Periodic Reporting for period 5 - MicroC (Agent-Based Modelling of Gene Networks to model clonal selection in the tumour microenvironment and predict therapeutic resistance)
Berichtszeitraum: 2023-08-01 bis 2024-11-30
The occurrence of therapeutic resistance is a major cause for the small effect on overall survival showed by targeted cancer therapies. Whilst experimental strategies to evaluate available treatments have been faced by an ever increasing number of possible combinations, computational approaches have been challenged by the lack of a framework able to model the multiple interactions encompassed by the three major factors affecting therapeutic resistance: selection of resistant clones, adaptability of gene signalling networks, and a protective and hypoxic tumour microenvironment.
Here I propose a novel modelling framework, Agent-Based Modelling of Gene Networks, which brings together powerful computational modelling techniques and gene networks. This combination allows biological hypotheses to be tested in a controlled stepwise fashion, and it lends itself naturally to model a heterogeneous population of cells acting and evolving in a dynamic microenvironment, which is needed to predict therapeutic resistance and guide effective treatment selection.
Using triple negative breast cancer (TNBC) as our initial testing case for this framework (15% of breast cancers, lacks validated), I propose to:
1. Develop a computational model of the TNBC tumour microenvironment using in-vitro and in-vivo, including patient-derived, models and data from clinical samples.
2. Validate the ability of the model to predict driver genes conferring a survival advantage to cancer cells under different microenvironmental conditions, focusing on those found in TNBC.
3. Predict combinations of druggable targets to tackle TNBC therapeutic resistance.
4. Select most effective drug combinations and proceed to pre-clinical validation in TNBC models.
Here I propose a novel modelling framework, Agent-Based Modelling of Gene Networks, which brings together powerful computational modelling techniques and gene networks. This combination allows biological hypotheses to be tested in a controlled stepwise fashion, and it lends itself naturally to model a heterogeneous population of cells acting and evolving in a dynamic microenvironment, which is needed to predict therapeutic resistance and guide effective treatment selection.
Using triple negative breast cancer (TNBC) as our initial testing case for this framework (15% of breast cancers, lacks validated), I propose to:
1. Develop a computational model of the TNBC tumour microenvironment using in-vitro and in-vivo, including patient-derived, models and data from clinical samples.
2. Validate the ability of the model to predict driver genes conferring a survival advantage to cancer cells under different microenvironmental conditions, focusing on those found in TNBC.
3. Predict combinations of druggable targets to tackle TNBC therapeutic resistance.
4. Select most effective drug combinations and proceed to pre-clinical validation in TNBC models.
We developed a new computer-based framework called Agent-Based Modelling of Gene Networks (ABM-GN), which allows us to simulate complex biological systems in 3D. This includes how cells interact with each other, respond to genetic changes (like mutations), and are influenced by external factors such as drug treatments or surrounding cells. The system also reflects the dynamic and varied nature of the tumour environment (Voukantsis et al., GigaScience, 2019).
Using ABM-GN, we created a virtual model of TNBC, our main case study. Simulations using this model showed that how cancer cells behave in their environment is closely tied to their genetic makeup. These findings highlight the importance of using models that reflect both the genetic and environmental complexity of cancer and support the idea of tailoring treatment to the unique features of each tumour (Jayathilake PG et al., PLoS Comput Biol., 2024). We then focused on resolving in detail one key feature of TNBC: a microenvironment characterized by low oxygen levels (hypoxia). To understand how TNBC adapt to hypoxia, we grew TNBC cells under hypoxic conditions for different periods (from 1 to 14 days) and collected time-series data under different nutrient levels. This was done across four cell lines: three TNBC subtypes (MDA-MB-468, HCC1806, MDA-MB-231) and the ER-positive MCF7 line, as a control. We also used ATAC-seq, single-cell RNA-seq, and spatial transcriptomics to study how cells adapted at both molecular and regulatory levels. Interestingly, we saw that adaptation to chronic hypoxia differed across cell types: MCF7 maintained gene expression patterns associated with acute hypoxia, indicating a slower or absent adaptive response; HCC1806 and MDA-MB-468 showed enrichment of pathways related to junction formation, cell-matrix adhesion, and extracellular matrix organization; MDA-MB-231 exhibited upregulation of carbohydrate catabolic pathways, suggesting metabolic adaptation. To explore this further, we repeated the hypoxia time-course in MDA-MB-231 cells at single-cell resolution, revealing a wide range of individual responses within the same cell line. To find genes that allow cancer cells survive hypoxia, we ran whole-genome CRISPR/Cas9 screens across these three TNBC cell lines. These experiments confirmed key pathways related to survival, but also revealed substantial differences between the cell lines. We validated many of these findings using clinical data and are now developing a specialized CRISPR library focused on hypoxia-related genes.
