Automated conditional screening of Escherichia coli knockout mutants in parallel adaptive fed-batch cultivations

In bioprocess development, the host and the genetic construct for a new biomanufacturing process are selected in the early developmental stages. This decision, made at the screening scale with very limited information about the performance of the selected cell factory in larger reactors, has a major influence on the performance of the final process. To overcome this, scale-down approaches are essential to run screenings that show the real cell factory performance at industrial like conditions. We present a fully automated robotic facility with 24 parallel mini-bioreactors that is operated by a model based adaptive input design framework for the characterization of clone libraries under scale-down conditions. The cultivation operation strategies are computed and continuously refined based on a macro-kinetic growth model that is continuously re-fitted to the available experimental data. The added value of the approach is demonstrated with 24 parallel fedbatch cultivations in a mini-bioreactor system with eight different Escherichia coli strains in triplicate. The 24 fed-batches ran under the desired conditions generating sufficient information to define the fastest growing strain in an environment with varying glucose concentrations similar to in dustrial scale bioreactors.


Introduction
Emerging technologies in robotic biolaboratories open new opportunities for both, High Throughput (HT) Screening and HT Bioprocess Development. Screening can be roughly divided into two stages; (i) the "clone library screening" (10 6 -10 12 candidates/factors), with yes/no experiments in Micro Well Plates (MWP) [1][2][3], and (ii) the stage known as "conditional screening" [4][5][6], the focus of this work. During the "conditional screening" a reduced number of candidate strains is tested with factors that significantly influence the performance at industrial scale (e.g. media, pH and temperature profiles, bioreactor heterogeneities, induction and feeding strategies [7][8][9][10][11][12]). These factors are known to affect the underlying nonlinear dynamics of the bioprocess and are par t of the very complex time-dependent interaction between the bioreactor environment and the cell factory. This highly nonlinear behavior makes it difficult to predict the effect of changes in the cultivating conditions and is responsible for the high failure rate in scale-up [13,14]. In order to overcome these challenges, experiments in conditional screening require highly advanced experimental setups able to: (i) operate as similar as possible to the industrial strategy (e.g. fed-batch or continuous cultivations), (ii) mimic the harsh conditions of industrial scale bioreactors as close as possible (e.g. growth limitation; bioreactor heterogeneities), and (iii) generate the maximal amount of information possible about the strain´s phenotype and its complex dynamic interaction with the process.
The technology to perform parallel experiments with advanced operation in fed -batch or continuous mode has recently become available [4,15,16]. Mini-BioReactors (MBR) integrated in Liquid Handling Stations (LHS) allow a large number of parallel cultivations while maintaining the properties of benchtop bioreactors. With working volumes of 2 -250 mL [17], geometric similarities to large-scale reactors [18], and high frequency measurements and analytics, MBRs have been used for process characterizations [15,[19][20][21] and scale-down studies [11,22] for up to 48 cultivations in parallel [23]. Such robotic facilities with automated cultivation control, sampling and at -line analytic operations [16,24] are very powerful systems that can accelerate bioprocess development, especially in combination with digital solutions for experiment planning [25][26][27][28], data acquisition [4,16]and realtime dynamic analysis [29,30]. The bottleneck is currently the lack of advanced computer aided tools to plan the experiments, operate the robots and build the necessary models and digital twins for scaleup and advanced process control. Because of limitations by the planning and operation capacity of humans much too often robots are on hold waiting for the next experiment to be planed, experimental campaigns need to be repeated because of failures that were not detected on time, and the same feeding strategy is used for strains with different characteristics . These are the main issues we address in the present work.
Initial attempts to solve these challenges have demonstrated the added value of model -based tools in terms of accelerating the development process and increasing robustness during scale-up [10,31,32]. Nevertheless, the existing solutions are mostly limited to single strain applications due to the complexity of the used mechanistic models and the difficulty to identify the parameters for a large number of strains at the same time [32,33]. Therefore, screening approaches often use simple blackbox models for the microorganisms, which do not allow a detailed comparison of their phenotypes. This contribution proposes an advanced conditional screening design framework that can interact with the robotic facility to run fed-batch like cultivations with feeding strategies tailored for each strain. To achieve this, (i) a model with a general macro-kinetic structure is defined with modelparameter ranges that can describe the phenotypes of all strains , and (ii) a parameter estimation is carried out for each strain to obtain a characteristic parameter set that uniquely describes it. By this we gain not only a robust and accurate prediction of the characteristics of each strain, but we also can easily quantify and confidently compare their performance. Finally, the method is applied in an online model calibration framework to adaptively define individual optimal feed start and feeding strategy. The framework provides all necessary parameters and actions t o define a wide range of alternative event triggers (e.g. depletion of glucose or consumption of acetate).
In summary, during the parallel cultivation the adaptive framework for conditional screening experiments recursively executes the following steps: (i) collection of cultivation data from the database, (ii) selection of an identifiable parameter (sub)set (PE regularization) for each strain, (iii) estimation of kinetic parameters for each clone, (iv) updating of the optimal feeding profiles for each clone, and (v) transfer of the new feeding profiles to the database ( Figure 1). As a proof of concept, parallel screening experiments with eight different strains including six knockout mutants of E. coli K-12 are conducted in 24 mini bioreactors. At the start of the experiment, virtually no information on the growth behaviour of all these strains was available. In this one experiment it was possible to identify 13 model parameters for all clones with sufficient accuracy. Figure 1: Illustration of the mode l calibration cycle in the adaptive frame work for conditional scre ening e xpe riments. On the Cultivation and analysis platform (consisting of two liquid handling stations, a mini-biore actor system) the cultivation of the clones is pe rformed, sample s are collected and autonomously analyzed. The ge ne rated online and at-line me asurements are sent to the ce ntral data storage (database). The mode l calibration cycle starts with the colle ction of all available data.
Base d on the me asurements the Sensitivity analysis is pe rformed, based on the re sults, the ide ntifiable parameters are selected, and non-ide ntifiable parameters are not considered/ fixe d in the subsequent parameter estimation. In the Parameter estimation, the ide ntifiable parameter subset is adjusted to fit the mode l to the me asurements. Based on the calibrated mode l, in the Feed calculation, the fe e d is calculate d according to pre viously de fine d crite ria and furthe r conve rted into corre sponding pulses with individual time s. The se time /pulse setpoints are stored in the Database and e xecuted dire ctly by the Cultivation and analysis platform.

