Patent Application
20230168662
The following description relates to processes for the production of chemical, pharmaceutical and/or biotechnological products. In particular, aspects of the application relate to optimising such processes as a whole in terms of overall performance.
Processes for the production of chemical, pharmaceutical and/or biotechnological products are complex procedures involving non-linear dynamics in the transformation of the inputs or ingredients into the final products. An example of such a process is upstream cell cultivation in drug manufacture. Design of this process may involve considerations about the resulting drug product (at large scale), about how to adjust the process to optimise the resulting drug product (at intermediate scale) or about identifying a cell line best able to do so (at small scale). At each scale, interactions between the organism, cultivation equipment, reagents, physical conditions and the timing of the steps of the process determine success or failure; therefore process design is critical.
However, determining how best to perform the production process for a particular organism, piece of cultivation equipment or set of reagents is challenging, because responses of cultivation to driving parameters is typically highly non-linear and the space of possible parameters regulating the production process is very large.
Conventionally, the optimisation of such processes has focused on identifying and optimising for, and then achieving a particular set-point, i.e. a target value for a single process output. Such a simplistic approach underestimates the complex interplay between the different factors characterizing the performance of the process. For example, maximising the product titre does not necessarily lead to the best way of performing the process. Additionally, the complex interplay between the different factors driving the outcome of the process is underestimated in conventional approaches.
Further, conventionally, available knowledge about a process, which may come from theoretical assumptions and previous executions, is taken into consideration only in a rudimental manner, e.g. selecting a priori a value for a parameter or a model for the physical and biological dynamics involved in the process. For example, in the design of experiments (DoE), the shape (e.g. linear or quadratic) of the response of the process to the driving parameters is usually selected a priori.
It is an object of the invention to optimise a process for the production of chemical, pharmaceutical and/or biotechnological products, wherein the process is regarded in its entirety and complexity. In particular, it is an object to account accurately for the plurality of factors affecting the execution and possible outcomes of the process and to duly consider the plural aspects characterizing the performance of the process. In this context, it is also an object to integrate effectively available knowledge about the process in the optimisation procedure.
The achievement of these objects in accordance with the invention is set out in the independent claims. Further developments of the invention are the subject matter of the dependent claims.
According to one aspect, a computer-implemented method of determining at least one recipe for a production process to produce a chemical, pharmaceutical and/or biotechnological product is provided. In particular, the determined recipe(s) may be the optimal recipe(s) in the sense that the production process is optimised in terms of a utility function, as will be explained below.
Examples of processes according to the present application are industrial processes, particularly biopharmaceutical processes. Other examples include research and development processes or scientific research.
Examples of inputs or ingredients for a production process according to the present application may include biological material, i.e. materials comprising a biological system, such as cells, cell components, cell products, and other molecules, as well as materials derived from a biological system, such as proteins, antibodies and growth factors. Further ingredients may include chemical compounds and various substrates.
Examples of inputs may include gasses and liquids. Gasses include any or all of air, oxygen, nitrogen, oxygen-enriched air and carbon dioxide. Liquids are typically:
A production process of the present application may involve chemical or microbiological conversion of material in conjunction with the transfer of mass, heat, and momentum. The process may include homogeneous or heterogeneous chemical and/or biochemical reactions. The process may comprise but is not limited to mixing, filtration, purification, centrifugation and/or cell cultivation. The production process may involve chemical or biological reactions that take time to complete, e.g. 6 hours for an E. coli microbial batch, 18 days for a mammalian cell cultivation and 60 days for a mammalian perfusion process.
In particular, “producing” a chemical, pharmaceutical and/or biotechnological product indicates any processing of the inputs, including but not limited to modifying a state of any of the inputs (e.g. changing the temperature, oxygen content etc. thereof), combining any of the inputs reversibly or irreversibly, and using the inputs for creating new material.
Possible products may include a transformed substrate, baker’s yeast, lactic acid culture, lipase, invertase, rennet. Further exemplary biopharmaceutical products that can be produced according to the techniques described in the present application include the following: recombinant and non-recombinant proteins, vaccines, gene vectors, DNA, RNA, mRNA, antibiotics, secondary metabolites, growth factors, cells for cell therapy or regenerative medicine, half-synthesized products (e.g. artificial organs). Various production systems may be used to facilitate the process, e.g. cell based systems such as animal cells (e.g. CHO, HEK, PerC6, VERO, MDCK), insect cells (e.g. SF9, SF21), microorganisms (e.g. E. coli, S. ceravisae, P. pastoris, etc.), algae, plant cells, cell free expression systems (cell extracts, recombinant ribosomal systems, etc.), primary cells, stem cells, native and gene manipulated patient specific cells, matrix based cell systems.
Exemplarily, the production process may be a batch process, in which a specific amount of feed medium for feeding an organism is provided as an initial condition and then a control period follows. The production process may be a fed batch process. The fed batch process may involve a culture in which a base medium supports initial cell culture and a feed medium is added to support further growth and/or product production, once the initial nutrients have been depleted. In other words, the fed batch process may involve a batch period followed by a feed period. The production process may be a perfusion process, in which a batch period is followed by a feed period with continual removal of e.g. the product, e.g. by filtration.
Exemplarily, the process may be a mammalian and/or microbial cultivation process.
Techniques described in the present application may be useful for bioreactor processes, and for processes carried out at other levels of production.
The production process is defined by a plurality of steps specified by one or more recipe parameters controlling an execution of the production process.
The production process is defined by the sequence of steps that is performed in order to produce the product and/or produce a particular outcome relating to the product. Some steps may occur simultaneously to one another and other steps may occur in a temporal sequence after one another. A step may correspond to an action carried out during the production process, wherein the action may be passive, such as waiting for an event to occur (such as an increase in oxygen levels due to inactivity of the culture), or active, such as causing an event to occur (e.g. stirring or adding a fluid) or setting values and/or profiles for a given quantity.
Exemplarily, steps that perform actions within a production process in a bioreactor may be typified by the following non-exhaustive list:
Further, the process may comprise step types that describe the flow of execution, e.g. specifying how/when an event occurs, such as: repeat a step or steps until a condition is met and/or for a specified number of iterations, choose between various options depending on state of bioreactor, wait until a condition becomes true (e.g. wait until dissolved oxygen rises to indicate end of batch phase before starting feed), perform a step or set of steps concurrently (e.g. perform pH control at the same time as temperature control).
