Individual modelling of haematotoxicity with NARX neural networks: A knowledge transfer approach.

Cytotoxic cancer therapy often results in dose-limiting haematotoxic side effects. Predicting an individual's risk is a major objective in precision medicine of cancer treatment. In this regard, patient heterogeneity presents a significant challenge. In this paper, we explore the use of hypothesis-free machine learning models based on recurrent nonlinear auto-regressive networks with exogenous inputs (NARX) as an approach to achieve this goal. Also, we propose a knowledge transfer approach to ameliorate the issue of sparse individual data, which typically hampers learning of individual networks. We demonstrate the feasibility of our approach based on a virtual patient population generated using a semi-mechanistic model of haematopoiesis and imposing different cytotoxic therapy scenarios on it. Employing different techniques of model optimisation, we derive robust and parsimonious individual networks with good generalisation performances. Moreover, we analyse in detail possible factors influencing the generalisation performance. Results suggest that our transfer learning approach using NARX networks can provide robust predictions of individual patient's response to treatment. As a practical perspective, we apply our approach to individual time series data of two patients with non-Hodgkin's lymphoma.

Individual hematotoxicity prediction of further chemotherapy cycles by dynamic mathematical models in patients with gastrointestinal tumors.

PURPOSE: Hematotoxicity is a common side-effect of cytotoxic gastrointestinal (GI) cancer therapies. An unsolved problem is to predict the individual risk therefore to decide on treatment adaptions. We applied an established biomathematical prediction model and primarily evaluated its predictive value in patients undergoing chemotherapy for GI cancers in curative intent. METHODS: In a prospective, observational multicenter study on patients with gastro-esophageal or pancreatic cancer (n = 28) receiving myelosuppressive adjuvant or neoadjuvant chemotherapy (FLO(T) or FOLFIRINOX), individual model parameters were learned based on patients' observed laboratory values during the first chemotherapy cycle and further external data resources. Grades of hematotoxicity of subsequent cycles were predicted by model simulation and compared with observed data. RESULTS: The most common high-grade hematological toxicity was neutropenia [19/28 patients (68%)]. For the FLO(T) regimen, individual grades of thrombocytopenia and leukopenia could be well predicted for cycles 2-4, as well as grades of neutropenia for cycle 2. Prediction accuracy for neutropenia in the third and fourth cycle differed by one toxicity grade on average. For the FOLFIRINOX-regimen, thrombocytopenia predictions showed a maximum deviation of one toxicity grade up to the end of therapy (8 cycles). Deviations of predictions were less than one degree on average up to cycle 4 for neutropenia, and up to cycle 6 for leukopenia. CONCLUSION: The biomathematical model showed excellent short-term and decent long-term prediction performance for all relevant hematological side effects associated with FLO(T)/FOLFIRINOX. Clinical utility of this precision-medicine approach needs to be further investigated in a larger cohort.

National and subnational short-term forecasting of COVID-19 in Germany and Poland during early 2021.

BACKGROUND: During the COVID-19 pandemic there has been a strong interest in forecasts of the short-term development of epidemiological indicators to inform decision makers. In this study we evaluate probabilistic real-time predictions of confirmed cases and deaths from COVID-19 in Germany and Poland for the period from January through April 2021. METHODS: We evaluate probabilistic real-time predictions of confirmed cases and deaths from COVID-19 in Germany and Poland. These were issued by 15 different forecasting models, run by independent research teams. Moreover, we study the performance of combined ensemble forecasts. Evaluation of probabilistic forecasts is based on proper scoring rules, along with interval coverage proportions to assess calibration. The presented work is part of a pre-registered evaluation study. RESULTS: We find that many, though not all, models outperform a simple baseline model up to four weeks ahead for the considered targets. Ensemble methods show very good relative performance. The addressed time period is characterized by rather stable non-pharmaceutical interventions in both countries, making short-term predictions more straightforward than in previous periods. However, major trend changes in reported cases, like the rebound in cases due to the rise of the B.1.1.7 (Alpha) variant in March 2021, prove challenging to predict. CONCLUSIONS: Multi-model approaches can help to improve the performance of epidemiological forecasts. However, while death numbers can be predicted with some success based on current case and hospitalization data, predictability of case numbers remains low beyond quite short time horizons. Additional data sources including sequencing and mobility data, which were not extensively used in the present study, may help to improve performance.

