Patient-specific Boolean models of signalling networks guide personalised treatments

  1. Arnau Montagud
  2. Jonas Béal
  3. Luis Tobalina
  4. Pauline Traynard
  5. Vigneshwari Subramanian
  6. Bence Szalai
  7. Róbert Alföldi
  8. László Puskás
  9. Alfonso Valencia
  10. Emmanuel Barillot
  11. Julio Saez-Rodriguez
  12. Laurence Calzone

  • Curated by eLife

    eLife logo

    Evaluation Summary:

    This paper presents a mathematical model for prioritizing drugs for prostate cancer patients based on signal network database. The manuscript is of broad interest to the field of oncology and precision medicine. The methodology developed is sophisticated and relevant to real patient prostate cancer data. The predictions from the model are validated in an experimental setting and provide suggestions for the personalisation of prostate cancer treatment. The study can serve as a roadmap for future development of predictive, personalized models.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #2 agreed to share their name with the authors.)

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Prostate cancer is the second most occurring cancer in men worldwide. To better understand the mechanisms of tumorigenesis and possible treatment responses, we developed a mathematical model of prostate cancer which considers the major signalling pathways known to be deregulated. We personalised this Boolean model to molecular data to reflect the heterogeneity and specific response to perturbations of cancer patients. A total of 488 prostate samples were used to build patient-specific models and compared to available clinical data. Additionally, eight prostate cell line-specific models were built to validate our approach with dose-response data of several drugs. The effects of single and combined drugs were tested in these models under different growth conditions. We identified 15 actionable points of interventions in one cell line-specific model whose inactivation hinders tumorigenesis. To validate these results, we tested nine small molecule inhibitors of five of those putative targets and found a dose-dependent effect on four of them, notably those targeting HSP90 and PI3K. These results highlight the predictive power of our personalised Boolean models and illustrate how they can be used for precision oncology.

Article activity feed

  1. Version published to 10.7554/elife.72626 on eLife
  2. eLife

    Author Response

    Reviewer #1 (Public Review):

    Summary:

    In this work, the authors develop a tool for personalising prostate cancer treatment using a Boolean model. The model is extremely complex and describes the regulation of invasion, migration, cell cycle, apoptosis, androgen and growth factors signalling in prostate cancer using 133 nodes (genes and our metrics) and 449 edges (regulation pathways. Using their model, they were able to grade the effect of combined treatments for each of the 488 patients for already-developed drugs and find several genes suitable for intervention in most of the 488 patients. The predications from their model could help develop a patient-tailored treatment that could boost success of pancreatic cancer treatments in clinical practice.

    Strengths:

    The authors clearly achieved their aims of predicative prostate cancer modelling and have added value to the field of prostate cancer personalisation.

    Calibrating and then validating predications of a model, as this work does, is a fundamental part of systems biology and mathematical modelling. By using a cell line to investigate predictions that AKT is the top hit for prostate cancer, validates the utility of their model and also shines a light on how useful models like this can be in oncology. The methodology in this paper provides a guide for future modelling work in this area.

    Providing a detailed Supplementary Information and additional links to the code and fundamental modelling platform publications, helps to provide readers with a tool that may be applied in other settings. However, while this is a strength of the publication, the model is extremely complex and relies heavily on readers spending time comprehending pre-published work and doesn't provide a single contained body of work.

    The methodology they are presenting could have significant impact on the field of cancer treatment, but would need to be testing clinically to validate that personalising treatment in this manner does improve outcomes.

    We thank the reviewer for these comments.

    Weaknesses:

    While it is a strength of this work that such a detailed, and complex model is developed for prostate cancer, and that the code is provided, the weakness of this work is that the model is not easily accessible, and a lot of the techniques used in model development feel brushed over. The work relies heavily on other works and does not provide detailed descriptions of the underlying algorithm, requiring readers to absorb knowledge from our places. This could be a challenge if an experimentalist wishing to implement this methodology in a different cancer treatment.

    We have summarised the main techniques on which this work relies upon in a dedicated section in the Supplementary Material (Appendix file) by describing small introductions to Boolean modelling, MaBoSS stochastic approach to Boolean models, and PROFILE methods.

    We have also provided the codes to reproduce the figures and the analyses. We tried to comment on the code files (e.g., Jupyter notebook) as much as possible to facilitate their use in different contexts.

