOSED entry has NVIDIA, TPU, or DGX-Cloud co-authorship. Academic labs without hardware co-marketers disclose less. 2. The $20k floor: Geneformer V2-104M is the cheapest fully-disclosed sc-FM training. Anything below that on the chart is either a POC or undisclosed. 3. The $5M ceiling: Evo2 + ESM-3 define the frontier. Both required vendor compute partnerships ~5 institutions can access. 4. Pathology FMs are mid-tier ($75k–$170k) and disclose more than sc-FMs.

§E. 2025–2026 evaluation papers catalog

The full reckoning canon, organized by what each paper evaluated. As of May 2026, Lane 4 (replication / critique) is the most-published lane of 2025–2026 by a wide margin.

# Paper Evaluated Method Headline finding Axis covered
E.1 Ahlmann-Eltze & Huber 2025 Nat Methods 6 sc-FMs (scGPT, Geneformer, scFoundation, GEARS, CPA + Universal baseline) inference + mean-of-training-perturbations linear baseline none beats linear baseline on Norman / Replogle / Jurkat held-out perturbations perturbation prediction
E.2 Kedzierska et al. 2025 Genome Biology scGPT, Geneformer, UCE + scFoundation zero-shot embedding eval; PCA + kNN baseline scFMs lose to PCA + kNN on cell-type and batch-correction in zero-shot zero-shot embeddings
E.3 Wenkel et al. 2025 Nat Methods sc-FMs vs PCA + linear-additive baseline proposes latent-additive + scGPT-embeddings as the new baseline floor current sc-FMs still don't beat the latent-additive baseline perturbation, w/ new baseline
E.4 Wu et al. 2025 Nat Methods (27 × 29 × 6) 27 methods × 29 datasets × 6 metrics comprehensive benchmark axis-by-axis failure decomposition — each FM family fails on different metrics perturbation (scaled)
E.5 Wu et al. 2025 Genome Biology (Oct 2025) 6 scFMs (Geneformer, scGPT, UCE, scFoundation, LangCell, scCello) gene-level + cell-level tasks, cell-ontology-grounded metrics "no single scFM consistently outperforms others across all tasks" task-dependence
E.6 Liu et al. 2026 Adv Sci (scEval) 10 scFMs × 8 tasks scEval framework "may not consistently outperform task-specific methods … challenges the necessity of developing FMs for single-cell analysis" paradigm-level
E.7 Parameter-free baseline 2026 (bioRxiv 2026.02) sc-FMs vs parameter-free representations direct successor to Ahlmann-Eltze parameter-free reps outperform sc-FMs on downstream benchmarks downstream tasks
E.8 PertEval-scFM ICML 2025 scFM embeddings standardized perturbation framework most models don't outperform simple baselines, particularly on strong / atypical perturbations perturbation (formal venue)
E.9 CellBench-LS (bioRxiv 2026.04) 7 scFMs + PCA / UMAP / scVI stratified low-supervision protocol FMs lead on cell-type recognition; classical methods stay competitive on gene-expression quantification low-supervision
E.10 Han et al. — real-world RNA-seq (bioRxiv 2026.04, industry) scFMs in pharma-relevant deployment real-world data integration robustness industry-grade evaluation finds robustness gaps real-world deployment
E.11 Cellular-dynamics zero-shot (bioRxiv 2026.03) zero-shot scFM embeddings on dynamics RNA-velocity / dynamics reconstruction scFMs fail to recover cellular dynamics in zero-shot RNA velocity / dynamics
E.12 Csendes scPerturBench (BM2 Lab) scGPT replication adversarial-split benchmark original scGPT split was leaky; cleaner splits expose failure replication w/ adversarial split
E.contrarian Foundation Models Improve Perturbation Response Prediction (bioRxiv 2026.02) sc-FMs with sufficient data re-evaluation with sufficient data, FMs DO significantly improve genetic + chemical perturbation predictions perturbation (counter-evidence)

Uncovered axes (Lane 4 frontier as of May 2026): (1) donor-split benchmarks — cross-donor generalization barely measured; (2) cross-tissue transfer — pretraining on tissue A, evaluating on tissue B; (3) time-resolved perturbation — 0h vs 6h vs 24h; (4) cancer-cell-line → primary-tumor transfer; (5) rare-cohort robustness (xVERSE-style: how does FM performance degrade with <100 cells per class?). Each is open and publishable at <$2k.

