Benchmarking biochemical networks generated by large language models

The ability of cells to grow or respond to external stimuli is fundamental to life, mediated by complex metabolic and signaling networks. Computational models of signaling networks have been widely used to understand their systems properties and guide new experiments (Karr et al., 2012), but they typically require substantial manual literature curation of network reactions and rate constants (Qiao et al., 2025). Similarly, metabolic network models provide a powerful framework for predicting cellular metabolism and substrate utilization, but their construction has historically required manual curation of stoichiometric coefficients, reaction reversibility, and metabolite–reaction connectivity.

Artificial intelligence (AI), and large language models (LLMs) in particular, have demonstrated strong performance in tasks that traditionally required labor-intensive knowledge curation. For example, general-purpose LLMs have scored highly on the U.S. Medical Licensing Examination (Singhal et al., 2025), law exams (Achiam et al., 2024), and Advanced Placement exams (Achiam et al., 2024). AI has also made substantial contributions in biology, including predicting protein structure (Abramson et al., 2024) and cell types (Theodoris et al., 2023), and is expected to play a central role in the development of comprehensive virtual cells (Bunne et al., 2024; Roohani et al., 2025). However, to our knowledge, LLM-generated predictive models of biochemical networks have not been reported or benchmarked. Here, we develop pipelines for LLM-generated models of signaling and metabolic networks, and we test whether general-purpose LLMs accurately predict network structure. Additionally, we assess the LLM-generated networks’ accuracy in predicting perturbation responses, as benchmarked against manually curated models and experimental evidence (Ryall et al., 2012; Zeigler et al., 2016; Tan et al., 2017).

We first established a computational pipeline to generate models of signaling networks from LLMs using a specified gene set (Figure 1A). For benchmarking, each gene set was derived from annotations of three manually curated, logic-based network models that have been extensively validated using experimental data (Ryall et al., 2012; Zeigler et al., 2016; Tan et al., 2017). We optimized a programmatic prompt that iteratively queries an LLM for directed interactions (activation or inhibition) among the gene products in the context of a particular phenotype (here ‘hypertrophy’, ‘fibroblast’, or ‘mechanosignaling’). We then extracted the returned directed interactions, examined network structure, and converted the network into a logic-based formalism to simulate network responses to perturbations (Clark et al., 2025). Each LLM was prompted 10 times per network to examine reproducibility (see methods for prompting parameters).

Figure 1 with 5 supplements see all

Download asset Open asset

Leveraging this pipeline, we tested the extent to which three general-purpose LLMs (GPT-5.2 Pro, OpenAI; Gemini 3.0 Pro-Preview, Google; and Claude Opus 4.6, Anthropic) could reconstruct the structure of the manually curated cardiomyocyte hypertrophy signaling network (Ryall et al., 2012), based on its gene annotations alone.

All three LLMs captured aspects of this manually curated network, but with varying coverage depending on the region of the network and the specific LLM (Figure 1B). In particular, GPT-5.2 Pro, Gemini 3.0 Pro-Preview, and Claude Opus 4.6 all performed relatively well in reconstructing upstream ligand-receptor interactions (e.g. EGF to EGFR) and well-established second-messenger cascades such as the β-adrenergic receptor–cAMP–CREB or phospholipase C–diacylglycerol–PKC pathways. They also accurately predicted the structure of highly conserved signaling axes such as PI3K-Akt-mTOR and calcium-calmodulin-calcineurin-NFAT. In contrast, the three LLMs were less able to reconstruct regulation of downstream transcription factors (e.g. cFOS, NFAT, cJUN, MEF2) and gene expression (SERCA, αMHC, βMHC, natriuretic peptides ANP/BNP) or cell area/hypertrophy. The lower reconstruction accuracy of downstream signaling may arise from cardiomyocyte-specific reactions. Overall, the three LLMs reconstructed 37.70–63.87% of network reactions in the cardiomyocyte hypertrophy network (Figure 1C).

