We aim to optimize therapeutic approaches (i.e., controlling toxicities while ensuring a maximal efficacy) in oncology using mechanistic and statistical modeling (see Figure 1). These therapeutic approaches include immunotherapy, radiotherapy, chemotherapy, targeted therapies and their planification: combinations, sequences, intensification – densification, dosing and scheduling. Specifically, our research will be organized along three main axes:
Of note, in the Research Priorities document released by the American Society of Clinical Oncology in February 2021, “Developing and Integrating Artificial Intelligence in Cancer Research”, “Identifying Strategies That Predict Response and Resistance to Immunotherapies” and “Optimizing Multimodality Treatment for Solid Tumors” are listed as top-priorities, which fit quite well with our 3 axes.
The project-team is based upon the development of model-driven clinical oncology as a means to optimize anticancer therapies. Despite continuous efforts to make available novel drugs beyond traditional cytotoxic chemotherapy (i.e., oral targeted therapies, biologics, immune checkpoint inhibitors), prognosis of outcome remains poor for many cancers. The dosing regimen of anticancer drugs given today remains largely empirical, because dose-finding studies are often performed using outdated, sub-optimal protocols (such as modified-Fibonacci dose-ranging protocols) or because concomitant administration is the rule when combining several drugs. Consequently, clinical oncologists struggle to refine the way they use the anticancer agents made at their disposal. For instance, it took several years of bedside practice to understand that paclitaxel in breast cancer patients should not be administrated using the officially approved 150 mg/m² every 3 weeks scheduling, but rather with an alternate 75 mg/m² weekly dosing 105. Similarly, multi-targets sunitinib is now given on a 50 mg two-weeks on / one-week off basis, rather than the officially approved four-weeks on / two-weeks off schedule 93. Elsewhere, several combinatorial strategies trials have failed to yield convincing results, mostly because of the lack of a strong rationale regarding the best way to sequence treatments 87. Globally, clinical oncology today is still all about finding the best way to treat patients ensuring an optimal efficacy / safety balance.
After having long been limited to cytotoxic chemotherapy (in addition to surgery and radiotherapy), the arsenal of anti-cancer agents has dramatically increased over the last two decades. Indeed, major advances in the understanding of cancer biology that occurred in 1) the discovery and quantification of (epi)genetic alterations leading to targeted therapy and 2) the realization of the importance of the non-cancer cell components of tumors, i.e., the tumor micro-environment and tumor immunity, have helped to identify novel targets. Drugs targeting the tumor vasculature (e.g., first-in-class bevacizumab, approved in the mid-2000’s 74) or tumor immunity (e.g., immune checkpoint inhibitors (ICI) such as first-in-class ipilimumab, approved in the early 2010’s 103) represent groundbreaking innovations in oncology. ICIs in particular are considered as game-changing drugs because diseases with once dismal prognosis (e.g., metastatic melanoma, non-small cell lung cancer (NSCLC), kidney cancer or head and neck cancer 125) now show 20-40% of 5-years survival. Nevertheless, these impressive results are limited to a minority of patients in a limited number of cancers. In addition, no validated biomarker predictive of response has yet been identified, thus highlighting how early prediction of response and probability of future relapse are a critical, unmet medical need. The encouraging yet still insufficient clinical results of ICIs have led current clinical oncology to consider combinations of such immunotherapies with preexisting anti-cancer modalities: radiation therapy 112, cytotoxic 99, targeted 117 or anti-angiogenic therapies 124. However, the near-infinite possibilities of combinations in terms of sequencing, dosing and scheduling challenge the ability of classical trial-and-error methods to find appropriate modes of combination 86.
In addition, day-to-day clinical decisions made by oncologists are based on a large amount of information, coming from: 1) their own knowledge integrating years of clinical practice combined to updated literature and 2) objective data coming from multiple sources (demographic data, radiology, functional imaging, molecular biology, histology, biomarkers, blood counts, etc…). The large amount of clinical and biological data generated now in clinical oncology is not properly analyzed, because of the lack of appropriate models to picture the complexity of longitudinal observations. Oncologists lack a comprehensive framework and numerical software that could support decision of therapeutic strategy (e.g., to treat or not? to what extent? with what treatment (surgery, radiotherapy, systemic therapy)? in what order? etc.), especially when their time dedicated to examination of a given patient case is limited (e.g., in multidisciplinary meetings (RCP), or individual consultations). Furthermore, modeling is the only way to retrieve similar characteristics from very different experimental conditions and clinical protocols.
To address these major issues, our project-team aims at:
We use non-clinical and clinical data related with the pharmacology of anti-cancer agents and medical monitoring of the disease status. The former includes pharmacokinetics (drug levels in plasma (patients) and full body pharmacokinetics (animal models)), pharmacodynamics (efficacy, safety), pharmacogenetics (i.e., constitutional genetic polymorphisms affecting drug transport and metabolizing enzymes), pharmacogenomics (i.e., molecular and genetic alterations affecting tumor cells). The latter include demographics, anatomical imaging (e.g., tumor sizes derived from CT scan or MRI), functional imaging (e.g., positron emission tomography), histopathology quantifications, biological variables (such as kidney and liver functions or blood counts), immuno-monitoring data (flow cytometry) or cell free DNA. We will especially rely on real-world data (also termed fragmentary data) collected from patient routine monitoring by our members with hospital activity.
Experimental data are generated by the experimental wet-lab group (involving COMPO team members: AR, RF, JC), relying on state-of-the-art experimental pharmacokinetic laboratory fully equipped to perform in vitro and in vivo explorations of drug metabolism, pharmacokinetics and experimental therapeutics in oncology, including bioanalytical support and fluorescence/bioluminescence monitoring in rodents with highly specialized staff. Clinical pharmacokinetics and pharmacogenetics data are generated by the clinical pharmacology group (JC, RF), relying on the expertise of the clinical pharmacokinetics laboratory of the La Timone University Hospital of Marseille, an FDA-labelled, ISO15189-labelled facility with state-of-the-art bioanalytical resources to assay any kind of drugs or drug metabolites in patients. Specific data regarding cancer biology and pharmacodynamics (immunomonitoring, pharmacogenomics) are generated in collaboration with other CRCM teams.
