Register now After registration you will be able to apply for this opportunity online.
Machine Learning with little data: PCE on agent-based model of osteoporosis and its treatments
Combine two exploding fields in computer science: machine learning and agent-based modelling.
Based on preclinical and in vitro studies of cell behaviour and cytokine reaction-diffusion and mechanical tests we have generated an in-house biofidelic agent-based model of the human skeleton and its response to diseases and their treatments. This model reproduces the effects of several widely used osteoporosis treatments on key parameters used to quantify fracture risk. This rule-based approach involves studying bone mechanobiology at the cell scale and extrapolating this to millions of cells at the tissue scale to understand the pharmacokinetics of treatments and identify possible new therapies and approaches to patient-specific treatment.
An alternative approach to in silico prediction of response to treatment is a supervised learning approach where we simply input baseline and follow-up bone scans to a CNN with twelve layers constructed using keras. We then attempt to dive into the black box and quantify what characteristics of the input govern the response of our model. The issue is the clinical data is not big enough to do this well so we use the agent-based model as input to the ML approach to construct a proxy model! This also helps us understand, validate and quantify the uncertainty in the agent-based model. To decide which runs of the agent-based model to use as input to the ML approach to construct the proxy model we use polynomial chaos expansion.
Osteoporosis is a skeletal condition characterized by decreased density (mass/volume) of normally mineralized bone. It is a major cause of morbidity in older people. _In silico_ simulations aim to provide fast, inexpensive, and ethical alternatives to prolonged and expensive animal and/or human trials for studying the deregulation of bone remodelling during osteoporosis and the effect of therapeutics [1]. Given the multiscale nature of biological systems, the integration of phenomena at different spatial and temporal scales has emerged to be essential in capturing mechanobiological mechanisms underlying bone remodelling processes [2]. In this context, the micro-multiphysics agent based (micro-MPA) models explicitly incorporate the complex cellular and molecular mechanisms causing metabolic bone diseases and the pathways involved in their treatments.
At the ETH Laboratory for Bone Biomechanics, we are developing a Python and C++ based micro-MPA model [3,4] to predict the evolution of bone as a result of different pathological conditions and their prescribed therapies. The micro-MPA model considers eight different types of cells namely osteocytes, osteoclasts, osteoblasts, pre-osteocytes, pre-osteoclasts, pre-osteoblasts, lining cells, and hematopoietic stem cells (HSCs). It also includes signalling molecules RANK-L, OPG, TGF-ß, sclerostin, and estrogen. The micro-MPA model is characterized by approximately 150 input parameters. More than two thirds of these parameters have no basis in literature as they have never been measured clinically while the remaining one third have values in clinical literature that span sometimes several orders of magnitude. Moreover, the parameter settings used vary across research groups, prohibiting the comparison of results. Therefore, there is a strong need to perform the Uncertainty Quantification (UQ) analysis of the proposed model. While simulations provide estimates of real-world values, UQ measures the uncertainty (or error) associated with these estimates [2].
The significance of computational model verification, validation, and uncertainty quantification (VVUQ) has been demonstrated in various biomedical fields such as systems biology [5], in-stent restenosis [6-8], cardiovascular haemodynamics [9], etc. The objective of this project is to analyse the sensitivity and uncertainty of the micro-MPA model outputs with respect to changes in inputs. To do so, the uncertainties in the selected model parameters will be considered. UQ Labs, the framework for UQ [10] at ETH Zurich, will be used to generate the set of input parameter combinations according to the assumed probability distribution and perform sensitivity analysis for the selected model parameters. The Sobol indices will then be obtained from post-processing as a measure of the sensitivity of the modelling outcome corresponding to the individual input parameters. As a result of the UQ study performed in this project, the model will predict the evolution of bone under different pathological conditions/therapies while appropriately considering the importance of all the model parameters and in turn increase the confidence in the predicted simulation results.
Sparse PCE on a database of iliac crest micro-CT scans and the corresponding virtual biopsies at different time points and in different disease and treatment scenarios generated using an agent-based model. The postdoc who started this project left to take a software engineering job at Max Planck. You will pick up where he left off. You will be supervised by Prof. Dr. Ralph Müller, h-index > 100, the most influential researcher in the bone biomechanics field worlwide, the developer of micro-computed tomography to analyse bone morphometrics and by his doctoral student Charles Ledoux, the president of the scientific staff association at ETH. You will also have an interim meeting and a final meeting with Prof. Marelli, one of the world's top experts on uncertainty quantification.