    The protein/genes in the model are not presented in a way that it can be easily validated as such, the complexity of such a Boolean model comes into question.

    We have listed all the proteins/genes of the model in SuppFile 1 with references for all the interactions of the network.

    For transparency, we have also described in the Appendix how we used information from all the different sources to construct the model in the section "Prior knowledge network construction".

    How sure are we in the model predications and are there are any potential weaknesses to modelling the network in such an extensive manner? For such a model like this, it is crucial to demonstrate its sensitivity to initial conditions and node additions/removals so some work could be done to demonstrate this so that the readers have an idea of how many over/under predications there might be in the model.

    For the sensitivity to initial conditions, we have tested some of them on the generic model in the Jupyter notebook (provided in the supplementary files) but have not done it systematically. The table of all the stable states can be computed exactly as it is done in the notebook (2460 fixpoints are found), and the simulations of MaBoSS clearly show that the proportions of some solutions (probability of model states) change depending on which input is ON. We have tested some conditions: all inputs random, all inputs at 0, growth factors ON (EGF, FGF, Androgen, and Nutrients ON), death signals ON (Carcinogen, Androgen, Acidosis, Hypoxia and TNFalpha ON) leading to very different outputs (Figure 3 for LNCaP and S22 for all 8 prostate cell lines). In fact, the MaBoSS simulation with all inputs random shows the existence of all possible, stable states as it explores the whole state transition graph: for all nodes, 50% of the trajectories will start at 0, and 50% will start at 1. Similarly, we tested the effect of some mutations on the generic model (e.g. mutation of p53, which reduces the probability to reach apoptosis). The aim of these simulations was to test the overall coherence in the model behaviour vs biological evidence as a first validation.

    As for automatic removal and addition of new nodes to assess the importance of each of them, we would recommend against it. Indeed, the model was built from the knowledge extracted from the literature, from databases (cf. Omnipath), discussions with experts, and results from data analyses. Removing nodes would mean that some nodes are considered less important, and adding new nodes would mean that some new findings were found that would justify a new addition.

    In addition, in this work, we need to balance the robustness of the model with the flexibility of being used to cover the different cell line personalisations. Thus, we do not want a highly robust wild type model that has extremely robust, few stable states but is unable to capture the different cell lines specifics. Nevertheless, we have partially covered this with our "High-Throughput mutant analysis of the LNCaP model" section in Appendix file (Section 6.1), where we study all the perturbations on one node and combinations of two nodes, let them be knock-outs (where a node is forced to be 0 throughout the simulation) or overexpression (forced to be 1). By using this analysis, we wanted to identify the fragility points of the mutants' models, but we did not perform this test to have a thorough robustness analysis. In any case, we found varying effects of these perturbations on the phenotype scores, and double perturbations having a greater effect than single ones.

    Finally, we have performed a perturbation on the stability of the logical rules. We have changed one and two logical gates from each logical rule of the LNCaP model and studied the effects on the phenotype scores. In short, we have changed an AND in OR and vice versa in each logical rule (level 1 with 372 simulations) or twice in the same rule (level 2 with 1263 simulations).

    Overall, we see that all of the most probable phenotypes are very robust to this kind of perturbation. Even the less stable phenotype, Invasion-Migration-Proliferation, only has ~3% of either level 1 or 2 perturbations that reduce this phenotype's probability to zero (Appendix File, Figure S30). Most of these perturbations were focused on HIF1, AR_ERG and p53 nodes (Appendix File, Figure S31).

    We added a sentence in the Methods section to explain this: "In addition, we found that the LNCaP model is very robust against perturbations of its logical rules, by systematically changing an AND for an OR gate or vice versa in all of its logical rules (Appendix File, Section 6.2, Figure S30 and S31)." and added Section 6.2 to the Appendix file titled "Robustness analysis of the logical model".

    As they test so many drugs and combination regimes it is also hard to extract information about which key drugs should be repurposed. It could be useful to the readers to have this spotlighted more in the model so that it is easily discernable.

    The complete study on the inhibition of all nodes of the LNCaP model can be found in the supplemental information (SuppFile 6 and Appendix file, Section "High-Throughput mutant analysis of the LNCaP model").