To optimize our virtual cells inner regulatory networks, which determine how the virtual cancer cells respond to different conditions, we generated and implemented new tools which can be used for broader application. These included: BioSWITCH, which turns static gene networks into dynamic models that can simulate how cells behave (Pavillet et al, https://doi.org/10.1101/2020.05.29.122200)(öffnet in neuem Fenster); a new evolutionary algorithm that helps reconstruct how genes regulate each other based on time-series data (Stranieri et al., doi: 10.1109/CIBCB58642.2024.10702181); INGRES, a tool that uses real RNA data from patient samples to personalize simulations and predict drug responses (Victori et al, https://doi.org/10.1101/2022.09.04.506528(öffnet in neuem Fenster)). Additionally, to make it easier to compare lab findings with clinical data, we built a software and protocol that maps gene signatures from experiments onto real patient datasets (Dhawan et al., Nat Protoc., 2019). Originally designed with validation of CRISPR screens in mind, it’s now widely used for RNA-seq, single-cell, and spatial data, including large meta-analyses (e.g. Di Giovannantonio et al., Cell Genom., 2025). As the project progressed, we also recognized the need to standardize the use of AI in biomedical research. This led to the development of RENOIR, a protocol designed to ensure robust and reproducible AI-based learning in the life sciences (Barberis et al., Sci Rep, 2024).
Finally, with the developed TNBC virtual model, we ran in-silico drug combination screens and identified promising treatment strategies. These are now being tested in collaboration with researchers at IFOM in Milan, where F.M. Buffa leads the AI and Systems Biology Lab. We are also partnering with computing experts at the Department of Computing Sciences, Bocconi University, Milan, where F.M. Buffa is Full Professor, to explore how ABM-GN can be combined with generative AI models to advance personalized medicine.
Using ABM-GN, we created a virtual model of TNBC, our main case study. Simulations using this model showed that how cancer cells behave in their environment is closely tied to their genetic makeup. These findings highlight the importance of using models that reflect both the genetic and environmental complexity of cancer and support the idea of tailoring treatment to the unique features of each tumour (Jayathilake PG et al., PLoS Comput Biol., 2024). We then focused on resolving in detail one key feature of TNBC: a microenvironment characterized by low oxygen levels (hypoxia). To understand how TNBC adapt to hypoxia, we grew TNBC cells under hypoxic conditions for different periods (from 1 to 14 days) and collected time-series data under different nutrient levels. This was done across four cell lines: three TNBC subtypes (MDA-MB-468, HCC1806, MDA-MB-231) and the ER-positive MCF7 line, as a control. We also used ATAC-seq, single-cell RNA-seq, and spatial transcriptomics to study how cells adapted at both molecular and regulatory levels. Interestingly, we saw that adaptation to chronic hypoxia differed across cell types: MCF7 maintained gene expression patterns associated with acute hypoxia, indicating a slower or absent adaptive response; HCC1806 and MDA-MB-468 showed enrichment of pathways related to junction formation, cell-matrix adhesion, and extracellular matrix organization; MDA-MB-231 exhibited upregulation of carbohydrate catabolic pathways, suggesting metabolic adaptation. To explore this further, we repeated the hypoxia time-course in MDA-MB-231 cells at single-cell resolution, revealing a wide range of individual responses within the same cell line. To find genes that allow cancer cells survive hypoxia, we ran whole-genome CRISPR/Cas9 screens across these three TNBC cell lines. These experiments confirmed key pathways related to survival, but also revealed substantial differences between the cell lines. We validated many of these findings using clinical data and are now developing a specialized CRISPR library focused on hypoxia-related genes.
To optimize our virtual cells inner regulatory networks, which determine how the virtual cancer cells respond to different conditions, we generated and implemented new tools which can be used for broader application. These included: BioSWITCH, which turns static gene networks into dynamic models that can simulate how cells behave (Pavillet et al, https://doi.org/10.1101/2020.05.29.122200)(öffnet in neuem Fenster); a new evolutionary algorithm that helps reconstruct how genes regulate each other based on time-series data (Stranieri et al., doi: 10.1109/CIBCB58642.2024.10702181); INGRES, a tool that uses real RNA data from patient samples to personalize simulations and predict drug responses (Victori et al, https://doi.org/10.1101/2022.09.04.506528(öffnet in neuem Fenster)). Additionally, to make it easier to compare lab findings with clinical data, we built a software and protocol that maps gene signatures from experiments onto real patient datasets (Dhawan et al., Nat Protoc., 2019). Originally designed with validation of CRISPR screens in mind, it’s now widely used for RNA-seq, single-cell, and spatial data, including large meta-analyses (e.g. Di Giovannantonio et al., Cell Genom., 2025). As the project progressed, we also recognized the need to standardize the use of AI in biomedical research. This led to the development of RENOIR, a protocol designed to ensure robust and reproducible AI-based learning in the life sciences (Barberis et al., Sci Rep, 2024).
Finally, with the developed TNBC virtual model, we ran in-silico drug combination screens and identified promising treatment strategies. These are now being tested in collaboration with researchers at IFOM in Milan, where F.M. Buffa leads the AI and Systems Biology Lab. We are also partnering with computing experts at the Department of Computing Sciences, Bocconi University, Milan, where F.M. Buffa is Full Professor, to explore how ABM-GN can be combined with generative AI models to advance personalized medicine.
The microC project has led to the discovery of new potential drug targets and treatment combinations for triple-negative breast cancer (TNBC), which are now being tested in follow-up studies. Even more significantly, it has developed a powerful new computer-based modelling platform, ABM-GN, that can simulate how cancer behaves and adapts within its environment and under different treatments. While initially applied to TNBC, ABM-GN is designed to be broadly applicable, offering a versatile tool to study many types of cancer and complex diseases. It opens a path to new possibilities for understanding disease progression and supporting the development of more precise, personalized therapies.