Theme — foundation-model

Items in corpus: 147

Papers

• How to build the virtual cell with AI — Bunne et al., Cell, 2024
  The canonical virtual-cell thesis — the goal everything else chases.
• The future of automated single-cell analysis — Rood, Regev et al., Cell, 2024
  Regev's reference-mapping framing of the same ambition.
• From modality-specific to compositional foundation models — Theis et al., Cell Systems, 2026
  Post-reckoning forward path — stop scaling monoliths.
• Generalist biological AI — modeling the language of life — Rao et al., Nature Biotechnology, 2026
  Position paper unifying the five FM families.
• Agentic AI and in silico team science — Li et al., Nature Biotechnology, 2026
  Position paper behind the agentic-FM intersection.
• Cellular Intelligence — Engineering Cell Fate, towards a foundation model for cell signaling — preprint, 2026
  Reframes virtual cells as signaling-FMs rather than transcriptome-FMs.
• Earliest 'linear baselines are competitive' warning — Boiarsky et al., NeurIPS workshop, 2023
  First warning shot — read it to see how early the signal was.
• Deep-learning predictions of gene expression don't generalize — Ahlmann-Eltze & Huber, Nature Methods, 2025
  THE canonical reckoning paper — start here.
• Limits of zero-shot foundation models in single-cell biology — Kedzierska et al., Genome Biology, 2025
  Extends the result to UCE and the zero-shot setting.
• 27 methods × 29 datasets × 6 metrics — Wu et al., Nature Methods, 2026
  First axis-by-axis failure decomposition.
• No single scFM consistently outperforms — Wu et al., Genome Biology, 2025
  6 scFMs, cell-ontology-grounded metrics.
• scEval — challenges the necessity of sc-FMs — Liu et al., Advanced Science, 2026
  Strongest 'is the paradigm worth it' framing.
• Parameter-free baseline beats sc-FMs — bioRxiv, 2026
  Cleanest post-reckoning headline; direct Ahlmann successor.
• Foundation Models Improve Perturbation Response — bioRxiv, 2026
  Contrarian voice — FMs DO improve with enough data.
• Virtual Cells Need Context, Not Just Scale — bioRxiv, 2026
  Names the theory: a causal-transportability problem (Pearl).
• Beyond Alignment — Synergistic Information Score — bioRxiv (Microsoft), 2026
  Microsoft's metric for whether a multimodal FM extracts genuinely cross-modal information.
• SAEs reveal organized biology but minimal regulatory logic — Simon & Zou, arXiv, 2026
  Mechanistic explanation of the reckoning, on Geneformer + scGPT.
• What do biological foundation models compute? — bioRxiv, 2026
  Synthesis across families — the wave's summary read.
• Generative Virtual Cells — Lewis & Zueco, ICLR Gen²-workshop, 2026
  Workshop proof (<$250 compute) — joint planner + world model under validation gating.
• sc-FM Perturbation Adapter — Maleki et al., ICLR LMRL workshop, 2026
  Drug-conditional adapter (<1% trainable params) on a frozen FM — generalises where CPA/GEARS/scGen don't.
• PEFT / LoRA recipe for sc-FMs — Hossain et al., arXiv 2412.13478, 2024
  Concrete PEFT recipe for single-cell FMs (Lane 2).
• The Rising Costs of Training Frontier AI Models — Cottier, Rahman et al., arXiv, 2024
  Economic backdrop for the compute-disclosure argument.
• Invariant Risk Minimisation — Arjovsky et al., arXiv, 2019
  Causal-objective method cited as basis for counterfactual pretraining.
• X-Cell — Scaling Causal Perturbation Prediction via Diffusion Language Models — Wang, Karimzadeh, Ravindra et al. (Bo Wang group), bioRxiv 2026.03.18.712807, 2026
  Diffusion LM for perturbation — direct rival to State / scGPT-perturb framing.
• Lingshu-Cell — generative cellular world model for transcriptome modelling — Han Zhang, Guo-Hua Yuan, Chaohao Yuan, Tingyang Xu, Tian Bian, Hong Cheng, Wenbing Huang, Deli Zhao, Yu Rong, preprint, 2026
  Chinese-industry-led generative cellular world model for transcriptome.
• Beyond Alignment — Synergistic Integration for Multimodal Cell FMs — preprint, 2026
  Argues multimodal cell-FMs need synergistic (not just aligned) integration.
• Cellular Intelligence — Engineering Cell Fate, towards a FM for cell signaling — preprint, 2026
  Reframes virtual cells as signaling-FMs.
• OmniCell — multimodal spatial FM — bioRxiv, 2025
  Multimodal spatial-transcriptomics FM.
• CLM-X — spatial-temporal cell-language model — bioRxiv, 2026
  Spatial + temporal context for cell language modelling.
• CELLama — cell language model — Advanced Science, 2026
  Alternative transformer-based cell FM.
• CancerFoundation — bioRxiv, 2024
  Cancer-domain-specific single-cell FM.
• TxGNN — knowledge-graph FM for therapeutics — Zitnik et al., Nature Medicine, 2024
  Graph FM bridging gene-level and patient-level AI.
• scPEFT — parameter-efficient fine-tuning for single-cell LLMs — He et al., Nature Machine Intelligence, 2025
  Adapter-style PEFT cuts scLLM fine-tuning to <4% params and beats zero-shot — practical answer to the reckoning.
• DeepRare — agentic system for rare-disease diagnosis with traceable reasoning — Zhao et al., Nature, 2026
  Multi-agent LLM diagnosis system that outperformed physicians on rare-disease cases — the agentic-clinical proof point.
• KEEP — knowledge-enhanced pathology vision-language foundation model — MAGIC-AI4Med (SJTU + Shanghai AI Lab), Cancer Cell, 2025
  Knowledge-graph-conditioned pathology VLM from Shanghai Jiao Tong — Chinese answer to PathChat/Virchow.
• DIA-BERT — pre-trained transformer for DIA proteomics — Tiannan Guo lab et al., Nature Communications, 2025
  Transformer pre-trained on 276M peptide precursors; +51% protein IDs vs DIA-NN. Westlake's bid for proteomics FM.
• DrugCLIP — contrastive protein-molecule virtual screening — NeurIPS 2023, 2023
  CLIP for protein-molecule retrieval — zero-shot virtual screen baseline that DrugReflector and successors build on.