The execution of the plurality of steps is controlled by one or more recipe parameters. Exemplarily, the plurality of steps may be specified by a plurality of recipe parameters, with each step being specified by one or more recipe parameters.
In other words, the production process may include (i.e. may be performed according to) at least one recipe parameter that has an influence on performance of the process (e.g., product titre, quality attributes) and the product produced by the process. The recipe parameters control the execution of the process in the sense that they influence the course of the process.
Specifically, the recipe parameters may be controllable, i.e. the values of the recipe parameters may be specifically adjusted prior to and/or during performing the process. In other words, at least one of the recipe parameters can be set e.g. by an operator or by a control system. In particular, the recipe parameters may be the ones describing and/or governing a state and/or behaviour of the equipment used for the production process, e.g. a bioreactor.
In the case of a bioreactor the recipe parameters may include but are not limited to:
Exemplarily, the recipe parameters may include (but are not limited to) any of the following: stir speed, sampling frequency and volume, total gassing rate, harvest time, glucose concentration in feed, timing, frequency or volume of feed, set points for concentration of dissolved oxygen or other gases, sheer stresses applied to cell population (e.g. from vigorous stirring), amino acid or other nutrient concentration in feed, rate of media replacement by perfusion, inoculum volume, inoculum concentration and trigger points for next stage of process (e.g. %O2), parameters defining feeding strategy (bolus vs continuous), temperature profile, pH profile, control approach (set points, P, I and D parameters in proportional-integral-derivative (PID) control), execution times and wait times, volumes and concentrations or various reagents, feeds or other liquids, fill volume, magnitude and timings of changes/shifts to any of the parameters or set points of variables. Profiles may be:
An exemplary recipe with recipe parameters A to X may be the following:
In particular, the recipe parameters may be given a constant, fixed value expressed by a given number. The values for the recipe parameters may be set at the beginning of the execution of the production process or may be dynamically determined during the execution of the production process. For example, set points and profiles may be dynamically determined to take into account events arising during the production process.
A given step in a recipe may be specified by a given recipe parameter in different ways. In some cases, the value assigned to the recipe parameter is directly used, as in the step “fill bioreactor with A L of media”. In other cases, the value assigned to the recipe parameter may be “processed”. For example, the recipe parameter may be used as argument of a mathematical expression and/or a conditional statement. The mathematical expression and/or a conditional statement may take one or more additional arguments beside the value of the recipe parameters. The other arguments may include, but are not limited to, a process variable (see below), time, a property of the machine onto which the recipe is deployed (e.g. bioreactor volume), or a combination thereof.
Exemplarily one step of the recipe may be “when the dissolved oxygen (DO) drops below 20%, then supplement oxygen”. The choice of the recipe parameter “20%” is predetermined, but how this manifests in a given run will vary. The same applies for a simulation run, in which the time at which the DO will drop below 20%, if it does so at all, may vary depending, among others, on any or a combination of the following factors: the process evolution information, the machine specification and the stochastic realisation (these factors will be discussed below).
A recipe comprises the plurality of steps defining the production process. In other words, a recipe is a structured representation of the production process, e.g. of the activity of a bioreactor. This means that the plurality of steps are expressed in a structured manner, e.g. in a format that can be interpreted by a machine.
As already explained above, there are steps indicating an action (passive or active) and steps controlling the flow, e.g. sequence (execute the contained steps in sequence), repeat (repeat the contained steps until a condition is met or a given number of times) and decision (depending on a condition, perform one step or another).
As discussed, the actual behaviour of the steps during (real or simulated) execution, i.e. what happens, is determined by one or more of the recipe parameters and, if applicable, their predetermined dependency to the response of the process variables, i.e. the status of the process.
Accordingly, the recipe comprises the plurality of steps as well as one or more values for the recipe parameters specifying the steps. In other words, the recipe provides a well-defined procedure that can be directly implemented when executing the production process and that dictates how the process equipment is controlled with time.
As explained above, the values of the recipe parameters in a recipe are predetermined, even if they may be dynamically fed to the recipe during execution (e.g. when they depend on the process variables). Conversely, a recipe template is a recipe in which at least one of the recipe parameters specifying the plurality of steps is a variable recipe parameter being variable and having no predetermined value at the outset.
The method comprises retrieving one or more recipe template composites, wherein a recipe template composite comprises one or more recipe templates.
As explained, a recipe template involves one or more degrees of freedom on how to execute the production process. Different values for the at least one variable recipe parameter result in different recipes being produced from a given recipe template. The difference may be in the values of the recipe parameters only or also in the sequence of steps, e.g. if a path of execution within a recipe template is contingent on a variable recipe parameter. Multiple recipe parameters or only one recipe parameter might be free within one recipe template.
Exemplary recipe parameters that would be free to vary (i.e. variable recipe parameters) may include but are not limited to one or more of the following:
Different recipe templates may be used at granularity of different organisms, clones, or production processes, deployment systems (e.g. disposable vs stainless steel reusable production bioreactors). Alternatively or in addition, a given recipe template may be used across a range of different organisms, clones, production processes or deployment systems, capitalising on the use of recipe parameters that are themselves functions of variables or parameters, for example, the bioreactor volume.
Recipe templates might also be shared between organisations or categorised on the web in a repository. They might be selected based on user feedback.
Recipe templates may be part of a library. For example, recipe templates may be stored in a structured data format (e.g. persisted by a serialiser in XML). Recipe templates could equally be stored in a database, spreadsheet, and stored locally, in an organisation file system, in a database within an organisation or on the cloud (remotely).
Further, each recipe template composite comprises one or more recipe templates. In the cases in which the recipe template composite comprises only one recipe template, the composite coincides with the recipe template. In other cases, the recipe template composite comprises a plurality of recipe templates, wherein the recipe templates may differ in the execution of one or more steps or may be identical. The recipe templates in the same recipe template composite may have the same variable recipe parameter(s), a subset of the variable recipe parameters may be shared between the constituent recipe templates, or none may be shared. The plural recipe templates in the composite are meant to be employed for parallel, sequential or overlapping executions of the production process, e.g. on plural bioreactors.
There may be several different motivations for having multiple executions of the production process in parallel, e.g. if a large amount of titre is needed and the outputs of the runs should be combined in order to obtain it, or in order to address the risk of one of the runs failing, or in order to obtain composite information from the multiple processes.
As explained below, the recipe template composites are used to simulate the production process. In the case of more than one recipe template composite, running a plurality of simulations respectively based on the plurality of recipe template composites enables the method to explore a wider range of available alternatives to find the optimal recipe(s), with respect to a case in which a single recipe template composite is considered.
