Pith. sign in

REVIEW 3 major objections 8 minor 49 references

PerturbPFN: Probing the Limits of Synthetic Priors in Drug Perturbation Modelling

T0 review · 3 major / 8 minor · reviewed 2026-07-30 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A model trained only on synthetic biology episodes can predict real drug responses in one forward pass while naming the targets and regulatory paths it thinks are responsible.

desk verdict Solid workshop-grade PFN for structured drug-effect prediction: real speed/interpretability trade-off, but Sci-Plex wins sit close to a prior that already looks like the eval protocol. read the letter →

arxiv 2607.23447 v1 pith:OQXBFWNY submitted 2026-07-26 cs.LG

Yuche Gao , José Miguel Hernández-Lobato , Siyuan Guo This is my paper · ORCID
classification cs.LG
keywords perturbationpredictionprior-datafittednetworksstructuralcausalmodelssingle-celldrugresponsesyntheticpriorsin-contextlearninggeneregulatoryunknown-targetinterventions
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Predicting how cells react to new small-molecule drugs is hard because the molecular targets are often unknown, the expression response is high-dimensional, and labs can only test a tiny slice of chemical space. PerturbPFN attacks this by never training on real perturbation outcomes. Instead it is pretrained entirely on synthetic episodes drawn from a hierarchical structural causal model that samples gene networks, sparse intervention targets, dose-dependent strengths, and noisy expression. At test time the model reads a few context experiments, infers a latent system graph plus query-specific targets and strengths, and rolls those interventions through an SCM decoder to produce the predicted response—all in a single forward pass with no gradient updates. On real single-cell drug data it matches specialized baselines on effect prediction (best distributional match among compared methods) while remaining far cheaper at inference than iterative causal fits, and it recovers non-collapsed intermediate estimates of graphs, targets, and strengths on held-out synthetic tasks. A sympathetic reader cares because the work shows that structured synthetic priors can amortize mechanism-aware perturbation reasoning instead of only black-box regression or per-dataset fitting.

What carries the argument

The hierarchical synthetic structural prior plus the graph–target–strength bottleneck: the model does not regress expression directly; it predicts a DAG, sparse targets, and strengths, then propagates additive shift interventions through an SCM decoder whose synthetic labels supervise every latent stage.

What would settle it

On the same Sci-Plex highest-dose holdout protocols, replace the biologically motivated prior with a clearly misspecified one (for example random dense graphs and unstructured targets) and check whether both effect-prediction metrics and recovery of intermediate targets/graphs collapse relative to the reported checkpoint.

Watch

Extended reading notes

Core claim

A single checkpoint trained only on prior-predictive synthetic SCM episodes can, via one in-context forward pass, competitively predict held-out highest-dose cellular responses on real Sci-Plex drug protocols while exposing interpretable intermediate estimates of the latent regulatory graph, sparse atomic targets, and intervention strengths—without any test-time parameter updates or real perturbation labels during training.

Load-bearing premise

The synthetic generators for gene networks, sparse drug targets, dose–response strengths, and expression noise must be close enough to real chemical perturbation biology that what the model learns transfers.

Editorial extensions

If this is right

  • Competitive unknown-target dose extrapolation is possible without real training labels or per-protocol optimization once a suitable synthetic prior exists.
  • Intermediate graph, target, and strength estimates remain usable for inspection and in-silico simulation (dose curves, combinations) rather than only end-to-end prediction.
  • Amortized structure inference from simulator-defined biological topology can transfer some regulatory signal to external GRN benchmarks even when the model is not trained to match curated annotations.
  • Inference cost after pretraining can drop to roughly one second per protocol while still beating slower instance-specific causal baselines on distributional accuracy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If richer empirically calibrated simulators close the prior gap, the same amortization recipe could extend from small marker-gene panels to larger transcriptomic subsystems composed locally.
  • Point-estimate bottlenecks leave open a natural next test: whether posterior samples over graphs and targets would improve both calibration of predictions and biological uniqueness of recovered mechanisms.
  • The competitive W2 with weaker mean-distance ranking suggests the decoder may capture distribution shape better than point means—worth checking whether that pattern holds on other chemical screens.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 8 minor

Summary. The manuscript introduces PerturbPFN, a prior-data fitted network for chemical perturbation response prediction. A transformer is trained entirely on synthetic episodes drawn from a hierarchical SCM prior — modular scale-free DAGs, truncated-geometric sparse targets, Hill-type signed shift interventions, and a zero-inflated log-normal observation model — to amortize inference of a latent graph, intervention targets, and strengths, which are then propagated through an SCM decoder. Evaluation has three parts: (i) in-prior latent recovery on fresh synthetic episodes (Table 1: target AUROC 0.93, graph AUROC 0.81, projected graph F1 0.33); (ii) the Sci-Plex highest-dose holdout benchmark of Schneider et al. (Table 2: best normalized W2, second overall behind RF X-Learner, ~1.1 s inference per protocol); and (iii) transfer of the graph head to the GeneRNIB 300BCG GRN benchmark (Table 3: best GS-F1, weak on regression-based metrics). The authors frame the contribution as a complementary trade-off — competitive prediction, low test-time cost, interpretable intermediates — rather than dominance, and the claims are mostly hedged accordingly.