Where to publish: Nature Methods, Genome Biology, Advanced Science, bioRxiv. RECOMB 2026 for methods-track. NeurIPS D&B for the data-side contribution.

§F. Institutional landscape

The compact table from talk page §3.11, then the extended dossier prose.

Tier Institute / Lab PI(s) Ships Why they matter
Academic FM builders Arc Institute (Palo Alto, non-profit, founded 2021) Patrick Hsu + Hani Goodarzi Evo2, STATE, Tahoe-100M (with Vevo) Largest non-DeepMind compute (2,048× H100 via NVIDIA DGX Cloud); sets the perturbation-atlas substrate
Mahmood Lab (Harvard / BWH / MGB) Faisal Mahmood UNI, UNI2-h, CONCH, PathChat, PathChat-DX, TITAN, CHIEF, HEST-1k Full pathology vertical stack; PathChat-DX = first FDA Breakthrough Designation for generative-AI pathology (Jan 2025)
Theodoris Lab (Gladstone / UCSF) Christina Theodoris Geneformer V1, V2-104M / V2-316M / _CLcancer (Dec 2024) Only academic sc-FM lab with full NVIDIA BioNeMo compute disclosure
Bo Wang Lab (U. Toronto / Vector / UHN) Bo Wang scGPT, scGPT-spatial, MedSAM Defined the sc-FM category; AACR 2026 plenary
Leskovec + Quake (Stanford / SNAP) Jure Leskovec + Steve Quake UCE (Universal Cell Embedding) Cross-species (8 species via ESM2-bridged tokenization)
Industrial FM builders Google DeepMind Žiga Avsec et al. AlphaGenome, AlphaFold 2 / 3, Med-Gemini Closed-weights + hosted-API model; TPU-scale; Isomorphic Labs is the commercial arm
EvolutionaryScale (ex-Meta FAIR, 2024) Alex Rives, Tom Sercu, Thomas Hayes ESM-3 (1.4B / 7B / 98B) Cleanest published FLOPs disclosure in biology FM (10²⁴)
Paige + MSK Thomas Fuchs (Paige) + Zimmermann / Vorontsov Virchow, Virchow2, Virchow2G, FullFocus First FDA 510(k)-cleared general-purpose pathology AI (Jan 2025)
Owkin (Paris / NYC, founded 2016) Industry consortium + academic partners Phikon, Phikon-v2, H-optimus-0, MOSAIC Only major industrial pathology player shipping open weights (Apache-2.0)
NVIDIA BioNeMo (infrastructure layer) (framework team) BioNeMo framework; Geneformer V2 recipe; Evo2 co-author Sets the compute-disclosure norm by making it a co-marketing artifact
Funders & substrates + models CZ Biohub + CZI (SF / Chicago / NY) Theofanis Karaletsos, Steve Quake, Angela Pisco CELLxGENE Census, Tabula Sapiens, Virtual Cell Platform, TranscriptFormer (Science 2025), rBio (2025) Started substrate-only; TranscriptFormer + rBio now ship both substrate and model. Breaks the §3.11 observation that no institute did both.
Critique anchors Ahlmann-Eltze / Huber (EMBL); Kedzierska / Lu (Oxford / MSR); Theis Lab (Helmholtz Munich) (multiple) The 2025 reckoning trio + Lotfollahi causal / compositional work Without these, §4 of the talk doesn't exist
2026 emergence BioMap Research + MBZUAI Le Song, Eric Xing VCHarness (autonomous virtual-cell builder), xTrimo PGLM The China/Middle-East-side competitor to CZ Biohub stack

F.1 Extended dossiers

  • Arc Institute — Palo Alto non-profit (founded 2021). Hosts Hsu Lab and Goodarzi Lab. Shipped: Evo (Science 2024), Evo2 (Nature 2026 with NVIDIA), STATE (bioRxiv 2025), Tahoe-100M (with Vevo, Feb 2025). Disclosure norm: foregrounds data scale; buries compute except where NVIDIA is a co-author. Strategic move 2025: launched "Virtual Cell Atlas" branding and co-sponsored the $50k Virtual Cell Challenge with CZ Biohub.