Reviews

• Single-cell foundation models — bringing AI into cell biology — Singh et al., Experimental & Molecular Medicine, 2025
  Clean mid-2025 review — the orientation read.
• AI virtual cell — review — Shugang Zhang, Huasen Jian et al., Briefings in Bioinformatics, 2026
  Mandarin-sphere review of the AIVC landscape.
• Learning beyond fine-tuning — Zheng et al., Nature Machine Intelligence, 2025
  Landscape of model-adaptation paradigms past fine-tuning.
• Genome language models — Consens et al., Nature Machine Intelligence, 2025
  Review of transformer architectures for DNA.
• Transformers vs state-space models — Tiezzi et al., Nature Machine Intelligence, 2025
  Transformer / recurrent / state-space crossroads.
• AI revolutionizes cellular metabolic pathway reconstruction (AIVM review) — Trends in Biochemical Sciences, 2026
  Names the metabolism-flavoured AIVC subfield (AI Virtual Metabolism) as a coherent agenda.

Tools

• scBERT — Yang et al., Nature Machine Intelligence, 2022
  BERT-style precursor that pre-dated the 2023 sc-FM wave.
• scGPT — Cui et al., Nature Methods, 2024
  Defined the category — genes + cells as tokens. The reckoning's main target.
• Geneformer / Geneformer V2 — Theodoris et al., Nature, 2023
  First atlas-pretrained transformer; rank-based tokenisation. V2 (Dec 2024) matched 316M at ⅓ the compute.
• Universal Cell Embedding (UCE) — Rosen et al., Nature Methods, 2024
  Cross-species sc-FM — bridges through ESM2 protein embeddings (8 species).
• scFoundation — Hao et al., Nature Methods, 2024
  Read-depth-aware attention; one of the four reckoning regulars.
• CellPLM — Wen et al., ICLR, 2024
  Cell-as-token sc-FM; beats scGPT/Geneformer on cell-typing cheaply.
• TranscriptFormer — Pearce et al., Science, 2025
  First generative cross-species sc-FM; CZ Biohub flagship (112M cells, 12 species).
• Nicheformer — Theis lab, Nature Methods, 2025
  Spatial-omics sc-FM with a niche-aware objective.
• scMulan
  Multitask single-cell language model; benchmarked in CellBench-LS.
• GenePT / scELMo
  NOT trained sc-FMs — frozen LLM embeddings of gene names + LR head (~$200) matching scGPT zero-shot.
• CellFM — Zhang et al., Nature Communications, 2025
  Another sc-FM from the 2025 wave.
• scPRINT — Kalfon et al., Nature Communications, 2025
  50M-cell pretraining aimed at robust zero-shot gene-network inference.
• scPRINT-2 — Kalfon, Peyré & Cantini, bioRxiv, 2025
  Next-gen cell FM AND benchmark suite in one preprint.
• xVERSE — transcriptomics-native sc-FM — bioRxiv, 2026
  First evidence the architectural choice is load-bearing (+17.9% over LM-derived sc-FMs).
• TxPert — multiple-knowledge-graph perturbation prediction — Wenkel et al., Nature Biotechnology, 2026
  The reckoning answering itself — Wenkel co-authored both critique and response.
• rBio — reasoning model trained on TranscriptFormer — CZ Biohub, 2025
  Agent that REASONS OVER a virtual cell as a verifier.
• VCHarness — autonomous virtual-cell builder — bioRxiv (BioMap), 2026
  Agent that BUILDS virtual-cell models end to end.
• Enformer — Nature Methods, 2021
  Pre-FM-era gene-expression-from-sequence model.
• AlphaGenome — Avsec et al. (DeepMind), Nature, 2025
  Variant-effect SOTA; replaces Enformer + Borzoi; 25/26 wins at 1-Mb context.
• Evo 2 — Brixi et al. (Arc / NVIDIA / Stanford / UCB / UCSF), Nature, 2026
  Only genomic FM with demonstrated in-context learning; 1M-token context, ~$5M training.