HTBD facility
The high throughput bioprocess development facility is composed of two liquid handling stations (Freedom Evo 200, Tecan, Switzerland; Microlab Star, Hamilton, Switzerland) and a minibioreactor system (48 BioReactor, 2mag AG, Munich, Germany). Both liquid handling stations are connected on hardware and software level to exchange samples, process and measurement information. A detailed description of the used hardware and software framework is given in Haby et al. 2019 [16].

Cultivation
Precultures were performed with EnPresso B (Enpresso GmbH, Berlin, Germany) medium with 9 U L -1 Reagent A at 37 °C in a 24 multi well Oxodish plate to keep the cells in the exponential growth phase (PreSens GmbH, Regensburg Germany). The main culture was started as batch at 37 °C with 5 g L -1 glucose. The initial batch phase was prolonged after 1 hour by an additional feed pulse to a final concentration of 5 g L -1 glucose. The stirrer speed was kept constant at 3000 rpm. After the end of the batch phase a fed-batch was started with a pulse-based glucose feeding every 5 min with a feed solution with 400 g L -1 of glucose dissolved in deionized water. The feeding rate was increased exponentially and switched to a constant feed when the maximum pulse volume of 22 µL was reached. In total the cultivations were carried out over 8 hours with fed-batch phases of 5.4 to 6.1 hours, depending on the length of the clone-specific batch phases. The µ set for the exponential feed was chosen to be 50 % of the model-predicted µmax value and was adapted in every modelling cycle for each clone. The volume of the feed pulses was determined on the basis of the calculated feed rate. All experiments were carried out as biological triplicates, each triplicate was run on three columns in the same row of the bioreactor system.