To what extent is the limited recall of LLM-generated networks specific to cardiomyocyte hypertrophy? To address this question, we used the same reasoning models to reconstruct network models of fibroblast activation (Zeigler et al., 2016; 91 nodes, 134 reactions) and mechanosignaling (Tan et al., 2017; 94 nodes, 125 reactions). Like the hypertrophy model, these networks were previously manually reconstructed from the literature, translated into predictive logic-based models, and had their predictions validated against substantial experiments from the literature not used to develop the model. Consistent with the hypertrophy network, LLM-generated models of fibroblast activation and mechanosignaling reconstructed 24.07–60.49% and 29.07–64.53% of their manually curated counterparts (Figure 1D and E).

We then asked whether the regional coverage of LLM-generated networks of fibroblasts and mechanosignaling mirrors that of cardiomyocytes. Indeed, the LLMs consistently reconstructed ligand-receptor interactions and highly conserved pathways such as cAMP and Ras/Raf/ERK, but did not consistently recall AT1R signaling, PDGFR-abl signaling, and downstream regulation of transcription factors, genes, and extracellular matrix (e.g. MMP-collagen subnetwork; Figure 1—figure supplement 1). Unlike hypertrophy and fibroblast networks, mechanosignaling is mediated by stretch-dependent proteins that are relatively less characterized. Indeed, the LLM-generated networks for mechanosignaling did not adequately enumerate direct mechanosensory reactions and also had limited coverage of gene regulation (Figure 1—figure supplement 2). Still, core conserved pathways for calcium regulation, regulation of protein synthesis, and MAPK signaling were reconstructed robustly.

Models of signaling networks are useful not just for their structure, but also for simulating the response to new perturbations. Prior to benchmarking, the hypertrophy, fibroblast, and mechanosignaling networks were automatically translated into logic-based differential equation models using the Netflux framework (Clark et al., 2025; Kraeutler et al., 2010). This allows network-wide prediction of dynamics in response to desired perturbations and does not require prior knowledge of rate constants or concentrations (Kraeutler et al., 2010). The predictions of these three logic-based models have been rigorously validated against experimental evidence not used to develop the model (Ryall et al., 2012; Zeigler et al., 2016; Tan et al., 2017; Khalilimeybodi et al., 2020). Using the same Netflux approach, we generated logic-based models based on all of the LLM-generated network structures shown in Figure 1.

Returning our focus to the hypertrophy network, we examined the ability of LLM-generated networks to predict the classic ‘fetal gene program’ gene expression signature in cardiomyocytes. The manually curated model accurately predicts that AngII and/or ISO increase the expression of ANP, BNP, GATA4, βMHC, sACT, and increase cell area/size, while decreasing expression of SERCA and αMHC. In contrast, the LLM-generated network models produce limited aspects of this signature, with the best performance in predicting ISO-dependent ANP in the GPT version (Figure 2A). We then tested whether this validation accuracy scales across all available manually curated experiments that cover a wider array of perturbations and measured outputs. Overall, for 114 validations where the manually curated network has a functional accuracy of 94.74%, LLM-generated networks have an accuracy of 6.14–33.33% (Figure 2B). We then repeated this systematic experimental validation in comparison to literature-based models of the fibroblast (82% validation against independent literature) and mechanosignaling (78% validation). LLM-generated models of the fibroblast (Figure 2C) and mechanosignaling (Figure 2D) networks exhibited validation accuracies similar to the hypertrophy model: 16.87–21.69% and 5.81–14.53%, respectively. Across all three networks, validation accuracy of LLM-generated models for perturbation responses is lower than their accuracy for reconstructing network structure.

Figure 2 with 1 supplement see all

Download asset Open asset

It is important to note that the functional validation simulations used LLM-generated signaling network models after pruning to the subset of reactions contained in the ground truth networks. This enabled direct mapping of LLM-predicted interactions onto the ‘AND’ formalism used in the logic-based network models, but also excluded predicted biological interactions that were generated by the LLMs but absent from the curated networks, which we tentatively classify as false positives (LLM⁺, ground-truth^-). When this constraint was removed and the full set of LLM-predicted interactions was used for functional validation, model performance did not substantially improve, except for the fibroblast network (Figure 2—figure supplement 1A-C). Additionally, the average size of the unconstrained LLM-generated networks was substantially larger than the manually curated network (Figure 2—figure supplement 1D).