Significance. If the results hold, the paper demonstrates that a 15M-parameter transformer trained for ~1.4 GPU-hours purely on prior-predictive synthetic SCM episodes can amortize structured perturbation inference — graph, sparse targets, strengths, and an SCM-decoded response — in a single forward pass at ~1 s per protocol, orders of magnitude cheaper than GIM/CondOT at comparable or better accuracy. The work is unusually transparent and falsifiable: the full prior hyperparameters (Tables 4-7), checkpoint-selection procedure, raw (non-normalized) protocol-level metrics (Fig. 2), and a clear loss to a random-forest baseline are all reported, and the Discussion candidly states that latent meaning is prior-determined. The GeneRNIB GS-F1 result hints that simulator-trained structure heads may carry transferable regulatory signal. The significance is bounded by the evaluation breadth: one real dataset, four same-class drugs, and a single low-corruption training configuration, so the paper currently establishes a proof of concept of prior-matched transfer rather than robust synthetic-to-real transfer.

major comments (3)
  1. [App. A, Highest-dose episodic protocol; App. D.2] App. A ('Highest-dose episodic protocol') vs. App. D.2: the training episodes replicate the evaluation protocol almost exactly — one held-out drug, a four-level geometric dose ladder (L=4 in both Table 5 and the benchmark), lower doses of the held-out drug in context, and Hill-type dose-response strengths baked into the prior (psi_{d,a,i} equation, App. A). Moreover, all four Sci-Plex drugs are same-class HDAC inhibitors (App. D.1), so the query is dose extrapolation along a monotone functional form the model was explicitly trained on, in a regime where the Naive nearest-dose baseline already scores 0.60 (Table 2). As constructed, the experiment does not distinguish 'the amortized inference transfers to real biology' from 'the prior was engineered to match this benchmark.' This is load-bearing for the contribution claim of competitive real-data prediction from a synthetic-only prior. A c
  2. [Table 5, App. A; §2.3] Table 5 (App. A): the final checkpoint uses sigma_obs = 0.001 and p0,max = 0.001, i.e., the zero-inflated log-normal observation model described in §2.3 as 'SERGIO noise injection' is parameterized but effectively switched off. Real Sci-Plex count data is heavily zero-inflated and overdispersed even after median-count normalization and log transform. Consequently the amortized inference procedure — including the context intervention memory (Eq. 7) and structure-evidence features (Eqs. 8-9) — has only ever been trained on essentially noiseless condition means, and both the in-prior recovery numbers (Table 1) and the W2 win (Table 2) may be conditional on this regime. The authors themselves flag prior misspecification as the key risk in the Discussion; this is its most concrete and testable instance. Please report a corruption-robustness sweep: train (or at minimum evaluate) checkpoints at
  3. [§4.1, Table 1; Abstract] The abstract states the evaluation covers 'effect prediction, target identification, and regulatory structure discovery,' but target identification is only evaluated in-prior (Table 1), where ground truth is the simulator's own latent — a result that establishes the bottleneck can represent its own prior, not that inferred targets are biologically meaningful. The Sci-Plex benchmark offers an unusually cheap real-data check: all four held-out drugs are HDAC inhibitors with known targets, and the model exposes per-query target gates I_j. Reporting whether inferred high-probability targets on the 12 protocols are enriched for HDAC-family genes (or known pathway annotations) would either substantiate or appropriately bound the 'interpretable intermediate estimates' claim. Without any such check, the interpretability contribution rests entirely on synthetic self-consistency.
minor comments (8)
  1. [Table 2; App. B] Table 2: the 'amortized' runtime (432.8 s/protocol) allocates the one-time 1.44 GPU-hour pretraining cost across exactly 12 protocols, so the figure is arbitrary in the number of protocols; with 100 protocols it would undercut the ICL column's own baselines. Report pretraining cost and ICL cost separately rather than as a per-protocol allocation.
  2. [§4.3; App. E] §4.3 and App. E: the auxiliary edge AUROC 0.9168 / AP 0.7570 are quoted without stating the reference network, link budget, or scorer; given that the official metrics (Table 3) show PerturbPFN well below Pearson correlation on R-Precision (0.225 vs 0.615), the provenance of these much stronger numbers needs full specification or the numbers should be removed.
  3. [§4.1, Table 1] Table 1 reports graph Proj. F1 = 0.326 and AUPRC = 0.298; these are above edge prevalence (0.0363) but modest in absolute terms, and the straight-through top-k gate F1 (0.6287) is cited in text but absent from the table. Please include the gate F1 in Table 1 and temper 'even stronger' / 'substantially above' phrasing accordingly.
  4. [§4.2] §4.2 / App. D.1: it should be stated in the main text, not only the appendix, that the benchmark's four drugs are all HDAC inhibitors; this materially conditions the generality of the real-data claims.
  5. [Figure 2] Figure 2: axis titles render as empty parentheses ('Mean Distance ( )', etc.) — units/labels appear to have been dropped; the x tick labels overlap the legend text. Please fix before publication.
  6. [§2.4, Eq. (6)] Eq. (6) and §2.4: the 'categorical bar distribution' for strengths is TabPFN-specific jargon used without citation or definition; define support, binning, and expectation extraction (Table 6 gives bins=81, support [-2,2], but the main text should cross-reference this).
  7. [Eq. (5); §2.2] The point-estimate approximation to the posterior predictive (Eq. 5) is acknowledged in the Discussion, but a brief note on why the ZILN decoder variance does not already capture latent uncertainty (it captures observation noise only) would help readers interpret the W2 results.
  8. [References; App. E.3] Ref. [21] (GeneRNIB) is a bioRxiv preprint; since the GRN transfer claims rest on this benchmark, note its version and maturity, and state whether the 300BCG score archive used for baselines is frozen or a living benchmark whose numbers may shift.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circular derivation: real Sci-Plex/GeneRNIB results are external; in-prior recovery is a standard same-simulator sanity check, not a forced prediction.