Sampling and Analytic
During the cultivations pH and DOT were measured online in the mini bioreactor system. Each column of the bioreactor system was sampled every 45 min in a sequential mode with a sampling interval of 15 min. Samples were inactivated directly with NaOH in 96 well plates at 4 °C on the deck of the robot until further processing. After 5 samplings the sampling plates were automatically transferred to the Hamilton robot for OD600, glucose and acetate measurements in 96 well plates as described earlier [16]. For the OD600 measurements, the samples were diluted to remain in the linear range. The dilution factor was adjusted between 20 and 100 over the course of the cultivation process. All OD600 values were multiplied by a correction factor of 2.62 to convert the values to a liquid height of 1 cm. Based on the OD600 measurements the dry cell weight of the biomass was calculated by multiplying the OD600 with 0.33 [34]. Due to the time-consuming sampling and analysis procedure, the values for biomass, glucose and acetate were written to the database with a delay of 0.25 -1.35 h for the biomass and 0.66 -2 h for glucose and acetate, respectively, depending on the column of the bioreactor system where the sample was taken.
In total, during the eight hours of cultivation per reactor 1440 values for DOT and pH, respectively, were collected, as well as 23 samples for biomass (OD600) and each 20 samples for glucose and acetate measurements. For each experiment, the parameter estimation had to consider 1503 measurements leading to a sensitivity matrix of 1503x4x18.

Computational methods
The E. coli macro-kinetic growth model consists of 5 ordinary differential equations describing biomass, glucose, acetate, oxygen, and enzymatic glucose release. The model contains 18 param eters from which 13 have been shown to vary with mutations and cultivation conditions, see [36] for details. Information on the procedure and numerical implementation for the parameter estimation are given in [11]. Cultivation time and data for the different sequential tasks are summarized in Table  1. All measurements used for the parameter estimation are available in the table S1.

Parameter estimation
The parameter estimation is formulated as the following optimization problem: Where the objective function formulated as: where , ( , ) are the simulated, and , are the corresponding measured states. The index = 1, … ,5 indicates the measured variables and the index = 1, … , indicates individual datapoints. For each state the sum of the squared differences between all measured and simulated datapoints is normalized by the number of datapoints.
All computations, i.e. the numerical solution of the dynamical model, the estimation of kinetic growth parameters, and the computation of optimal cultivation conditions are written in MATLAB (The MathWorks, Inc., Natick, Massachusetts, USA). The parameter estimation is solved with the interior-point algorithm using the wrapper from MATLAB. The states of the model and their sensitivities are computed using CVODE available in the SUNDIALS Toolbox [37]. Initial values, lower and upper bounds of the parameter estimation are based on expert´s knowledge and summarized in table S2. The PE is regularized using the Subset Selection method described by Lopez et al. embedded in the optimization. The algorithm implements a stepwise forward selection of parameters to be included in the estimation problem based on the dyna mical parameter sensitivities. Identifiable parameters are selected by a ranking of all parameters according to linear independence and an analysis of the matrix rank condition of the sensitivity matrix.

Feed calculation
The exponential feed was calculated using the standard fed-batch equation [39] which was adapted to consider a pulse based profile. It is computed as : where ( ℎ −1 ) represents the feed rate at time point and µ (ℎ −1 ) the targeted specific growth rate. 0 is the initial feed rate and = 0 the time of the feed start. Since the feed in a fed-batch process is the only major volume changing factor, volume changes due to sampling is neglected at this point, the volume change could be described as The pulse volume is calculated as ] is the Yield coefficient of glucose per biomass, [ −1 ] the glucose concentration in the feed solution, 0 [ ] the biomass concentration and 0 the volume at the feed start. Volume manipulations by the pipetting robot (e.g. volume balancing, sampling, base addition for pH control) are considered in the feed calculation apart of the equations above. Biomass and volume for the calculation of 0 (eq. 6) were estimated by simulations based on the current parameter set. The end of the batch phase was defined as the time point where the predicted glucose and acetate concentrations were below 0.02 g L -1 . If the acetate consumption was slow, the feed was started anyway no later than 45 min after the depletion of glucose.

Results
Eight different E. coli K-12 clones were cultivated in parallel with an industrial process-relevant feeding design consisting of batch, exponential fed-batch and constant feed phases. The feed is applied as pulses to expose the cells to inhomogeneities similar to those in large-scale bioreactors.

