Skip to content

The paper map — how the literature connects, and how to read it

Explainer page for the FM-to-Virtual-Cells talk. The talk cites ~110 papers; this page is the map — an interactive network of the ~50 load-bearing ones, colour-coded by category, with edges that say how each paper relates to the others. Below the map is a systematic reading order: the sequence to actually work through them in.

How to read the map

Every node is a paper. Colour = what kind of paper it is. Edges = how papers relate, and there are only four edge types:

  • builds on / lineage (grey) — B is a descendant or successor of A.
  • evaluates / critiques (red dashed) — B tested A and reported how it did.
  • responds to (green) — B is a direct answer to A.
  • frames / explains (purple dotted) — B gives A a theory, a mechanism, or a position.

The layout is not force-directed — it is deliberately arranged. The reckoning sits at the centre because almost everything else points at it: the model papers it critiqued, the responses it provoked, the theory that explains it. The other seven categories ring it. Ahlmann-Eltze 2025 is the single hub node — the paper the whole corpus orbits.

Interactive — hover any node for a one-line "why read this"; drag to pan, scroll to zoom. The dense red-dashed bundle in the middle is the visual thesis of the whole talk: ~13 independent evaluation papers all converging on the same handful of single-cell FMs.

The systematic reading order

Read the corpus in seven passes. Each pass is one region of the map. You do not need all ~50 — the bold papers in each pass are the minimum spine; the rest are depth.

Pass 1 — Why the field exists (position / framing, purple)

Start with the ambition, so the reckoning lands as a disappointment rather than a curiosity.

  1. Bunne et al. 2024 Cell — "How to build the virtual cell with AI." The canonical thesis. Everything downstream is chasing this.
  2. Roohani et al. 2025 Cell — the Virtual Cell Challenge: what "success" is supposed to look like, operationalized as a Turing test.
  3. Rood, Hupalowska & Regev 2024 Cell — Regev's reference-mapping framing of the same goal.
  4. Rao et al. 2026 Nat Biotech — "Generalist biological AI." The position paper that unifies all five FM families — read it to see why the other-family FMs (right side of the map) belong in this story at all.
  5. Li et al. 2026 Nat Biotech — agentic AI + in silico team science; the framing behind Pass 6.
  6. Singh et al. 2025 Exp Mol Med — a clean orientation review if you want one survey before the primary literature.

Pass 2 — What got built (single-cell FM model papers, blue)

The models the reckoning will later test. Read these for architecture, not results.

  1. scGPT (Cui et al. 2024 Nat Methods) — genes + cells as tokens; defined the category. This is the node the reckoning converges on — read it first and most carefully.
  2. Geneformer (Theodoris et al. 2023 Nature) — rank-based tokenization; the other universal reckoning target.
  3. UCE (Rosen et al. 2024) — cross-species, bridged through ESM-2 (note the grey edge to the protein-FM cluster).
  4. scFoundation, CellPLM, scBERT — the rest of the lineage; skim for how each varies the tokenization.
  5. TranscriptFormer (Pearce et al. 2025 Science) — the generative cross-species model; note it has a green edge — it is also a response (Pass 5).
  6. STATE (Arc), Nicheformer (spatial) — read if perturbation/spatial is your angle.
  7. scPRINT (Kalfon et al. 2025 Nat Commun) and its successor scPRINT-2 (bioRxiv 2025.12) — a sc-FM line aimed at gene-network inference; scPRINT-2 ships its own benchmark suite, so it carries a green edge into the reckoning (Pass 3).

Pass 3 — The reckoning (critique / benchmark, red — the centre)

This is the heart of the talk. Read in roughly this order — it is a chronological argument.

  1. Ahlmann-Eltze & Huber 2025 Nat Methods — THE paper. A one-line linear baseline beats every published sc-FM on perturbation prediction. Start here; everything else extends it.
  2. Boiarsky 2023 + Csendes scPerturBench — the earlier warnings Ahlmann-Eltze built on (grey edges into the hub).
  3. Kedzierska et al. 2025 Genome Biology — extends the result to the zero-shot setting.
  4. Wenkel et al. 2025 Nat Methods — proposes the stronger latent-additive baseline. Remember this name — it reappears in Pass 5.
  5. Wu Nat Methods 2026, Wu Genome Biology 2025, Liu scEval 2026 — the corpus broadens past perturbation to most cell-level tasks.
  6. Parameter-free baseline, PertEval-scFM, CellBench-LS, Han et al., cellular-dynamics — the 2026 wave; each opens a new evaluation axis. Skim unless one matches your project.

→ The dedicated evaluation papers catalog is the deep version of this pass.

Pass 4 — The pushback (contrarian + theory, orange)

The reckoning is not a settled consensus. Read these to hold the honest May-2026 position.

  1. Foundation Models Improve Perturbation Response (2026) — the contrarian voice: with enough data, FMs do improve. Note its green edges into the reckoning hub.
  2. Virtual Cells Need Context, Not Just Scale (2026) — names the theory: a causal-transportability problem in Pearl's sense. The purple "frames" edge.
  3. Beyond Alignment / SIS (Microsoft 2026) — a multimodal-FM evaluation metric nobody has applied broadly yet (a project lead).

Pass 5 — The architectural response (green)

What the field is building because of the reckoning.

  1. xVERSE (2026) — transcriptomics-native architecture; first evidence the design choice is load-bearing.
  2. TxPert (Wenkel et al. 2026 Nat Biotech) — knowledge-graph-driven perturbation prediction. Follow its green edge to Wenkel 2025 in Pass 3: the reckoning is answering itself — the same author wrote the critique and the response.
  3. MAP (2026) — knowledge-driven zero-shot prediction for unprofiled drugs.

Pass 6 — Agents built on the models (brown)

The 2026 convergence: agentic AI and FMs are complementary, not competing.

  1. rBio — reasons over a virtual cell (trained on TranscriptFormer — follow the grey edge).
  2. VCHarness (2026) — builds virtual-cell models autonomously.
  3. CellVoyager (2026 Nat Methods) — analyzes single-cell data autonomously.

→ The agentic-meets-foundation explainer is the deep version of this pass.

Pass 7 — Why the models fail, mechanistically (interpretability, teal)

Optional, but it is the satisfying ending — the mechanism behind Pass 3.

  1. Adams et al. 2025 PNAS — the protein-FM SAE paper that started the interpretability wave.
  2. Simon & Zou 2026 — SAEs show sc-FMs encode cell-type and pathway features but minimal regulatory logic. Follow its purple edge to Ahlmann-Eltze: this is the mechanistic explanation of the reckoning.
  3. SAE-on-scGPT, SAE synthesis 2026 — independent confirmation and the cross-family synthesis.

Aside — the other four FM families (grey)

The talk is single-cell-centric, but it places sc-FMs against pathology, genomic, and protein FMs. Read one representative each if you want the full five-family picture: AlphaGenome (genomic), ESM-3 (protein), Virchow2 (pathology). They connect to the map through the generalist position paper (Rao 2026) and through UCE's ESM-2 bridge. The model glossary has one-liners for all of them.

The shortest honest path

If you only read six papers, read the spine: Bunne 2024 (the goal) → scGPT + Geneformer (what got built) → Ahlmann-Eltze 2025 (the reckoning) → FMs Improve Perturbation (the pushback) → xVERSE (the response). That is the talk's Act 1 in six citations.

Where to go next

Last updated 2026-05-14.