Theme — perturbation-prediction

Items in corpus: 39

Papers

• scPerturBench replication — Csendes et al., BM2 Lab preprint, 2024
  Showed the original scGPT split was leaky.
• Deep-learning predictions of gene expression don't generalize — Ahlmann-Eltze & Huber, Nature Methods, 2025
  THE canonical reckoning paper — start here.
• latent-additive is the new baseline floor — Wenkel et al., Nature Methods, 2025
  Proposed the stronger baseline current FMs still don't beat.
• PertEval-scFM — standardised evaluation framework — ICML, 2025
  Formal venue stamp on the perturbation critique.
• Foundation Models Improve Perturbation Response — bioRxiv, 2026
  Contrarian voice — FMs DO improve with enough data.
• 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.
• Causal representation learning (Lopez/Hsu) — Nature Methods, 2025
  Causal-rep exemplar for pretraining objectives that target causality.
• 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.
• CellFluxRL — Biologically-Constrained Virtual Cell Modeling via RL — Wu, Su, Zhang, Sui, Lundberg, Fox, Yeung-Levy, preprint, 2026
  RL-driven virtual-cell modelling with biological constraints — Stanford/CZI/Lundberg axis.
• SCALE — Scalable Conditional Atlas-Level Endpoint transport for virtual-cell perturbation — preprint, 2026
  Optimal-transport / endpoint framing of perturbation prediction at atlas scale.
• TxGNN — knowledge-graph FM for therapeutics — Zitnik et al., Nature Medicine, 2024
  Graph FM bridging gene-level and patient-level AI.
• Evaluating single-cell perturbation response models is far from straightforward — Heidari, Karimpour, Srivatsa & Montazeri, bioRxiv, 2026
  Second front in the perturbation-evaluation reckoning — Wasserstein/Energy distances mislead in high-dim gene space.
• VCWorld — biological world model for virtual cell simulation — Shuangjia Zheng lab et al., arXiv, 2025
  White-box LLM-driven simulator: stepwise mechanistic reasoning, SOTA on drug-perturbation benchmarks. Anti-black-box alternative.

Tools

• scGPT — Cui et al., Nature Methods, 2024
  Defined the category — genes + cells as tokens. The reckoning's main target.
• scFoundation — Hao et al., Nature Methods, 2024
  Read-depth-aware attention; one of the four reckoning regulars.
• STATE (Arc Institute) — bioRxiv, 2025
  First production virtual cell at Tahoe-100M scale.
• 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.
• MAP — knowledge-driven perturbation framework — bioRxiv, 2026
  Zero-shot perturbation prediction for unprofiled drugs.
• CPA — Compositional Perturbation Autoencoder — Lotfollahi et al., Nature Cell Biology, 2023
  Autoencoder baseline the reckoning repeatedly tests FMs against.
• GEARS — Roohani et al., Nature Biotechnology, 2024
  Graph-based perturbation-effect prediction with a gene-gene KG — the pre-FM SOTA.
• CINEMA-OT
  Optimal-transport perturbation-effect disentanglement.
• PerturbHD — Stanford
  Stanford perturbation method — only a name so far.
• CellFluxV2
  Successor to CellFluxRL — name only.
• CRADLE-VAE — compositional active learning — Bunne et al.
  Closed-loop virtual-cell active learning prototype.
• scPoli — single-cell policy learning — Lotfollahi et al.
  Policy-based perturbation prediction; the scArches lineage.
• Pertpy — end-to-end framework for perturbation analysis — Heumos, Ji, May, Theis et al., Nature Methods, 2025
  scverse-blessed perturbation toolbox — distances, perturbation spaces, harmonisation. The canonical analysis layer.
• scpFormer
  scpFormer — name only; transformer for single-cell perturbation, no paper yet.

Benchmarks

• PerturBench — Altos Labs
  Altos Labs perturbation-prediction benchmark.
• scPerturBench — BM2-Lab
  Benchmark with adversarial perturbation/cell-type splits.
• Compositional-generalisation benchmark (Hetzel et al.) — NeurIPS LMRL, 2024
  Compositional generalisation benchmark referenced in supplementary §C.3.

Datasets

• Replogle whole-genome Perturb-seq — Replogle et al., Cell, 2022
  Standard perturbation test set.
• Norman combinatorial Perturb-seq — Norman et al., Science, 2019
  Substrate for compositional-generalisation benchmarks.

Organisations

• Bo Wang Lab (U. Toronto / Vector / UHN) — Bo Wang
  Defined the sc-FM category (scGPT); agentic-AI pivot 2026 (X-Cell).

People

• Mohammad Lotfollahi — Wellcome Sanger
  scArches, scPoli, CPA, CRADLE-VAE — the perturbation-and-adaptation stack.
• Constantin Ahlmann-Eltze — Isomorphic Labs (ex-EMBL)
  Linear-baseline-2025 first author; recently moved to Isomorphic.

Competition

• CZ Biohub Virtual Cell Challenge 2025
  First community virtual-cell benchmark, $50k prize.
• Echoes of Silenced Genes — Myllia Biotechnology cell challenge — Myllia, 2026
  Biotech-run challenge focused on perturbation prediction.

Move

• Turbine raises $25M — 2026
  Validates commercial market for VC-style perturbation simulators.

---

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