  1. self definitional [§4.1 In-Prior Latent Recovery; Table 1; cf. §2.5 training objective]
    "We first evaluate whether PerturbPFN recovers the latent variables represented by its structured bottleneck on 1,024 fresh episodes drawn from the synthetic prior. Because the latent graph, intervention targets, strengths, and response distributions are known by construction, this evaluation isolates in-prior amortized inference... Table 1 shows that all components of the latent bottleneck remain non-collapsed."

    Mild only: multi-task losses directly supervise Ĝ, Î, ψ̂ and outcomes on synthetic labels from the same hierarchical prior later used for ‘recovery’ metrics. Non-collapse partly restates that supervised heads still work on held-out draws of that prior. The paper frames this as a pre-transfer sanity check, not as an external biological prediction, so it does not force the Sci-Plex or GeneRNIB claims.

full rationale

PerturbPFN’s central empirical claims do not reduce to fitted real-data inputs or to self-definitional equations. The checkpoint is trained only on hierarchical synthetic SCM episodes (graphs, sparse targets, Hill-type strengths, ZILN observations) and is applied to Sci-Plex highest-dose holdouts and GeneRNIB 300BCG without test-time gradients or real-response supervision (§2.3–2.5, §4.2–4.3). Effect metrics (MD, r, W2) and GRN metrics are computed against held-out biological data the model never optimized on, so they are not ‘predictions’ of quantities used to set parameters. In-prior latent recovery (§4.1, Table 1) evaluates the same simulator family used for training; that is methodologically self-referential as a sanity check that the graph–target–strength bottleneck did not collapse under its own prior, but the paper labels it as such and does not treat those numbers as external biological discovery. Protocol alignment (synthetic highest-dose episodes mirroring Sci-Plex holdout structure; near-noiseless observation settings in Table 5) is a prior-design / transfer-validity concern, not circularity by construction: no real-data coefficient is fitted and then re-reported as a prediction. Self-citations (e.g. related PFN/causal work, modular GRN generators) appear as background, not as uniqueness theorems that force the Sci-Plex ranking. Overall circularity is negligible.