Additional performance metrics, including confusion matrices and calibration analyses for each signaling network, are provided in Figure 1—figure supplements 3–5. Notably, Gemini generally exhibited substantially higher precision than Claude or GPT in predicting signaling network reactions. F1 scores were relatively consistent among the evaluated LLMs.

Reactions generated by the LLMs but absent from the ground-truth signaling networks were classified as ‘false positives’ for the purpose of calculating performance metrics. This classification should be interpreted cautiously, as some LLM-predicted/ground-truth-negative reactions may represent true biomolecular interactions or regulatory relationships that were not captured in the manually curated reference models. To assess the impact of retaining these additional predictions, we evaluated the functional performance of the full LLM-generated signaling networks without pruning LLM-predicted/ground-truth-negative reactions. This analysis revealed only marginal improvements, and in some cases reduced accuracy, for the hypertrophy and mechanosignaling networks (Figure 2—figure supplement 1A and C). In contrast, the unrefined fibroblast network models exhibited substantially greater functional validation performance than the pruned reconstructions, which retained only LLM-predicted/ground-truth-positive reactions (Figure 2—figure supplement 1B). The Claude- and GPT-generated signaling networks contained more connections than the corresponding manually curated networks – this increased network density may have contributed to their improved functional performance within the fibroblast context (Figure 2—figure supplement 1D).

Expanding the scope of our LLM-benchmarking, we asked how effectively these models could reconstruct a metabolic network of core Escherichia coli (E. coli) metabolism (Orth et al., 2010). The reconstruction was guided by the genes represented in the gene–protein–reaction (GPR) associations of the reference network. We adapted our prompting approach to generate the information needed to define metabolic reactions, including substrates, products, stoichiometry, and reversibility. These outputs were used to assemble LLM-generated metabolic network models for comparison against the manually curated reference network, as described in the Methods.

There is notable heterogeneity in the regional coverage of this metabolic network by the different LLMs. Generally, both Claude and Gemini were able to accurately reconstruct the glycolysis and tricarboxylic acid cycle (TCA) while GPT failed to capture a few specific reactions in each process, including the phosphoglycerate mutase and kinase (PGM and PGK, respectively). Additionally, GPT failed to reconstruct the lactate dehydrogenase and lactate transport reactions. Gemini and Claude performed poorly in recreating the malate and fumarate transport reactions feeding into the TCA cycle. Overall, the LLM-generated metabolic network models demonstrated markedly greater coverage of the ground-truth core E. coli metabolic network than the predicted signaling networks achieved for their respective ground-truth networks. Specifically, the LLM-generated models reconstructed 63.63–90.90% of reactions within the manually curated core E. coli network (Figure 3C).

Figure 3

Download asset Open asset

Despite substantial improvement in coverage of the ground truth reactions, the LLM-generated metabolic networks models largely underperformed in the substrate utilization analysis. Indeed, the majority of the networks yielded infeasible flux solutions or predicted no growth on all of the substrates (Figure 3D). Interestingly, a sparse subset of the Gemini- and Claude-generated models achieved strong substrate utilization prediction performance, with accuracies approaching the score of the ground-truth model (Figure 3E).

In conclusion, we developed pipelines that use general-purpose LLMs to generate predictive models of signaling and metabolic networks. We benchmarked the signaling network pipeline across three LLMs against large-scale, manually curated, and experimentally validated network models. Across cardiac hypertrophy, fibroblast, and mechano-signaling networks, LLM-generated models exhibit 24.07-64.53% reconstruction of reactions from the manually curated signaling models. Reconstruction accuracy is high for ligand-receptor interactions and well-conserved pathways, but lower for cell-specific gene regulation or biophysical regulation.