The method further comprises retrieving process evolution information that describes the time evolution of process variable(s) that describe a state of the production process, wherein the process evolution information comprises evolution parameter(s) and at least one of the evolution parameter(s) is a variable evolution parameter being variable and having no predetermined value at the outset.
The process variables describe the state of the production, which is of course at least partly determined by the behaviour of the equipment, as well as by the specifics of the equipment and the inputs (e.g. organism) used. The values of these process variables may be intrinsic to the production process and not directly adjustable. However, they may be indirectly adjustable by modifying factors that affect them, such as the recipe parameters. Further, process variables may influence each other.
The values of the process variables change during the production process. The measured value of a process variable at a given time during the process (either executed in the real world or simulated) is usually referred to as “process value”.
For example, the process variables for a production process in a bioreactor may be classified as:
The process evolution information characterises how the one or more process variables change with time, including initial conditions for the process variables. In particular, the process evolution information may comprise relations empirically derived from previous executions of the production process and/or equations derived by theoretical model about the evolution of the production process. The process evolution information may comprise the explicit dependence of one or more variables on time together with relations linking the process variables to each other.
Process evolution information may have a number of constituent parts e.g. process evolution information related to cell dynamics, to bioreactor dynamics and/or to chemical reactions occurring within the bioreactor.
The process evolution information may also describe events related and going beyond what may be strictly considered to be the production process per se. In particular, the process evolution information may also describe what happens outside a bioreactor e.g. time evolution of samples taken, or in a secondary reactor vessel, or in a downstream processing facility (e.g. a purification unit), and/or in a piece of analytics equipment.
In one example, the cell culture constituent of the process evolution information is modelled using a hierarchical set of ordinary differential equations describing “cell processes”, where any individual cell process describes, amongst other things, the dependencies of that cell process, i.e. how its rate depends on pH, DO, temperature, and concentrations of various nutrients, and the results of that process, i.e. the change that occurs in cell count, titre, pH, DO etc. as a result of that process being active. It is generally noted that a “set” can comprise one or more elements.
A given cell process A is allowed to depend on one or more driving processes, X, Y ... such that the rate of A is computed and then multiplied by either the sum or product of the rates of X, Y. A given cell process may depend on no driving cell process or any number of driving cell processes, and a given cell process may drive no other cell process or any number of other cell processes.
In a very simple case, for example, where there is just one dependency (on temperature) and one consequence (cell growth), this amounts to solving the differential equation:
where p is cell density, t time, rg maximum growth rate, T temperature, Topt optimum growth temperature, Tsens indicates the sensitivity of the growth to the temperature, and N(x,s) denotes the value of a normal distribution with standard deviation s at x.
A more typical case would have considerably more dependencies (e.g. dependency on key nutrients) and consequences (e.g. reduction in DO due to cell activity, elevation of temperature in the case of microbial cell activity, reduction in quantities of nutrients etc). In addition, there may be a number of such cell processes with an additive effect e.g. constitutive growth, death due to toxin presence, product accumulation, etc.
The process evolution information for the cell growth then amounts to describing the “cell processes” variables and dependencies on each other. This may take several forms:
These data might then be stored in several ways:
The data might further then be stored and accessed:
In addition to the cell culture, physical dynamics of the process are described in the process evolution information. This comprises how the process variables are related to each other at a point in time and how each process variable evolves over time.
As with the cell culture model, the model for physical dynamics may be stored as XML, database tables, and so on and the substrate for the data could be DVD, CDROM, hard disk etc., and the location of the data local or distinct, and the distribution of the data in one place or distributed.
To summarize, the data representing the process evolution information may be retrieved according to a plurality of different implementations.
The process evolution information is determined by one or more evolution parameters. These evolution parameters (which may be also referred to as “model parameters”) are the factors that characterize how the one or more process variables depend on time and/or on other process variables. The evolution parameters are the parameters that appear in the empirically derived relations and/or theoretical equations. In particular, the evolution parameters may have values that do not depend on time, i.e. that do not vary with time.
Referring to the exemplary differential equation discussed above for cell growth, p (the cell density) and T (the temperature) represent process variables, while rg (maximum growth rate), Topt (optimum growth temperature) and Tsens (sensitivity of the growth to the temperature) represent evolution parameters.
More generally, evolution parameters may exemplarily include any of the following: growth rates, rate of apoptosis or lysis or other change of state, rate of production of inducer or inhibitors, rates or production or consumption of nutrients or products, sensitivities or other descriptive response parameters (e.g. critical values, capacities etc.) to external factors such as pH/pressure/dissolved oxygen (DO)/temperature/CO2/ultraviolet or visible light or other chemical or physical factors, sensitivities or other descriptive response parameters to nutrients or chemical gradients, cytokine concentration, growth rate variability (in the context of a stochastic differential equation), oxygen consumption by cell type distribution, crowding/capacity constants (e.g. max viable cell density), decay or degradation rates of nutrients, products or other, rates of change, relative or absolute magnitudes of changes or shifts, timing or control parameters of any of the above.
In particular, the external factors to which the sensitivities or other descriptive parameters relate may include factors that are set by recipe parameters. For example, the growth of cell may depend on pH, which may be set according to a corresponding recipe parameter.
As explained, the process evolution information provides a model of how the production process behaves on the basis of biological, physical, chemical etc. considerations about the phenomena involved in the production process. Such a model is necessary to simulate the production process, as will be discussed below, and acquire valuable information from the simulations on how to run the actual process.
If the evolution parameters of the process evolution information have a fixed, predetermined or pre-determinable value, a single view on the science behind the process is taken. However, the production processes are rather complex and such an assumption may unduly influence the outcome of the process simulation. In other words, the determination of the optimal recipe(s) may be affected by the specific choice for the values of the evolution parameters.
Accordingly, in order to determine more accurately the optimal recipe(s) by mitigating the effect of the assumed process evolution information, at least one of the evolution parameters is a variable evolution parameter being variable and having no predetermined value at the outset. Exemplarily, only a proper subset of the evolution parameters may be variable. In other examples, all evolution parameters may be left free to vary.
The presence of one or more variable evolution parameters entails a plurality of sets of values for the set of evolution parameters. For example, if there is one variable evolution parameter in a set of three evolution parameters and n different values v1, ..., vn are assigned to the variable evolution parameter, there will be n sets of values: (v1, fa, fb), (v2, fa, fb), ..., (vn, fa, fb), with fa and fb being the fixed values for the other two evolution parameters. The number of different values considered for any of the evolution parameters may differ from the number considered for another evolution parameter. In reference to the likelihood function below, the number of different values characterised or evaluated may be pre-determined or dynamically determined until adequate accuracy is reached, depending on the precise algorithm used.
