Skip to content

Foundation Models in Biology — 2026 State of Play

A cross-vault synthesis of every foundation-model (FM) dossier in this corpus. Links resolve to per-tool deep dives in the AACR / ICLR / ISBI / Single-Cell-Genomics vaults.

The corpus distinguishes 15 net-new dossiers written for this synthesis (single-cell, pathology, genomic, protein, multimodal) from 12 existing dossiers that received an FM-comparison deepening pass in May 2026. The 27 entries cover the FMs that AACR 2026 sessions and ICLR/ISBI 2026 papers either build on or directly evaluate.

At a glance

Family Anchor model(s) 2026 SOTA Load-bearing benchmark Open question
Single-cell scGPT, Geneformer scGPT-spatial (Feb 2025); Geneformer-V2-104M_CLcancer SC-Arena, PerturBench, scPerturBench, Open Problems v2 Can any sc-FM beat mean-of-training-perturbations on perturbation prediction?
Pathology UNI, Prov-GigaPath, CHIEF Virchow2 / UNI2-h (Jan 2025) Campanella et al. 2025 clinical benchmark; HEST-1k; BACH; CRC polyp Where does the scale curve plateau? (Virchow2G at 1.85B is the current ceiling)
Genomic / DNA Evo2, Enformer AlphaGenome (Nature 2026) — 25/26 variant-effect wins gnomAD-pathogenic, ENCODE/GTEx tracks, GUE/GUE+, NT 18-task Is ICL still a viable genomic capability after AlphaGenome's track-prediction win?
Protein ESM-2 / ESM-3, AlphaFold3 ESM-3 (Science 2025; 98B params) Long-context folding; novel-binder design hit rates Generative protein design vs structure-prediction — which yields wet-lab hits faster?
Multimodal BioMedCLIP, PathChat, LLaVA-Med PathChat-2 (Nature 2024 extension; >1M instructions) PMC-VQA, SLAKE, VQA-RAD, PathQABench Will FDA-cleared generative pathology copilots (PathChat DX got Breakthrough Designation Jan 2025) ship in 2026?

The two biggest stories of 2025-2026

1. The linear-baseline reckoning for single-cell FMs

Three independent 2025 papers — Ahlmann-Eltze & Huber, Nature Methods, Kedzierska et al. Genome Biology, and Wenkel et al. Nature Methods — showed that scGPT, Geneformer, scFoundation, and UCE fail to beat trivial linear baselines (mean-of-training-perturbations or latent-additive) on perturbation-prediction tasks that the original papers reported as wins. PerturBench's latent-additive + scGPT-embeddings baseline is now the floor every new claim must clear.

Consequences:

  • AACR 2026 has zero posters with scGPT/Geneformer/UCE in the title (the pathology-FM stack appears in 18 posters by contrast).
  • The SC-Arena ICLR 2026 paper is now a scorecard, not a contestant.
  • Geneformer V2 inverts the scale-wins narrative: the cancer-domain V2-104M_CLcancer checkpoint (14M cancer cells, 104M params) matches or beats the 316M deep model. Domain-specific pretraining beat raw scale.

This is the clearest discipline correction in single-cell ML since 2023. The frame moving forward: any sc-FM that wants to claim utility for AACR-style downstream tasks must publish (a) cell-type classification, (b) drug-perturbation prediction, and (c) cross-tissue transfer numbers all against the LA-on-scGPT baseline.

2. The Virchow rebalancing in pathology FMs

The Nature 2024 ordering (Prov-GigaPath → UNI → CHIEF) flipped in 2025:

  • Virchow2 (Paige+MSK, 632M ViT-H, 3.1M WSIs) took first place on the Campanella et al. 2025 Nature Communications clinical benchmark.
  • UNI2-h released Jan 14 2025 — 681M ViT-H/14, 200M+ tiles, added IHC — and now leads on BACH (1.0 BAcc) and several Mahmood-stack downstream tasks.
  • CHIEF has shipped no v2 in 2025. AGPL-3.0 + gated weights is now a real adoption tax versus Apache-2.0 alternatives, explaining its near-zero AACR 2026 traction.
  • PathChat / PathChat DX received FDA Breakthrough Device Designation in January 2025 — the first generative-AI pathology tool to do so.
  • Hibou (HistAI — not Owkin) is the only Apache-2.0 top-tier pathology FM. Owkin's analogue is Phikon / Phikon-v2.

The corollary: pathology FMs are the only family in this corpus with FDA-track regulatory activity in 2025-2026. Single-cell, genomic, and protein FMs remain pre-clinical-tool.

Family deep dives

Single-cell FMs

Five major models. Two existed pre-2025; three landed in 2024-2025 alongside the linear-baseline critique.

Model Architecture Params Pretraining corpus Public weights Page
scGPT Encoder-decoder transformer 51M 33M cells ✅ MIT scgpt.md
Geneformer Encoder-only BERT 10M / 104M / 316M 30M / 95M cells (V2) ✅ Apache-2.0 geneformer.md
UCE Encoder-only ViT-style 650M ~36M cells / 8 species ✅ MIT uce.md
scFoundation xTrimoGene encoder 100M 50M+ human cells ✅ Apache-2.0 scfoundation.md
GET Multi-modal transformer ~50M scATAC + scRNA, ~1.3M cells ✅ MIT get.md
CellPLM Cell-as-token transformer ~80M scRNA, paired spatial ✅ MIT cellplm.md

Non-FM but adjacent: Cell2Location — the Bayesian incumbent for spatial deconvolution, now threatened by scGPT-spatial + CellPLM as pretrained alternatives.