Clinical data and additional biomarker data are collected from either clinical trials or real-world studies performed by hospital pharmacists and oncologists of the joint-team (RF, JC, SS, LG, XM) and their residents, or by other medical oncologist partners. We have strong collaborations and ongoing projects with pediatrics (Pr N. Andre), hematology (Dr G. Venton), nuclear medicine (Pr D. Taïeb) and radiotherapy (Pr L. Padovani). Importantly, the project-team is located near the INCa-labeled center for early clinical trials (CLIP2), thus facilitating the data collection and later, the implementation of modeling approaches in early clinical trials.
In addition, we also rely on publicly available data from online databases such as the TCGA (genomic data), the TCIA (imaging data) or data from clinical trials.
Our primary objective is centered on the improvement of therapeutic strategies in oncology. Nevertheless, this brings novel methodological challenges requiring developments at the formal level within the generic field of modeling biological and pharmaco-patho-physiological systems. Difficulties to take into account include: the longitudinal profile of the quantities of interest; measurement uncertainty (requiring statistical considerations); difficulties in sampling the real processes leading to scarcity of the observed data and large inter-individual variability. Many specific problems in life science systems are very different from that encountered from physical modeling in industrial applications (e.g., mechanical engineering or energy).
To summarize our methodology, we are interested in modeling the dynamics of pharmaco-oncological processes (mechanistic modeling) and their inter-individual variability (statistical modeling). Our intended methodological contributions are: 1) to invent novel mechanistic models for complex physiological processes able to describe the effect of therapeutic intervention, 2) to design appropriate statistical frameworks for parameter estimation and description of inter-individual variability, 3) to test and validate the models against experimental and clinical data, 4) to combine state-of-the art machine learning (ML) methods with mechanistic models to integrate large dimension data.
Mechanistic models are defined here as mathematical constructs that describe physiological variables (e.g., plasma drug concentration, tumor size, or biomarkers) and their dynamics based on physical and biological principles (e.g., law of mass conservation). They describe the time profiles of the variables of interest by means of ordinary or partial differential equations (ODEs and PDEs, respectively) and are thus deterministic.
The main challenge of the modeling exercise is to find the appropriate balance between the degree of integration of biological phenomena (model complexity) and granularity of the data available (i.e., sampling time resolution, observed variables, spatio-temporal or only temporal measurements) ensuring the feasibility of parameter estimation. Indeed, cancer biology is extremely complex, involving processes at multiple temporal and spatial scales (intra- and inter-cellular, tissular, organism). It is thus tempting to build intricate models integrating as many phenomena as possible. Along these lines, the last decades have witnessed the proliferation of multiple such complex models. We see two shortcomings to this approach. First, in contrast with models of physical phenomena, the parameters of biological models are often not directly measurable and thus have to be estimated from fitting the models to experimental or clinical data. Therefore, their number has to be commensurate to the available data in order to ensure identifiability. Unfortunately, many complex models from mathematical oncology have too many parameters to be reasonably identified and have thus had a limited application in terms of biological insights or clinical applications. Second, complex, multiscale models are characterized by a reductionist point-of-view whereby general phenomena could be explained by decomposing them into elementary pieces. However, corresponding elementary experiments would not be suitable for quantification of the several homeostatic mechanisms involved in the whole real process. Thereby, we do not adhere to this reductionist vision and for modeling purposes we rather adopt a holistic approach considering the process as an indivisible whole.
To avoid the above-mentioned caveats, our methodology always starts from: 1) a clinically relevant medical problem but more importantly 2) the data available to build models.
In several instances, the mechanistic models are ordinary differential equations (ODEs). This is the case for the simplest type of experimental data that we generate, i.e., tumor growth kinetics. Departing from previous works establishing models for untreated experimental growth, we are now actively engaged into designing pharmacokinetics (PK)/ pharmacodynamics (PD) models of the effect of multiple therapies. These models have to account for the specificities of the drug delivery (e.g., nanoparticles), the biological effect of the treatments (e.g., cytotoxics, antiangiogenics or immunotherapies) and resistance to the therapy (either innate or acquired). The resulting models are novel nonlinear ODEs that need to be validated against the data and, when necessary, theoretically studied for their qualitative behavior. With the advent of immunotherapies, there has been a regain of interest to modeling tumor-immune interactions. Again, despite a wide literature on the subject, very few models have been validated against empirical data. A methodological objective is to establish and validate such models, including effect of immunotherapies.
Description of other phenomena are more adapted to partial differential equations (PDE) models. For instance, following an approach initiated by Iwata et al. 108, structured PDE models can be written for description of a population of metastases (see 4.3.2). Indeed, at the organism scale, cancer diseases are often characterized by a generalized (metastatic) state. However, few modeling efforts are currently focused on this aspect. The only validated models in large cohorts for systemic disease concern the sum of largest diameters as defined by the RECIST criteria 83. We aim to go beyond this state of the art by: 1) providing models of coupled tumor growth with interactions (and quantification of inter-lesion variability) 79, 80 and more importantly 2) developing models accounting not only for growth of the tumors, but also dissemination (birth of new lesions).
To date, most of the available and collectable clinical data about tumor growth and response to therapy consist of scalar data, often even limited to lesion diameters or sum of diameters. This is why we primarily focus our efforts on developing kinetic models of such data, the novelty coming from integrating other longitudinal biological data (e.g., from blood counts). Nevertheless, imaging data are now increasingly accessible and recent advances in image analysis allow the automatic segmentation of lesions make it possible to quantify the spatial shape and texture of tumors without a prohibitive cost for radiologists. This opens the way to develop spatially distributed PDE models of tumor dynamics. The existing models have largely remained unconfronted to data, apart from notable exceptions from the Inria MONC 90 and EPIONE 88 teams, as well as the Swanson 75 and Yankeelov 104 groups. Radiomics approaches quantifying heterogeneity in the images could bring additional information. We will rely on existing or establish collaborations with other dedicated Inria teams (EPIONE, MONC) for such purpose. COMPO would ideally bridge the gap between clinical studies and the Inria ecosystem. Finally, PDE models are also well adapted to describe intra-tumor drug penetration and we have recently developed such models for description of intra-tumor fluid flow and transport of antibody nano-conjugates 126.