Similarly, we benchmarked the LLMs’ performance in generating a metabolic network using experimental data and a manually curated reference model. Although the LLMs achieved greater structural recovery of the core E. coli metabolic network relative to the signaling networks, these reconstructions did not translate into robust functional performance as assessed by substrate utilization analysis. This discrepancy may reflect gaps in key reconstructed metabolic pathways and aligns with findings from the signaling network models, where incomplete reconstruction of critical pathways likely limited the accuracy of simulated perturbation responses. The ability of LLMs to reconstruct core E. coli metabolism may reflect the depth of existing literature on this model system, and similar approaches may perform worse for metabolic networks that are less extensively characterized.

This benchmarking does not yet test the ability of LLMs to reconstruct aspects of signaling that are not represented in literature-curated networks, such as isoform-specific regulation (George et al., 2020) or liquid-liquid phase separation (Zhang et al., 2020). Such mechanisms are not yet adequately characterized for global inclusion in either literature-curated or LLM-generated models.

Future efforts could build on the pipelines developed in this work. Improving the performance of LLM-generated network models may require hybrid manual–LLM workflows or specialized AI models tailored to biological network reconstruction. Another promising direction would be to implement an agentic framework in which some agents propose plausible signaling reactions between genes, while others serve in a reviewer role to critique and approve candidate reactions. Such structured critique paradigms have shown promise in AI-oriented research contexts (Swanson et al., 2025). As LLMs and other AI models improve in their ability to synthesize and connect biological knowledge derived from experimental studies, developing rigorous platforms for organizing and evaluating this information will be increasingly important.

The code used in this project is freely available on GitHub at https://github.com/saucermanlab/LLM-network-generation (copy archived at Tewari, 2026).

1. 1. Abramson J
   2. Adler J
   3. Dunger J
   4. Evans R
   5. Green T
   6. Pritzel A
   7. Ronneberger O
   8. Willmore L
   9. Ballard AJ
   10. Bambrick J
   11. Bodenstein SW
   12. Evans DA
   13. Hung C-C
   14. O’Neill M
   15. Reiman D
   16. Tunyasuvunakool K
   17. Wu Z
   18. Žemgulytė A
   19. Arvaniti E
   20. Beattie C
   21. Bertolli O
   22. Bridgland A
   23. Cherepanov A
   24. Congreve M
   25. Cowen-Rivers AI
   26. Cowie A
   27. Figurnov M
   28. Fuchs FB
   29. Gladman H
   30. Jain R
   31. Khan YA
   32. Low CMR
   33. Perlin K
   34. Potapenko A
   35. Savy P
   36. Singh S
   37. Stecula A
   38. Thillaisundaram A
   39. Tong C
   40. Yakneen S
   41. Zhong ED
   42. Zielinski M
   43. Žídek A
   44. Bapst V
   45. Kohli P
   46. Jaderberg M
   47. Hassabis D
   48. Jumper JM

   (2024) Accurate structure prediction of biomolecular interactions with alphaFold 3

   Nature 630:493–500.

   https://doi.org/10.1038/s41586-024-07487-w

   • PubMed
   • Google Scholar
2. Preprint

   1. Achiam J
   2. Adler S
   3. Agarwal A
   4. Ahmad L
   5. Akkaya I
   6. OpenAI

   (2024) GPT-4 technical report

   arXiv.

   https://doi.org/10.48550/arXiv.2303.08774

   • Google Scholar
3. 1. Bunne C
   2. Roohani Y
   3. Rosen Y
   4. Gupta A
   5. Zhang X
   6. Roed M
   7. Alexandrov T
   8. AlQuraishi M
   9. Brennan P
   10. Burkhardt DB
   11. Califano A
   12. Cool J
   13. Dernburg AF
   14. Ewing K
   15. Fox EB
   16. Haury M
   17. Herr AE
   18. Horvitz E
   19. Hsu PD
   20. Jain V
   21. Johnson GR
   22. Kalil T
   23. Kelley DR
   24. Kelley SO
   25. Kreshuk A
   26. Mitchison T
   27. Otte S
   28. Shendure J
   29. Sofroniew NJ
   30. Theis F
   31. Theodoris CV
   32. Upadhyayula S
   33. Valer M
   34. Wang B
   35. Xing E
   36. Yeung-Levy S
   37. Zitnik M
   38. Karaletsos T
   39. Regev A
   40. Lundberg E
   41. Leskovec J
   42. Quake SR