As mentioned, the process evolution information is used to simulate the production process. Running a plurality of simulations corresponding to a plurality of sets of values for the evolution parameters enables the method to explore a wider range of available alternatives to find the optimal recipe(s), with respect to a case in which only a set of values is used for the evolution parameters.
Indeed, conventionally, insights into biological, physical and chemical dynamics are frequently expressed in terms of a single model of choice and further refined through determining the best fit parameters for that model of choice. However, this over-represents certainty about the science behind the mechanisms, and under-represents variation and uncertainty around the most likely explanation for observed data.
It should be noted that having variable evolution parameters does not only enable one to account for uncertainty in the model parameters but also for uncertainty in the choice of model itself. Indeed, the process evolution information may comprise a plurality of competing models, each having different evolution parameters. In a merely illustrative example, there may be two competing models that describe the evolution of process variables X and Y, wherein the first model comprises equations
and the second model comprises equations
Thus, the first model has evolution parameters (a, b) and the second model has evolution parameters (c, d), and the variable evolution parameters of the process evolution information may be (a, b, c, d). By assigning values (a′, b′, 0, 0) or (0, 0, c′, d′) the first or the second model, respectively, may be selected, so that the production process may be simulated according to different dynamics. In some examples, if no value is assigned to a given evolution parameter, it may be assumed it is zero by default. Accordingly, choosing different sets of values for the evolution parameters may, in some cases, mean choosing different models for the production process.
The values for the one or more variable evolution parameters may be assigned on the basis of a likelihood function, as will be explained below.
The method further comprises retrieving a likelihood function for the evolution parameters, wherein the likelihood function provides a likelihood that value(s) for the evolution parameter(s) are feasible for the production process. In other words, the likelihood function provides a measure of how adequately a certain value for an evolution parameter can describe the production process as modelled by the process evolution information. Said otherwise, the likelihood function maps possible values for evolution parameters to the likelihood that these values correctly describe the process evolution, i.e. the time evolution of process variable(s) that describe the state of the production process.
In one example, the likelihood function may be the rectangular distribution, meaning that all values are equally feasible. In other examples, the likelihood function may be different from the rectangular distribution and may describe which value(s) is/are more feasible than others. In particular, the likelihood function may be a multivariate probability distribution, in the case of a plurality of evolution parameters.
The likelihood function may be a function of all evolution parameters, i.e. variable evolution parameters and non-variable or fixed evolution parameters. In this case, the likelihood function may be considered as being the product of the marginal distribution for the variable evolution parameters and the marginal distribution for the fixed evolution parameters, wherein the latter is a delta distribution, possibly multi-dimensional, peaked on the value(s) actually assumed by the fixed evolution parameter(s). Otherwise, the likelihood function may be more simply expressed only as a function of the variable evolution parameter(s).
Possible shapes of the likelihood function include, but are not limited to, any of the following:
The likelihood function may be stored in a database and stored locally, in a local or an organisation file system, in a database within an organisation or on the cloud (remotely).
The likelihood function describes and can be visualised as a likelihood space. By extension, the likelihood function may be referred to as “likelihood space”, “probability space” or “knowledge space”.
The likelihood function may be determined on the basis of prior knowledge about the values of the evolution parameters. This prior knowledge may be derived from any of the following sources and any combinations thereof: data from literature, personal experience of a user, data based on theoretical considerations, data from completed runs of the same production process, data from ongoing runs of the same production process, data from completed or ongoing runs of production processes related to the production process in question. These options will be discussed singularly in the following. The term “run” is herein used for actual execution of the production process in the real world, while “simulation run” is used for the simulated execution thereof.
Generally, ways in which prior knowledge may be expressed include, but are not limited to:
In some examples, the prior knowledge may have to be interpreted in the light of the process evolution information in order to determine the likelihood function. Accordingly, the likelihood function may be computed based on the prior knowledge and process evolution information describing the production process in question and/or other processes.
Deriving the likelihood function from the prior knowledge may be a completely automated process or may require at least partially input from a user. Exemplarily, the user input may be required when the prior knowledge itself is based on the experience of the user (e.g. the semi-qualitative description example above). Additionally or alternatively, the user input may be required to determine how exactly the prior knowledge should be used in deriving the likelihood function, e.g. the user may choose a distribution shape for fitting some data and/or for expressing a given assumption.
In one particular example, the chemical, pharmaceutical and/or biotechnological product may be a first product and retrieving the likelihood function may comprise:
In other words, according to this particular example, a transfer of process knowledge between processes that differ from each other, e.g. because of recipe, cell type or media used, is made possible. The use of relatedness information can be particularly useful if there is no previous knowledge available about a given process/product. Further, even if previous knowledge is available, it helps to characterize the likelihood space more accurately, that is by modulating the likelihood landscape and, potentially, for example, increasing the set of evolution parameters’ values which have zero or vanishingly small likelihood.
The akin production process is a production process and, as such, the above explanation about the nature and the features of a production process is equally applicable. However, the akin production process is different from the production process. In some cases the difference may be intrinsic to the process, e.g. the two processes may differ at least in the inputs or ingredients used therein and/or in the products and/or in the scales at which the processes are performed. In other cases, the difference may be due to one or more “external” factors, such as the use of different equipment (including analytical equipment), wherein the equipment may be different also because e.g. it has been upgraded/modified in the course of time. Accordingly, the first product may or may not differ from the second product.
The akin data are data measured and/or derived from a run of the akin production process and/or are related in another way to the akin production process. The run of the akin production process may have been already completed, i.e. the akin production process may have been already performed in its entirety, or the run may still be ongoing, i.e. the akin production process may have been only partially performed.
The akin data may comprise, but are not limited to, data related to process variables of the akin production process and/or evolution parameters derived from the akin production process (e.g. by assuming given process evolution information for the akin production process). Exemplarily, the akin data may comprise the trajectories for the process variables of the akin production process. A trajectory corresponds to a time-based profile of values recordable during the execution (real or simulated) of a production process. Additionally or alternatively, the akin data may comprise a prior (e.g. in the form of a probability distribution) about the values of the evolution parameter(s).