Statistical models are defined here as mathematical constructs that describe the stochastic sources of variability in the data. They comprise both: 1) classical statistical models defining the functional and probabilistic relationship explicitly and 2) machine learning (ML) algorithms highly based on the data alone (e.g., tree-based models and associated ensemble methods or support vector machines) 101. We use such models for the following purposes: defining appropriate frameworks for parameter estimation; quantitative testing of biological hypothesis; addressing interindividual variability (using nonlinear mixed-effects (NLME) modeling); and building predictive models.
NLME – also termed the population approach in PK/PD modeling, or hierarchical modeling 110 – consists in assuming a statistical distribution of the parameters of the structural (often mechanistic) model, in order to describe longitudinal observations within a population of individuals. Instead of estimating individual parameters on a subject per subject basis – leading to identifiability issues in sparse data situations characteristic of longitudinal measurements in oncology – all data can be pooled together and a joint likelihood is obtained. Likelihood maximization becomes more complex than for classical nonlinear regression, nevertheless this problem has already been addressed by means of algorithms such as the deterministic first-order conditional expansion (FOCE) algorithm 111 or the stochastic approximation of the expectation-maximization algorithm (SAEM) 92. These algorithms are implemented in widely used software in the PMX community such as NONMEM® (Icon) or Monolix® (Lixoft), or R packages (e.g., saemix or nlme). Once the population distribution is estimated, empirical Bayes estimates (EBEs) can be derived for estimation of individual parameters. We also use the language Stan that implements state-of-the art Bayesian methods 84.
Departing from a general distribution of the parameters (often assumed log-normal) with quantified but unexplained interindividual variability, covariates are incorporated to explain this variability and build predictive models. This is traditionally done by means of linear models (possibly up to a functional transformation). However, with the increase in number of such covariates, the traditional tools and algorithms are limited. We thus develop advanced covariate models in NLME incorporating ML algorithms. Such methods require novel contributions. A possible lead is to first identify the EBEs and then use ML algorithms to predict these from the covariates 118. In other cases, ensemble models could be built from the heterogeneous sources of data, integrating one sub-model from EBEs identified from early data. Another, more challenging, avenue would be to adapt the parameter estimation algorithms like SAEM to include ML models in the covariate part.
In addition, because few data have been available longitudinally so far (i.e., small number of quantities measured at each time point), the current use of NLME relies on models with a small number of output variables. In this respect, modern clinical oncology studies bring new modeling and statistical challenges because many more quantitative data are collected at each time point (e.g., hundreds of variables from immuno-monitoring or possibly tens of thousands from circulating DNA, or radiomics features from imaging). Defining high-dimensional ODE models describing all the physiologically meaningful variables becomes intractable, therefore new methods are required. A possible avenue is to have a sequential approach, using first ML methods to reduce the dimension, then model the reduced number of variables. Another, more challenging, avenue would be to perform the two tasks (dimension reduction and temporal modeling) at the same time, and include this in an NLME framework for population estimation. The first part could be done using tools from unsupervised learning such as auto-encoders.
Following the availability of longitudinal tumor measurements, recent developments in the field of NLME have concerned joint modeling 94. This consists in modeling the longitudinal kinetics of a biomarker (e.g., tumor size) together with censored time-to-event data (e.g., overall survival) in a single step. Promising results have been obtained so far and we intend to develop methods beyond the state-of-the art in this area. This includes, in connection with above: 1) extension to models with emergence of new metastases, 2) integration of high-dimensional covariates and 3) high-dimensional longitudinal data. This Bayesian integration of data for updated survival predictions could lead to high impact results, as demonstrated by a recent publication in Cell 109.
Finally, we intend to bring the use of established ML tools to address concrete clinical problems emerging from the data collected in routine or clinical trials. Indeed, such data is so far analyzed using traditional statistical methods. ML algorithms could bring added value for predicting efficacy or toxicity from demographic, clinical and biological data.
The project-team is based upon generating experimental and clinical data to identify and test the models, and to provide proof-of-concept studies so as to validate the model-based dosing and scheduling prior to transposing them in patients. Historically, experimental therapeutics in oncology has relied on a wide variety of in vitro and in vivo models mimicking human cancer disease. In oncology, hundreds of in vitro models using cancer cell lines cultivated following 2D or 3D (spheroids) fashion, plus more sophisticated models with cancer cells enriched with fibroblasts or endothelial cells 119, eventually leading to complex organoids 98. Similarly, almost all kind of tumors can be tested in vivo, mostly in small rodents. In oncology, in vivo models are mostly based upon xenografting human tumors from established cell lines or from patient biopsies (patient-derived xenografts or PDX) so as to better mimic human pharmacology when testing active compounds next. To achieve this, several strains of immune-compromised mice have been successfully developed. Because immune checkpoint inhibitors do not exert direct anti-proliferative activity on cancer cells but are rather expected to harness tumor immunity, human xenografts in immuno-compromised mice is not anymore a suitable model. This has led investigators to shift towards immune-competent syngeneic mice models. Non-clinical experiments with drug candidates in immunotherapy mostly focus on deciphering the pharmacology of the targeted pathways, assessing the cytokine release potential, studying receptor occupancy, by using models the most likely to mimic tumor immunity in human. More sophisticated animal models such as human knock-in mice, immuno-avatar, hemato-lymphoid humanized mice or immune-PDX mice have been developed 122 (i.e., allowing to test immune checkpoint inhibitors in mice models combining human xenograft with relevant, humanized immunity and stroma cells) have been made available as well. Beyond generating data on efficacy such as reduction in primary tumor mass or metastatic spreading, experimental models help providing as well in depth knowledge on human and animal target cells, in vitro and in vivo concentration-effect studies, search for biomarkers, plus the most comprehensive knowledge on animal vs. human differences on dose – exposure – effects relationships and finally drug distribution throughout the body, target expression, affinity of target-binding and intrinsic efficacy, duration and reversibility of the effects. In particular, animal drug metabolism and pharmacokinetics (i.e., exploration of liver metabolism and distribution / absorption processes using in vitro or in vivo dedicated models) help understanding the disposition and distribution of the drug in the body throughout time, especially its ability to target tumor tissues (i.e., in vivo distribution in tumor-bearing mice) and help understanding sources of pharmacokinetic variability. All this information requires state-of-the-art techniques for measuring drugs and drug metabolites into biological fluids in tissues, such as fluorescence-imaging, high-performance liquid chromatography or liquid-chromatography-mass spectrometry bioanalysis. Our team has proven track records in the field of experimental therapeutics in oncology, with two PhDs on developing anticancer nanoparticles in breast and colorectal cancer 97, 121 plus experiments on model-driven way to combine anti-angiogenics with cytotoxics in breast and lung cancers 107, 114, 123, model-driven determination of alternate dosing in neuroblastoma, or methodological studies on monitoring tumor growth 115.
