Predicting transcriptional interactions with GrnPFN

arXiv:2611.14559 · q-bio.MN · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "Predicting transcriptional interactions with GrnPFN".

Marcus: Here's the gist of the paper. Predicting transcriptional interactions with GrnPFN presents a method to predict nonadditive transcriptional responses to combinations from single-perturbation profiles without needing measured combinations in the target screen This model demonstrates that pretraining on simulated regulatory networks enables prediction of nonadditive transcriptional interactions across perturbation modalities when experimental measurements are scarce The core concept of GrnPFN GrnPFN is a prior-data fitted network that predicts nonadditive transcriptional responses to combinations from single-perturbation profiles It is pretrained on simulated screens generated from random gene-regulatory networks and incorporates measured combinations through in-context learning without retraining The model learns a transferable interaction prior from simulated screens, which allows it to predict the residual response beyond the additive effects of individual perturbations Model architecture and training signal The GrnPFN architecture arranges single perturbations, measured pairs, and query pairs as a grid of populations by genes It uses a shared transformer encoder that applies attention over genes and attention over populations with membership biases The model supports zero-shot prediction from single-perturbation and control profiles, in-context prediction conditioned on measured pairs, and fine-tuning on measured pairs The hybrid prior adds a screen-specific component to the network-derived interaction Performance advantages over baselines GrnPFN outperformed gradient boosting trained on other screens when no target-screen combinations were available, with a pooled cosine similarity difference of 0.049 (97.5% CI, 0.038–0.060) This advantage extended to an additional screen combining CRISPR interference with a drug, demonstrating the transferability of the simulation-trained prior across perturbation modalities GrnPFN also outperformed kernel ridge regression fitted to the same measured pairs when combination data were limited Generalizability and limitations GrnPFN generalized beyond perturbations represented in measured combinations, exceeding kernel ridge by 0.091 for unseen pairs at the 15% budget However, as more combination measurements became available, screen-specific regression models generally improved and are comparable to GrnPFN The relative advantage diminishes as screen-specific data accumulate The results position simulation-trained in-context learning as a complement to experimental measurement in sparse combinatorial screens Biological interpretation and evaluation rigor Directly predicting the interaction residual provides a biologically meaningful measure of nonadditive transcriptional responses by isolating deviations from the additive effects of individual perturbations To minimize bias from shared sampling errors, independent cells were used for model inputs, training targets, and test targets The findings suggest that transferable relationships between single effects and interactions can be easier to capture than the screen-specific structure needed for effective adaptation The practical value lies in reducing the dependence on a large initial combination training set while leaving room for fitted models to become preferable as a screen grows Materials and methods details Datasets used include K562 CRISPR activation, THP-1 Cas13 knockdown, and drug screens in A172, T98G, and U87MG The simulation prior was generated from random gene-regulatory networks containing 4–15 unmeasured pathway nodes and approximately 220–400 target genes The hybrid prior added a screen-specific component to the network-derived interaction, defined as RAB = RGRA B + ρ

f(zA, zB)P + (mA ⊙ mB)Q: Training involved using one simulated screen per training episode with a probability of 0.65 for query pairs The main benchmark included CRISPR activation in K562 cells, Cas13 knockdown in THP-1 cells, and drug combinations in A172, T98G, and U87MG The principal score was cosine similarity between predicted and observed interaction vectors over the 100 genes with the largest absolute centered observed interaction The model was equivariant to gene and population permutations and symmetric under exchange of pair members The final model contained 347,521 parameters without gene-identity embeddings Data and code availability Source datasets are available from Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The melanoma Perturb-CITE-seq dataset is available through scPerturb and Broad Single Cell Portal study SCP1064 The project repository is https://github.com/haihuilab/GrnPFN The paper is 2611.14559 References The references listed are 1 through 23 The references are Dixit A, Parnas O, Li B, Chen J, Fulco CP, Jerby-Arnon L, et al.

Ines: Perturb-Seq: Dissecting Molecular Circuits with Scalable Single-Cell RNA Profiling of Pooled Genetic Screens The references are Replogle JM, Saunders RA, Pogson AN, Hussmann JA, Lenail A, Guna A, et al.