Enhance your programming skills by applying some of the most searched for skills on the job market to a very concrete problem with a visual output representing changes in bone structure over time. Look at how your C++ and python code impacts the changes in the behaviour of the osteoclasts and osteoblasts remodelling the bone surface. Learn standard parallel processing techniques in MPI and openMP and how these impact the weak and strong scaling efficiency of your runs on Piz Daint at the Swiss National Supercomputer, the 6th fastest computer in the world!
Most importantly there is a very motivating aspect to doing medical research. Of all people 50 and above, one third of women and one fifth of men will suffer a fragility fracture in their remaining lifetime. The Physiome Project and the Virtual Physiological Human project and other large-scale efforts to model the human body are recognized as a paradigm shift in personalized medicine that could put a significant dent in the tens of millions of fragility fracture that occur every year and the associated horrific and unacceptable reduction in quality of life.
[1] Ledoux, C., Boaretti, D., Sachan, A., Müller, R. and Collins, C.J., (2022). Clinical Data for Parametrization of In Silico Bone Models Incorporating Cell-Cytokine Dynamics: A Systematic Review of Literature. Frontiers in bioengineering and biotechnology, 10.
[2] Hambli, R., Katerchi, H., Benhamou, C.L., (2011) Multiscale methodology for bone remodelling simulation using coupled fnite element and neural network computation. Biomech Model Mechan 10:133–145. doi: https://doi.org/10.1007/s10237-010-0222-xhttps://doi.org/10.1016/j.bone.2019.07.024
[3] Tourolle, D. (2019). A Micro-scale Multiphysics Framework for Fracture Healing and Bone Remodelling ETH Library. ETH Zurich Res. Collect. doi:10.3929/ethzb-000364637
[4] Tourolle, D., Dempster, D. W., Ledoux, C., Boaretti, D., Aguilera, M., Saleem, N., et al. (2021). Ten-Year Simulation of the Effects of Denosumab on Bone Remodeling in Human Biopsies. JBMR plus 5, 5. doi:10.1002/JBM4.10494
[5] Marino, S., Hogue, I. B., Ray, C. J., and Kirschner, D. E. (2008). A Methodology for Performing Global Uncertainty and Sensitivity Analysis in Systems Biology. J. Theor. Biol. 254, 178–196. doi:10.1016/j.jtbi.2008.04.011
[6] Nikishova, A., Veen, L., Zun, P., and Hoekstra, A. G. (2018). Uncertainty Quantification of a Multiscale Model for In-Stent Restenosis. Cardiovasc. Eng. Tech. 9, 761–774. doi:10.1007/s13239-018-00372-4
[7] Nikishova, A., Veen, L., Zun, P., and Hoekstra, A. G. (2019). Semi-intrusive Multiscale Metamodelling Uncertainty Quantification with Application to a Model of In-Stent Restenosis. Phil. Trans. R. Soc. A. 377, 20180154. doi:10.1098/rsta.2018.0154
[8] Ye, D., Nikishova, A., Veen, L., Zun, P., and Hoekstra, A. G. (2021a). Non-intrusive and Semi-intrusive Uncertainty Quantification of aMultiscale In-Stent Restenosis Model. Reliability Eng. Syst. Saf. 214, 107734. doi:10.1016/j.ress.2021.107734
[9] Fleeter, C. M., Geraci, G., Schiavazzi, D. E., Kahn, A.M., and Marsden, A. L. (2020). Multilevel and Multifidelity Uncertainty Quantification for Cardiovascular Hemodynamics. Computer Methods Appl. Mech. Eng. 365, 113030. doi:10.1016/j.cma.2020.113030
[10] Marelli S. and Sudret, B., (2014) UQLab: A Framework for Uncertainty Quantification in MATLAB, In The 2nd International Conference on Vulnerability and Risk Analysis and Management (ICVRAM 2014), University of Liverpool, United Kingdom, July 13-16, pp. 2554–2563.
