Mots-Clés
artificial intelligence
inductive logic programming
future health
radiobiology
Description
PhD proposal 2022: Computational dynamical modeling of Tumor response to radiation therapy to find a precise and accurate treatment for glioblastoma using logic programming
Context
External beam radiation therapy is one of the most common techniques of therapy against cancer using exposure to ionizing radiation (essentially photon from X-rays) to control tumor growth. The advance of high precision radiotherapy techniques has opened doors to innovative treatments with either modulation of physical parameters of irradiation devices, such as radiation dose, dose rate and energy of the rays, or the combination with chemical drugs radiosensitizing the tumor. These protocols may reduce biochemical recurrence percentages after radiotherapy. However, numerous uncertainties concerning the optimal schedule in terms of tumor control and toxicity risk still exist and cannot be estimated by pricey and long-term clinical trials because of the high number of modulated parameters.
Aim
Computational models may help unravel the complexity of these dose effect relationships. Especially hybrid modeling as developped in the MeForBio team consists in discrete models capturing not only chronological but also chronometrical information (i.e., durations) about the evolution of interacting biological entities. Such models will allow to create, at a limited cost, infinite virtual tumors on which different irradiations protocols can be simulated. Nevertheless, in silico modeling faces several challenges. The variety of complex biological mechanisms characterizing cancer and the response to radiotherapy i) may be described with different formalisms involving multiple parameters, ii) take place at different spatiotemporal scales and iii) have an high level of interdependence.
The goal of this PhD thesis is the development of an integrative in silico dynamical model of glioblastoma (GBM) tumor growth and response to various modalities of radiotherapy, which is the core of interest of Team PETRY of CRCINA. Inductive logic programming methods, like the Learning From Interpretation Transition (LFIT) framework developped in the MeForBio team at LS2N, will help us to design an explainable model of such a complex biological system.
Previous works
Team PETRY, expert in radiobiology, already experimented the benefits of computational models for the understanding of cell behaviors. It previously proposed in collaboration with Berkeley University (Berkeley, CA, USA), a 2D and 3D cellular automaton [Pau2017] integrating O2 diffusion and radiation effect on both endothelial and tumor cells. In addition, it introduced mitotic catastrophe instead of arbitrary delayed cell- death. It was validated on human prostate tumors transplanted in mice and then used to generate tumor control probability (TCP) curves for different radiotherapy protocols (2, 3, 4, 6, 8 and 10 Gy per fraction). Furthermore, Team PETRY in collaboration with LTSI (Rennes University, France), developed a multiscalein silico approach was implemented in C++ using the Multi-formalism Models and Simulation Library (M2SL) [Sos2021] which integrated the following mechanisms taking place at different temporal and spatial scales: (i) dynamic oxygenation, (ii) angiogenesis, division of (iii) tumor cells (iv) healthy cells, implementing their cycle, and (v) cycle-phase-and-oxygen dependent response to irradiation, considering mitotic death.
These previous works, based on discrete automata, were developped by hand, meaning that the models were developped using human expertise. A major challenge reside now in being able to process the large amount of data generated by radiotherapy experiments to build an understandable dynamic model of the underlying system. That is where inductive logic programming, and especially the Learning From Interpretation Transition (LFIT) framework, seems promising.
Learning the dynamics of systems with many interactive components becomes more and more important due to many applications, e.g., multi-agent systems, robotics and bioinformatics. In bioinformatics, learning the dynamics of biological systems can correspond to the identification of the influence of genes and can help to understandtheirinteractions. Totacklethisproblem,werelyonaframeworkthatisthecoreofacollaboration between the MeForBio team at Centrale Nantes/LS2N and the Inoue Lab at National Institute of Informatics (Tôkyô, Japan) since 2014: learning from interpretation transition (LFIT), which automatically constructs a model of the dynamics of a system from the observation of its state transitions. The following figure shows this learning process.
Given some raw data, like time-series data of gene expression, a discretization of those data in the form of state transitions is assumed. From those state transitions, according to the semantics of the system dynamics, different inference algorithms that model the system as a logic program have been proposed. The semantics of system dynamics can indeed differ with regard to the synchronism of its variables, the determinism of its evolution and the influence of its history. In the LFIT framework we propose several modeling and learning algorithms to tackle those different semantics. To date, the following systems have been tackled: memory-less consistent systems with Boolean variables [Ino2013], systems with memory [Rib2015b], non-consistent systems [Mar2015] and their multi-valued extensions [Rib2015c, Mar2016]. All those methods are dedicated to discrete systems or assume an abstraction of time series data as discrete transitions. In [Rib2017], we proposed a method that allows to deal with continuous time series data, the abstraction itself being learned by the algorithm. It allows, for example, to learn the gene expressions levels of expression alongside their interactions.