The relatedness information (also referred to simply as “relatedness” in the following) provides a quantitative indication of similarity between the production process and the akin production process. As explained, the two processes are different, but they may still hold some degree of similarity (even in the aspects that differentiate them), so that information extracted from the akin production process may provide insights also concerning the production process. The relatedness is a way to give the appropriate weight to the information from the akin production process, i.e. to express how relevant it is for the production process in question. The relatedness may also express the absence of similarity.
The relatedness may in general pertain to, but is not limited to, similarity between:
In particular, the similarity between the processes, which is quantified by the relatedness, may reside in the process dynamics, which is modelled by the process evolution information, as discussed above. Accordingly, in particular, the relatedness may be a function linking one or more evolution parameters associated with the two processes. In other words, the similarity between the two processes is reflected in the fact that the value of an evolution parameter is not independent from the value of a relevant akin evolution parameter. Thus, the relatedness may indicate the probability that an evolution parameter pa takes a certain value relative to a given value for the relevant akin evolution parameter pb.
Exemplarily, the process evolution information that can model the two processes may at least partly overlap so that the sets of evolution parameters for both processes have a non-zero intersection. In some cases, the process evolution information may be identical.
The relevant akin evolution parameter may be the equivalent akin evolution parameter, i.e. the same evolution parameter, e.g. the growth rate r, so that pa = ra and pb = rb. In other cases, the relevant akin evolution parameter may be different, e.g. inhibitor concentration capacity k, so that pa = ra and pb = kb. In other words, the similarity may be significant only for pairs of equivalent evolution parameters (former case) or also for pairs of different evolution parameters (latter case).
The same applies for a plurality of evolution parameters pa1, pa2 .... paN and relevant akin evolution parameters pb1, pb2, ... pbN. In the case in which the similarity is significant only for equivalent evolution parameters, the relatedness function R(pa1, pa2 .... paN, pb1, pb2, ... pbN) may be reduced to the product R(pa1, pb1) * R(pa2, pb2) * ... * R(paN, PbN).
Relatedness between processes may be estimated a priori by virtue of knowledge of the entities concerned (e.g. 5 different clones of Chinese hamster ovary (CHO) cells are known a priori to have higher similarity than a clone each of CHO, E. coli, P. pastoris, human embryonic kidney (HEK) and G. graminis). For example, a given, constant relatedness Rc may be assumed for cell-lines for a given molecule and organism, 0.1 *Rc between cell-lines for different molecules for the same organism, and 0.01*Rc between different organisms.
Alternatively or additionally, relatedness may also be determined a posteriori given data concerning dynamics of the entities concerned e.g. by Bayesian inference.
Ways in which relatedness may be expressed include (but are not limited to):
Retrieving the relatedness information may comprise receiving a user input selecting the relatedness information, e.g. a distribution shape and/or parameters for the distribution shape.
Using the akin data and the relatedness information, the likelihood function may be computed. For example, the likelihood function L(pa) for evolution parameter pa may be obtained as follows:
wherein R(pa, pb) is the relatedness information and L′(pb) is the likelihood function for the akin evolution parameter pb obtained from the akin data.
If the akin data comprise only trajectories, L′(pb) may be replaced by L′(pb|pvarb), wherein pvarb is the tuple (or vector) of the process variables values for the akin production process. L′(pb|pvarb) may exemplarily be estimated using Bayes’ theorem, the akin data about the process variables and assuming no prior or a broad prior. Alternatively, the prior may also be part of the akin data and originate e.g. from theoretical models. In particular, in order to estimate L′(pb|pvarb) as mentioned, fitting of process evolution information that models the akin production process to the akin data may be performed.
As mentioned, the akin data relate to at least one akin production process. In some examples, there may be a plurality of akin production processes and a plurality of sets of akin data may be correspondingly retrieved. Similarly, relatedness information indicating the similarity between each of the akin production processes and the production process may be retrieved and the likelihood function may be computed using the akin data and relatedness information relative to each akin production process.
Alternatively or additionally, retrieving the likelihood function may comprise retrieving current data related to an ongoing run of the production process and computing the likelihood function based on the current data.
Alternatively or additionally, retrieving the likelihood function may comprises retrieving historical data related to one or more past runs of the production process and computing the likelihood function based on the historical data.
Constraints on the values of the evolution parameters may be obtained on the basis of knowledge acquired by already running the production process. It can be considered that the production process has a maximal similarity, i.e. it is identical, with itself. In this perspective, considering data relating to the production process itself may be considered a special case of considering an akin production process, with relatedness equal to one. For example, more precisely, R(pa, pb) in the equation (2) above would be a Dirac delta function δ(pa - pb).
Accordingly, the historical data and the current data, similarly to the akin data, are data measured and/or derived from a run of the production process and/or are related in other way to the production process. The run of the production process may have been already completed, in the case of the historical data, or the run may still be ongoing, in the case of the current data.
The historical/current data may in particular comprise numerical values of the process variables of the production process, e.g. as a function of time, i.e. the trajectories. In some examples the likelihood function may be computed as the posterior probability L(p|pvar), with p the tuple of evolution parameters and pvar the tuple of process variables. Using Bayesian inference, the posterior probability is given by
with L(pvar|p) indicating the probability of observing certain values for the process variables given a set of values for the evolution parameter(s), L(p) being the prior probability and L(pvar) the marginal likelihood.
L(pvar|p) may be determined from the values of the process variables by fitting them to the process evolution information. Optionally, the historical/current data may also comprise the prior about the evolution parameter(s), L(p), or the prior may be input e.g. by a user in response to a prompt. L(pvar) may also be determined from the values of the process variables. Alternatively, since L(pvar) is a normalization factor, it may be neglected when the goal is optimisation, as in the case of the method herein described.
Any of the various sources discussed above may be combined for computing the likelihood function, so that the likelihood function may be computed based on the akin data in combination with relatedness and/or the current data and/or the historical data.
It should be noted that the likelihood function may be a function with an analytical expression or may be a function whose value is estimated numerically, e.g. numerical optimisation or via Monte Carlo methods. In cases where the likelihood function is at least in part determined by consideration of akin data, and/or by fitting to a trajectory of process variable(s), the function will typically be non-analytic.
The likelihood function is a way to take into account any prior knowledge about the values of the evolution parameters in a systematic and accurate way. To summarize, the likelihood function may be based on knowledge obtained from current and/or past runs of the production process in question and/or akin production processes. In addition or alternatively, the likelihood function may be based on knowledge coming from scientific literature and/or theoretical assumptions. In some examples, the likelihood function may be derived from all these sources.