Marcus: Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq The references are Srivatsan SR, McFaline-Figueroa JL, Ramani V, Saunders L, Cao J, Packer J, et al.

Ines: Massively multiplex chemical transcriptomics at single-cell resolution The references are Norman TM, Horlbeck MA, Replogle JM, Ge AY, Xu A, Jost M, et al.

Paper summary: Ines: To summarize this paper, "Predicting transcriptional interactions with GrnPFN," they propose a prior-data fitted network designed to predict nonadditive transcriptional responses when you only have single-perturbation profiles. They claim that pretraining on simulated screens allows the model to capture interaction structure across different perturbation types, like CRISPR activation or RNA knockdown and drug combinations.

Marcus: What matters for us is that they show this model beats a gradient boosting model trained on other screens when there are no measured combinations available in the target screen, showing a pooled cosine similarity difference of zero point zero four nine over ninety-seven percent confidence interval.

Yuki: It also extends this transferability to different types of perturbations, demonstrating that the simulation-trained prior works across modalities, even when combining CRISPR interference with a drug.

Ines: The authors are using separate cells for their model inputs, training targets, and test targets when evaluating GrnPFN in single-cell screens covering things like K562 CRISPR activation or THP1 Cas13 knockdown and drug combinations in A172, T98G, and U87MG.

Marcus: They're looking at the residual response beyond what you would expect from just adding up the individual effects of each perturbation. They frame this as directly targeting the nonadditive component of the transcriptional change.

Yuki: So, if you only have single-perturbation data and some sparse combination measurements, GrnPFN is positioned as a tool that can fill in those gaps by leveraging learned prior knowledge.

Conclusion: Ines: So, looking at "Predicting transcriptional interactions with GrnPFN" by Haihui Zhang and colleagues, the main implication is that we can predict complex biological outcomes without needing massive experimental combination datasets upfront.

Marcus: It suggests that focusing on building strong prior knowledge from simulated biological networks can be a way to make predictive models robust even when real-world experimental measurements are very limited.

Yuki: For us, this means we don't always need a huge library of interaction data to start modeling how different genetic or chemical signals influence a cell. We can use those simulated structures to build a baseline understanding of the system's potential complexity.

Ines: It positions simulation-trained in-context learning as a really useful complement when experimental measurement is sparse, but they are clear that as you gather more combination measurements, screen-specific regression models tend to get better and become comparable to GrnPFN.

Marcus: So the practical value seems to be reducing the initial dependence on training a massive combination dataset while still allowing those fitted models to improve as more data becomes available for that specific screen.

Department of Orofacial Sciences, School of Dentistry, University of California San Francisco

q-bio.MN

Submitted: 2026-11-05

Updated: 2026-11-05

Code: https://github.com/haihuilab/GrnPFN

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 88/100

The gist: Predicting transcriptional interactions with GrnPFN presents a method to predict nonadditive transcriptional responses to combinations from single-perturbation profiles without needing measured…

Key concepts

GrnPFN
A prior-data fitted network designed to predict how genes react when multiple perturbations (like drugs or genetic changes) are applied together. It learns interaction rules from simulated data and then uses this learned knowledge to predict the nonadditive effects of real experimental combinations.
Nonadditive Transcriptional Responses
This refers to effects where combining two individual perturbations does not equal the sum of their individual effects. GrnPFN specifically aims to model these deviations, which are crucial for understanding complex biological systems like gene regulation.
Transferable Interaction Prior
The knowledge learned by the model from simulating many different random gene-regulatory networks. This prior acts as a general rulebook that allows the model to make accurate predictions about new, unseen combinations of perturbations, even if those specific combinations were not directly tested in the training data.
Hybrid Prior
A combination of two learning sources: a network-derived interaction prediction and a screen-specific component. This hybrid approach improves accuracy by combining the general rules learned from simulations with specific knowledge gained from actual experimental measurements.