The different steps of patient therapeutic management by clinicians consist mostly of: diagnosis, estimation of the extension of the disease, choice of therapy and evaluation of the therapy (efficacy, toxicity). This axis is specifically concerned with such clinical problems, apart from the pharmacological aspects addressed in the other axes.
In this axis, we aim to develop mathematical and statistical models and methods able to process this information to bring added value by inferring hidden parameters and provide simulations and predictions about the past and future behavior of the disease.
In the short-term (4 years), our research projects are: (1) modeling large-scale longitudinal data from immuno-oncology for prediction of response to immune checkpoint inhibition (QUANTIC and TGI-ML projects), (2) developing clinically relevant mathematical models of metastasis and (3) modeling the kinetics of clinical biomarkers.
This axis fits within the “population approach” introduced in the 80’s by L. Sheiner 121 and aims at gathering exhaustive information about the multiple sources for variability in response in patients (including but not limited to drug-drug interactions, pharmacogenetics, and cormorbidities affecting renal and liver functions), build specific mechanistic models including relevant covariates and determine the PK/PD relationships of drugs used in oncology-hematology or for treating solid tumors. This covers cytotoxics, oral targeted therapies, biologics or immune checkpoint inhibitors. The overall goal is to achieve precision and personalized drug administration, i.e., the right dosing and scheduling regimen for the right patient.
In the short-term (4 years), our research projects will be focused on the following objectives: 1) predict response or toxicity variability dependent on pharmacogenetic (PGx) and pharmacogenomic data, 2) assess PMX of anticancer agents such as biologics, including ICI, and 3) develop physiologically-based PK models of nanoparticles distribution. Together, these objectives will allow to gain insights in the variability in drug response that depends on PK (1, 2 and 3) or germinal genetic alterations (1).
Our hypothesis is that so many attempts to combine drugs fail not because the underlying pharmacological concepts are wrong (such as immunogenic cell death triggered by cytotoxics or radiation therapy, or increase in T cells infiltration with anti-angiogenics) but because these combinations probably require fine tuning in terms of dosing, scheduling and sequencing, whereas in practice all the drugs are given the same day. The goal of this second axis is therefore to shift from current empirical and suboptimal combinatorial regimen to model-informed designs to best combine drugs and therapeutic approaches so as to maximize efficacy while controlling toxicities. To do so, we will rely on our pioneering work about model-driven scheduling in early phase trials for combination of cytotoxic agents in metastatic breast cancer (MODEL1 trial) 102, 113 and metronomic vinorelbine in lung cancers (MetroVino trial) 78, 96. Leveraging the unique multidisciplinary aspect of our team, we implement a fully translational approach going from experimental therapeutics, PMX and quantitative systems pharmacology, to clinical trials either in early (phase I/II) or late (phase III) settings. Of note, our group has already an expertise in developing mathematical models determining the best sequencing between chemotherapy and anti-angiogenics.
In the short-term (4 years), our research projects will be focused on providing model-informed designs for combining ICI with: (1) cytotoxics, (2) an experimental immunoliposome and (3) radiotherapy.
We plan to achieve two main things on the modeling side: 1) the development of effective numerical tools (either as web applications or as part of simulation software) and 2) the empirical validation of the models.
For 1), this includes a tool able to predict response to ICI
monotherapy in NSCLC from baseline and early response data . We will work with our industrial partners (in particular,
HalioDX from the PIONeeR consortium), to transfer the tool for
commercial use. Second, we plan to have a validated
numerical tool able to predict metastatic relapse from clinical
biomarkers at diagnosis, using our mechanistic model. This model will
integrate the effect of adjuvant therapy (hormonotherapy or cytotoxic
therapy) and will be able to simulate the long-term impact of
alternative treatments (e.g., number of cycles to be administered to
prevent distant relapse). It will have been validated from our local
databases and will be implemented in a clinically-usable online tool
such as
PREDICT 127.
The main difference with this tool will be the ability to
mechanistically simulate the effect of therapy. We plan to have
initiated a larger initiative at the national or European level to
collect large data bases, validate further the predictive power
of the model, and refine its structure if required. We will also extend
this tool to other pathologies that share the same problematic
(diagnosis at early-stage, important probability of future distant
relapse) such as kidney cancer, following our initial work from
preclinical
data 76, 81, 91.
The pharmacometric models that will have been developed will also be implemented as clinically effective dose
adaptation numerical tools directly usable to personalize the dose and
scheduling of multiple anti-cancer agents not only by the clinicians and
pharmacists from our group, but also by others, at least at a regional
level.
For 2), our strategy of validation is the following. First, during the
development phase, a proportion of the dataset (usually, 30%) is left
aside unused for establishment of the model and initial calibration, and
then used as a test set. When the sample size is too small (n retrospective, external data sets. We
have for instance initiated a collaboration with Dr C. Scherer
(Clermont-Ferrand) to validate our metastatic prediction model on an
external database of 3061 patients. The third step is to validate the
added value of the model-based approach compared with the standard of
care and is a long-term rather than mid-term objective.
In axis 2, in addition to standard drugs, developing tools for similarly
better understanding the sources of therapeutic and PK variability and
understanding the PK/PD relationships of cell therapy in
oncology such as CAR-T cell therapies, is a challenging task.
The challenge with CAR-T cells is that first, developing bioanalytical
tools to monitor them in patients is not trivial, and second, little but
nothing is known regarding their PK properties and possible sources
impacting on PK/PD relationships such as disease status or immune status
of the patient. We aim at developing both a platform to monitor
CAR-T cells and future cell therapies in plasma and mechanistic models
to describe the disposition of these new therapies in the body.