Parallel cultivation
The length of the batch phase varied between the and lasted 1.65 h for E. coli W3110 (the fastest growing clone) and 1.86 h for E. coli BW25113 ΔglcB (the slowest growing clone). After the end of the batch phase the feed was automatically started. Due to the pulse nature of the feed procedure the feed start is visible through the oscillating DOT values (see Error! Reference source not found. a). These oscillations, as well as the glucose at -line data proof that glucose limitation was maintained during the fed-batch phase in all cultivations. Furthermore, no significant acetate accumulation was observed (Error! Reference source not found. b). The cultivations show a low variance between triplicates which is obvious from the online DOT and pH profiles as well as from the automatically analysed glucose and acetate values. Nine glucose data points were detected as possible outliers (at 6.14 and 7.14 hours). However, no technical issues were found to explain the sudden drift. In the case of E. coli W3110, the biomass of one triplicate was also lower, due to oxygen limitation. This could mean that the higher glucose concentration would indicate overfeeding. As expected, the pH decreased during the batch phase and started to increase after glucose depletion (typically caused by acetate consumption). During each glucose pulse cycle, perturbation of pH is observed which is caused by the transient production of acetic acid (Error! Reference source not found. c). Finally, a small increase in the pH was observed after the switch to constant pulse based feed.

Prediction of batch and feed start
The first model calibration cycle (cf. Figure 1) was initiated after 1.4 hours of batch cultivation. During the batch phase the feed start time and initial biomass were re-computed using the updated model parameters after the first measurement. The end of batch was defined as the time point at which glucose as well as the acetate (produced during overflow growth) were depleted. Therefore, the fed-batch phase in our cultivations started purposely later compared to typical fed-batch processes which are mostly started when glucose is depleted, and the DO signal increases. Note that 7 of 19 feeding was started only when acetate had been metabolised. This prevents possible overfeeding with glucose by co-metabolism of the remaining acetate and thus allowed a higher process stability.   Figure 3a illustrates the outcome of the model calibration cycle during the batch phase, at the example of the E. coli BW25113 ΔglcB cultivations data (grey cross) and simulations after model calibration (blue line). It is obvious that the first parameter estimate indicates for this strain a slower growth compared to the initial parameter set. However, with every model calibration cycl e, the computed growth rate (µ max) increased from 0.36 h -1 at t1 to 0.58 h -1 at t2 and up to 0.82 h -1 at the third shown model calibration cycle. The fit to the cultivation data is improved with each modelling cycle and the trend of the cultivation is well represented, at least after the third modelling cycle.

Preprints
In addition, due to the underestimated µ max, the first model calibration cycle failed to propose the end of the batch phase properly. An accurate estimation of the specific glucose consumption rate is only reached after the glucose had been used up, but then the estimation is very precise. Although, the end of the batch phase is equally estimated in the third model calibration cycle and in the initial unadjusted model (black dashed lines, Figure 3a . This is because of differences in the production and consumption rate of acetate resulting in different starting times of the fed-batch phase. Based on the DOT profiles, acetate was consumed after 2.5 hours; this also corresponds well with the at -line measurements of ( Figure 2, supplementary table S1). Re sults for computed fe eding profile s after the sequential task 1-3 as cumulative volume .
The predicted end of the batch phase is very close to the observed one in all cultivations even after the second model calibration and 1.5 hours of cultivation (Table 2). Due to minor variations in the initial biomass concentrations the calculated batch end differs from clone to clone, already with the initial model and with an equal parameter set. For some cultivations the time of glucose depletion was predicted with an accuracy of less than one minute (E. coli BW25113 ΔgatZ). In the worst case the time of glucose depletion was predicted 22.8 min too late (E. coli BW25113 ΔaceA). A missed batch (a) (b) end and even a short starvation phase could lead in unwanted metabolic reactions by the strain and can influence the process and product quality. However, in this cultivation triplicate, one cultivation can be considered as outlier (Figure 2a) and a difference of 22 min is still in an operational range. Due to operational reasons the model calibration with all clones was maintained. The mean difference between the observed and predicted time points for glucose depletion is 6.9 min for the calibrated model after 1.5 hours and thus better compared to the initial model with a mean prediction error of 7.3 min.
Complete consumption of acetate is only observed for five of the eight strains. For all these strains the adjusted model predicts the acetic acid consumption better compared to the initial model, with the exception of E. coli BW25113 Δomp. Complete consumption of acetic acid was not observed for three clones, because of the time depending restrictions in the feed (maximum tolerance between end of glucose depletion and feed start, see section 0). However, for these three clones the initial model predicted a faster and the adjusted model a slower acetic acid consumption rate. The times of the first feed pulse are summarized in Table 2 (Feed start), the predicted end of batch and the first pulse may differ due to technical reasons (delay in computation or first pulses are calculated with 0 µL due as the minimal pipetting volume restrictions).