Until recently, the systems that LFIT handles were restricted to synchronous deterministic dynamics, i.e., all variables update their values at the same time and, for each state of the system, there is only one possible next state. However, many other dynamics exist in the field of logical modeling, in particular the asynchronous and generalized semantics which are of deep interest to model biological systems. In [Rib2018], we propose a method that learns the dynamics of the system independently of its dynamics semantics. Here, we model discrete multi-valued systems as logic programs in which each rule represents that a variable possibly takes some value at the next state, extending the formalism introduced in [Mar2015, Rib2015c, Mar2016]. The multi- valued logic representation used in [Rib2018] is based on annotated logics. Here, to each variable corresponds a domain of discrete values. In a rule, a literal is an atom annotated with one of these values. It allows us to represent annotated atoms simply as classical atoms and thus to remain in the normal logic program semantics. This modeling allows us to represent non-determinism and to propose an extension of LFIT in the form of a semantics free algorithm to learn from discrete multi-valued transitions, regardless of their update schemes. We showed from theoretical results that our new algorithm can learn systems dynamics from both synchronous (deterministic or not), asynchronous and general semantics transitions.
The current PhD thesis will invite us to extend the LFIT framework in such a way that we can process dynamical data like the evolution of cells (including the information about the division and parent-child relations). This appears to be doable by choosing carefully the litterals that allow us to model the problem.
Conclusions and perspectives.
Integration and simulation of the tumor cell response to innovative treatments of radiotherapy represent a challenge to better estimate the numerous uncertainties concerning the optimal schedule in terms of tumor control. The thesis will design computational models that will allow to build the most accurate propositions of the most efficient radiotherapy protocols. Using this model, we intend to propose an optimized and personalized treatment to allow the best anti-tumor efficiency, which cannot be tested by translational or clinical studies.
Profile of expected applicants
Applicants should hold a Master degree in Computer Science or Bioinformatics.
Type of funding
If successful, the application will be granted a 36-months funding granted by the AIBy4 project, one national funded program for research in AI: http://aiby4.ls2n.fr
References
[Ino2014] K. Inoue, T. Ribeiro, and C. Sakama. Learning from interpretation transition. Machine Learning, 94(1):51-79, 2014.
[Kif1992] M. Kifer and V. Subrahmanian. Theory of generalized annotated logic programming and its applications. Journal of Logic Programming, 12(4):335-367, 1992.
[Mar2016] D. Martinez, G. Alenya, C. Torras, T. Ribeiro, and K. Inoue. Learning relational dynamics of stochastic domains for planning. In Proceedings of the 26th International Conference on Automated Planning and Scheduling, 2016.
[Mar2015] D. Martinez Martinez, T. Ribeiro, K. Inoue, G. Alenya Ribas, and C. Torras. Learning probabilistic action models from interpretation transitions. In Proceedings of the Technical Communications of the 31st International Conference on Logic Programming (ICLP 2015), pages 1-14, 2015.
[Pau2017] Paul-Gilloteaux P, Potiron V, Delpon G, Supiot S, Chiavassa S, Paris F, Costes SV. Optimizing radiotherapy protocols using computer automata to model tumor cell death in function of oxygen diffusion processes. Sci Rep. 2017. 23;7(1):2280. doi: 10.1038/s41598-017-01757-6.
[Rib2015a] T. Ribeiro and K. Inoue. Learning prime implicant conditions from interpretation transition. In Inductive Logic Programming, pages 108-125. Springer, 2015.
[Rib2015b] T. Ribeiro, M. Magnin, K. Inoue, and C. Sakama. Learning delayed influences of biological systems. Frontiers in Bioengineering and Biotechnology, 2:81, 2015.
[Rib2015c] T. Ribeiro, M. Magnin, K. Inoue, and C. Sakama. Learning multi-valued biological models with delayed influence from time-series observations. In 2015 IEEE 14th International Conference on Machine Learning and Applications (ICMLA), pages 25-31, Dec 2015.
[Rib2018] T.Ribeiro,S.Tourret,M.Folschette,M.Magnin,D.Borzacchiello,F.Chinesta,O.Roux,andK.Inoue.Inductive learning from state transitions over continuous domains. In N. Lachiche and C. Vrain, editors, Inductive Logic Programming, pages 124-139, Cham, 2018. Springer International Publishing.
[Sos2021] Sosa-Marrero C, De Crevoisier R, Hernandez A, Fontaine P, Rioux-Leclercq N, Mathieu R, Fautrel A, Paris F and Acosta O. Towards a reduced in silico model predicting biochemical recurrence after radiotherapy in prostate cancer. IEEE Transactions on Biomedical Engineering. 2021. Accepted
[Ulm2017] Ulman V, Maška M, Magnusson KEG, Ronneberger O, Haubold C, Harder N, Matula P, Matula P, Svoboda D, Radojevic M, Smal I, Rohr K, Jaldén J, Blau HM, Dzyubachyk O, Lelieveldt B, Xiao P, Li Y, Cho SY, Dufour AC, Olivo-Marin JC, Reyes-Aldasoro CC, Solis-Lemus JA, Bensch R, Brox T, Stegmaier J, Mikut R, Wolf S, Hamprecht FA, Esteves T, Quelhas P, Demirel O, Malmström L, Jug F, Tomancak P, Meijering E, Muñoz-Barrutia A, Kozubek M, Ortiz-de-Solorzano C. An objective comparison of cell-tracking algorithms. Nat Methods. 2017. 14(12):1141-1152. doi: 10.1038/nmeth.4473.