The nanoPBPK model will be extrapolated to humans and used to
determine the specifications of an optimal nanosystem in order to
penetrate solid tumors such as pancreatic tumors. The rationally
designed nanosystem will be evaluated in vitro and in vivo
in order to validate the approach. The nanoPBPK model will be interfaced
to become a software and be shared with the scientific
community. To achieve this goal, a partnership with ESQlabs, the company
developing the opensource PBPK platform PKSim, has already been approved
by both sides. The nanoPBPK model will be combined with pharmacodynamic
modeling describing the effect of the loaded anticancer drug on tumor
growth and metastases spread, on the immune system, and on dose-limiting
toxicities.
Our mid-term objective in axis 3 is to assist the design of
scheduling regimen for combinatorial treatments in early phase clinical
trials, which represent an important clinical challenge of the next 10
years. To do so, we will benefit from our close connection to the
INCa-labeled AP-HM's center for early phase clinical trials (CLIP2). Our
aim is to design model-based, individualized and adaptive
scheduling regimen that depend on the monitoring of the disease
evolution. We plan to run phase I/II trials based on the model
recommendations. Depending on our achievements and success in phase I/II
trials our mid-term goal would be to lead a prospective,
randomized, phase III trial comparing a model-based adaptive regimen to
the standard of care for combination of immune checkpoint inhibition
with chemotherapy and/or anti-angiogenic and targeted therapy. According
to our team expertise, the target malignancies would be primarily lung
cancer (LG) and head and neck cancer (SS).
At long-term, we globally wish to have established a worldwide leader
position in the fields of quantitative mathematical oncology
and PMX, as well as the pharmacokinetics of
nanoparticles. We hope that this would translate into the achievement
of three goals: (1) the development of software effectively used
for clinical decision-making and dosing adjustment (estimated
achievable), (2) the initiation of prospective, phase III
clinical trials comparing model-guided therapy versus standard of care
(highly challenging), (3) clinical trials of nanosystems
designed by our group (estimated achievable). In addition, we foresee
several avenues both in terms of modeling opportunities and
applications.
Our short-term program is devoted to the development of new models and
their confrontation to empirical data. The mid-term program will focus
on the validation and refinement of these models. In the long-term, we
foresee that this will bring novel questions in terms of mathematical
analysis of the models. For instance, metastatic modeling will
establish validated models for tumor-tumor interactions, including
immune-mediated interactions. In turn, this leads to
nonlinear, size-structured, renewal PDEs. Study of the
asymptotic behavior of such equations is non-trivial.
More generally, we expect that physiologically structured PDEs
(psPDE) can become relevant to practical modeling in oncology, from two
types of data: flow cytometry and single-cell sequencing. Flow
cytometry is currently becoming of increasing relevance to characterize
multiple populations of cells, for instance in the context of
immuno-oncology. In the QUANTIC project (10.2), we are starting to
interact with such data, only by means of scalar quantities so far.
However, the structure of this data is to have, for each cell, a
quantitative measure (e.g., a surface marker). Measuring these in a
population of millions or billions of cells makes it adapted to modeling
by such psPDE. Similarly, single-cell sequencing is a technique by which
every cell of a population (e.g., in a tumor) is sequenced, thus having
mutation information. In turn, this allows to quantify subclones in the
population. Such data has already generated fascinating results, for
instance in the study of metastatic development theories 100, 106. Although
evolutionary modeling is a wide field with established groups (M. Nowak
or F. Michor in Harvard, C. Curtis in Stanford, T. Graham at the CRUK),
few groups are modeling dynamical data at single-cell
resolution. To this regard, the theoretical work initiated by J.
Clairambault and B. Perthame (Inria MAMBA) suggesting to use psPDEs to
model evolution in cancer cell populations could be
appropriate 85. Parameter estimation
in such models is a challenging task 95 and
data assimilation from flow cytometry or single-cell sequencing data
sets would represent an important avenue. Dynamical data can be provided
by circulating tumor DNA and we have already initiated contacts
with an important clinical and biological study in Marseille on this
topic (the SCHISM study, PIs: SS and F. Fina). The recent developments
of technology enabling spatial resolution of single-cell sequencing also
paves the way to exciting avenues in terms of modeling
116.
We will also build models that can optimize the effectiveness of treatments incorporating new criteria (other than the evolution of tumor mass) of diagnostic and therapeutic evaluation, especially those we have forged around the information provided by functional imaging (T80) computational algorithm time at which 80% of FDG is metabolized 77, 89, 120.
A general, challenging, long-term objective, would be to run
prospective clinical trials in which a model-informed arm would
be compared to the standard of care. In the model-informed arm,
therapeutic decision would be based on the recommendation of the model.
This applies to the models developed in all axes. For instance, in
breast cancer, the number of cycles of chemotherapy would be adapted
based on the model indication (axis 1), decision of the maximum
tolerated dose in the treatment of leukemia patients would be based on
the PGx/PK/PD model (axis 2) or the combination scheduling regimen would
be given by model calculations (axis 3).
Regarding nanosystems initially designed by our group based on the nanoPBPK modeling, they will also be tested in early clinical trials. Our group will drive the design of these based on simulations performed with the nanoPBPK and pharmacodynamic model, in order to guarantee the highest chances of success while ensuring patients safety. In particular, nanoparticles specifically transporting cytoxics could replace standard systemic myelo-ablation in hematopoietic stem cell transplantation, a risky strategy with frequent life-threatening, when not lethal, toxicities. Because of the fully controlled distribution phase in the body, nanoparticles encapsulating several drugs could thus be implemented in the preparative regimens for allogeneic stem cell transplantation in leukemia or myeloid malignancies.
In addition, several groups predict that in addition to standard drugs
or biologics, or rising gene therapy and cell therapy strategies, new
devices such as nanobots will be developed to treat cancers.