Osteoporosis is a skeletal condition characterized by decreased density (mass/volume) of normally mineralized bone. It is a major cause of morbidity in older people. _In silico_ simulations aim to provide fast, inexpensive, and ethical alternatives to prolonged and expensive animal and/or human trials for studying the deregulation of bone remodelling during osteoporosis and the effect of therapeutics [1]. Given the multiscale nature of biological systems, the integration of phenomena at different spatial and temporal scales has emerged to be essential in capturing mechanobiological mechanisms underlying bone remodelling processes [2]. In this context, the micro-multiphysics agent based (micro-MPA) models explicitly incorporate the complex cellular and molecular mechanisms causing metabolic bone diseases and the pathways involved in their treatments.
At the ETH Laboratory for Bone Biomechanics, we are developing a Python and C++ based micro-MPA model [3,4] to predict the evolution of bone as a result of different pathological conditions and their prescribed therapies. The micro-MPA model considers eight different types of cells namely osteocytes, osteoclasts, osteoblasts, pre-osteocytes, pre-osteoclasts, pre-osteoblasts, lining cells, and hematopoietic stem cells (HSCs). It also includes signalling molecules RANK-L, OPG, TGF-ß, sclerostin, and estrogen. The micro-MPA model is characterized by approximately 150 input parameters. More than two thirds of these parameters have no basis in literature as they have never been measured clinically while the remaining one third have values in clinical literature that span sometimes several orders of magnitude. Moreover, the parameter settings used vary across research groups, prohibiting the comparison of results. Therefore, there is a strong need to perform the Uncertainty Quantification (UQ) analysis of the proposed model. While simulations provide estimates of real-world values, UQ measures the uncertainty (or error) associated with these estimates [2].
The significance of computational model verification, validation, and uncertainty quantification (VVUQ) has been demonstrated in various biomedical fields such as systems biology [5], in-stent restenosis [6-8], cardiovascular haemodynamics [9], etc. The objective of this project is to analyse the sensitivity and uncertainty of the micro-MPA model outputs with respect to changes in inputs. To do so, the uncertainties in the selected model parameters will be considered. UQ Labs, the framework for UQ [10] at ETH Zurich, will be used to generate the set of input parameter combinations according to the assumed probability distribution and perform sensitivity analysis for the selected model parameters. The Sobol indices will then be obtained from post-processing as a measure of the sensitivity of the modelling outcome corresponding to the individual input parameters. As a result of the UQ study performed in this project, the model will predict the evolution of bone under different pathological conditions/therapies while appropriately considering the importance of all the model parameters and in turn increase the confidence in the predicted simulation results.
Sparse PCE on a database of iliac crest micro-CT scans and the corresponding virtual biopsies at different time points and in different disease and treatment scenarios generated using an agent-based model. The postdoc who started this project left to take a software engineering job at Max Planck. You will pick up where he left off. You will be supervised by Prof. Dr. Ralph Müller, h-index > 100, the most influential researcher in the bone biomechanics field worlwide, the developer of micro-computed tomography to analyse bone morphometrics and by his doctoral student Charles Ledoux, the president of the scientific staff association at ETH. You will also have an interim meeting and a final meeting with Prof. Marelli, one of the world's top experts on uncertainty quantification.
Enhance your programming skills by applying some of the most searched for skills on the job market to a very concrete problem with a visual output representing changes in bone structure over time. Look at how your C++ and python code impacts the changes in the behaviour of the osteoclasts and osteoblasts remodelling the bone surface. Learn standard parallel processing techniques in MPI and openMP and how these impact the weak and strong scaling efficiency of your runs on Piz Daint at the Swiss National Supercomputer, the 6th fastest computer in the world!
Most importantly there is a very motivating aspect to doing medical research. Of all people 50 and above, one third of women and one fifth of men will suffer a fragility fracture in their remaining lifetime. The Physiome Project and the Virtual Physiological Human project and other large-scale efforts to model the human body are recognized as a paradigm shift in personalized medicine that could put a significant dent in the tens of millions of fragility fracture that occur every year and the associated horrific and unacceptable reduction in quality of life.