Assumptions & free parameters 6 free parameters · 6 assumptions · 2 invented entities

The central transfer claim rests on a large synthetic generative story and many hand-chosen prior/architecture knobs, not on a short list of standard theorems. Real-data success is an empirical bet that this prior family is close enough. Invented machinery is the hierarchical prior plus the amortized bottleneck; free parameters are numerous simulator and training choices rather than a few fitted physical constants.

free parameters (6)
  • Synthetic graph prior knobs (α0, σs, αt, βwithin, λdist, module counts, edge-weight params) = e.g. α0=-2.5, βwithin=1.35, λdist=1.5, p in [40,60]
    Table 4 lists multiple hand-set generators for modular DAG topology; these define what 'structure' the model learns to recover and are not fit to real GRNs during training.
  • Perturbation prior knobs (pstop, E[|I|], Hill n range, Emax lognormal, dose ladder) = pstop=1/3, E[|I|]≈3, n in [2,5], L=4 doses
    Table 5 sets sparse target-count, dose-response, and episode protocol; they shape the amortized target/strength heads.
  • Observation corruption (σobs, p0,max) and cells per condition = σobs=0.001, p0,max=0.001, C=128
    Final checkpoint uses near-noiseless ZILN (σobs=0.001, p0,max=0.001, C=128), a deliberate low-corruption choice that eases latent recovery versus realistic scRNA-seq noise.
  • Loss weights and curriculum (λG, λI, λψ, focal/pos weights, structure pretrain 2500 + joint steps) = λG/I/ψ=1/2/2; selected step 3100
    Table 7 multi-task weights and two-stage curriculum are hand-tuned; checkpoint selected at step 3100 on a development set.
  • Inference graph density ρG and top-k target gate k=3 = k=3; ρG from synthetic prior
    Hard DAG extraction calibrates to prior-derived density; decoder uses straight-through top-k=3 targets—both fix sparsity independently of real data.
  • Architecture width/depth (14.98M params, dims in Table 6) = hidden 192/256, 81 strength bins on [-2,2]
    Encoder/decoder sizes and strength bar bins are design choices that affect capacity and reported performance.
assumptions (6)
  • domain assumption Cellular responses factor through a latent DAG SCM with local mechanisms and additive shift interventions on sparse targets (Eqs. 1–4, §2.1–2.2).
    Standard causal-ML modeling choice for perturbations; real regulatory systems may be cyclic, context-dependent, or non-additive.
  • domain assumption A regime descriptor γ induces interventions only via amortized maps to sparse I and strengths ψ, not arbitrary high-dimensional mechanisms (Eq. 6).
    Structural inductive bias of the bottleneck; misspecification if drugs act through diffuse or unmeasured mechanisms.
  • ad hoc to paper Biologically motivated synthetic generators (modular GRNs, Hill dose-response, SERGIO-like ZILN) are adequate prior-predictive training fuel for real transfer (§2.3, App. A).
    Core PFN bet of this paper; not independently validated as the correct real-data generative model.
  • domain assumption NOTEARS-style continuous acyclicity plus greedy density-calibrated projection yields a usable episode graph (Eqs. 10–11).
    Common structure-learning device; enforces DAG-ness that real GRNs may violate.
  • ad hoc to paper Point estimates of G, I, ψ suffice to approximate the posterior predictive (Eq. 5) without representing posterior uncertainty.
    Stated approximation in §2.2; Discussion flags missing uncertainty quantification.
  • domain assumption Transformer in-context conditioning on synthetic episodes amortizes Bayesian-like inference for new real episodes without gradients (PFN paradigm).
    Inherited from PFN literature cited in §3; empirical, not proved for this prior.
invented entities (2)
  • PerturbPFN hierarchical synthetic structural prior (episode GRN + drug-family target propensities + Hill strengths + paired-noise rollout)
    purpose: Supply supervised labels for graphs, targets, strengths, and outcomes that real data lack, enabling multi-task PFN pretraining.
    Custom simulator stack assembled from cited GRN/expression ideas but configured specifically for this model; transfer success is the only external check.
  • Graph–target–strength bottleneck with intervention memory and SCM decoder rollout
    purpose: Replace direct high-dimensional response regression with interpretable latent interventions propagated on an inferred DAG.
    Architectural invention of the paper; intermediates are validated in-prior and only indirectly via predictive metrics on real data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PerturbPFN: Probing the Limits of Synthetic Priors in Drug Perturbation Modelling." pith.science (2026). https://pith.science/paper/OQXBFWNY

@misc{pith2026260723447,
  author       = {Pith},
  title        = {Pith review of: PerturbPFN: Probing the Limits of Synthetic Priors in Drug Perturbation Modelling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OQXBFWNY}},
  note         = {Machine review of arXiv:2607.23447}
}
read the original abstract

Predicting cellular responses to unseen chemical perturbations is challenging due to unknown targets and mechanisms, high-dimensional expression responses, and limited experimental coverage of the large small-molecule design space. We propose PerturbPFN, a PFN-style amortized model for unknown-target perturbation prediction under a hierarchical synthetic structural prior. Instead of directly regressing high-dimensional expression responses, PerturbPFN infers a latent system graph, sparse atomic intervention targets, and intervention strengths, then propagates their effects through an SCM decoder. The model is trained entirely on prior-predictive synthetic episodes generated from biologically motivated graph and expression simulators, enabling structured in-context learning without test-time gradient updates. We evaluate PerturbPFN on both real single-cell perturbation data and synthetic benchmarks, covering effect prediction, target identification, and regulatory structure discovery. Our results show that PerturbPFN offers a complementary trade-off to specialized baselines, achieving competitive perturbation prediction with low inference cost while exposing interpretable intermediate estimates of targets, strengths, and system structure.