Nanobots are entities which are not designed to interact with standard
pharmacological targets or genes like current anticancer treatments, but
could fix the cancer cell, either by providing a missing protein, or
ultimately trigger a mechanic cell-death using radiation, thermal wave,
or by disrupting cell membrane. These new entities should exhibit
totally new pharmacokinetics, because they are unlikely to be
metabolized in the liver or to be cleared by the kidney or the biliary
tract. Therefore, new models for PK/PD should be developed,
because neither behavior in the body nor intrinsic mechanisms of action
are known yet. In addition, the issue of nanosafety with such devices
will be particularly critical and will require extensive PMX resources,
to predict long-term effects or to keep under control the mechanism of
action. Should such nanobots be developed, the COMPO joint-team should
develop specific resources to monitor their fate in the body, specific
resources to write equations describing nanobots/body, nanobots/immune
system, and nanobots/cancer cells interactions, plus new global models
encapsulating all the interactions, and pharmacodynamics impact of such
devices.
Depending on the expertise we gain in the PK/PD knowledge of those cell therapies as part of the mid-term objectives, optimizing combinatorial strategies to such cell therapies will be another long-term objective.
Last, pharmaco-economic studies including impact of quality of
life, increase in both tolerability and efficacy will be performed to
determine whether PMX-based dosing is a cost-saving strategy.
For instance, by refining the scheduling of immune checkpoint
inhibitors, such as determining, using modeling and simulation
strategies, when the plasma concentration of the drug reaches the
threshold in trough levels necessary to ensure maximal target
engagement, it will be possible to customize the frequency of the
administrations, with possible strong impact on treatment costs. By
using real-life data, we propose to use dedicated models to quickly
define alternate and more appropriate dosing and/or scheduling
with newly marketed anticancer agents, either chemotherapy, targeted
therapies or biologics including immune checkpoint inhibitors.
The COMPO research team's projects all focus on a serial of complementary and inter-related domains described in an itemized fashion below:
Due to its unique composition including medical oncologists, clinical pharmacologists and mathematical modelers, COMPO is at stake with important social challenges: oncology healthcare and innovation in drug development. The software and results developed by COMPO are devoted to these challenges and aim to be directly used by medical and pharmaceutical oncologists or by the biotech and pharmaceutical industry to help drug development and biomarker discovery.
To give a few examples:
The package compoEDA aims to provide a comprehensive exploratory analysis of data from clinical studies in oncology. These studies commonly investigate biological markers able to reveal and distinguish different tumor profiles, in order to early adapt the therapeutic strategy for patients.
The objective of this software is to provide a simplified tool for both computational scientists and clinical researchers to easily generate agraphical results and automatic reports containing the following analyses:
Available features:
Available data:
This package provides multiple functions to perform machine learning analysis using the `tidymodels` framework. Tasks include: feature selection, plot feature importances, train, corss-validate or apply supervised machine learning algorithms (classification or survival analyses), evaluate metrics of predictive performances, compute learning curves.
Initial development was part of the `stats_pioneer` package (also called `pioneerPackage`) and `ml.tidy` evolved as a standalone package only in February 2023.
This software was built to analyse the PIONeeR (Precision Immuno-Oncology for advanced Non-small cell lung cancer patients with PD-(L) 1 ICI Resistance) data. PIONeeR is a prospective, multicenter study with primary objective being to validate the existence of a hypothetical immune profile explaining resistance to immunotherapy in non-small cell lung cancer patients.
It initially integrated preprocessing, exploratory data analysis, visualization, statistical analysis, feature selection, machine learning and results generation and reporting. Since, exploratory data analysis, visualization and statistical analysis have been promoted to the COMPO-level `compoEDA` package and feature selection and machine learning to the COMPO-level `ml.tidy` package.
This software corresponds to the very first step of the data analysis, which is the preprocessing, and the very last: generation of results. Some of its functions aim at:
Machine learning (ML) for prediction of early progression in 2+ line patients:
- Addition and development of a python package python-based analyses:
Development of the first version of the sofware and several updates.
This sofware was used and will be described in a work to be published in 2024. The application was the use of this model to describe the dynamics of the brain metastasis disease in small cell lung cancer patient with or without prophylactic cranial irradiation (PCI) and study the impact of PCI on the overall survival of the patients as well as the progression of the brain disease.
Development of the first version of the sofware and several updates.
This sofware is works in combinaison with the metamats R package and was used and will be described in a work to be published in 2024 (see the BIL entry of metamats R package for more details).
SChISModeling aims to analyze SChISM data (Size CfDNA Immunotherapies Signature Monitoring). SChISM is a clinical study that introduces an innovative approach to quantify circulating free DNA in cancer patients treated with immunotherapy. The study's objective is to early predict response to immunotherapy in patients at an advanced/metastatic stage according to these quantitative cfDNA data.
This software corresponds to the very first step of the data analysis, which is the statistical analysis. Some of its functions aim at:
Funding and data: Roche pRED
Publication: Submitted to Clinical Cancer Research (preprint in medrxiv 58), Oral communications 34, 35, 33, 36, 37
Existing survival prediction models rely only on baseline or tumor kinetics data and lack machine learning integration. We introduce a novel kinetics-machine learning (kML) model that integrates baseline markers, tumor kinetics and four on-treatment simple blood markers (albumin, CRP, lactate dehydrogenase and neutrophils).
Developed for immune-checkpoint inhibition (ICI) in non-small cell lung cancer on three phase 2 trials (533 patients), kML was validated on the two arms of a phase 3 trial (ICI and chemotherapy, 377 and 354 patients).
It outperformed the current state-of-the-art for individual predictions with a test set c-index of
Critically, kML predicted the success of the phase 3 trial using only 25 weeks of on-study data (predicted HR
Funding: Inria-Inserm Phd Grant
Publication: Published in ”Computer Methods and Programs in Biomedicine” 15 and AACR communication in 2022 82
Estimating the risk of metastatic relapse is a major challenge to decide adjuvant treatment options in early-stage breast cancer (eBC). To date, distant metastasis-free survival (DMFS) analysis mainly relies on classical, agnostic, statistical models (e.g., Cox regression). Instead, we propose here to derive mechanistic models of DMFS.
The present series consisted of eBC patients who did not receive adjuvant systemic therapy from three datasets, composed respectively of 692 (Bergonié Institute), 591 (Paoli-Calmettes Institute, IPC), and 163 (Public Hospital Marseille, AP-HM) patients with routine clinical annotations. The last dataset also contained expression of three non-routine biomarkers. Our mechanistic model of DMFS relies on two mathematical parameters that represent growth (
We found that Ki67 and Thymidine Kinase-1 were associated with
Overall, we demonstrate that our novel method combining mechanistic and advanced statistical modeling is able to unravel the biological roles of clinicopathological parameters from DMFS data.