  • Mahmood Lab (Harvard / BWH / MGB) — PI Faisal Mahmood. Ships a fleet rather than a single FM: UNI / UNI2-h, CONCH, PathChat / PathChat-DX, TITAN, CHIEF, HEST-1k. Disclosure norm: opaque on compute; HF gating tightened Jan 2025. Strategic move 2025: PathChat-DX FDA Breakthrough Designation — first generative-AI pathology tool. AACR 2026: ISBI 2026 keynote on the three-tier FM stack.

  • Theodoris Lab (Gladstone / UCSF) — PI Christina Theodoris. Shipped: Geneformer V1 (Nature 2023), Geneformer V2-104M / V2-316M / V2-104M_CLcancer. Disclosure norm: the only academic sc-FM lab with full compute disclosure via NVIDIA BioNeMo. Only reproducible academic sc-FM training run. Strategic move: domain curation matched the 316M general-domain model at ⅓ the parameters.

  • Bo Wang Lab (U. Toronto / Vector / UHN) — PI Bo Wang. Ships: scGPT, scGPT-spatial, MedSAM, BulkRNABERT. Disclosure norm: opaque on scGPT training compute (50× uncertainty band). Strategic move 2025: explicit agentic-AI pivot — Wang's AACR 2026 talk is framed around clinical agents.

  • Leskovec + Quake (Stanford / SNAP) — Jure Leskovec + Steve Quake. Shipped: UCE, CRADLE-VAE, scGenePT. Disclosure norm: UCE neither preprint nor Nat Methods version discloses training compute — least transparent of major sc-FMs.

  • Google DeepMind — life-sciences team under Kohli + Hassabis. Shipped: AlphaFold 2, AlphaFold 3, AlphaGenome, Med-PaLM / Med-Gemini. Disclosure norm: discloses architecture and benchmarks; rarely full compute. Closed weights + free hosted API non-commercial. The AlphaFold 3 controversy pressured Jan 2026 AlphaGenome code release.

  • EvolutionaryScale — founded 2024 by ex-Meta FAIR protein team (Rives, Sercu, Hayes). Shipped: ESM-3. Disclosure norm: cleanest compute disclosure in biology FM (10²⁴ FLOPs as a single number). License-tiered (1.4B Cambrian NC; 7B / 98B Forge-gated).

  • Paige.AI + MSK — commercial spinout (founded 2018) + Memorial Sloan Kettering (Thomas Fuchs lab). Shipped: Virchow (arXiv 2408.00738), Virchow2, Virchow2G, FullFocus FDA 510(k) Jan 2025. Disclosure norm: clean compute disclosure (512× V100 32GB, ~$170k). License: CC-BY-NC academic; commercial requires Paige license.

  • Owkin — Paris/NYC biotech (founded 2016, ~$300M Series C 2024). Shipped: Phikon, Phikon-v2, H-optimus-0, MOSAIC. Disclosure norm: federated-learning-first; weights MIT / Apache-2.0. Only major industrial pathology player shipping open weights. AACR 2026: H-optimus-0 in poster #1444.

  • NVIDIA BioNeMo — infrastructure team within NVIDIA Healthcare. Not an FM builder per se, but the only public reproduction recipe for Geneformer V2, silent co-author behind Evo2, the Lilly + NVIDIA $1B AI Co-Innovation Lab (JPM 2026). Disclosure norm: treats compute disclosure as co-marketing artifact.

  • CZ Biohub + CZI — Theofanis Karaletsos heads AI for Science; Angela Pisco leads single-cell. Funds + ships: CELLxGENE Census, Tabula Sapiens, Virtual Cell Platform, TranscriptFormer (Science 2025), rBio (Qwen-2.5-3B post-trained with GRPO using TranscriptFormer as verifier). Strategic posture (revised 2025–2026): started substrate-only; TranscriptFormer + rBio mean CZ Biohub now ships both.

  • Ahlmann-Eltze + Huber Group (EMBL Heidelberg) — Constantin Ahlmann-Eltze (now Isomorphic Labs) + Wolfgang Huber. Authored the 2025 Nat Methods linear-baseline paper that retired the sc-FM perturbation leaderboard. Ahlmann-Eltze's EMBL → Isomorphic transition is itself a market signal — the most influential FM critic was hired by Alphabet's biology arm.