   (2024) How to build the virtual cell with artificial intelligence: priorities and opportunities

   Cell 187:7045–7063.

   https://doi.org/10.1016/j.cell.2024.11.015

   • PubMed
   • Google Scholar
4. 1. Cao Y
   2. Mu H
   3. Guo J
   4. Liu H
   5. Zhang R
   6. Liu W
   7. Xian M
   8. Liu H

   (2020) Metabolic engineering of Escherichia coli for the utilization of ethanol

   Journal of Biological Research-Thessaloniki 27:1.

   https://doi.org/10.1186/s40709-020-0111-0

   • PubMed
   • Google Scholar
5. 1. Clark AP
   2. Chowkwale M
   3. Paap A
   4. Dang S
   5. Saucerman JJ

   (2025) Logic-based modeling of biological networks with Netflux

   PLOS Computational Biology 21:e1012864.

   https://doi.org/10.1371/journal.pcbi.1012864

   • PubMed
   • Google Scholar
6. 1. Ebrahim A
   2. Lerman JA
   3. Palsson BO
   4. Hyduke DR

   (2013) COBRApy: constraints-based reconstruction and analysis for python

   BMC Systems Biology 7:74.

   https://doi.org/10.1186/1752-0509-7-74

   • PubMed
   • Google Scholar
7. 1. George SA
   2. Kiss A
   3. Obaid SN
   4. Venegas A
   5. Talapatra T
   6. Wei C
   7. Efimova T
   8. Efimov IR

   (2020) p38δ genetic ablation protects female mice from anthracycline cardiotoxicity

   American Journal of Physiology. Heart and Circulatory Physiology 319:H775–H786.

   https://doi.org/10.1152/ajpheart.00415.2020

   • PubMed
   • Google Scholar
8. 1. Karp PD
   2. Paley S
   3. Caspi R
   4. Kothari A
   5. Krummenacker M
   6. Midford PE
   7. Moore LR
   8. Subhraveti P
   9. Gama-Castro S
   10. Tierrafria VH
   11. Lara P
   12. Muñiz-Rascado L
   13. Bonavides-Martinez C
   14. Santos-Zavaleta A
   15. Mackie A
   16. Sun G
   17. Ahn-Horst TA
   18. Choi H
   19. Juenemann R
   20. Knudsen CNM
   21. Covert MW
   22. Collado-Vides J
   23. Paulsen I

   (2025) The EcoCyc database (2025)

   EcoSal Plus 13:eesp00192024.

   https://doi.org/10.1128/ecosalplus.esp-0019-2024

   • PubMed
   • Google Scholar
9. 1. Karr JR
   2. Sanghvi JC
   3. Macklin DN
   4. Gutschow MV
   5. Jacobs JM
   6. Bolival B Jr
   7. Assad-Garcia N
   8. Glass JI
   9. Covert MW

   (2012) A whole-cell computational model predicts phenotype from genotype

   Cell 150:389–401.

   https://doi.org/10.1016/j.cell.2012.05.044

   • PubMed
   • Google Scholar
10. 1. Khalilimeybodi A
    2. Paap AM
    3. Christiansen SLM
    4. Saucerman JJ

    (2020) Context-specific network modeling identifies new crosstalk in β-adrenergic cardiac hypertrophy

    PLOS Computational Biology 16:e1008490.

    https://doi.org/10.1371/journal.pcbi.1008490

    • PubMed
    • Google Scholar
11. 1. King ZA
    2. Lu J
    3. Dräger A
    4. Miller P
    5. Federowicz S
    6. Lerman JA
    7. Ebrahim A
    8. Palsson BO
    9. Lewis NE

    (2016) BiGG Models: A platform for integrating, standardizing and sharing genome-scale models

    Nucleic Acids Research 44:D515–D522.

