Skip to content

Adjacent methods — the catalog

Resource section for the FM-to-Virtual-Cells talk. The talk is about foundation models, but the problems it cares about — perturbation prediction, generative cell modelling, evaluation, cheap model adaptation, the architecture debate — are also worked on by people not building FMs. This is the catalog of those papers: ~55 algorithm, benchmark, and methodology papers that sit next to the talk's argument without being FM papers, curated entirely from the talk's own supplementary bibliography (§E, §H.3, §H.10) plus the lane and track dossiers (§B, §C).

How to read this catalog

  • It is the index, not the deep-dive. Eight of these papers have full dossier pages — see the table below. The rest are one-liner entries here.
  • It deliberately excludes foundation models. Those live in the model glossary and the paper map. This catalog is the methods next door.
  • Some entries also appear in the paper map or the evaluation papers catalog — where they do, the entry says so and points there for depth.
  • Grounding: every entry is curated from the talk's existing bibliography. Where the source bibliography gives a direct link, it is reproduced; where it gives only a citation, the entry is text-only by design rather than inventing a URL.

The dossier pages

The eight papers worth a full page — what the method is, how it connects, where to read more:

Page Paper Theme
PerturbFate Xu et al., Nature 2026 Perturbation-prediction methods
Flow matching in computational biology Morehead et al., Nat Mach Intell 2026 Generative methods
Learning beyond fine-tuning Zheng et al., Nat Mach Intell 2025 Adaptation methodology
Genome language models Consens et al., Nat Mach Intell 2025 Architecture reviews
Protein-LM explainability Hunklinger & Ferruz, Nat Mach Intell 2026 Mechanistic interpretability
Transformers vs state-space models Tiezzi et al., Nat Mach Intell 2025 Architecture reviews

A. Perturbation-prediction & cell-modelling methods

Non-FM methods aimed directly at the talk's central task — predicting or mapping what a perturbation does to a cell.

  • CPA — Compositional Perturbation Autoencoder. Lotfollahi et al. 2023 Nat Cell Biol. The autoencoder baseline the reckoning repeatedly tests FMs against; one of the six methods in Ahlmann-Eltze & Huber.
  • GEARS. Roohani et al. 2024 Nat Biotechnol. Graph-based perturbation-effect prediction using a gene-gene knowledge graph — the pre-FM state of the art, and another standard reckoning baseline.
  • PerturbFate. Xu et al. 2026 Nature. Perturbation screening that maps convergent regulators of melanoma drug resistance — the assay-side counterpart to prediction. → dossier.
  • TxPert. Wenkel et al. 2026 Nat Biotech. Multiple-knowledge-graph perturbation prediction — "the reckoning answering itself." Also a paper-map node.
  • MAP — knowledge-driven framework. bioRxiv 2026.02. Zero-shot perturbation prediction for unprofiled drugs via biology-specific inductive bias (supplementary §C.4).
  • sc-FM Perturbation Adapter. Maleki et al., ICLR 2026 LMRL workshop. A drug-conditional adapter (<1% trainable params) on a frozen FM backbone — generalises to unseen drugs where CPA / GEARS / scGen cannot (supplementary §B.2).
  • CINEMA-OT. Optimal-transport perturbation-effect disentanglement, cited as a causal-objective exemplar in supplementary §C.2.
  • Lopez / Hsu 2025 Nat Methods — causal representation learning. A causal-representation approach cited as the exemplar for new pretraining objectives that target causality (supplementary §C.2).

B. Evaluation & critique — the reckoning corpus

The 2025–2026 evaluation papers are themselves non-FM method papers — benchmark designs and critiques. They have a dedicated catalog and most are paper-map nodes; listed compactly here for completeness.

→ Full depth: the evaluation papers catalog.

C. Benchmarks & datasets

The substrates and leaderboards the methods above are trained and judged on.