We compare forecasts of weekly case and death numbers for COVID-19 in Germany and Poland based on 15 different modelling approaches. These cover the period from January to April 2021 and address numbers of cases and deaths one and two weeks into the future, along with the respective uncertainties. We find that combining different forecasts into one forecast can enable better predictions. However, case numbers over longer periods were challenging to predict. Additional data sources, such as information about different versions of the SARS-CoV-2 virus present in the population, might improve forecasts in the future.

On the Parametrization of Epidemiologic Models-Lessons from Modelling COVID-19 Epidemic.

Numerous prediction models of SARS-CoV-2 pandemic were proposed in the past. Unknown parameters of these models are often estimated based on observational data. However, lag in case-reporting, changing testing policy or incompleteness of data lead to biased estimates. Moreover, parametrization is time-dependent due to changing age-structures, emerging virus variants, non-pharmaceutical interventions, and vaccination programs. To cover these aspects, we propose a principled approach to parametrize a SIR-type epidemiologic model by embedding it as a hidden layer into an input-output non-linear dynamical system (IO-NLDS). Observable data are coupled to hidden states of the model by appropriate data models considering possible biases of the data. This includes data issues such as known delays or biases in reporting. We estimate model parameters including their time-dependence by a Bayesian knowledge synthesis process considering parameter ranges derived from external studies as prior information. We applied this approach on a specific SIR-type model and data of Germany and Saxony demonstrating good prediction performances. Our approach can estimate and compare the relative effectiveness of non-pharmaceutical interventions and provide scenarios of the future course of the epidemic under specified conditions. It can be translated to other data sets, i.e., other countries and other SIR-type models.

Individual prediction of thrombocytopenia at next chemotherapy cycle: Evaluation of dynamic model performances.

AIMS: Thrombocytopenia is a common major side-effect of cytotoxic cancer therapies. A clinically relevant problem is to predict an individual's thrombotoxicity in the next planned chemotherapy cycle in order to decide on treatment adaptation. To support this task, 2 dynamic mathematical models of thrombopoiesis under chemotherapy were proposed, a simple semimechanistic model and a comprehensive mechanistic model. In this study, we assess the performance of these models with respect to existing thrombocytopenia grading schemes. METHODS: We consider close-meshed individual time series data of 135 non-Hodgkin's lymphoma patients treated with 6 cycles of CHOP/CHOEP chemotherapies. Individual parameter estimates were derived on the basis of these data considering a varying number of cycles per patient. Parsimony assumptions were applied to optimize parameter identifiability. Models' predictability are assessed by determining deviations of predicted and observed degrees of thrombocytopenia in the next cycles. RESULTS: The mechanistic model results in better agreement of model prediction and individual time series data. Prediction accuracy of future cycle toxicities by the mechanistic model is higher even if the semimechanistic model is provided with data of more cycles for calibration. CONCLUSION: We successfully established a quantitative and clinically relevant method for assessing prediction performances of biomathematical models of thrombopoiesis under chemotherapy. We showed that the more comprehensive mechanistic model outperforms the semimechanistic model. We aim at implementing the mechanistic model into clinical practice to assess its utility in real life clinical decision-making.

Integration of mathematical model predictions into routine workflows to support clinical decision making in haematology.

BACKGROUND: Individualization and patient-specific optimization of treatment is a major goal of modern health care. One way to achieve this goal is the application of high-resolution diagnostics together with the application of targeted therapies. However, the rising number of different treatment modalities also induces new challenges: Whereas randomized clinical trials focus on proving average treatment effects in specific groups of patients, direct conclusions at the individual patient level are problematic. Thus, the identification of the best patient-specific treatment options remains an open question. Systems medicine, specifically mechanistic mathematical models, can substantially support individual treatment optimization. In addition to providing a better general understanding of disease mechanisms and treatment effects, these models allow for an identification of patient-specific parameterizations and, therefore, provide individualized predictions for the effect of different treatment modalities. RESULTS: In the following we describe a software framework that facilitates the integration of mathematical models and computer simulations into routine clinical processes to support decision-making. This is achieved by combining standard data management and data exploration tools, with the generation and visualization of mathematical model predictions for treatment options at an individual patient level. CONCLUSIONS: By integrating model results in an audit trail compatible manner into established clinical workflows, our framework has the potential to foster the use of systems-medical approaches in clinical practice. We illustrate the framework application by two use cases from the field of haematological oncology.

Modeling individual time courses of thrombopoiesis during multi-cyclic chemotherapy.