Data: early-stage NSCLC patients with brain metastases, Multidisciplinary oncology and therapeutic innovation department, APHM
Publication: 14
Intracranial progression after curative treatment of early-stage non-small cell lung cancer (NSCLC) occurs from 10 to 50% and is difficult to manage, given the heterogeneity of clinical presentations and the variability of treatments available.
The objective of this study was to develop a mechanistic model of intracranial progression to predict survival following a firstbrain metastasis (BM) event occurring at a time
Data included early-stage NSCLC patients treated with a curative intent who had a BM as the first and single relapse site (N = 31). We propose a mechanistic mathematical model able to derive computational markers from primary tumor and BM data at
The model was able to correctly describe the number and size of metastases at
This may help lung cancer physicians to guide and personalize the management of NSCLC patients with intracranial oligoprogression.
Data: Preclinical data of primary tumor and metastatic growth in 128 mice.
Publication: under revision at PLoS Computational Biology 59
Clinical trials involving systemic neoadjuvant treatments in breast cancer aim to shrink tumors prior to surgery while simultaneously allowing for controlled evaluation of biomarkers, toxicity, and suppression of distant (occult) metastatic disease. Yet such trials are rarely preceded by preclinical testing involving surgery.
Here we used a mouse model of spontaneous metastasis after surgical removal to develop a predictive mathematical model of neoadjuvant treatment response to sunitinib, a receptor tyrosine kinase inhibitor (RTKI).
Longitudinal data consisted of measurements of presurgical primary tumor size and postsurgical metastatic burden in 128 mice (104 for model training, 24 for validation), following variable neoadjuvant treatment schedules over a 14-day period. A nonlinear mixed-effects modeling approach was used to quantify inter-animal variability. Machine learning algorithms were applied to investigate the significance of several biomarkers at resection as predictors of individual kinetics. Biomarkers included circulating tumor- and immune-based cells (circulating tumor cells and myeloid-derived suppressor cells) as well as immunohistochemical tumor proteins (CD31 and Ki67).
Our simulations showed that neoadjuvant RTKI treatment inhibits primary tumor growth but has little efficacy in preventing (micro)-metastatic disease progression after surgery. Surprisingly, machine-learning algorithms demonstrated only limited predictive power of tested biomarkers on the mathematical parameters. These results suggest that presurgical modeling might be an effective tool to screen biomarkers prior to clinical trial testing.
Mathematical modeling combined with artificial intelligence techniques represent a novel platform for integrating preclinical surgical metastasis models in outcome prediction of neoadjuvant treatment.
Funding: Laënnec Institute.
Publication: 39
Survival analysis is an essential tool for the study of health data. An inherent component of such data is the presence of missing values. In recent years, researchers proposed new learning algorithms for survival tasks based on neural networks. Here, we studied the predictive performance of such algorithms coupled with different methods for handling missing values on simulated data that reflect a realistic situation, i.e., when individuals belong to unobserved clusters. We investigated different patterns of missing data. The results show that, without further feature engineering, no single imputation method is better than the others in all cases. The proposed methodology can be used to compare other missing data patterns and/or survival models. The Python code is accessible via the package survivalsim.
Data: PK and PD data on breast cancer bearing mice, Paris-Sud.
Commercial Paclitaxel (PTX) formulations such as Taxol are associated with adverse drug toxicities related to the added emulsionizer. Therefore, PTX formulations have been explored as they could increase efficacy and have improved pharmacokinetic (PK) properties. Recently our partners have developed a new polymer prodrug of PTX for which we proposed that metronomic dosing could result in an effective and tolerable anticancer treatment that is simultaneously convenient for patients as it allows for subcutaneous (SC) administration.
The objective of this study was to develop a PK/PD model of a novel polymer prodrug subcutaneously injectable to predict the best administration scheme.
We developed a population PK model and calibrated it with experimental data of a novel PTX-Polymer from a separate study using Monolix software.
The model was able to correctly describe treatment pharmacokinetic and efficacy profiles. Pharmacokinetic was described using a two-compartment model with an additional compartment for the polymer groupe. We selected the reduced Gompertz model as the optimal model to represent unperturbed tumor growth using available pharmacodynamic (PD) data. Efficacy was best described using Simeoni model. We discovered that adding drug resistance of the treatment efficacy was necessary to correctly fit the data. Our simulations show that administrating the prodrug daily is a promising strategy for further preclinical studies.
We developed a PK/PD model describing preclinical data of a novel anticancer polymer prodrug and validated that metronomic dosing could result in greater efficacy.
Pancreatic ductal adenocarcinoma (PDAC) has been widely studied at multiomics level. However, little is known about its specific ubiquitination, a major post-translational modification (PTM). As PTMs regulate the final function of any gene, we decided to establish the ubiquitination profiles of 60 PDAC.
We used specific proteomic tools to establish the ubiquitin dependent proteome (ubiquitinome) of frozen PDXs (Patients' derived xenographs). Then, we performed bioinformatics analysis to identify the possible associations of these ubiquitination profiles with tumour phenotype, patient survival and resistance to chemotherapies. Finally, we used proximity ligation assays (PLA) to detect and quantify the ubiquitination level of one identified marker.
We identified 38 ubiquitination site profiles correlating with the transcriptomic phenotype of tumours and four had notable prognostic capabilities. Seventeen ubiquitination profiles displayed potential theranostic marker for gemcitabine, seven for 5-FU, six for oxaliplatin and thirteen for irinotecan. Using PLA, we confirmed the use of one ubiquitination profile as a drug-response marker, directly on paraffin embedded tissues, supporting the possible application of these biomarkers in the clinical setting.
These findings bring new and important insights on the relationship between ubiquitination levels of proteins and different molecular and clinical features of PDAC patients. Markers identified in this study could have a potential application in clinical settings to help to predict response to chemotherapies thereby allowing the personalization of treatments.
Atezolizumab is an anti-PDL1 approved for treating lung cancer. A threshold of 6 μg/mL in plasma has been associated with target engagement. The extent to which patients could be overexposed with the standard 1200 mg q3w dosing remains unknown.