    https://doi.org/10.1093/nar/gkv1049

    • PubMed
    • Google Scholar
12. 1. Kraeutler MJ
    2. Soltis AR
    3. Saucerman JJ

    (2010) Modeling cardiac β-adrenergic signaling with normalized-Hill differential equations: comparison with a biochemical model

    BMC Systems Biology 4:157.

    https://doi.org/10.1186/1752-0509-4-157

    • PubMed
    • Google Scholar
13. 1. Liang H
    2. Ma X
    3. Ning W
    4. Liu Y
    5. Sinskey AJ
    6. Stephanopoulos G
    7. Zhou K

    (2021) Constructing an ethanol utilization pathway in Escherichia coli to produce acetyl-CoA derived compounds

    Metabolic Engineering 65:223–231.

    https://doi.org/10.1016/j.ymben.2020.11.010

    • PubMed
    • Google Scholar
14. 1. Orth JD
    2. Fleming RMT
    3. Palsson BØ

    (2010) Reconstruction and use of microbial metabolic networks: the core Escherichia coli metabolic model as an educational guide

    EcoSal Plus 4:1128.

    https://doi.org/10.1128/ecosalplus.10.2.1

    • PubMed
    • Google Scholar
15. 1. Pedregosa F
    2. Varoquaux G
    3. Gramfort A
    4. Michel V
    5. Thirion B

    (2011)

    Scikit-learn: machine learning in python

    JMLR 12:2825–2830.

    • Google Scholar
16. 1. Qiao L
    2. Khalilimeybodi A
    3. Linden-Santangeli NJ
    4. Rangamani P

    (2025) The evolution of systems biology and systems medicine: from mechanistic models to uncertainty quantification

    Annual Review of Biomedical Engineering 27:425–447.

    https://doi.org/10.1146/annurev-bioeng-102723-065309

    • PubMed
    • Google Scholar
17. 1. Roohani YH
    2. Hua TJ
    3. Tung P-Y
    4. Bounds LR
    5. Yu FB
    6. Dobin A
    7. Teyssier N
    8. Adduri A
    9. Woodrow A
    10. Plosky BS
    11. Mehta R
    12. Hsu B
    13. Sullivan J
    14. Ricci-Tam C
    15. Li N
    16. Kazaks J
    17. Gilbert LA
    18. Konermann S
    19. Hsu PD
    20. Goodarzi H
    21. Burke DP

    (2025) Virtual cell challenge: toward a turing test for the virtual cell

    Cell 188:3370–3374.

    https://doi.org/10.1016/j.cell.2025.06.008

    • PubMed
    • Google Scholar
18. 1. Ryall KA
    2. Holland DO
    3. Delaney KA
    4. Kraeutler MJ
    5. Parker AJ
    6. Saucerman JJ

    (2012) Network reconstruction and systems analysis of cardiac myocyte hypertrophy signaling

    The Journal of Biological Chemistry 287:42259–42268.

    https://doi.org/10.1074/jbc.M112.382937

    • PubMed
    • Google Scholar
19. 1. Singhal K
    2. Tu T
    3. Gottweis J
    4. Sayres R
    5. Wulczyn E
    6. Amin M
    7. Hou L
    8. Clark K
    9. Pfohl SR
    10. Cole-Lewis H
    11. Neal D
    12. Rashid QM
    13. Schaekermann M
    14. Wang A
    15. Dash D
    16. Chen JH
    17. Shah NH
    18. Lachgar S
    19. Mansfield PA
    20. Prakash S
    21. Green B
    22. Dominowska E
    23. Agüera Y Arcas B
    24. Tomašev N
    25. Liu Y
    26. Wong R
    27. Semturs C
    28. Mahdavi SS
    29. Barral JK
    30. Webster DR
    31. Corrado GS
    32. Matias Y
    33. Azizi S
    34. Karthikesalingam A
    35. Natarajan V

    (2025) Toward expert-level medical question answering with large language models

    Nature Medicine 31:943–950.