Figures

Figures reproduced from arXiv: 2607.23447 by the authors.

Figure 1
Figure 1. PerturbPFN pretraining pipeline. Synthetic episodes provide latent references for the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Raw effect-prediction performance on the 12-protocol highest-dosage holdout benchmark [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 3 linked inside Pith

  1. [1]

    Computational approaches streamlining drug discovery.Nature, 616(7958):673–685, 2023

    Anastasiia V Sadybekov and Vsevolod Katritch. Computational approaches streamlining drug discovery.Nature, 616(7958):673–685, 2023

    2023

  2. [2]

    Predicting causal effects in large-scale systems from observational data.Nature methods, 7(4):247–248, 2010

    Marloes H Maathuis, Diego Colombo, Markus Kalisch, and Peter Bühlmann. Predicting causal effects in large-scale systems from observational data.Nature methods, 7(4):247–248, 2010

    2010

  3. [3]

    Joint causal inference from multiple contexts.Journal of machine learning research, 21(99):1–108, 2020

    Joris M Mooij, Sara Magliacane, and Tom Claassen. Joint causal inference from multiple contexts.Journal of machine learning research, 21(99):1–108, 2020

    2020

  4. [4]

    Gen- erative intervention models for causal perturbation modeling

    Nora Schneider, Lars Lorch, Niki Kilbertus, Bernhard Schölkopf, and Andreas Krause. Gen- erative intervention models for causal perturbation modeling. InForty-second International Conference on Machine Learning, 2025

    2025

  5. [5]

    Causal machine learning for single-cell genomics.Nature Genetics, 57(4): 797–808, 2025

    Alejandro Tejada-Lapuerta, Paul Bertin, Stefan Bauer, Hananeh Aliee, Yoshua Bengio, and Fabian J Theis. Causal machine learning for single-cell genomics.Nature Genetics, 57(4): 797–808, 2025

    2025

  6. [6]

    Predicting cellular responses to novel drug perturbations at a single-cell resolution.Advances in Neural Information Processing Systems, 35:26711–26722, 2022

    Leon Hetzel, Simon Boehm, Niki Kilbertus, Stephan Günnemann, Fabian Theis, et al. Predicting cellular responses to novel drug perturbations at a single-cell resolution.Advances in Neural Information Processing Systems, 35:26711–26722, 2022

    2022

  7. [7]

    Learning single- cell perturbation responses using neural optimal transport.Nature methods, 20(11):1759–1768, 2023

    Charlotte Bunne, Stefan G Stark, Gabriele Gut, Jacobo Sarabia Del Castillo, Mitch Levesque, Kjong-Van Lehmann, Lucas Pelkmans, Andreas Krause, and Gunnar Rätsch. Learning single- cell perturbation responses using neural optimal transport.Nature methods, 20(11):1759–1768, 2023

    2023

  8. [8]

    Predicting cellular responses to complex perturbations in high-throughput screens.Molecular systems biology, 19(6):MSB202211517, 2023

    Mohammad Lotfollahi, Anna Klimovskaia Susmelj, Carlo De Donno, Leon Hetzel, Yuge Ji, Ignacio L Ibarra, Sanjay R Srivatsan, Mohsen Naghipourfar, Riza M Daza, Beth Martin, et al. Predicting cellular responses to complex perturbations in high-throughput screens.Molecular systems biology, 19(6):MSB202211517, 2023

    2023

Show all 49 references
  1. [9]

    Learning independent causal mechanisms

    Giambattista Parascandolo, Niki Kilbertus, Mateo Rojas-Carulla, and Bernhard Schölkopf. Learning independent causal mechanisms. InInternational Conference on Machine Learning, pages 4036–4044. PMLR, 2018

    2018

  2. [10]

    Combinatorial prediction of therapeutic perturbations using causally inspired neural networks.Nature Biomedical Engineering, pages 1–18, 2025

    Guadalupe Gonzalez, Xiang Lin, Isuru Herath, Kirill Veselkov, Michael Bronstein, and Marinka Zitnik. Combinatorial prediction of therapeutic perturbations using causally inspired neural networks.Nature Biomedical Engineering, pages 1–18, 2025

    2025

  3. [11]