Terminology

Summary

Predicting transcriptional interactions with GrnPFN presents a method to predict nonadditive transcriptional responses to combinations from single-perturbation profiles without needing measured combinations in the target screen This model demonstrates that pretraining on simulated regulatory networks enables prediction of nonadditive transcriptional interactions across perturbation modalities when experimental measurements are scarce

The core concept of GrnPFN

GrnPFN is a prior-data fitted network that predicts nonadditive transcriptional responses to combinations from single-perturbation profiles It is pretrained on simulated screens generated from random gene-regulatory networks and incorporates measured combinations through in-context learning without retraining The model learns a transferable interaction prior from simulated screens, which allows it to predict the residual response beyond the additive effects of individual perturbations

Model architecture and training signal

The GrnPFN architecture arranges single perturbations, measured pairs, and query pairs as a grid of populations by genes It uses a shared transformer encoder that applies attention over genes and attention over populations with membership biases The model supports zero-shot prediction from single-perturbation and control profiles, in-context prediction conditioned on measured pairs, and fine-tuning on measured pairs The hybrid prior adds a screen-specific component to the network-derived interaction

Performance advantages over baselines

GrnPFN outperformed gradient boosting trained on other screens when no target-screen combinations were available, with a pooled cosine similarity difference of 0.049 (97.5% CI, 0.038–0.060) This advantage extended to an additional screen combining CRISPR interference with a drug, demonstrating the transferability of the simulation-trained prior across perturbation modalities GrnPFN also outperformed kernel ridge regression fitted to the same measured pairs when combination data were limited

Generalizability and limitations

GrnPFN generalized beyond perturbations represented in measured combinations, exceeding kernel ridge by 0.091 for unseen pairs at the 15% budget However, as more combination measurements became available, screen-specific regression models generally improved and are comparable to GrnPFN The relative advantage diminishes as screen-specific data accumulate The results position simulation-trained in-context learning as a complement to experimental measurement in sparse combinatorial screens

Biological interpretation and evaluation rigor

Directly predicting the interaction residual provides a biologically meaningful measure of nonadditive transcriptional responses by isolating deviations from the additive effects of individual perturbations To minimize bias from shared sampling errors, independent cells were used for model inputs, training targets, and test targets The findings suggest that transferable relationships between single effects and interactions can be easier to capture than the screen-specific structure needed for effective adaptation The practical value lies in reducing the dependence on a large initial combination training set while leaving room for fitted models to become preferable as a screen grows

Materials and methods details

Datasets used include K562 CRISPR activation, THP-1 Cas13 knockdown, and drug screens in A172, T98G, and U87MG The simulation prior was generated from random gene-regulatory networks containing 4–15 unmeasured pathway nodes and approximately 220–400 target genes The hybrid prior added a screen-specific component to the network-derived interaction, defined as RAB = RGRA B + ρ [f(zA, zB)P + (mA ⊙ mB)Q] Training involved using one simulated screen per training episode with a probability of 0.65 for query pairs The main benchmark included CRISPR activation in K562 cells, Cas13 knockdown in THP-1 cells, and drug combinations in A172, T98G, and U87MG The principal score was cosine similarity between predicted and observed interaction vectors over the 100 genes with the largest absolute centered observed interaction The model was equivariant to gene and population permutations and symmetric under exchange of pair members The final model contained 347,521 parameters without gene-identity embeddings

Data and code availability

Source datasets are available from Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The melanoma Perturb-CITE-seq dataset is available through scPerturb and Broad Single Cell Portal study SCP1064 The project repository is https://github.com/haihuilab/GrnPFN The paper is 2611. p. 0

Where the paper says it

Zhang@ucsf. edu (HZ) Short title: Predicting interactions in sparse perturbation screens 1 arXiv:2611.