The method further comprises retrieving a mapping that maps the one or more process variables onto one or more performance indicators. The mapping associates the value(s) for the process variable(s) to value(s) for at least one performance indicator. In some examples, the mapping maps the time series (or trajectories) of the one or more process variables onto one or more performance indicators.
A performance indicator is a quantity that describes a characteristic of one or more outputs of the production process, in an exemplary case, the product itself. In particular, the performance indicator may be a measurable quantity related to a chemical, biological, physical attribute of the one or more outputs.
Examples of performance indicators include but are not limited to final titre, rate of increase of titre, total feed used, quality of the product, running time of experiment or bioreactor, time for downstream process to complete, duration of cell survival or cell activity, maximum growth rate or density of cell population, concentration of impurities including but not limited to by-products of cell metabolism and cell death, robustness or repeatability of process, robustness of product (rate of degradation of product in controlled conditions over time), robustness of cell line (stability of genetic code over generations and/or ability of cells to withstand particular conditions e.g. pH extreme or high shear), cost of substrates, disposables materials or other costs.
Exemplarily, performance indicators like the total feed used, the maximum cell density and the rate of increase of titre require the trajectories of the process variables, while other performance indicators, e.g. the final titre, are mapped from a single value of a process variable, e.g. at the end point.
One or more process variables may be mapped to one performance indicator. This means that the value of the performance indicator is influenced by the value(s) of the one or more process variables. Further, a given process variable may be relevant for, i.e. influence, more than one performance indicator. In some examples, a process variable may be itself a performance indicator, so that the mapping is the identity. In other examples, the mapping may be different from the identity.
If the state of the production process is characterized by a plurality of process variables, the mapping may map all process variables or only a proper subset thereof.
In some examples, e.g. for the cases in which a recipe template composite comprises more than one recipe template, the mapping may take a plurality of values for a single process variable, wherein the values come from parallel or multiple runs (real or simulated) of the production process.
The mapping may depend on how the production process is performed, namely on the recipe used for the production process. Accordingly, the mapping may need not only the process variable(s) but also the recipe parameters as an input. Alternatively, the method may comprise retrieving a plurality of mappings, each corresponding to a different recipe.
The mapping may be based e.g. on previous runs of the production process and/or on theoretical predictions. The mapping may be stored in a database, spreadsheet, and stored locally, in an organisation file system, in a database within an organisation or on the cloud (remotely). It should be noted that mapping may be a function with an analytical expression or may be a function estimated numerically, e.g. via Monte Carlo methods.
The method further comprises retrieving a utility function for at least one of the performance indicators, wherein the utility function provides a desirability value associated to the performance indicator(s). In other words, the utility function quantifies the overall worth or benefit of the performance of the production process on the basis of the value(s) of the performance indicator(s).
In the case of a single performance indicator, the utility function may simply quantify how good or how bad a specific value of the performance indicator is. For example, if the performance indicator is the titre, the utility function may be a monotonic function, indicating that larger values of the titre are preferred to smaller values. If there are more performance indicators, the utility function takes into account the interplay of the different performance indicators in assessing the overall worth of the performance. For example, if the performance indicators are the titre and the quality of the product, a lower titre associated with a very high quality may still be desirable. In some examples, the utility function may be constant within a range, and largely negative outside the range (i.e. a strong requirement for “robustness”).
The worth of the performance may be evaluated according to one or more various factors, including, but not limited to, efficiency, yield, safety, reliability, costs. In particular, the utility function may be defined depending on a specific, desired goal for the determination of the optimal recipe. Indeed, the production process may be optimised with respect to many different aspects. Examples of goals may include, but are not limited to, optimising a given performance indicator, evaluating which clone among a set of clones is most suitable for the production of a molecule, compensating for unexpected conditions, scaling the production process.
Even if the mapping maps the process variable(s) to more than one performance indicator, the utility function may be a function of only a subset of the performance indicators. Further, the utility function may be a function of performance indicators that are not derived from process variables, such as the overall running time of the production process. In other words, the utility function takes as argument at least one performance indicator derived from a process variable via the mapping (which, in some cases, may be simply the identity function) and may optionally take as argument also one or more performance indicators not derived from a process variable. It should be noted that the utility function may be a function with an analytical expression or may be a function estimated numerically, e.g. via Monte Carlo methods.
In some examples, e.g. for the cases in which a recipe template composite comprises more than one recipe template, the utility function may take a plurality of values for a single performance indicator, wherein the values come from parallel and/or multiple runs (real or simulated) of the production process.
As explained so far, the output of the utility function depends directly on the performance indicator(s), and, since the performance indicator(s) can be linked to the process variable(s) via the mapping, it depends indirectly on the process variable(s). Further, since the process variables depend on the process evolution information, specifically the evolution parameters, the output of the utility function also depends indirectly on the evolution parameters. Since the values for the variable evolution parameters are based on the likelihood function, the output of the utility function also depends indirectly on the likelihood function.
To summarize, the method requires a plurality of inputs, namely the one or more recipe template composites, the process evolution information, the likelihood function (which is based, optionally, on relatedness information), the mapping and the utility function. These inputs are used, as explained in the following, to simulate execution of the production process and select, based on the results of the simulation, the optimal recipe template composite with the optimal value(s) for the variable recipe parameter(s) among the plurality of recipe template composites and the plurality of possible values for the variable recipe parameter(s), respectively.
The method comprises
The method comprises a series of nested loops, wherein some loops may reduce to a single iteration in some cases. The outermost loop involves considering different recipe template composites, while the next loop involves considering different options for the values of the variable recipe parameter(s).
The different iterations of each loop may be performed sequentially, i.e. one after the other. Alternatively, the iterations may be performed in parallel. Generally, the execution of the method may comprise a combination of parallel and sequential executions of the iterations of the loops. Accordingly, “looping” as used herein merely indicates considering each one of a plurality of possibilities when performing the simulation and does not necessarily indicate a succession in time. Hence, the one or more repetitions of any one of the steps described below may occur at the same time as the “first” execution of the step and/or may occur in a temporal sequence.
For a given recipe template composite a first optimisation step is performed, which comprises providing a set of input values for the at least one variable recipe parameter in the given recipe template composite.