    Predicting transcriptional outcomes of novel multigene perturbations with gears.Nature Biotechnology, 42(6):927–935, 2024

    Yusuf Roohani, Kexin Huang, and Jure Leskovec. Predicting transcriptional outcomes of novel multigene perturbations with gears.Nature Biotechnology, 42(6):927–935, 2024

    2024

  4. [12]

    scgpt: Towards building a foundation model for single-cell multi-omics using generative ai.bioRxiv, 2023

    Haotian Cui, Chloe Wang, Hassaan Maan, Kuan Pang, Fengning Luo, and Bo Wang. scgpt: Towards building a foundation model for single-cell multi-omics using generative ai.bioRxiv, 2023

    2023

  5. [13]

    Chen and James Zou

    Yiqun T. Chen and James Zou. Genept: A simple but hard-to-beat foundation model for genes and cells built from chatgpt.bioRxiv, 2023

    2023

  6. [14]

    In-silico biologi- cal discovery with large perturbation models.arXiv preprint arXiv:2503.23535, 2025

    Djordje Miladinovic, Tobias Höppe, Mathieu Chevalley, Andreas Georgiou, Lachlan Stuart, Arash Mehrjou, Marcus Bantscheff, Bernhard Schölkopf, and Patrick Schwab. In-silico biologi- cal discovery with large perturbation models.arXiv preprint arXiv:2503.23535, 2025

    2025 arXiv

  7. [15]

    Transformers can do bayesian inference

    Samuel Müller, Noah Hollmann, Sebastian Pineda Arango, Josif Grabocka, and Frank Hutter. Transformers can do bayesian inference. InInternational Conference on Learning Representa- tions, 2022

    2022

  8. [16]

    Accurate predictions on small data with a tabular foundation model.Nature, 637(8045):319–326, 2025

    Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tabular foundation model.Nature, 637(8045):319–326, 2025. 10

    2025

  9. [17]

    Do-pfn: In-context learning for causal effect estimation

    Jake Robertson, Arik Reuter, Siyuan Guo, Noah Hollmann, Frank Hutter, and Bernhard Schölkopf. Do-pfn: In-context learning for causal effect estimation. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  10. [18]

    Tabicl: A tabular foundation model for in-context learning on large data

    Jingang Qu, David Holzmüller, Gaël Varoquaux, and Marine Le Morvan. Tabicl: A tabular foundation model for in-context learning on large data. InInternational Conference on Machine Learning, pages 50817–50847. PMLR, 2025

    2025

  11. [19]

    Tabpfn-2.5: Advancing the state of the art in tabular foundation models.arXiv preprint arXiv:2511.08667, 2025

    Léo Grinsztajn, Klemens Flöge, Oscar Key, Felix Birkel, Philipp Jund, Brendan Roof, Benjamin Jäger, Dominik Safaric, Simone Alessi, Adrian Hayler, et al. Tabpfn-2.5: Advancing the state of the art in tabular foundation models.arXiv preprint arXiv:2511.08667, 2025

    2025 arXiv

  12. [20]

    Massively multiplex chemical transcriptomics at single-cell resolution.Science, 367(6473): 45–51, 2020

    Sanjay R Srivatsan, José L McFaline-Figueroa, Vijay Ramani, Lauren Saunders, Junyue Cao, Jonathan Packer, Hannah A Pliner, Dana L Jackson, Riza M Daza, Lena Christiansen, et al. Massively multiplex chemical transcriptomics at single-cell resolution.Science, 367(6473): 45–51, 2020

    2020

  13. [21]

    genernib: a living benchmark for gene regulatory network inference.bioRxiv, pages 2025–02, 2025

    Jalil Nourisa, Antoine Passemiers, Jeremie Kalfon, Marco Stock, Berit Zeller-Plumhoff, Ro- brecht Cannoodt, Christian Arnold, Alexander Tong, Jason Hartford, Mihai G Netea, et al. genernib: a living benchmark for gene regulatory network inference.bioRxiv, pages 2025–02, 2025

    2025

  14. [22]

    Gene regulatory network structure informs the distribution of perturbation effects.PLOS Computational Biology, 21(9):e1013387, 2025

    Matthew Aguirre, Jeffrey P Spence, Guy Sella, and Jonathan K Pritchard. Gene regulatory network structure informs the distribution of perturbation effects.PLOS Computational Biology, 21(9):e1013387, 2025

    2025

  15. [23]

    Bartlett, Melodie Li, Chenyu Song, Yuche Gao, and Qiulin Huang

    Thomas E. Bartlett, Melodie Li, Chenyu Song, Yuche Gao, and Qiulin Huang. Inferring gene- regulatory networks using epigenomic priors.iScience, 29:115165, 2026. doi: 10.1016/j.isci. 2026.115165