  • Theis Lab (Helmholtz Munich) — PI Fabian Theis. Long-running single-cell methods group; built scvi-tools, scArches, framework vocabulary every sc-FM paper inherits. The methodological reference class for the entire field. Author of the 2026 Cell Systems compositional-FM Perspective.

  • Zitnik Lab (Harvard Med) — PI Marinka Zitnik. Ships TxGNN, TDC-2. Strategic posture: graph-FM-for-clinic — the bridge between sc-FMs (gene-level) and agentic-clinical-AI (patient-level).

  • Aviv Regev @ Genentech — Head of gRED. Not shipping FMs as PI, but the agenda-setter for "causal foundation models of cells" (Rood + Hupalowska + Regev 2024 Cell). Genentech under Regev is the largest pharma buyer for academic sc-FM tools.

§G. Cross-references

G.1 The 5 gaps × 9 lanes × 9 tracks matrix

The connective tissue between "what's broken" and "what we'd work on."

Gap (§5) Apply via §B lane Innovate via §C track
5.1 Donor diversity B7 — FM-aided application on under-represented cohort; B8 — synthetic augmentation for rare cohorts C4 — architecture with donor-conditioning priors; C9 — transportability benchmark with donor-split
5.2 Cross-species B3 — domain FM on non-human-cell corpus C7 — cross-species transfer / phylogenetic priors (post-TranscriptFormer); C4 — cross-species equivariance priors
5.3 Causal vs correlational B4 — linear-baseline audit; B9 — active-learning loop targets causal direction C2 — causality-targeting pretraining; C6 — causal benchmarks; C9 — causal transportability
5.4 Compute asymmetry B6 — wrapper tool that democratizes access; B8 — synthetic-data engine reduces wet-lab cost C1 — interpretability tells us what the compute bought
5.5 Evaluation honesty B5 — benchmark curation; B4 — replication catalog (now 11 papers, see §E) C3 — compositional benchmarks; C6 — causal benchmarks; C8 — synergistic-info evaluation; C9 — transportability

G.2 The three §9 pitches × lanes × tracks × buyer

The talk's payoff conversation — three concrete project pitches the group could take on at <$25k each.

Pitch §6 lane §7 track Compute Time Buyer archetype Output
A Pathology FM interpretability + clinical UQ B7 C1 + C5 <$3k 12 mo Clinical-AI (Paige, PathAI, Tempus, Modella → AZ) NeurIPS Mech-Interp + Nat Methods Resource + AACR poster
B Compositional perturb benchmark + linear-baseline re-audit (cancer) B4 C3 <$2k 8 mo AI-native biotech (Recursion, Insitro, Latent, Vevo) NeurIPS D&B + Nat Methods
C Rare-cancer FM-aided subtyping + small domain FM B7 + B3 (applied focus) <$25k 18–24 mo Pharma rare-disease (Pfizer, Sanofi, AZ, Vertex) AACR poster + Clin Cancer Res + Nat Methods

G.3 Pitch detail

Pitch A — Pathology FM interpretability + clinical UQ. Take UNI2-h (or Virchow2 — pick based on license), train sparse autoencoders on its residual-stream activations over a TCGA slice, cluster features and map them to histology grammar (tumor regions, stromal compartments, immune infiltrate patterns). Then layer a post-hoc Bayesian or conformal-prediction head and evaluate calibration on TCGA → CPTAC distribution shift. Output: NeurIPS Mech-Interp workshop (interpretability), Nature Methods Resource (UQ), AACR-Annual-style clinical poster. Why us: builds on existing TCGA familiarity; no wet-lab required.

Pitch B — Compositional perturbation benchmark + linear-baseline re-audit on cancer perturb-seq. Rebuild scGPT + Geneformer + UCE evaluation pipelines on Norman 2019 + Replogle 2022 + Tahoe-100M with cancer-specific cell-line subsets. Add the Ahlmann-Eltze linear baseline and the Wenkel latent-additive baseline. Then design a formal compositional split (A+B held out when A and B seen separately) and prove a theoretical lower bound on what additivity achieves. Output: NeurIPS D&B (benchmark), Nature Methods (linear-baseline replication in cancer context), maybe a second Genome Biology paper. Why us: leverages the AACR-corpus context; cancer-specific findings are differentiating.