14559

References

The references listed are 1 through 23 The references are Dixit A, Parnas O, Li B, Chen J, Fulco CP, Jerby-Arnon L, et al. Perturb-Seq: Dissecting Molecular Circuits with Scalable Single-Cell RNA Profiling of Pooled Genetic Screens The references are Replogle JM, Saunders RA, Pogson AN, Hussmann JA, Lenail A, Guna A, et al. p. 1

Where the paper says it

Introduction Cellular responses to combined perturbations can depart from the sum of their individual effects, revealing interactions that are not apparent from single perturbations alone. Single-cell genetic and chemical screens measure these responses at transcriptome scale [1–3], and combinatorial studies have characterized interactions involving genes, RNA perturbations and drugs [4–7].

Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq The references are Srivatsan SR, McFaline-Figueroa JL, Ramani V, Saunders L, Cao J, Packer J, et al. p. 5
Where the paper says it

This coupling can produce high apparent accuracy without recovery of the underlying interaction. In a melanoma Perturb-CITE-seq dataset [20], a predictor that assumed the knockout effect disappeared under interferon-𝛾 achieved a median cosine similarity of 0.

Massively multiplex chemical transcriptomics at single-cell resolution The references are Norman TM, Horlbeck MA, Replogle JM, Ge AY, Xu A, Jost M, et al. p. 1
Where the paper says it

Introduction Cellular responses to combined perturbations can depart from the sum of their individual effects, revealing interactions that are not apparent from single perturbations alone. Single-cell genetic and chemical screens measure these responses at transcriptome scale [1–3], and combinatorial studies have characterized interactions involving genes, RNA perturbations and drugs [4–7].

Exploring genetic interaction manifolds constructed from rich single-cell phenotypes The references are Horlbeck MA, Xu A, Wang M, Bennett NK, Park CY, Bogdanoff D, et al. p. 1
Where the paper says it

Introduction Cellular responses to combined perturbations can depart from the sum of their individual effects, revealing interactions that are not apparent from single perturbations alone. Single-cell genetic and chemical screens measure these responses at transcriptome scale [1–3], and combinatorial studies have characterized interactions involving genes, RNA perturbations and drugs [4–7].

Mapping the Genetic Landscape of Human Cells The references are Wessels H-H, Méndez-Mancilla A, Hao Y, Papalexi E, Mauck WM III, Lu L, et al. Efficient combinatorial targeting of RNA transcripts in single cells with Cas13 RNA Perturb-seq The references are McFaline-Figueroa JL, Srivatsan S, Hill AJ, Gasperini M, Jackson DL, Saunders L, et al. p. 1
Where the paper says it

Introduction Cellular responses to combined perturbations can depart from the sum of their individual effects, revealing interactions that are not apparent from single perturbations alone. Single-cell genetic and chemical screens measure these responses at transcriptome scale [1–3], and combinatorial studies have characterized interactions involving genes, RNA perturbations and drugs [4–7].

Multiplex single-cell chemical genomics reveals the kinase dependence of the response to targeted therapy The references are Roohani Y, Huang K, Leskovec J. p. 1
Where the paper says it

Introduction Cellular responses to combined perturbations can depart from the sum of their individual effects, revealing interactions that are not apparent from single perturbations alone. Single-cell genetic and chemical screens measure these responses at transcriptome scale [1–3], and combinatorial studies have characterized interactions involving genes, RNA perturbations and drugs [4–7].

Predicting transcriptional outcomes of novel multigene perturbations with GEARS The references are Lotfollahi M, Klimovskaia Susmelj A, De Donno C, Hetzel L, Ji Y, Ibarra IL, et al. p. 1
Where the paper says it

Combined genetic and chemical perturbations can elicit nonadditive transcriptional responses, yet the vast number of possible combinations limits experimental characterization. Predicting these interactions is particularly challenging when few or no combinations have been measured.

Predicting cellular responses to complex perturbations in high-throughput screens The references are Klein D, Fleck JS, Bobrovskiy D, Zimmermann L, Becker S, Palma A, et al. p. 1
Where the paper says it

Introduction Cellular responses to combined perturbations can depart from the sum of their individual effects, revealing interactions that are not apparent from single perturbations alone. Single-cell genetic and chemical screens measure these responses at transcriptome scale [1–3], and combinatorial studies have characterized interactions involving genes, RNA perturbations and drugs [4–7].