Here, we monitored atezolizumab peak and trough levels in 27 real-world patients with lung cancer as part of routine therapeutic drug monitoring. Individual pharmacokinetic (PK) parameters were calculated using a population approach and optimal dosing-intervals were simulated with respect to the target trough levels.
No patient had plasma levels below 6 μg/mL. The results showed that the mean trough level after the first treatment was 78.3 ± 17 μg/mL, that is, 13 times above the target concentration. The overall response rate was 55.5 percent. Low-grade immune-related adverse events was observed in 37 percent of patients. No relationship was found between exposure metrics of atezolizumab (i.e., minimum plasma concentration, maximum plasma concentration, and area under the curve) and pharmacodynamic end points (i.e., efficacy and toxicity). Further simulations suggest that the dosing interval could be extended from 21 days to 49 up to 136 days (mean: 85.7 days, i.e., q12w), while ensuring plasma levels still above the 6 μg/mL target threshold. This observational, real-world study suggests that the standard 1200 mg q3w fixed-dose regimen of atezolizumab results in significant overexposure in all the patients. This was not associated with increased side effects. As plasma levels largely exceed pharmacologically active concentrations, interindividual variability in PK parameters did not impact efficacy.
Our data suggest that dosing intervals could be markedly extended with respect to the target threshold associated with efficacy.
Successful immunotherapy is restricted to some cancers only, and combinatorial strategies with other drugs could help to improve their efficacy. Here, we monitor T cells in NSCLC model after treatment with cytotoxics (CT) and anti-vascular endothelial growth factor (VEGF) drugs, to understand when immune checkpoint inhibitors should be best associated next.
In vivo study was performed on BALB/c mice grafted with KLN205 cells. Eight treatments were tested including control, cisplatin and pemetrexed as low (LD CT) and full (MTD CT) dose as single agents, flat dose anti-VEGF and the association anti-VEGF + CT. Full immunomonitoring was performed by flow cytometry on tumor, spleen and blood over 3 weeks.
Immunomodulatory effect was dependent upon both treatments and time. In tumors, combination groups shown numerical lower Treg cells on Day 21. In spleen, anti-VEGF and LD CT group shown higher CD8/Treg ratio on Day 7; on Day 14, higher T CD4 were observed in both combination groups. Finally, in blood, Tregs were lower and CD8/Treg ratio higher, on Day 14 in both combination groups. On Day 21, CD4 and CD8 T cells were higher in the anti-VEGF + MTD CT group.
Anti-VEGF associated to CT triggers notable increase in CD8/Tregs ratio. Regarding the scheduling, a two-week delay after using anti-VEGF and CT could be the best sequence to optimize antitumor efficacy.
Azacitidine (Aza) is a mainstay of treatment for patients with acute myeloid leukaemia (AML) ineligible for induction chemotherapy and other high-risk myelodysplastic syndromes (MDS). Only half of patients respond, and almost all will eventually relapse. There are no predictive markers of response to Aza. Aza is detoxified in the liver by cytidine deaminase (CDA).
Here, we investigated the association between CDA phenotype, toxicity and efficacy of Aza in real-world adult patients.
Median overall survival (OS) was 15 months and 13 months in AML and high-risk MDS patients respectively. In addition, our data suggest that delaying Aza treatment was not associated with lack of efficacy and should not be considered a signal to switch to an alternative treatment. Half of the patients had deficient CDA activity (i.e. <2 UA/mg), with a lower proportion of deficient patients in MDS patients (34 percent) compared to AML patients (67 percent). In MDS patients, CDA deficiency correlated with longer landmark OS (14 vs. 8 months; p = 0.03), but not in AML patients.
Taken together, our data suggest that CDA is an independent covariate and may therefore be a marker for predicting clinical outcome in MDS patients treated with Aza.
Azacitidine (Vidaza®, AZA) is a mainstay for treating acute myeloid leukemia (AML) in patients unfit for standard induction and other myelodysplastic syndromes (MDS). However, only half of the patients usually respond to this drug and almost all patients will eventually relapse. Predictive markers for response to AZA are yet to be identified. AZA is metabolized in the liver by a single enzyme, cytidine deaminase (CDA). CDA is a ubiquitous enzyme coded by a highly polymorphic gene, with subsequent great variability in resulting activities in the liver. The quantitative determination of AZA in plasma is challenging due the required sensitivity and because of the instability in the biological matrix upon sampling, possibly resulting in erratic values.
We have developed and validated following EMA standards a simple, rapid, and cost-effective liquid chromatography-tandem mass spectrometry method for the determination of azacitidine in human plasma.
After a simple and rapid precipitation step, analytes were successfully separated and quantitated over a 5-500 ng/mL range. The performance and reliability of this method were tested as part of an investigational study in MDS/AML patients treated with standard azacitidine (75 mg/m2 for 7 days a week every 28 days).
Overall, this new method meets the requirements of current bioanalytical guidelines and could be used to monitor drug levels in MDS/AML patients.
kML 2.0
NanoImmuno
CetuxiMAX
IMHOTEP
MOIO
PEMBOV
REZOLVE
DROP
VIDAZAML
VENETACIBLE
Morpheus-Lung
PDC-LUNG-101
NCT03708328
PROPEL
TRIDENT-1
NCT04721015
PIONeeR
ELDERLY
NCT01817192
SAVIMMUNE
NCT04042558
NIPINEC
RESILIENT
Canopy-A
NCT04350463
TROPION-LUNG01
MERMAID-1
NCT03840902
CARMEN-LC03
SKYSCRAPER-03
NCT03899155
NCT03798535
IMbrella A
BEAT-meso
SPECTA
TROPION-Lung05
SKYSCRAPER-06
PERSEE
SAPPHIRE
Nivothym
ELEVATE HNSCC
ROMANE
XRAY VISION
Iintune-1
AHEAD-MERIT
CODEBREAK IGR
ADAPTABLE
MATISSE
DESTINY-Lung04
KRYSTAL-12
PECATI
A2A-005
Delivir
BEAMION-Lung 1
ELEVATE Lung & UC
IDE397-001
CA099-003
Anne Rodallec