Pitch C — FM-aided rare-cancer subtyping + small domain FM. Pick a rare-cancer cohort the public pathology-FM training corpora missed (rhabdomyosarcoma, ATC, NET, mesothelioma, etc.). First publish a frozen-encoder application paper (like AACR #2758 PAX3/7::FOXO1 in rhabdomyosarcoma — Lane 7). Then continual-pretrain a Geneformer-V2-104M-style domain FM on the cohort's matched scRNA-seq + a 100k-cell rare-cancer reference. Output: AACR poster + Clinical Cancer Research or npj Precision Oncology (Year 1), then Nature Methods domain-FM paper (Year 2). Why us: most defensible if we have a clinical collaborator with a rare-cancer cohort.

§H. Appendix carryover

H.1 Likely Q&A questions + scripted answers

Q: Are FMs ready for clinical deployment? A: In pathology, yes — PathChat DX has FDA Breakthrough Designation, and Virchow2 / UNI2-h are in active clinical-grade evaluation. In single-cell biology, no — the linear-baseline reckoning means we don't yet have a single-cell FM whose predictions can be trusted for clinical decisions. Genomic FMs (AlphaGenome) are intermediate: predict variant effects well in benchmark settings but not deployed against clinical decision-making at scale.

Q: Why doesn't more compute solve this? A: It might for pathology — Virchow2G at 1.85B is the latest scale push and we don't yet know if the curve is plateauing. For single-cell, the Geneformer-V2-104M_CLcancer result suggests more compute on general data is not the bottleneck. Domain curation and evaluation are.

Q: What about agentic AI / LLM-driven pipelines? A: MedAgentGym (ICLR 2026) is the strongest demonstration so far: 72k-task sandboxed code-exec gym + Med-Copilot-7B that orchestrates biomedical analyses. Agentic AI is the interface layer on top of the FMs — it doesn't fix underlying FM accuracy problems but makes them more accessible. CZ Biohub's rBio (2025) is the first reasoning model trained on virtual-cell simulations; BioMap's VCHarness (2026) is the first autonomous virtual-cell-builder agent.

Q: Will any of these be available with permissive licenses? A: Mixed picture. Single-cell FMs (scGPT, Geneformer, UCE, scFoundation, CellPLM, GET, TranscriptFormer) are mostly MIT/Apache-2.0. Pathology FMs are mostly CC-BY-NC with the Hibou and Owkin Phikon/H-optimus exceptions (Apache-2.0). Protein FMs split (ESM-2 open, ESM-3 partially gated). Genomic FMs mostly open except AlphaGenome. Track licenses carefully — UNI's Jan 2025 HF gating shift is the cautionary tale.

Q: What benchmark would I actually trust? A: For single-cell FMs: PerturBench latent-additive + scGPT-embeddings baseline is the floor; clear it first, then report your task. For pathology: Campanella et al. 2025 Nature Communications multi-task panel. For genomic: gnomAD-pathogenic + ENCODE/GTEx variant-effect benchmark. For protein: CASP15/16 + retro-validated novel-binder hit rates.

Q: What about the contrarian voice in 2026? A: Foundation Models Improve Perturbation Response Prediction (bioRxiv 2026.02.18) argues that with sufficient data, FMs do improve perturbation prediction. Honest framing: the reckoning was current scFMs trained on current data with current evals fail to clear linear baselines; whether the FM paradigm fails is separate and still open. Virtual Cells Need Context (bioRxiv 2026.02) argues it's a causal transportability problem, not a capacity problem.

H.2 Datasets, weights, code — concrete starting points

Datasets: - Tahoe-100M — 100M cells × 1,100 drugs × 50 cell lines (Vevo / Arc Institute) - CellxGene Census — ~50M+ standardized cells, the canonical scRNA atlas - Replogle Perturb-seq — gold-standard whole-genome perturbation dataset - Open Problems for Single Cell — benchmark hub - HEST-1k — 1k WSI spatial transcriptomics benchmark

Weights: - scGPT (HuggingFace) — MIT - Geneformer V2 (HuggingFace) — Apache-2.0 - UCE — MIT - scFoundation — Apache-2.0 - TranscriptFormer — open - UNI / UNI2-h (HuggingFace) — CC-BY-NC, gated - Virchow (HuggingFace) — CC-BY-NC - Hibou (HuggingFace) — Apache-2.0 - Evo2 — Apache-2.0 - ESM-3 (1.4B) — Cambrian Non-Commercial

Code / benchmarks: - PerturBench — single-cell perturbation benchmark - scPerturBench — adversarial split benchmark - Open Problems — community benchmark hub

The bibliography is organized into 12 buckets. Bold = read first. URLs are direct. Several algorithm and review papers that are adjacent to the talk's topic — perturbation prediction, generative modelling, the architecture debate — without being FM papers have their own dossiers in Adjacent methods.