As already mentioned, a set of input values may comprise one or more elements. If the recipe template composite comprises only one recipe template and the recipe template comprises only one variable recipe parameter, then the set of input values comprises only one input value. In the other cases, the set of input values comprises more than one value. For example, if the recipe template composite comprises two recipe templates, each having one variable recipe parameter (which may be the same quantity for both, e.g. pH set point), the set of input values comprises two values, one for the variable recipe parameter in the first recipe template and one for the variable recipe parameter in the second recipe template. If the recipe template composite comprises a single recipe template with a plurality of variable recipe parameters, the set of input values comprises a corresponding plurality of input values.
The input value(s) may be an educated guess based on process knowledge and could be part of a set of possible values linked to the specific recipe template composite. For example, each recipe template composite may comprise one or more candidate values together with a testing range for each value, so that an input value may be chosen as any value lying in an interval around a candidate value. The candidate value(s) may, thus, be retrieved together with the recipe template composites. Alternatively, the input value(s) may be supplied by a user.
Once a set of input values has been assigned to the variable recipe parameter(s), a second optimisation step providing a utility score is performed. The details of the second optimisation step will be discussed below; in any case the utility score is a numerical value associated with a certain recipe template composite and a certain set of input values. The second optimisation step is repeated for at least another set of input values, in some cases for a plurality of other sets of input values. The various sets of input values are different from each other in at least one input value.
To each set of input values, the second optimisation step associates a utility score. Since the second optimisation step is performed for a plurality of sets of input values, as a result a plurality of utility scores is obtained. The repetition of the second optimisation step constitutes the second outermost loop, nested in the potential outermost loop over the recipe template composites, which is discussed below. The number of different sets of input values, i.e. the number of iterations for the second optimisation step, may be a fixed, predetermined number or may be set each time the method is performed, e.g. by a user, or maybe determined computationally, e.g. by the iterations required to reach convergence of an algorithm.
Besides performing the second optimisation step for different sets of input values, the first optimisation step comprises identifying the optimal utility score from the plurality of utility scores. Each value of the utility score, as mentioned, is associated to a set of input values. Said otherwise, the value of the utility score is dependent, among other things, on the value(s) assigned to the variable recipe parameter(s) in the selected recipe template composite. Thus, the utility score may be seen as a function of the one or more variable recipe parameters, hereafter called “score function”. In the case of a plurality of variable recipe parameters, the score function is a multi-variable function. The score function may be evaluated by means of numerical methods on the basis of the plurality of discrete values actually computed.
The score function is optimised to find the optimal set of input values for the variable recipe parameters. Finding the optimal value(s) for the variable recipe parameter(s) is an optimisation problem in which the score function is maximised or minimised. Depending on the number of variable recipe parameters, the optimisation problem may be a multi-dimensional problem. Depending on how the utility function as well as the quantities derived from it and discussed below (e.g. utility tally, utility value) are defined, the optimal utility score may be the highest or the lowest utility score. Examples of optimisation algorithms are the Nelder-Mead method and the steepest descent method.
The result of optimising is that the set of input values that give the optimal (e.g. highest score) for the selected recipe template composite is found. The optimal utility score and the set of input values associated thereto may exemplarily be stored, e.g. in a storage device such as a hard disk.
If only one recipe template composite was retrieved, the outermost loop “collapses” to a single iteration and the first optimisation step is performed only once. Accordingly, the set of input values that yields the optimal utility score identified within that first optimisation step, or “optimal set of input values”, together with the recipe template composite is chosen as the at least one recipe for the production process. Indeed, by definition, a recipe template with values assigned to its variable recipe parameter(s) becomes a recipe. In this way, the optimal way of executing the production is determined, namely by determining the recipe(s), i.e. recipe template composite plus set of input values, that optimize the score function.
If there is more than one recipe template composite, the first optimisation step is repeated for each of them, so that an optimal utility score is determined for each recipe template composite. Once the iteration over all recipe template composites is over, i.e. once the outermost loop has completed, the best optimal utility score among the optimal utility scores corresponding to the plurality of recipe template composites is identified. Again, the best optimal utility score may be the highest or lowest optimal utility score, depending on how the utility function is defined. The optimal utility scores may form a discrete set, so that the best optimal utility score may be identified simply by comparing the optimal utility scores against each other. The best optimal utility score and the recipe template composite associated thereto may exemplarily be stored, e.g. in a storage device such as a hard disk.
Also in this case, the recipe template composite and the set of input values that are associated with the best optimal utility score are selected as the recipe(s) to be used for the production process.
If the recipe template composite comprises only one recipe template, then only one recipe is determined as the optimal recipe for running the production process. If the recipe template composite comprises more recipe templates, then a plurality of recipes are identified for parallel, sequential or overlapping runs of the production process.
For example, a recipe template composite (A) may comprise two recipe templates for two parallel runs on bioreactors, wherein according to one recipe template bolus feed is used and according to the other gradual feed is used. Another recipe template composite (B) may comprise two identical recipe templates with bolus feed with volume dependent on glucose measurements. For both composites the variable recipe parameter in all recipe templates is the time of harvest. According to the method disclosed herein, it may be determined that it is best to have two parallel runs using the recipe templates from composite A rather than B, and in particular with harvest times of 15 days for the bolus feed and of 18 days for the gradual feed.
To summarize, the final result provided by the method is at least one recipe for running the production process, wherein the recipe(s) can be considered optimal on the basis of one or more specific criteria, which are expressed by the utility function. In the following, the path from the utility function to the utility score will be explained. It should be noted that the term “first/second/third/fourth optimisation step” does not necessarily indicate that an optimisation process is being performed within that step. Rather, each of these steps contributes to the optimisation achieved by the method as a whole, namely finding the optimal recipe(s).
As mentioned above, the second optimisation step provides a utility score, which is obtained by potentially looping over different values for the variable evolution parameters. In particular, the second optimisation step comprises:
Every time the second optimisation step is performed, the recipe template composite and the set of input values for the variable recipe parameter(s) are “fixed”. Within the second optimisation step, the repetition of the third optimisation step represents the third outermost loop, nested within the second outermost loop over the sets of input values.
The third outermost loop is a loop over possible sets of likely values for the variable evolution parameter(s), however this loop involves only optionally a plurality of iterations, namely only if there is more than one suitable set of likely values. In some cases, only one set of likely values for the variable evolution parameter(s) is considered and, thus, the third outermost loop “collapses” to a single iteration.