    Because of the size of the model, we chose to filter the full list of nodes with the list of existing drugs and their targets. Thus, Table 1 gathers the drugs we discuss in this article along with the node that they target. We also studied a selection of combinations of drugs, as depicted in Section "Experimental validation of drugs in LNCaP" of Results. In that section, we focused on the combinations that reduced Proliferation and/or increased Apoptosis. For completeness sake, we provide all the combinations of all the drugs from Table 1 in Appendix File, Figures S34 and S35, and their Bliss score in Appendix File, Figures S36 and S37. Furthermore, the code to reproduce these in our GitHub repository: https://github.com/ArnauMontagud/PROFILE_v2/blob/main/Gradient%20inhibition%20of%20nodes/data_analysis.R

    We could have identified the nodes from Table 1 on the figure of the network (main text Figure 1), but we decided against it because the figure is already hard to read, and colours were added to specify the signalling pathways that are included.

    Suggestions:

    Another way to validate the cohort level predications could have been to examine the efficacy of the predicted personalised protocols, or sensitive parts of the Boolean network, in a new prostate cancer patient cohort. Do we see the same sensitive pathways if we examine a different cohort of prostate cancer patients?

    We thank the reviewer for this suggestion. Indeed we are working on using this pipeline in other cancers and in other studies.

    One of the topics that we think can facilitate the use of this methodology is on optimising its runtime and portability. Thus, we are currently working on having a containerised, HPC semi-automatic workflow to reduce the time and optimise the efforts to get results using (almost) any published model and (almost) any omics data.

    In terms of the reproducibility of the results and as we say in the discussion section of the main text, there is a kind of effect size on this type of study. You may find that for a specific patient, their conclusions are not in line with what is expected, but when you analyse at the level of groups of patients, these outliers dampen off.

    Reviewer #2 (Public Review):

    Montagud et al. present a very successful experiment - modeling feedback loop: the authors develop a Boolean model of the major signaling pathways deregulated in prostate cancer, use molecular data from patient samples to personalise this model, use drug response of cell lines to validate the model, predict 15 actionable interventions based on the model, and test nine of these interventions, confirming four.

    The premise of the work is well-supported by prior work by the team and the wider community. The methods are sound, well integrated and thoroughly documented, with one notable omission. The process through which the logic functions of the nodes were determined/decided is not described. The Appendix file indicates "The model is completed by logical rules (or functions), which assign a target value to each node for each regulator level combination.". The interested reader would want to know what information is used and what considerations are the basis of these assignments, and what would change if an assignment were different.

    The manuscript makes a number of testable predictions of actionable single and combinatorial therapeutic interventions for prostate cancer. Equally important, the combination of information and methodologies used in this paper offers a roadmap for future development of predictive and personalised models. Such models are much needed in precision oncology.

    We thank the reviewer for these encouraging comments.

    Reviewer #3 (Public Review):

    This paper tries to establish a model for drug (and combination) selection for individual prostate cancer patient based on a prior signal network knowledge base and genomic/transcriptomic profiling data. This is of great clinical potential. However, whether this approach could be robustly applied in clinic is not validated. Limited validation using cell line is provided. Most tumors have complex structure including tumor cells and surrounding microenvironment. The model is mainly built from onco-signaling pathways. The contribution of microenvironment including immunity is unclear.

    The focus of this model is intracellular only. We explored the interplay between signaling pathways that may be linked to tumorigenesis. We only consider the microenvironment effect as indirect and in no way comprehensive. For instance, we have not considered any immune cells or the effect of the metabolism.

    Nevertheless, we are building on top of this work a multiscale model where we can include different cell types, such as immune cells, and drug-related pharmacodynamics.

  3. eLife

    Evaluation Summary:

    This paper presents a mathematical model for prioritizing drugs for prostate cancer patients based on signal network database. The manuscript is of broad interest to the field of oncology and precision medicine. The methodology developed is sophisticated and relevant to real patient prostate cancer data. The predictions from the model are validated in an experimental setting and provide suggestions for the personalisation of prostate cancer treatment. The study can serve as a roadmap for future development of predictive, personalized models.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #2 agreed to share their name with the authors.)