Feed and fed-batch
During the fed-batch phase the size of the feed pulses is re-computed during each model calibration cycle. Based on the new parameter set the maximal glucose uptake rate was determined as basis for the new feeds. With the exception of E. coli BW25113 ΔglcB (Error! Reference source not found. (h)), the first feed rate (grey bars) was higher than the following calculated feed pulses. However, the second applied feed rates for E. coli BW25113 ΔompT, E. coli BW25113 ΔfliA and E. coli BW25113 ΔgatZ (Error! Reference source not found. (c), (d) and (f)) were close to the initial feed rates but were reduced in the later model calibration cycles. In the case of E. coli BW25113 ΔglcB the second feed is somewhat higher than the later one, which is reflected in both the initial feed rate and in the slope of the feed (all feed pulses are summarised in Error! Reference source not found.). Feed pulses are calculated by the optimisation algorithm for each strain and applied to all biologi cal triplicates. In this way, eight different feeding rates were calculated, and 24 cultivations were carried out in parallel.

Parameter estimation
During all model calibration cycles the model parameters are estimated on the basis of all available data, i.e. all data which were collected from the start of the cultivations to the actual time point. For all strains, the measurements and dynamics of cultivation are well represented in the simulation of the calibrated model as illustrated in Figure 5 for the strain E. coli BW25113 ΔglcB (last modelling cycle, for the other strains see supplementary figures S1 -7). In contrast to the calibrated model, the initial model overestimated the biomass formation. This trend could be observed for all strains. The DOT measurements indicate a slower glucose uptake than predicted. A lower specific glucose uptake rate was calculated in the first two modelling calibration cycles compared to the later ones (see Figure 3). The lack of the glucose measurement results in the first two model calibration cycles is caused by the time delay in the at -line analytics. The prediction accuracy of acetic acid is increased in the batch and fed-batch phase after model calibration. The cultivation dynamics are well fitted. The parameters to be adjusted in each model calibration cycle are selected by the included subset selection. The parameters Kap, kLa and qm, are not adjusted in one model calibration cycle; Ko, Ksq, Yofm and Yoresp are only partly selected for parameter estimation (Figure 6, all parameter sub-sets are shown the Supplementary Table S2). Regularization of parameter estimation using a subsets selection method [38] was used to ensure a meaningful parameter set. Monte Carlo simulations have shown to give a good insight into the actual, non -linear parameter distribution [40] and were therefore performed to get a better understanding of the parameter correlation.  The correlation between all parameters is very weak. Only, Kaq and Ksq showed a correlation with qAmax. Kaq and Ksq are the affinity constants for acetate and glucose uptake, respectively, and a dependence to the maximal acetate uptake rate (qAmax) cannot be avoided in the model. Parameter distribution as well as pairwise correlation of the adapted parameters for E. coli BW25113 ΔglcB (last model calibration cycle) is summarized in Figure 7. The high significance of each parameter is indicated by the narrow distribution and low variation for the most important model parameters (Table 3), especially for Error! Bookmark not defined.the parameters for Yem, qSmax and Yosresp. Normal distribution is given for all parameters except for Yam. This parameter is quite close to the lower bound of the previously defined solution space. It is noted that this situation should be avoided as it might reduce the accuracy of the parameter estimates. The parameter distribution of all other strains at the last modelling cycle are disposed in the supplementary Figures S09-S16.  In the present work, eight strains were examined in 24 successful cultivations. The end of glucose uptake was in part predicted with small errors of less than one minute, thanks to the iterative model calibration cycle. The feed start was automatic and in an operable acceptable time window using the dynamic process redesign as defined in the model calibration cycle. The parameter sets estimated are always unique and with a physiological meaning even with very little data in the initial phase of this study, e.g. the first 3 hours. This is ensured by the built -in subset selection and is proofed by the Monte Carlo simulations made afterwards.