Position papers — start here

The 2025 critique trio — the reckoning

Single-cell FM technical papers

  • Cui et al. 2024 Nat Methods — scGPT — defined the field shape.
  • Theodoris et al. 2023 Nature — Geneformer — first atlas-pretrained transformer.
  • Geneformer-V2 HuggingFace card (Dec 2024) — domain-curated _CLcancer matches 316M at ⅓ params.
  • Rosen, Brbić et al. 2024 — UCE — cross-species via protein-language tokens.
  • Hao et al. 2024 Nat Methods — scFoundation.
  • Yang et al. 2022 Nat Mach Intell — scBERT.
  • Wen et al. 2024 ICLR — CellPLM.
  • Pearce et al. 2025 Science — TranscriptFormer — first generative cross-species sc-FM.
  • Adduri et al. 2025 bioRxiv — STATE (Arc Institute).
  • Schaar et al. 2025 Nat Methods — Nicheformer — spatial-omics sc-FM.
  • Zhang et al. 2025 Nat Commun — CellFM.
  • Jiang & Xie 2026 — xVERSE (bioRxiv 2026.04) — transcriptomics-native.
  • Kalfon et al. 2025 Nat Commun — scPRINT — pretrained on 50M cells for robust zero-shot gene-network inference.
  • Kalfon, Peyré & Cantini 2025 — scPRINT-2 (bioRxiv 2025.12) — next-generation cell FM and a benchmark suite.

Pathology FM technical papers

  • Chen et al. 2024 Nat Medicine — UNI.
  • Vorontsov et al. 2024 Nat Medicine — Virchow.
  • Zimmermann et al. 2024 arXiv — Virchow2 / Virchow2G.
  • Wang et al. 2024 Nature — CHIEF.
  • Xu et al. 2024 Nature — Prov-GigaPath.
  • Lu et al. 2024 Nat Medicine — CONCH.
  • Lu et al. 2024 Nature — PathChat.
  • Nechaev et al. 2024 arXiv — Hibou-L / Hibou-B — Apache-2.0 top-tier option.
  • Filiot et al. 2024 — Phikon-v2.
  • Saillard et al. 2024 Nat Medicine — H-Optimus-0.
  • Ding et al. 2025 Nat Med — TITAN.
  • Xu et al. 2025 Nat Commun — mSTAR.
  • Yan et al. 2025 npj Digit Med — PathOrchestra.
  • Ma et al. 2026 Nat Biomed Eng — GPFM.
  • Campanella et al. 2025 Nat Communications — clinical-grade benchmark.
  • Janowczyk et al. 2025 Nat Med — first real-world clinical UNI deployment.

Genomic / DNA FM technical papers

Protein FM technical papers

  • Hayes et al. 2025 Science — ESM-3.
  • Lin et al. 2023 Science — ESM-2 / ESMFold.
  • Abramson et al. 2024 Nature — AlphaFold 3.
  • Jumper et al. 2021 Nature — AlphaFold 2.
  • Watson et al. 2023 Nature — RFdiffusion.
  • Madani et al. 2023 Nat Biotechnol — ProGen2.
  • Notin et al. 2023 NeurIPS — ProteinGym benchmark.

Virtual cell-specific work

  • Bunne et al. 2024 Cell — virtual cell perspective (canonical).
  • CZ Biohub Virtual Cell Program — program page.
  • Roohani, Hsu et al. 2025 — Tahoe-100M — first large-scale perturbation atlas framed as virtual-cell training data.
  • Arc Institute 2025 — Virtual Cell Atlas + STATE.
  • CZ Biohub Virtual Cell Challenge 2025 — first community benchmark, $50k prize.
  • Lotfollahi et al. 2023 Nat Cell Biol — CPA (Compositional Perturbation Autoencoder).
  • Roohani et al. 2024 Nat Biotechnol — GEARS.
  • Cheng et al. (BioMap, Le Song / Eric Xing) 2026 bioRxiv — VCHarness — autonomous virtual-cell builder.
  • CZ Biohub 2025 — rBio — reasoning model trained on TranscriptFormer.