  4. eLife

    Reviewer #1 (Public Review):

    Summary:

    In this work, the authors develop a tool for personalising prostate cancer treatment using a Boolean model. The model is extremely complex and describes the regulation of invasion, igration, cell cycle, apoptosis, androgen and growth factors signalling in prostate cancer using 133 nodes (genes and our metrics) and 449 edges (regulation pathways. Using their model, they were able to grade the effect of combined treatments for each of the 488 patients for already-developed drugs and find several genes suitable for intervention in most of the 488 patients. The predications from their model could help develop a patient-tailored treatment that could boost success of pancreatic cancer treatments in clinical practice.

    Strengths:

    The authors clearly achieved their aims of predicative prostate cancer modelling and have added value to the field of prostate cancer personalisation.
    Calibrating and then validating predications of a model, as this work does, is a fundamental part of systems biology and mathematical modelling. By using a cell line to investigate predictions that AKT is the top hit for prostate cancer, validates the utility of their model and also shines a light on how useful models like this can be in oncology. The methodology in this paper provides a guide for future modelling work in this area.
    Providing a detailed Supplementary Information and additional links to the code and fundamental modelling platform publications, helps to provide readers with a tool that may be applied in other settings. However, while this is a strength of the publication, the model is extremely complex and relies heavily on readers spending time comprehending pre-published work and doesn't provide a single contained body of work.
    The methodology they are presenting could have significant impact on the field of cancer treatment, but would need to be testing clinically to validate that personalising treatment in this manner does improve outcomes.

    Weaknesses:

    While it is a strength of this work that such a detailed, and complex model is developed for prostate cancer, and that the code is provided, the weakness of this work is that the model is not easily accessible, and a lot of the techniques used in model development feel brushed over. The work relies heavily on other works and does not provide detailed descriptions of the underlying algorithm, requiring readers to absorb knowledge from our places This could be a challenge if an experimentalist wishing to implement this methodology in a different cancer treatment.
    The protein/genes in the model are not presented in a way that it can be easily validated as such, the complexity of such a Boolean model comes into question. How sure are we in the model predications and are there are any potential weaknesses to modelling the network in such an extensive manner? For such a model like this, it is crucial to demonstrate its sensitivity to initial conditions and node additions/removals so some work could be done to demonstrate this so that the readers have an idea of how many over/under predications there might be in the model.
    As they test so many drugs and combination regimes it is also hard to extract information about which key drugs should be repurposed. It could be useful to the readers to have this spotlighted more in the model so that it is easily discernable.

    Suggestions:

    Another way to validate the cohort level predications could have been to examine the efficacy of the predicted personalised protocols, or sensitive parts of the Boolean network, in a new prostate cancer patient cohort. Do we see the same sensitive pathways if we examine a different cohort of prostate cancer patients?

  5. eLife

    Reviewer #2 (Public Review):

    Montagud et al. present a very successful experiment - modeling feedback loop: the authors develop a Boolean model of the major signaling pathways deregulated in prostate cancer, use molecular data from patient samples to personalize this model, use drug response of cell lines to validate the model, predict 15 actionable interventions based on the model, and test nine of these interventions, confirming four.

    The premise of the work is well-supported by prior work by the team and the wider community. The methods are sound, well integrated and thoroughly documented, with one notable omission. The process through which the logic functions of the nodes were determined/decided is not described. The Appendix file indicates "The model is completed by logical rules (or functions), which assign a target value to each node for each regulator level combination.". The interested reader would want to know what information is used and what considerations are the basis of these assignments, and what would change if an assignment were different.

    The manuscript makes a number of testable predictions of actionable single and combinatorial therapeutic interventions for prostate cancer. Equally important, the combination of information and methodologies used in this paper offers a roadmap for future development of predictive and personalized models. Such models are much needed in precision oncology.

  6. eLife

    Reviewer #3 (Public Review):

    This paper tries to establish a model for drug (and combination) selection for individual prostate cancer patient based on a prior signal network knowledge base and genomic/transcriptomic profiling data. This is of great clinical potential. However, whether this approach could be robustly applied in clinic is not validated. Limited validation using cell line is provided. Most tumors have complex structure including tumor cells and surrounding microenvironment. The model is mainly built from onco-signaling pathways. The contribution of microenvironment including immunity is unclear.

  7. Version published to 10.1101/2021.07.28.454126v1 on bioRxiv