• Nucleotide Transformer — InstaDeep
  Early DNA FM pretrained across many genomes for variant and regulatory tasks.
• DNABERT-2
  BPE-tokenised DNA FM; efficiency-focused DNABERT successor.
• HyenaDNA
  State-space (non-transformer) genomic FM for very long context at single-nucleotide resolution.
• Caduceus
  Bidirectional, reverse-complement-equivariant state-space DNA FM.
• ESM-2 / ESMFold — Science, 2023
  Open protein LM; UCE's cross-species bridge; still the academic workhorse.
• ESM-3 — Hayes et al. (EvolutionaryScale), Science, 2025
  98B-param multimodal protein FM; generated esmGFP de novo; cleanest compute disclosure.
• UNI — Nature Medicine, 2024
  Original Mahmood-lab pathology tile encoder.
• UNI2-h — Mahmood Lab, 2025
  Mahmood-stack flagship tile encoder (681M); HF-gated since Jan 2025.
• Virchow2 / Virchow2G — Zimmermann et al. (Paige + MSK), arXiv, 2024
  Current pathology-FM SOTA and most hardware-transparent; 632M ViT-H/14 on ~2B tiles / ~225k patients.
• CHIEF
  Pathology FM for cancer detection and prognosis across tissue types.
• Prov-GigaPath — Microsoft + Providence
  Whole-slide pathology FM trained on real-world clinical slides.
• Phikon / Phikon-v2 — Owkin
  Owkin's pathology FMs; notable as Apache-2.0 exceptions in a license-restrictive family.
• H-optimus-0 — Bioptimus
  Open-weights pathology tile encoder; permissive-license exception.
• Hibou
  Open pathology FM (Apache-2.0).
• CONCH — Mahmood Lab
  Contrastive vision-language pathology FM.
• TITAN — Mahmood Lab
  Slide-level model on top of the UNI2-h tile encoder.
• BioMedCLIP
  Biomedical image–text contrastive model on large-scale figure–caption pairs.
• Med-Gemini — Google
  Google's medical multimodal model family for clinical reasoning and image Q&A.
• VCHarness (GenBio AI) — GenBio AI
  Genbio's autonomous virtual-cell builder agent — see the paper too.
• AROMA — Tencent
  Tencent-flagged virtual-cell project — name only.
• mSTAR — multimodal spatial-transcriptomics aggregation FM — Xu et al., Nature Communications, 2025
  Spatial pathology FM aggregating ST + WSI signals.
• GPFM — Gigapixel Pathology Foundation Model — Ma et al., Nature Biomedical Engineering, 2026
  Large-scale gigapixel pathology FM.
• PathPT — pathology transformer — Nature Communications, 2026
  Pathology FM variant in the 2026 wave.
• PathOrchestra — ensemble pathology FM — Yan et al., npj Digital Medicine, 2025
  Ensemble approach across pathology FMs.
• Evo (v1) — Nguyen et al., Science, 2024
  First genomic FM with demonstrated in-context learning (Evo2 successor).
• DNABERT — Ji et al., Bioinformatics, 2021
  Original BERT-style genomic FM.
• LangCell / LangStem
  LLM-based cell annotation tool.
• GET — gene expression toolkit
  sc-FM-based gene-expression analysis toolkit.
• CytoVerse — browser-native sc-FM via ONNX — bioRxiv, 2026
  In-browser sc-FM inference; pushes inference to where the data is.
• signifinder — sc-FM wrapper for Bioconductor — Calura lab, U Padua
  Bioconductor interface to sc-FMs.
• NVIDIA BioNeMo Framework
  Sets compute-disclosure norm; co-marketed with Geneformer V2.
• DrugReflector — 2026
  2026 AI drug-screening foundation model picked up in industry coverage — provenance still firming up.
• EvoLen
  EvoLen — name only; protein/evolution model, watch the literature.
• scpFormer
  scpFormer — name only; transformer for single-cell perturbation, no paper yet.
• scFOCAL
  scFOCAL — name only.