CellFlow enables generative single-cell phenotype modeling with flow matching The references are Ahlmann-Eltze C, Huber W, Anders S. p. 2
Where the paper says it

Informed modeling incorporates biological information in a form suited to the inference task. For example, scBIOT can use cell-identity annotations to guide optimal-transport alignment across single-cell datasets [13].

Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines The references are Miller HE, Mejia GM, Leblanc FJA, Swain B, Wang B, de Lima Camillo LP. p. 1
Where the paper says it

Assessing their contribution requires distinguishing prediction of the full expression response from prediction of the interaction itself. Recent benchmarks have shown that simple baselines can be competitive with deep-learning models [11], while calibration-aware analyses have identified settings in which more 2

Deep learning perturbation models can outperform baselines on calibrated metrics The references are Zhang H, Qin P. p. 1
Where the paper says it

Assessing their contribution requires distinguishing prediction of the full expression response from prediction of the interaction itself. Recent benchmarks have shown that simple baselines can be competitive with deep-learning models [11], while calibration-aware analyses have identified settings in which more 2

Informed, scalable and unified transport couplings for singlecell alignment and downstream inference with scBIOT The references are Müller S, Hollmann N, Pineda Arango S, Grabocka J, Hutter F. p. 2
Where the paper says it

Informed modeling incorporates biological information in a form suited to the inference task. For example, scBIOT can use cell-identity annotations to guide optimal-transport alignment across single-cell datasets [13].

Transformers Can Do Bayesian Inference The references are Hollmann N, Müller S, Eggensperger K, Hutter F. p. 2
Where the paper says it

However, regression models fitted to measured pairs have limited evidence from which to learn this component when combinations are sparsely sampled. Informed modeling incorporates biological information in a form suited to the inference task.

TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second The references are Sextro M, Kłos W, Dernbach G. MapPFN: Learning Causal Perturbation Maps in Context The references are Gao Y, Hernández-Lobato JM, Guo S. PerturbPFN: Probing the Limits of Synthetic Priors in Drug Perturbation Modelling The references are Palla G, Hillsley A, Kim Y-J, Royer LA. p. 2
Where the paper says it

Predicting nonadditive transcriptional responses to combined perturbations is essential for understanding how molecular interventions interact and for guiding the rational design of combinatorial therapies. Recent applications include MapPFN, which learns from simulated gene-regulatory networks (GRNs) to predict genetic perturbation responses [17], and PerturbPFN, which uses synthetic structural models for chemical perturbation prediction [18].

Tabular Foundation Models Are Competitive Cellular Perturbation Predictors Across Biological Scales The references are Frangieh CJ, Melms JC, Thakore PI, Geiger-Schuller KR, Ho P, Luoma AM, et al. p. 2
Where the paper says it

Recent applications include MapPFN, which learns from simulated gene-regulatory networks (GRNs) to predict genetic perturbation responses [17], and PerturbPFN, which uses synthetic structural models for chemical perturbation prediction [18]. General-purpose tabular foundation models have also been evaluated on single-perturbation responses [19].

Multimodal pooled Perturb-CITE-seq screens in patient models define mechanisms of cancer immune evasion The references are Xu Z, Lu Z, Ugurbil A, Abdulraouf A, Liao A, Zhang J, et al. p. 5
Where the paper says it

This coupling can produce high apparent accuracy without recovery of the underlying interaction. In a melanoma Perturb-CITE-seq dataset [20], a predictor that assumed the knockout effect disappeared under interferon-𝛾 achieved a median cosine similarity of 0.

Mapping convergent regulators of melanoma drug resistance by PerturbFate The references are Stuart T, Butler A, Hoffman P, Hafemeister C, Papalexi E, Mauck WM III, et al. Comprehensive Integration of Single-Cell Data The references are Wolf FA, Angerer P, Theis FJ. p. 5
Where the paper says it

The main prediction benchmark used frozen model weights trained entirely on simulated data. Separating cell roles enables evaluation of interaction prediction Interaction targets contain the negative of each single-perturbation effect.