    https://doi.org/10.1038/s41591-024-03423-7

    • PubMed
    • Google Scholar
20. 1. Swanson K
    2. Wu W
    3. Bulaong NL
    4. Pak JE
    5. Zou J

    (2025) The Virtual Lab of AI agents designs new SARS-CoV-2 nanobodies

    Nature 646:716–723.

    https://doi.org/10.1038/s41586-025-09442-9

    • PubMed
    • Google Scholar
21. 1. Tan PM
    2. Buchholz KS
    3. Omens JH
    4. McCulloch AD
    5. Saucerman JJ

    (2017) Predictive model identifies key network regulators of cardiomyocyte mechano-signaling

    PLOS Computational Biology 13:e1005854.

    https://doi.org/10.1371/journal.pcbi.1005854

    • PubMed
    • Google Scholar
22. Software

    1. Tewari J

    (2026) LLM-network-generation, version swh:1:rev:b9d78b5a3f2406a415d82db4cb1245dbbe6422cd

    Software Heritage.

    https://archive.softwareheritage.org/swh:1:dir:d6bda79b3734bdb0fc737bb1406689612b0677c2;origin=https://github.com/saucermanlab/LLM-network-generation;visit=swh:1:snp:a8e512cfac4adec17d95ad09150bf00fb9a26946;anchor=swh:1:rev:b9d78b5a3f2406a415d82db4cb1245dbbe6422cd
23. 1. Theodoris CV
    2. Xiao L
    3. Chopra A
    4. Chaffin MD
    5. Al Sayed ZR
    6. Hill MC
    7. Mantineo H
    8. Brydon EM
    9. Zeng Z
    10. Liu XS
    11. Ellinor PT

    (2023) Transfer learning enables predictions in network biology

    Nature 618:616–624.

    https://doi.org/10.1038/s41586-023-06139-9

    • PubMed
    • Google Scholar
24. 1. Zeigler AC
    2. Richardson WJ
    3. Holmes JW
    4. Saucerman JJ

    (2016) A computational model of cardiac fibroblast signaling predicts context-dependent drivers of myofibroblast differentiation

    Journal of Molecular and Cellular Cardiology 94:72–81.

    https://doi.org/10.1016/j.yjmcc.2016.03.008

    • PubMed
    • Google Scholar
25. 1. Zhang JZ
    2. Lu TW
    3. Stolerman LM
    4. Tenner B
    5. Yang JR
    6. Zhang JF
    7. Falcke M
    8. Rangamani P
    9. Taylor SS
    10. Mehta S
    11. Zhang J

    (2020) Phase separation of a PKA regulatory subunit controls cAMP compartmentation and oncogenic signaling

    Cell 182:1531–1544.

    https://doi.org/10.1016/j.cell.2020.07.043

    • PubMed
    • Google Scholar

Author details

1. Jeevan Tewari

   Department of Biomedical Engineering, University of Virginia, Charlottesville, United States

   Contribution

   Conceptualization, Formal analysis, Investigation, Methodology, Writing – original draft, Writing – review and editing

   Contributed equally with

   Benjamin W Dahl

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-6786-3917

2. Benjamin W Dahl

   Department of Biomedical Engineering, University of Virginia, Charlottesville, United States

   Contribution

   Conceptualization, Investigation, Methodology, Writing – review and editing

   Contributed equally with

   Jeevan Tewari

   Competing interests

   No competing interests declared

3. B Adam Bates

   Department of Biomedical Engineering, University of Virginia, Charlottesville, United States

   Contribution

   Investigation, Writing – review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0009-0000-7607-6237

4. Jason A Papin

   Department of Biomedical Engineering, University of Virginia, Charlottesville, United States

   Contribution

   Supervision, Project administration, Writing – review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-2769-5805

5. Jeffrey J Saucerman

   Department of Biomedical Engineering, University of Virginia, Charlottesville, United States

   Contribution

   Conceptualization, Supervision, Project administration, Writing – review and editing

   For correspondence

   jsaucerman@virginia.edu

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-9464-8374

• 496

  views

• 25

  downloads

• 0

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.