Datasets

• CELLxGENE Census
  Aggregated single-cell corpus most sc-FMs pretrain on.

Organisations

• CZI / Biohub / Virtual Biology Initiative — Priscilla Chan
  The single largest philanthropic backer of the virtual-cell vision.
• Arc Institute — Patrick Hsu & Hani Goodarzi
  Largest non-DeepMind compute (~2,048× H100); built STATE, Evo2, Tahoe-100M.
• CZ Biohub — Theofanis Karaletsos, Steve Quake, Angela Pisco
  Substrate + model shipper (TranscriptFormer, rBio); breaks substrate-only stereotype.
• Mahmood Lab (Harvard / BWH / MGB) — Faisal Mahmood
  Full pathology vertical stack: UNI → TITAN → CONCH → PathChat; PathChat-DX FDA Breakthrough.
• Theodoris Lab (Gladstone / UCSF) — Christina Theodoris
  Only sc-FM lab with full compute disclosure (NVIDIA BioNeMo).
• Bo Wang Lab (U. Toronto / Vector / UHN) — Bo Wang
  Defined the sc-FM category (scGPT); agentic-AI pivot 2026 (X-Cell).
• Leskovec + Quake (Stanford / SNAP) — Jure Leskovec & Steve Quake
  Cross-species sc-FM (UCE) — 8 species via ESM2 bridge.
• Google DeepMind
  TPU-scale; closed weights + hosted API; Isomorphic Labs is the commercial arm.
• EvolutionaryScale (ex-Meta FAIR) — Alex Rives, Tom Sercu, Thomas Hayes
  ESM-3 with the cleanest compute disclosure (10²⁴ FLOPs) in biology FMs.
• Paige.AI + MSK — Thomas Fuchs
  First FDA 510(k)-cleared general-purpose pathology AI (Jan 2025); Virchow line.
• Owkin (Paris / NYC)
  Only major industrial pathology player shipping open weights (Phikon, Apache-2.0).
• Theis Lab (Helmholtz Munich) — Fabian Theis
  Methodological reference class; author of 2026 compositional-FM perspective.
• BioMap Research + MBZUAI — Le Song & Eric Xing
  China/Middle-East virtual-cell competitor (VCHarness, xTrimo).
• Isomorphic Labs
  DeepMind's commercial drug-discovery arm; bought Ahlmann-Eltze.
• Recursion Pharmaceuticals
  AI-native biotech using FMs for drug discovery.
• Latent Labs — Simon Kohl + DeepMind alumni
  DeepMind-alumni virtual-cell startup.
• Modella AI
  Acquired by AstraZeneca for pathology FM stack (JPM 2026).