BACKGROUND: Thrombocytopenia is a major side-effect of cytotoxic cancer therapies. The aim of precision medicine is to develop therapy modifications accounting for the individual's risk. METHODOLOGY/PRINCIPLE FINDINGS: To solve this task, we develop an individualized bio-mechanistic model of the dynamics of bone marrow thrombopoiesis, circulating platelets and therapy effects thereon. Comprehensive biological knowledge regarding cell differentiation, amplification, apoptosis rates, transition times and corresponding regulations are translated into ordinary differential equations. A model of osteoblast/osteoclast interactions was incorporated to mechanistically describe bone marrow support of quiescent cell stages. Thrombopoietin (TPO) as a major regulator is explicitly modelled including pharmacokinetics and-dynamics of TPO injections. Effects of cytotoxic drugs are modelled by transient depletions of proliferating cells. To calibrate the model, we used population data from the literature and close-meshed individual data of N = 135 high-grade non-Hodgkin's lymphoma patients treated with CHOP-like chemotherapies. To limit the number of free parameters, several parsimony assumptions were derived from biological data and tested via Likelihood methods. Heterogeneity of patients was explained by a few model parameters. The over-fitting issue of individual parameter estimation was successfully dealt with a virtual participation of each patient in population-based experiments. The model qualitatively and quantitatively explains a number of biological observations such as the role of osteoblasts in explaining long-term toxic effects, megakaryocyte-mediated feedback on stem cells, bi-phasic stimulation of thrombopoiesis by TPO, dynamics of megakaryocyte ploidies and non-exponential platelet degradation. Almost all individual time series could be described with high precision. We demonstrated how the model can be used to provide predictions regarding individual therapy adaptations. CONCLUSIONS: We propose a mechanistic thrombopoiesis model of unprecedented comprehensiveness in both, biological mechanisms considered and experimental data sets explained. Our innovative method of parameter estimation allows robust determinations of individual parameter settings facilitating the development of individual treatment adaptations during chemotherapy.

Predicting time to castration resistance in hormone sensitive prostate cancer by a personalization algorithm based on a mechanistic model integrating patient data.

BACKGROUND: Prostate cancer (PCa) is a leading cause of cancer death of men worldwide. In hormone-sensitive prostate cancer (HSPC), androgen deprivation therapy (ADT) is widely used, but an eventual failure on ADT heralds the passage to the castration-resistant prostate cancer (CRPC) stage. Because predicting time to failure on ADT would allow improved planning of personal treatment strategy, we aimed to develop a predictive personalization algorithm for ADT efficacy in HSPC patients. METHODS: A mathematical mechanistic model for HSPC progression and treatment was developed based on the underlying disease dynamics (represented by prostate-specific antigen; PSA) as affected by ADT. Following fine-tuning by a dataset of ADT-treated HSPC patients, the model was embedded in an algorithm, which predicts the patient's time to biochemical failure (BF) based on clinical metrics obtained before or early in-treatment. RESULTS: The mechanistic model, including a tumor growth law with a dynamic power and an elaborate ADT-resistance mechanism, successfully retrieved individual time-courses of PSA (R(2) = 0.783). Using the personal Gleason score (GS) and PSA at diagnosis, as well as PSA dynamics from 6 months after ADT onset, and given the full ADT regimen, the personalization algorithm accurately predicted the individual time to BF of ADT in 90% of patients in the retrospective cohort (R(2) = 0.98). CONCLUSIONS: The algorithm we have developed, predicting biochemical failure based on routine clinical tests, could be especially useful for patients destined for short-lived ADT responses and quick progression to CRPC. Prospective studies must validate the utility of the algorithm for clinical decision-making.

Complex pattern of interleukin-11-induced inflammation revealed by mathematically modeling the dynamics of C-reactive protein.

Inflammation underlies many diseases and is an undesired effect of several therapy modalities. Biomathematical modeling can help unravel the complex inflammatory processes and the mechanisms triggering their emergence. We developed a model for induction of C-reactive protein (CRP), a clinically reliable marker of inflammation, by interleukin (IL)-11, an approved cytokine for treatment of chemotherapy-induced thrombocytopenia. Due to paucity of information on the mechanisms underlying inflammation-induced CRP dynamics, our model was developed by systematically evaluating several models for their ability to retrieve variable CRP profiles observed in IL-11-treated breast cancer patients. The preliminary semi-mechanistic models were designed by non-linear mixed-effects modeling, and were evaluated by various performance criteria, which test goodness-of-fit, parsimony and uniqueness. The best-performing model, a robust population model with minimal inter-individual variability, uncovers new aspects of inflammation dynamics. It shows that CRP clearance is a nonlinear self-controlled process, indicating an adaptive anti-inflammatory reaction in humans. The model also reveals a dual IL-11 effect on CRP elevation, whereby the drug has not only a potent immediate influence on CRP incline, but also a long-term influence inducing elevated CRP levels for several months. Consistent with this, model simulations suggest that periodic IL-11 therapy may result in prolonged low-grade (chronic) inflammation post treatment. Future application of the model can therefore help design improved IL-11 regimens with minimized long-term CRP toxicity. Our study illuminates the dynamics of inflammation and its control, and provides a prototype for progressive modeling of complex biological processes in the medical realm and beyond.