SCANPY: large-scale single-cell gene expression data analysis The references are Chernozhukov V, Chetverikov D, Demirer M, Duflo E, Hansen C, Newey W. p. 3
Where the paper says it

These results demonstrate that a model trained on simulated regulatory dynamics can improve transcriptional interaction prediction at the beginning of combinatorial screens, when empirical combination data are most limited. Results GrnPFN learns a transferable interaction prior from simulated screens We defined the effect of perturbation 𝑝, ∆𝑝, as its mean change in log-normalized gene expression relative to control cells.

Double/debiased machine learning for treatment and structural parameters The references are Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: Machine Learning in Python The references are Owen AB, Eckles D. p. 2
Where the paper says it

For perturbation prediction, biological assumptions can be encoded in a distribution of simulated datasets used for pretraining. Prior-data fitted networks (PFNs) learn predictive structure from datasets sampled from such a prior and subsequently condition on observations from a new dataset through in-context learning [14–16].

Bootstrapping data arrays of arbitrary order The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 The references are scPerturb and Broad Single Cell Portal study SCP1064 The references are Gene Expression Omnibus under GSE133344, GSE213957, GSM7056151 (within GSE225775), GSE291147, and GSE280506 <ref:2607. p. 6
Where the paper says it

or testing within a fold. This design separated the single-perturbation and control-cell noise across the three roles while retaining the available depth for each combination.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Prior-data fitted networks can predict nonadditive transcriptional responses to combinations from single-perturbation profiles even when no combinations have been measured in a target screen. This enables zero-shot prediction without requiring measured pairs in the target screen, outperforming gradient boosting and kernel ridge regression when combination data are limited.

  2. The GrnPFN architecture uses a hybrid prior that adds a screen-specific interaction component, specifically R AB = R GRN AB + rho [f(z A, z B)P + (m A ⊙ m B)Q], to learn predictive structure from simulated regulatory dynamics. This prior allows the model to draw on relationships learned before application and incorporate new observations through in-context learning without retraining.

  3. The model can be evaluated across multiple perturbation modalities, demonstrating that pretraining on simulated regulatory networks enables prediction of nonadditive transcriptional interactions across perturbation modalities when experimental measurements are scarce. This transferability is evidenced by GrnPFN's performance on an additional CRISPR interference drug screen not included in the pretraining data.

  4. Predictive uncertainty can identify more accurate subsets, with the most confident third having higher cosine similarity than the least confident third. Retaining a smaller fraction of the most confident pairs generally increased mean accuracy, suggesting that focusing model adaptation on high-confidence predictions is beneficial when measurement budgets are low.

  5. The model supports zero-shot prediction from single-perturbation profiles and in-context prediction conditioned on measured pairs, as well as fine-tuning on measured pairs. This versatility allows the system to immediately provide an initial interaction prior and revise it with empirical measurements as data becomes available.

Abstract

Combinatorial therapies offer opportunities to improve treatment efficacy and overcome therapeutic resistance, but their rational design requires understanding how combined interventions interact beyond their individual effects. Combined genetic and chemical perturbations can elicit nonadditive transcriptional responses, yet the vast number of possible combinations limits experimental characterization. Predicting these interactions is particularly challenging when few or no combinations have been measured. Here we present GrnPFN, a prior-data fitted network that predicts nonadditive transcriptional responses to combinations from single-perturbation profiles. Pretrained on simulated screens generated from random gene-regulatory networks, GrnPFN incorporates measured combinations through in-context learning without retraining. We evaluated GrnPFN in single-cell screens spanning CRISPR activation, RNA knockdown and drug combinations, using separate cells to construct model inputs, training targets and test targets. Without measured combinations in the target screen, GrnPFN outperformed gradient boosting trained on other screens. With limited combination measurements, it outperformed kernel ridge regression fitted to the same measured pairs. These advantages extended to an additional screen combining CRISPR interference with a drug. Our results demonstrate that pretraining on simulated regulatory networks enables prediction of nonadditive transcriptional interactions across perturbation modalities when experimental measurements are scarce.

Sources

Related papers