[1] Ledoux, C., Boaretti, D., Sachan, A., Müller, R. and Collins, C.J., (2022). Clinical Data for Parametrization of In Silico Bone Models Incorporating Cell-Cytokine Dynamics: A Systematic Review of Literature. Frontiers in bioengineering and biotechnology, 10. [2] Hambli, R., Katerchi, H., Benhamou, C.L., (2011) Multiscale methodology for bone remodelling simulation using coupled fnite element and neural network computation. Biomech Model Mechan 10:133–145. doi: https://doi.org/10.1007/s10237-010-0222-xhttps://doi.org/10.1016/j.bone.2019.07.024 [3] Tourolle, D. (2019). A Micro-scale Multiphysics Framework for Fracture Healing and Bone Remodelling ETH Library. ETH Zurich Res. Collect. doi:10.3929/ethzb-000364637 [4] Tourolle, D., Dempster, D. W., Ledoux, C., Boaretti, D., Aguilera, M., Saleem, N., et al. (2021). Ten-Year Simulation of the Effects of Denosumab on Bone Remodeling in Human Biopsies. JBMR plus 5, 5. doi:10.1002/JBM4.10494 [5] Marino, S., Hogue, I. B., Ray, C. J., and Kirschner, D. E. (2008). A Methodology for Performing Global Uncertainty and Sensitivity Analysis in Systems Biology. J. Theor. Biol. 254, 178–196. doi:10.1016/j.jtbi.2008.04.011 [6] Nikishova, A., Veen, L., Zun, P., and Hoekstra, A. G. (2018). Uncertainty Quantification of a Multiscale Model for In-Stent Restenosis. Cardiovasc. Eng. Tech. 9, 761–774. doi:10.1007/s13239-018-00372-4 [7] Nikishova, A., Veen, L., Zun, P., and Hoekstra, A. G. (2019). Semi-intrusive Multiscale Metamodelling Uncertainty Quantification with Application to a Model of In-Stent Restenosis. Phil. Trans. R. Soc. A. 377, 20180154. doi:10.1098/rsta.2018.0154 [8] Ye, D., Nikishova, A., Veen, L., Zun, P., and Hoekstra, A. G. (2021a). Non-intrusive and Semi-intrusive Uncertainty Quantification of aMultiscale In-Stent Restenosis Model. Reliability Eng. Syst. Saf. 214, 107734. doi:10.1016/j.ress.2021.107734 [9] Fleeter, C. M., Geraci, G., Schiavazzi, D. E., Kahn, A.M., and Marsden, A. L. (2020). Multilevel and Multifidelity Uncertainty Quantification for Cardiovascular Hemodynamics. Computer Methods Appl. Mech. Eng. 365, 113030. doi:10.1016/j.cma.2020.113030 [10] Marelli S. and Sudret, B., (2014) UQLab: A Framework for Uncertainty Quantification in MATLAB, In The 2nd International Conference on Vulnerability and Risk Analysis and Management (ICVRAM 2014), University of Liverpool, United Kingdom, July 13-16, pp. 2554–2563.
• Review the literature on the micro-MPA model for osteoporosis and UQ implementation in the field of biomedicine
• Explore our micro-MPA model for osteoporosis to gain insight into the python and C++ implementation
• Assist in creating the list of input and output parameters of micro-MPA model as required by the UQLab partners
• Perform micro-MPA simulations with input sets determined using the latin hypercube sampling method by UQLab partners
• Format the simulation output into the excel sheet required for Bayesian inversion analysis
• Discuss the output of Bayesian inversion analysis performed by the UQLab partners
• Present the work performed to the entire LBB (20-minute presentation)
• Submit a final project report prepared as per the ETH template
• Review the literature on the micro-MPA model for osteoporosis and UQ implementation in the field of biomedicine • Explore our micro-MPA model for osteoporosis to gain insight into the python and C++ implementation • Assist in creating the list of input and output parameters of micro-MPA model as required by the UQLab partners • Perform micro-MPA simulations with input sets determined using the latin hypercube sampling method by UQLab partners • Format the simulation output into the excel sheet required for Bayesian inversion analysis • Discuss the output of Bayesian inversion analysis performed by the UQLab partners • Present the work performed to the entire LBB (20-minute presentation) • Submit a final project report prepared as per the ETH template
Charles Ledoux | MEng Chemical Eng.
PhD Candidate Prof. Dr. Ralph Müller | www.bone.ethz.ch
Institute for Biomechanics (IfB) ETH Zurich
GLC H 29 | Gloriastrasse 37 | CH-8006 Zürich
Charles Ledoux | MEng Chemical Eng. PhD Candidate Prof. Dr. Ralph Müller | www.bone.ethz.ch Institute for Biomechanics (IfB) ETH Zurich GLC H 29 | Gloriastrasse 37 | CH-8006 Zürich