    2026

  16. [24]

    Sergio: a single-cell expression simulator guided by gene regulatory networks.Cell systems, 11(3):252–271, 2020

    Payam Dibaeinia and Saurabh Sinha. Sergio: a single-cell expression simulator guided by gene regulatory networks.Cell systems, 11(3):252–271, 2020

    2020

  17. [25]

    Mappfn: Learning causal perturbation maps in context.arXiv preprint arXiv:2601.21092, 2026

    Marvin Sextro, Weronika Kłos, and Gabriel Dernbach. Mappfn: Learning causal perturbation maps in context.arXiv preprint arXiv:2601.21092, 2026

    2026 arXiv

  18. [26]

    A meta-learning approach to bayesian causal discovery

    Anish Dhir, Matthew Ashman, James Requeima, and Mark van der Wilk. A meta-learning approach to bayesian causal discovery. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  19. [27]

    Dags with no tears: Continuous optimization for structure learning.Advances in neural information processing systems, 31, 2018

    Xun Zheng, Bryon Aragam, Pradeep K Ravikumar, and Eric P Xing. Dags with no tears: Continuous optimization for structure learning.Advances in neural information processing systems, 31, 2018

    2018

  20. [28]

    scgen predicts single-cell perturbation responses.Nature methods, 16(8):715–721, 2019

    Mohammad Lotfollahi, F Alexander Wolf, and Fabian J Theis. scgen predicts single-cell perturbation responses.Nature methods, 16(8):715–721, 2019

    2019

  21. [29]

    Disentanglement of single-cell data with biolord.Nature Biotechnology, 42(11):1678–1683, 2024

    Zoe Piran, Niv Cohen, Yedid Hoshen, and Mor Nitzan. Disentanglement of single-cell data with biolord.Nature Biotechnology, 42(11):1678–1683, 2024

    2024

  22. [30]

    Modelling cellular perturbations with the sparse additive mechanism shift variational autoencoder.Advances in Neural Information Processing Systems, 36:1–12, 2023

    Michael Bereket and Theofanis Karaletsos. Modelling cellular perturbations with the sparse additive mechanism shift variational autoencoder.Advances in Neural Information Processing Systems, 36:1–12, 2023

    2023

  23. [31]

    Supervised training of conditional monge maps.Advances in Neural Information Processing Systems, 35:6859–6872, 2022

    Charlotte Bunne, Andreas Krause, and Marco Cuturi. Supervised training of conditional monge maps.Advances in Neural Information Processing Systems, 35:6859–6872, 2022

    2022

  24. [32]

    Permutation-based causal structure learning with unknown intervention targets

    Chandler Squires, Yuhao Wang, and Caroline Uhler. Permutation-based causal structure learning with unknown intervention targets. InConference on Uncertainty in Artificial Intelligence, pages 1039–1048. PMLR, 2020

    2020

  25. [33]

    Bacadi: Bayesian causal discovery with unknown interventions

    Alexander Hägele, Jonas Rothfuss, Lars Lorch, Vignesh Ram Somnath, Bernhard Schölkopf, and Andreas Krause. Bacadi: Bayesian causal discovery with unknown interventions. In International Conference on Artificial Intelligence and Statistics, pages 1411–1436. PMLR, 2023. 11

    2023

  26. [34]

    Learning genetic perturbation effects with varia- tional causal inference.PLOS Computational Biology, 22(2):e1013194, 2026

    Emily Liu, Jiaqi Zhang, and Caroline Uhler. Learning genetic perturbation effects with varia- tional causal inference.PLOS Computational Biology, 22(2):e1013194, 2026

    2026

  27. [35]

    Tabiclv2: A better, faster, scalable, and open tabular foundation model.arXiv preprint arXiv:2602.11139, 2026

    Jingang Qu, David Holzmüller, Gaël Varoquaux, and Marine Le Morvan. Tabiclv2: A better, faster, scalable, and open tabular foundation model.arXiv preprint arXiv:2602.11139, 2026

    2026

  28. [36]

    Causalpfn: Amortized causal effect estimation via in-context learning.arXiv preprint arXiv:2506.07918, 2025

    Vahid Balazadeh, Hamidreza Kamkari, Valentin Thomas, Benson Li, Junwei Ma, Jesse C Cresswell, and Rahul G Krishnan. Causalpfn: Amortized causal effect estimation via in-context learning.arXiv preprint arXiv:2506.07918, 2025

    2025

  29. [37]

    Turner, and Mark van der Wilk

    Anish Dhir, Cristiana Diaconu, Valentinian Mihai Lungu, James Requeima, Richard E. Turner, and Mark van der Wilk. Estimating interventional distributions with uncertain causal graphs through meta-learning. InThe Thirty-ninth Annual Conference on Neural Information Process- ing...