People

• Aviv Regev — Genentech
  Agenda-setter for causal foundation models in cell biology.
• Christina Theodoris — Gladstone Institutes
  Geneformer creator; one of the few sc-FM PIs with full compute disclosure.
• Bo Wang — U. Toronto / Vector
  scGPT PI; 2026 pivot to agentic AI (X-Cell).
• Faisal Mahmood — Harvard
  UNI, CONCH, PathChat, TITAN — most complete vertical pathology stack.
• Jure Leskovec — Stanford
  UCE co-PI; graph + sc-FM bridge.
• Theofanis Karaletsos — CZI
  CZI Head of AI for Science — runs the virtual-cell program.
• Fabian Theis — Helmholtz Munich
  Compositional-FM Perspective author; scvi-tools / Nicheformer.
• Le Song / Eric Xing — BioMap + MBZUAI
  VCHarness; autonomous virtual-cell direction from China.
• Marinka Zitnik — Harvard Med
  TxGNN, TDC-2 creator; KG-FM-therapeutics axis.
• Xi Peng — Xlearning Lab, Sichuan University
  Xlearning lab PI — multimodal-learning angle into virtual-cell territory.
• Jianhua Yao — Tencent AI Lab
  scTranslator PI, AIMBE fellow — Tencent's bio-FM lead.

Commentary

• Grow AI virtual cells — three data pillars and closed-loop learning — Cell Research (editorial), 2026
  Editorial framing closed-loop learning as the missing ingredient.
• Epoch AI — estimating training compute — Epoch AI blog
  FLOPs ≈ 6 × params × tokens heuristic the talk's cost estimates use.
• From data chaos to precision medicine — Schönhuth, Nature Machine Intelligence, 2025
  Short framing commentary on the data-to-clinic gap.
• AI builds models of biology — Marx, Nature Methods, 2024
  Accessible field overview.
• Owl Posting — Biology + ML essays — Abhishaike Mahajan
  Long-form bio+ML essays — one of the better field-walkthrough writers.
• Decoding Bio — weekly biotech + AI — Substack
  Weekly biotech + FM-paper digest.
• Asimov Press
  Long-form bio writing.
• Arc Institute News
  Updates on STATE, Evo2, virtual-cell atlas.
• M2D2 — Molecular Modelling & Drug Discovery — Valence Labs & Mila
  Weekly reading group + podcast on ML for discovery.
• The Gradient — ML-for-science long-form
  Curated ML + science writing.
• Latent Space podcast — Swyx & Alessio
  AI podcast that books biology-FM founders.
• Eric Topol's Substack — biomedical AI — Substack
  Clinical-AI perspective from a cardiologist.
• preLights — preprint commentary — Company of Biologists
  Community preprint commentary on bio papers.

Event

• Westlake Autumn Symposium — AIVC (Tiannan Guo)
  Westlake's annual AIVC symposium — gathers the China AIVC scene.
• CZ Biohub Virtual Cell Program launch — 2024
  Multi-billion-dollar non-dilutive commitment to virtual cells.

Move

• Merck + Mayo Clinic collaboration — 2026
  Feb 2026 collaboration — pharma + clinical data axis around AI models.
• Noetik + GSK collaboration — 2026
  Noetik (spatial-omics AIVC startup) lands a GSK deal.
• UCSD + DARPA virtual-cell programme — 2026
  DARPA money flowing to UCSD for virtual-cell work.
• JPM 2026 — Lilly + NVIDIA $1B AI Co-Innovation Lab — 2026
  Largest pharma–AI infrastructure commitment, JPM Day 1.
• JPM 2026 — AstraZeneca acquires Modella AI — 2026
  First big-pharma acquisition of an AI-FM firm.

---

Regenerated 2026-05-15 from a 277-item corpus by resourcelib-views.