Generative virtual-cell models (a separate sub-class): - Generative Virtual Cells (Lewis & Zueco, ICLR 2026) — jointly-updated planner + world model - STATE (Arc Institute, 2025) — frozen Tahoe-100M snapshot

Pathology FMs

Model Architecture Pretraining corpus Public weights License Page
UNI / UNI2-h ViT-L/16 → ViT-H/14 100M tiles → 200M+ tiles ✅ (gated 2025) CC-BY-NC uni.md
Prov-GigaPath DINOv2 ViT-G/14 1.4B tiles from 171K WSIs ✅ CC-BY-NC prov-gigapath.md
CHIEF ViT-L + CTransPath 60K WSIs ✅ (gated) AGPL-3.0 chief.md
Virchow / Virchow2 DINOv2 ViT-H 1.5M / 3.1M WSIs ✅ Virchow only CC-BY-NC virchow.md
Hibou-B / Hibou-L DINOv2 ViT 1.14M slides ✅ Apache-2.0 hibou.md
PathChat / PathChat-2 Vision-language (LLaVA-style) 457K instructions (PathChat-2: >1M) ❌ (FDA Breakthrough) Proprietary pathchat.md

Synthesis context: Mahmood pathology FM keynote (ISBI 2026) and FM-pathology-traction landscape.

Genomic / DNA FMs

Model Architecture Context Params Tokenization Page
Nucleotide Transformer Encoder-only 6 kb 50M–2.5B 6-mer (vocab 4,104) nucleotide-transformer.md
DNABERT-2 Encoder + ALiBi + FlashAttn 128 bp (pre-tr) 117M BPE (vocab 4,096) dnabert-2.md
HyenaDNA Hyena implicit-conv (no attention) up to 1M 6.6M–1.6B Single-nucleotide (vocab 12) hyenadna.md
Evo2 StripedHyena SSM up to 1M 7B / 40B Byte-level genomic-icl-evo2.md
AlphaGenome U-Net + transformer bottleneck 1 Mb ~450M (vendor) Single-nucleotide alphagenome.md

The line dividing this family: track-prediction transformers (Enformer, AlphaGenome) vs generative DNA LMs (Evo2, HyenaDNA). AlphaGenome wins 25/26 on regulatory variant-effect evaluations, but Evo2 holds unique ICL ground because you cannot prompt a track-prediction model.

Protein + multimodal FMs

Model Modality Notes Page
ESM-3 Protein (3-track: seq + structure + function) 98B params; generated esmGFP de novo (36% identity to avGFP) esm-3.md
Proteina Complexa Atomistic protein binder design 63.5% PDGFR hit rate; first de novo carbohydrate binders proteina-complexa.md
TEA De novo protein with epitope conditioning ICLR 2026 tea-de-novo-protein.md
BioMedCLIP Vision-language (biomedical) PMC-15M (15.3M image-text pairs); MIT licensed biomedclip.md

Virtual cells — why the problem is hard

A "virtual cell" is shorthand for a learned model that can predict any cell's gene expression under any perturbation. Three orthogonal problems block it:

  1. Evaluation — most "virtual cell" papers report SOTA on tasks that turn out to be dominated by linear baselines once you run the right control. SC-Arena is the post-2025 attempt at honest scorekeeping.
  2. Training data — Tahoe-100M (Vevo / Arc Institute, 2025) is the largest perturbation atlas, but coverage is still drug-skewed and donor-shallow.
  3. Generation vs prediction — do you (a) condition a generative model on a perturbation token and sample, or (b) train a deterministic head to predict perturbed expression? Lewis & Zueco's Generative Virtual Cells bets on (a); Arc's STATE bets on (b).

What success looks like in 2026: an FM that beats latent-additive + scGPT-embeddings on perturbation prediction held out by perturbation, donor, and cell type. As of May 2026, no public model has cleared all three splits.

Open questions for 2026

  1. Single-cell: Will any FM beat the LA-on-scGPT baseline on the post-2025 splits? (If not, the field needs a new pretraining objective, not more cells.)
  2. Pathology: Does Virchow2G (1.85B) signal that pathology FM scale still pays, or did UNI2-h's H/14 + 200M tiles already hit the curve plateau?
  3. Genomic: Is ICL still useful for genomics now that AlphaGenome wins 25/26 on variant effects without prompting? Or is Evo2's ICL territory niche?
  4. Protein: Will ESM-3's de novo design hit rate hold up at wet-lab scale, or is esmGFP a single anecdote?
  5. Clinical: Will any second generative-pathology copilot (after PathChat DX) reach FDA Breakthrough Designation in 2026?

How to use this page

  • Linked from individual conference vaults: each dossier above sits in its host vault (AACR / ICLR / ISBI). This page is the cross-vault map.
  • Linked from the Speed Read: the 0:35–1:15 segment of the 90-minute path covers the highest-value FM dossiers.
  • Backed by primary sources: every numeric claim in the dossiers ends with a citation (Nature / Nature Methods / Science / ICLR / NeurIPS / arXiv / bioRxiv / HuggingFace model card).

If you're tracking a specific FM family, start at the family-deep-dive table above and follow the tool links. If you're tracking the discipline-level debate, the linear-baseline reckoning and Virchow rebalancing sections summarize 2025-2026's two biggest course-corrections.