  • Replogle et al. 2022 Cell — whole-genome Perturb-seq; a standard perturbation test set.
  • Norman et al. 2019 Science — combinatorial Perturb-seq; the substrate for compositional-generalization benchmarks (supplementary §C.3).
  • Open Problems v2 — the community single-cell benchmark suite.
  • PerturBench — Altos Labs perturbation-prediction benchmark.
  • scPerturBench — BM2-Lab benchmark with adversarial perturbation/cell-type splits.
  • CZ Biohub Virtual Cell Challenge 2025 — the first community virtual-cell benchmark, $50k prize.
  • Roohani, Hsu et al. 2025 — Tahoe-100M — the first large-scale perturbation atlas framed as virtual-cell training data.
  • Tabula Sapiens Consortium 2022 Science — multi-organ human single-cell reference atlas.
  • CELLxGENE Census — the aggregated single-cell corpus most sc-FMs pretrain on.
  • HuBMAP 2023 Nature — the human tissue / spatial reference atlas.
  • HEST-1k — the histology-with-spatial-transcriptomics benchmark dataset.
  • CASP15 / CAPRI — the community structure- and complex-prediction assessments.
  • ProteinGym. Notin et al. 2023 NeurIPS — the protein variant-effect benchmark.
  • Hetzel et al. 2024 NeurIPS LMRL — compositional-generalization benchmark cited in supplementary §C.3.

D. Generative & structure-prediction methods

The generative-modelling lineage behind generative virtual cells and protein design.

  • Flow matching in computational biology. Morehead et al. 2026 Nat Mach Intell. The generative technique under generative virtual cells. → dossier.
  • AlphaFold 2. Jumper et al. 2021 Nature. The structure-prediction breakthrough the whole protein-modelling wave builds on.
  • RFdiffusion. Watson et al. 2023 Nature. Diffusion-based de novo protein design — the generative-design exemplar.
  • ProGen2. Madani et al. 2023 Nat Biotechnol. Generative protein-sequence modelling, pre-ESM-3.
  • scDiffusion. A diffusion model for single-cell expression generation, cited as a Lane 8 data-augmentation exemplar (supplementary §B.8).

E. Mechanistic interpretability methods

The toolkit for asking what a trained model actually encodes — the mechanism behind the reckoning.

F. Adaptation & training methodology

How a pretrained model is adapted, and what the training run costs — the methods backbone of the small-lab lanes.

G. Architecture reviews

The architecture debate behind Cause 2 of the reckoning — "the transformer was borrowed from NLP."

H. Uncertainty, causality & evaluation methodology

The methodological scaffolding behind the post-reckoning evaluation tracks (supplementary §C).

  • Virtual Cells Need Context, Not Just Scale 2026 — frames the generalization failure as a causal-transportability problem (Pearl). Also a paper-map node.
  • Beyond Alignment — Synergistic Information Score 2026 — Microsoft's metric for whether a multimodal FM extracts genuinely cross-modal information (supplementary §C.8).
  • Arjovsky et al. 2019 — Invariant Risk Minimization — the causal-objective method cited as the basis for counterfactual pretraining (supplementary §C.2).
  • van Amersfoort et al. 2024 — deep deterministic UQ — an uncertainty-quantification method for the clinical-grade-FM track (supplementary §C.5).
  • Angelopoulos & Bates 2023 — conformal prediction — the distribution-free uncertainty method paired with the UQ track (supplementary §C.5).

How this section relates to the rest of the talk prep

  • The model glossary and paper map cover the foundation models and the literature about them.
  • The evaluation papers catalog covers the reckoning corpus in depth (Theme B above is its index-level summary).
  • This catalog covers methods next door — algorithm, benchmark, and methodology papers that inform the same problems without being FM papers.
  • Anything here that turns out to be load-bearing for the argument graduates into the paper map; until then it lives here as a resource.

Last updated 2026-05-14. ~55 papers, curated from supplementary §E / §H.3 / §H.10 / §B / §C.