    2026

  30. [38]

    Foundation models for causal inference via prior-data fitted networks.arXiv preprint arXiv:2506.10914, 2025

    Yuchen Ma, Dennis Frauen, Emil Javurek, and Stefan Feuerriegel. Foundation models for causal inference via prior-data fitted networks.arXiv preprint arXiv:2506.10914, 2025

    2025

  31. [39]

    Transfer learning enables predictions in network biology.Nature, 618(7965):616–624, 2023

    Christina V Theodoris, Ling Xiao, Anant Chopra, Mark D Chaffin, Zeina R Al Sayed, Matthew C Hill, Helene Mantineo, Elizabeth M Brydon, Zexian Zeng, X Shirley Liu, et al. Transfer learning enables predictions in network biology.Nature, 618(7965):616–624, 2023

    2023

  32. [40]

    Stack: In-context learning of single-cell biology.bioRxiv, 2026

    Mingze Dong, Abhinav Adduri, Dhruv Gautam, Christopher Carpenter, Rohan Shah, Chiara Ricci-Tam, Yuval Kluger, Dave P Burke, and Yusuf H Roohani. Stack: In-context learning of single-cell biology.bioRxiv, 2026

    2026

  33. [41]

    Metalearners for estimating heterogeneous treatment effects using machine learning.Proceedings of the national academy of sciences, 116(10):4156–4165, 2019

    Sören R Künzel, Jasjeet S Sekhon, Peter J Bickel, and Bin Yu. Metalearners for estimating heterogeneous treatment effects using machine learning.Proceedings of the national academy of sciences, 116(10):4156–4165, 2019

    2019

  34. [42]

    ppcor: an r package for a fast calculation to semi-partial correlation coefficients

    Seongho Kim. ppcor: an r package for a fast calculation to semi-partial correlation coefficients. Communications for statistical applications and methods, 22(6):665, 2015

    2015

  35. [43]

    Grnboost2 and arboreto: efficient and scalable inference of gene regulatory networks.Bioinformatics, 35(12):2159–2161, 2019

    Thomas Moerman, Sara Aibar Santos, Carmen Bravo González-Blas, Jaak Simm, Yves Moreau, Jan Aerts, and Stein Aerts. Grnboost2 and arboreto: efficient and scalable inference of gene regulatory networks.Bioinformatics, 35(12):2159–2161, 2019

    2019

  36. [44]

    Fast and accurate inference of gene regulatory networks through robust precision matrix estimation.Bioinformatics, 38(10): 2802–2809, 2022

    Antoine Passemiers, Yves Moreau, and Daniele Raimondi. Fast and accurate inference of gene regulatory networks through robust precision matrix estimation.Bioinformatics, 38(10): 2802–2809, 2022

    2022

  37. [45]

    Scenic: single-cell regulatory network inference and clustering.Nature methods, 14(11):1083–1086, 2017

    Sara Aibar, Carmen Bravo González-Blas, Thomas Moerman, Vân Anh Huynh-Thu, Hana Imrichova, Gert Hulselmans, Florian Rambow, Jean-Christophe Marine, Pierre Geurts, Jan Aerts, et al. Scenic: single-cell regulatory network inference and clustering.Nature methods, 14(11):1083–1086, 2017

    2017

  38. [46]

    Causal de finetti: On the identification of invariant causal structure in exchangeable data.Advances in Neural Information Processing Systems, 36:36463–36475, 2023

    Siyuan Guo, Viktor Tóth, Bernhard Schölkopf, and Ferenc Huszár. Causal de finetti: On the identification of invariant causal structure in exchangeable data.Advances in Neural Information Processing Systems, 36:36463–36475, 2023

    2023

  39. [47]

    MIT press, 2000

    Peter Spirtes, Clark N Glymour, and Richard Scheines.Causation, prediction, and search. MIT press, 2000

    2000

  40. [48]

    Emergence of scaling in random networks.science, 286(5439):509–512, 1999

    Albert-László Barabási and Réka Albert. Emergence of scaling in random networks.science, 286(5439):509–512, 1999

    1999

  41. [49]

    Position: The future of bayesian prediction is prior-fitted

    Samuel Müller, Arik Reuter, Noah Hollmann, David Rügamer, and Frank Hutter. Position: The future of bayesian prediction is prior-fitted. InInternational Conference on Machine Learning, pages 81861–81875. PMLR, 2025. 12 Table 4: Main hyperparameters of the synthetic graph prior...

    2025

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.