Personalizing oncology treatments by predicting drug efficacy, side-effects, and improved therapy: mathematics, statistics, and their integration.

Despite its great promise, personalized oncology still faces many hurdles, and it is increasingly clear that targeted drugs and molecular biomarkers alone yield only modest clinical benefit. One reason is the complex relationships between biomarkers and the patient's response to drugs, obscuring the true weight of the biomarkers in the overall patient's response. This complexity can be disentangled by computational models that integrate the effects of personal biomarkers into a simulator of drug-patient dynamic interactions, for predicting the clinical outcomes. Several computational tools have been developed for personalized oncology, notably evidence-based tools for simulating pharmacokinetics, Bayesian-estimated tools for predicting survival, etc. We describe representative statistical and mathematical tools, and discuss their merits, shortcomings and preliminary clinical validation attesting to their potential. Yet, the individualization power of mathematical models alone, or statistical models alone, is limited. More accurate and versatile personalization tools can be constructed by a new application of the statistical/mathematical nonlinear mixed effects modeling (NLMEM) approach, which until recently has been used only in drug development. Using these advanced tools, clinical data from patient populations can be integrated with mechanistic models of disease and physiology, for generating personal mathematical models. Upon a more substantial validation in the clinic, this approach will hopefully be applied in personalized clinical trials, P-trials, hence aiding the establishment of personalized medicine within the main stream of clinical oncology.

An integrated disease/pharmacokinetic/pharmacodynamic model suggests improved interleukin-21 regimens validated prospectively for mouse solid cancers.

Interleukin (IL)-21 is an attractive antitumor agent with potent immunomodulatory functions. Yet thus far, the cytokine has yielded only partial responses in solid cancer patients, and conditions for beneficial IL-21 immunotherapy remain elusive. The current work aims to identify clinically-relevant IL-21 regimens with enhanced efficacy, based on mathematical modeling of long-term antitumor responses. For this purpose, pharmacokinetic (PK) and pharmacodynamic (PD) data were acquired from a preclinical study applying systemic IL-21 therapy in murine solid cancers. We developed an integrated disease/PK/PD model for the IL-21 anticancer response, and calibrated it using selected "training" data. The accuracy of the model was verified retrospectively under diverse IL-21 treatment settings, by comparing its predictions to independent "validation" data in melanoma and renal cell carcinoma-challenged mice (R(2)>0.90). Simulations of the verified model surfaced important therapeutic insights: (1) Fractionating the standard daily regimen (50 µg/dose) into a twice daily schedule (25 µg/dose) is advantageous, yielding a significantly lower tumor mass (45% decrease); (2) A low-dose (12 µg/day) regimen exerts a response similar to that obtained under the 50 µg/day treatment, suggestive of an equally efficacious dose with potentially reduced toxicity. Subsequent experiments in melanoma-bearing mice corroborated both of these predictions with high precision (R(2)>0.89), thus validating the model also prospectively in vivo. Thus, the confirmed PK/PD model rationalizes IL-21 therapy, and pinpoints improved clinically-feasible treatment schedules. Our analysis demonstrates the value of employing mathematical modeling and in silico-guided design of solid tumor immunotherapy in the clinic.

Critical-point analysis for three-variable cancer angiogenesis models.

We perform critical-point analysis for three-variable systems that represent essential processes of the growth of the angiogenic tumor, namely, tumor growth, vascularization, and generation of angiogenic factor (protein) as a function of effective vessel density. Two models that describe tumor growth depending on vascular mass and regulation of new vessel formation through a key angiogenic factor are explored. The first model is formulated in terms of ODEs, while the second assumes delays in this regulation, thus leading to a system of DDEs. In both models, the only nontrivial critical point is always unstable, while one of the trivial critical points is always stable. The models predict unlimited growth, if the initial condition is close enough to the nontrivial critical point, and this growth may be characterized by oscillations in tumor and vascular mass. We suggest that angiogenesis per se does not suffice for explaining the observed stabilization of vascular tumor size.