Benchmarks & datasets

  • Replogle et al. 2022 Cell — Whole-genome Perturb-seq.
  • Norman et al. 2019 Science — combinatorial Perturb-seq.
  • Open Problems v2 — openproblems.bio.
  • PerturBench / scPerturBench / PertEval-scFM / CellBench-LS — see §E.
  • Tabula Sapiens Consortium 2022 Science.
  • CELLxGENE Census.
  • HuBMAP 2023 Nature.
  • HEST-1k.
  • CASP15 / CAPRI.

Mechanistic interpretability — the 2025–2026 wave

  • Adams et al. 2025 PNAS — SAEs uncover features in protein language model representations.
  • Simon & Zou 2026 arXiv 2603.02952 — SAEs reveal organized biology but minimal regulatory logic in sc-FMs.
  • bioRxiv 2025 — SAEs Reveal Interpretable Features in Single-Cell FMs (independent confirmation on scGPT).
  • Hibou-LP team 2024–2025 arXiv 2407.10785 — first pathology-FM SAE.
  • bioRxiv 2026 — "What Do Biological Foundation Models Compute?" — synthesis across families.
  • Hunklinger & Ferruz 2026 Nat Mach Intell — "Towards the explainability of protein language models" — review of interpretability methods for protein FMs. → resource page.

Compute / cost methodology

Industry / strategy / clinical context

Cross-vault index

H.4 People to follow

  • Christina Theodoris (Gladstone / UCSF) — Geneformer creator. Lab · bio.
  • Bo Wang (U. Toronto / Vector / UHN) — scGPT PI. WangLab · X @BoWang87.
  • Mohammad Lotfollahi (Wellcome Sanger / Open Targets) — scArches, scPoli. Lab.
  • Aviv Regev (Genentech) — Head of gRED; HCA co-founder. bio.
  • Patrick Hsu (Arc Institute) — Evo2 + STATE co-PI. Hsu Lab @ Arc · patrickhsu.com.
  • Faisal Mahmood (Harvard / BWH / MGB) — UNI, CONCH, PathChat, TITAN. faisal.ai.
  • Jure Leskovec (Stanford / SNAP) — UCE co-PI. stanford.edu/~jure.
  • Sara Mostafavi (UW Allen School) — deep-learning models of differential gene expression. Lab.
  • Constantin Ahlmann-Eltze (Isomorphic Labs, ex-Huber/EMBL) — 2025 linear-baseline first author. const-ae.name.
  • Wolfgang Huber (EMBL Heidelberg) — Bioconductor + statistics-of-genomics elder. Huber Group.
  • Theofanis Karaletsos (CZI, Head of AI for Science) — runs virtual-cell program + rBio. CZI tech blog.
  • Kasia Z. Kedzierska (Oxford / MSR) — zero-shot critique first author. Scholar.
  • Hani Goodarzi (Arc / UCSF) — STATE co-lead. Goodarzi Lab.
  • Fabian Theis (Helmholtz Munich) — author of 2026 Cell Systems compositional-FM Perspective. Theis Lab.
  • Le Song / Eric Xing (BioMap Research + MBZUAI) — VCHarness; the autonomous virtual-cell-builder direction.
  • Marinka Zitnik (Harvard Med) — TxGNN, TDC-2. Zitnik Lab.

H.5 Newsletters / blogs / podcasts

H.6 Live benchmarks / leaderboards

H.7 Recorded talks / lecture series

  • M2D2 Talks (Valence + Mila) — ~200 archived ML-for-drug-discovery seminars.
  • CZI Virtual Cells Platform — workshops & webinars — recorded sessions on FM training, benchmarks, evaluation.
  • scverse community talks — annual conference recordings.
  • Simons Institute "ML & Statistics for Single-Cell Genomics" — public lectures.
  • Broad Institute MIA seminars — bio-ML talks.
  • NeurIPS / ICLR Generative & Experimental Perturbations workshops — annual recordings.