Discussion
In this study we presented a computational framework able to design and operate parallel E. coli cultivations without human supervision. The results demonstrate that a robust operation tailored to each specific clone is possible through an adaptive input design. Undesired experimental conditions (e.g. overfeeding and starvation) are avoided while sufficient infor mation to allow for a confident discrimination of the strains is generated. Both, start time and feed rate were accurately predicted for each one of the eight strains, using feedback information from the online and at -line measurements during the cultivation. This is essential in an experimental facility aimed to perform screening cultivations for clones whose phenotype is not known beforehand. The relevance of an adaptive and specific experimental design can be seen in this case study. As illustrated in Error! Reference source not found., despite the fact the strains characteristics differ only minimally from each other, an experiment with a fixed start time and feeding rate would have violated im portant experimental constrains. Additionally, the use of a macro-kinetic grow th model that describes the main extracellular dynamics of E. coli was shown to be sufficient. In average the predicted feed start differs by less than 10 min for the optimal one, which is in an acceptable range and is mainly caused by unobserved disturbances in the system. If necessary, the mismatch can be further reduced by increasing the frequency of model adaptations.
As expected, the parameter variance in general decreases with every model calibration cycle. However, some parameters could not or only with insufficient confidence be identified. This hampered a distinction of some essential parameters as like the maximal acetate uptake rate (qAmax). Despite this, the quantification of the reliability of the outputs as presented by Anane et al [41] allows the differentiation of the main parameter, e.g. the maximal glucose uptake (qSmax) with statistical significance (see Error! Reference source not found.).
Nevertheless, statements on the performance of the clones can be made using the model parameters. The final results show that ∆ompT has the largest qSmax value and ∆gatZ the lowest. Furthermore, the parameter identifiability can be increased in future applications using methods for Optimal Experimental Design (OED) [24,42].
The macro-kinetic model used in this study is clearly insufficient to describe the complex nonlinear dynamics of the system. Still, we overcome this issue by the adaptive nature of the framework since a proper prediction within the current horizon is sufficient to assure a robust operation of the cultivations. The variations on the parameters caused by intracellular changes in the metabolic machinery together with heterogeneous mutations in the population [43] are shown in ( Figure 6). Therefore, an iterative recalculation of the feed is necessary to cope with disturbances in the experiments and inaccuracies in the model. Figure 9 Mode l unce rtainty based on parameter standard de viation: Monte Carlo simulation: Results of 1000 parame ter e stimates based on in-silico data. In silico data we re generated based on the last data se t for E. coli BW25113 ΔglcB and by random σ of 0.15 for biomass, glucose, acetate, and a σ of 0.05 for DOT.
The frequency of the parameter estimation was defined based on the availability of at -line data (biomass, glucose and acetate) and as expected, the at -line data are decisive to achieve model identifiability. Still, the results show that especially parameters related to glucose consumption can be identified using only the online DOT signal. This shows that, even though in a significantly limited manner, the framework can also be used to increase the robustness of robotic facilities that do not have embedded at-line analytics. The glucose consumption rate seems to be observable from the DOT signal by which a reduced version of the macro-kinetic model could be used to build an observer based feeding control [44]. Finally, we also demonstrated that the length of the batch phase is essential to assure sufficient data before the start of the feeding so as to allow a reliable operation of the following phases.

Conclusions
The operation of robotic experiments with mult iple fed-batch cultivations in parallel is very challenging even for skilled operators, since many decisions and tasks are needed at the same time. In this work we present an adaptive framework for conditional screening for parallel fed-batch experiments aiming to identify the best candidate strain for industrial scale biomanufacturing. We demonstrate that the use of a macro-kinetic growth model in an adaptive framework using online and at-line data information in a feedback loop is necessary to: 1. design a specific strategy for each different strain of the screening experiment 2. increase the robustness of the robotic operation against experimental disturbance, and 3. give an approximation of the reliability of the simulation results with respect to production scale performance.
To our knowledge, this is the first successful model-based operation of 24 fed-batch cultivations with as much as eight different strains in parallel including its characterization. The results clearly demonstrate the capabilities of the framework to increase the efficiency of combined mini -bioreactor systems with liquid handling stations to drastically reduce the experimental time, efforts, and failure rate in High Throughput Bioprocess Development. Table s S1 all e xpe rimental me asurements used for the modelling cycle s; Table S2: all parameter sets; Figure S1-7

Funding:
The authors acknowle dge financial support by the Ge rman Fe de ral Ministry of Education and Re se arch (BMBF) within the Europe an program EraSysApp (grant number: 031L0018A, Le anprot project), which is managed by the Project Management Age ncy Jülich (PTJ).