H.8 Communities / Discord / Slack

  • scverse Zulip — Scanpy / AnnData / squidpy / scvi-tools developer chat. scverse.zulipchat.com.
  • Open Problems community — GitHub Discussions + Slack. github.com/openproblems-bio.
  • Polaris Hub Slack — drug-discovery ML benchmark community.
  • Hugging Face Biology / Bio-ML — model cards + discussion. huggingface.co/biology.
  • EleutherAI #biology channel — open-research; Discord invite via eleuther.ai.

H.9 Timing cheat sheet

The talk in 12 slides — speaker's eye view

# Slide Source
1 Title + operational definition of virtual cell Act 1 opening
2 Why now — 2024–25 inflection (CZ Biohub, Tahoe-100M, PathChat FDA) Act 1
3 The scatter plot — 2023 paradigm → 2024 ambition → 2025 reckoning → 2026 response Act 1
4 Today's institutional landscape table Act 2
5 The 5 FM families Act 2
6 The 2025 reckoning — linear baseline + contrarian + theoretical underpinning + new architectures Act 1 close
7 The 5 gaps mapped to lanes and tracks Act 3 open
8 The 9 lanes table + decision tree + combinations Act 3 part A
9 The 9 tracks table + unifying frame Act 3 part B
10 Commercial — 3 buyer archetypes + per-pitch buyer map Act 4 open
11 The 3 pitches table Act 4 close
12 AACR 2026 as the live evaluation surface Act 5
Section Time Slides Discussion pivot
Act 1 — the 2023–2026 arc + scatter plot 10 min 3 "Where on the plot would our work go?"
Act 2 — today's landscape 5 min 2 "Which family is closest to our existing work?"
Act 3 — 9 lanes + 9 tracks 15 min 5 "Lane 1 + Lane 7 are the AACR posters; Lane 8 + Lane 9 are the 2026 frontier. Which of our datasets fits?"
Act 4 — commercial + pitches 10 min 3 "Of the three pitches, which has the clearest non-academic check writer?"
Group discussion 15 min 2 The whole point
Act 5 — close 3 min 1 —
Buffer 4 min — —
Total 62 min 16 —

If 75 min: expand group discussion to 25 min. If 45 min: cut Act 3 to lanes-only (skip tracks), trim Act 4 to per-pitch buyer map only.

H.10 2026 updates — the 37-paper catalog

Carried over from talk-page §11.6, organized as in the original. Bold = new in this revision wave. Each is annotated with what changed because of it.

New benchmarks confirming or extending the 2025 reckoning

Position / framing papers

The virtual-X wave — beyond the single cell

New agentic-AI-for-biology systems

New single-cell / spatial foundation models

New pathology foundation models

Mechanistic interpretability — the wave that broke between drafts

Genomic FM follow-ups

Net effect on the talk (May 2026 revision)

  • Act 1 thesis gets sharper: Liu + Theis generalize past perturbation. The question is no longer "do scFMs predict perturbations" but "does the monolithic-FM paradigm earn its compute at all."
  • Act 2 institutional landscape revised: CZ Biohub now ships both substrate and model (TranscriptFormer, rBio). Breaks the §3.11 substrate-vs-model split.
  • Act 1 reckoning evidence base is now ten papers, not three.
  • Act 1 gains the contrarian voice: FMs Improve Perturbation (bioRxiv 2026.02).
  • Act 1 gains theoretical underpinning: Virtual Cells Need Context — causal transportability problem, structural not capacity-bounded.
  • Act 3 Tracks 7, 8, 9 are net-new in 2026: cross-species, SIS, causal transportability.
  • Act 4 commercial framing notes: CZ Biohub (TranscriptFormer + rBio) and BioMap (VCHarness) as the two fully-open / open-research stacks.
  • Act 3 Track 4 gains TxPert — the reckoning's own authors (Wenkel) now publishing knowledge-graph-driven methodological responses. The reckoning → response loop is closing.
  • Act 5 closing now has a virtual-X-is-forking point: virtual cell (this talk) → virtual embryo (Cao/Lu/Qiu 2026) → virtual organ (Zhou 2026 lungs + digital-organ Nat Methods 2025).
  • Agentic ↔ FM convergence is now three-way: rBio reasons, VCHarness builds, CellVoyager analyzes. Plus the Li et al. 2026 Nat Biotech "in silico team science" position paper frames the whole intersection.

Pointers back to the main talk

Last updated: 2026-05-13.