If the recipe template composite comprises more than one recipe template, in some cases, the same likely values may be used for the variable evolution parameters when simulating, irrespective of the specific recipe template being considered. Accordingly, each set of likely values may comprise one likely value if there is only one variable evolution parameter, or a plurality of likely values if there are more variable evolution parameters. In other cases, different likely values may be used when simulating using different recipe templates. Accordingly, even when there is only one variable evolution parameter, each set of likely values may comprise a plurality of likely values corresponding to the plurality of recipe templates in the recipe template composite.
The set of likely values is selected based on the likelihood function, which can incorporate different types of knowledge and can assume different forms, as explained above. In other words, the one or more likely values are points drawn from the likelihood space.
In one example, the likely value(s) may be the maximum likelihood estimate(s) according to the likelihood function, i.e. the value(s) that maximize the likelihood function. In this case, finding the likely value(s) for the variable evolution parameter(s) is a maximisation problem in which the likelihood function is maximised by systematically choosing values from within an allowed set and computing the value of the likelihood function. Depending on the number of variable evolution parameters, the maximisation problem may be a multi-dimensional problem. Numerical methods, such as Monte-Carlo simulations, or algorithms, e.g. the Nelder-Mead method, may be used to solve the maximisation problem. In some cases, it may be necessary to use multiple starting values for maximisation, as the likelihood space being explored may be highly non-linear and may exhibit multiple maxima.
According to this example, there may be only one set of likely values that maximises the likelihood function, in which case the third outermost loop is not an actual loop, or there may be more than one set of likely values that maximises the likelihood function, in which case the third optimisation step is repeated for each set.
In another example, the likely value(s) may be sampled from the likelihood space, wherein e.g. the sampling may satisfy given constraints, such as that the likelihood function assumes a value larger than a specific predetermined or predeterminable threshold. The sampling may alternatively be stochastic. Also according to these examples, there may be one or more sets of likely values. If more than one set of likely values is considered, the number of different sets of likely values, i.e. the number of iterations for the third optimisation step, may be a fixed, predetermined number or may be set each time the method is performed, e.g. by a user, or determined dynamically, for example, on the basis of the variability in outcome from successive samples.
The result of the third outermost loop is to associate a utility tally, which will be explained later, to each set of likely values. Besides the potential loop over different sets of likely values, the second optimisation step comprises computing the utility score based on the one or more utility tallies.
The utility tally is a numerical value associated to a certain recipe template composite, a certain set of input values and a certain set of likely values for the variable evolution parameter(s). In a certain sense, the utility tally can be seen as a function (the utility tally function) of these three factors and, when computing the utility score, the dependency on the set of likely values is “eliminated”. Specifically, this is achieved by combining different utility tallies, each weighted according to the probability that the values for the evolution parameters correctly describe the process evolution, namely the likelihood function.
In one example, the utility score US (for a certain recipe template composite and a certain set of input values) may be obtained as
wherein UT is the utility tally function, p represents the vector (or tuple) of the evolution parameters and L is the likelihood function. Equation (4) is a schematic equation in which other details may have been omitted to illustrate the basic concept for computing the utility score, namely that an integral is performed over the space of the evolution parameters. The utility tally function is not a function formulated as an analytic expression, but is rather evaluated numerically by considering the different utility tallies associated to different sets of likely values.
If only one set of likely values is considered, p′, it may be interpreted as the likelihood function being a Dirac delta function centred on p′, so that US=UT(p′).
According to the above equation, the utility score may be considered the expected value of the utility tally function. In other examples, higher moments of the utility tally function may be considered, such as the variance.
To summarize, the second optimisation step provides a utility score for a given recipe template composite and a given set of input values for the variable recipe parameter(s), which is then optimised as described above. Each utility score may exemplarily be stored, e.g. in a storage device such as a hard disk.
As mentioned above, the third optimisation step provides a utility tally, which is obtained by potentially looping over different stochastic realizations. In particular, the third optimisation step comprises:
Every time the third optimisation step is performed, the recipe template composite, the set of input values for the variable recipe parameter(s) and the set of likely values for the variable evolution parameter(s) are “fixed”. Within the third optimisation step, the repetition of the fourth optimisation step represents the fourth outermost loop, nested within the third outermost loop over the sets of likely values.
The fourth outermost loop is a loop that creates multiple versions of a simulation (the generation of the simulation is part of the fourth optimisation step) with the same nominal inputs, in order to account for potential variabilities in the production process. However, this loop involves only optionally a plurality of iterations, namely only if stochasticity is taken into consideration. If variability shall not be accounted for, the fourth optimisation step is performed only once and, thus, the fourth outermost loop “collapses” to a single iteration.
The stochasticity in the production process may be due to equipment or machine variability and/or to process evolution variability. Indeed, the way the production process is simulated or executed in real life is dictated by the recipe with its recipe parameters. However, in reality, the recipe needs to be implemented by a process equipment and the recipe parameters get “translated” into equipment parameters (also referred to as “machine parameters” or “setup parameters”). The machine parameters describe how a specific equipment setup carries out the recipe steps according to the recipe parameters. For a given equipment setup, the “mapping” between the recipe parameters and the machine parameters is fixed. However, unlike the recipe parameters, the machine parameters may be affected by stochasticity.
Ways in which the equipment may introduce stochasticity into the production process may include, but are not limited to:
An exemplary correspondence between some recipe parameters and machine parameters is reported in the table below:
The stochasticity that would affect the actual run of the production process can be introduced in the simulation run by considering the mapping of the recipe parameters onto machine parameters when simulating. For example, stochasticity could be introduced by using stochastic differential equations in the simulation or by sampling from a distribution during simulation, the distribution itself being determined by the machine parameters. The case in which stochasticity is due to machine variability is discussed further below in the context of setup specification.
As already mentioned, additionally or alternatively, stochasticity may be intrinsic to the process evolution itself. In this case, the process evolution information may be stochastic. For example, a stochastic model for cell growth may be:
wherein the evolution parameters a, rg,c and rg,s (respectively drift in growth rate, mean initial growth rate and variation in initial growth rate) drive the stochasticity and wherein W is a Wiener process, ρ is cell density, and N is the normal distribution with specified mean and standard deviation.
Accordingly, in the third optimisation step, the fourth optimisation step (which will be discussed below), i.e. the generation of the simulation, is performed more than once if a stochasticity flag is set. The stochasticity flag signals to the computer system performing the method that the fourth optimisation step needs to be repeated.
The stochasticity flag may be set before the method starts to be performed or may be set on the fly. It may be a flag that indicates how many times the fourth optimisation step shall be repeated, e.g. a non-Boolean flag. Alternatively, the flag may be Boolean, i.e. only indicating whether a repetition must occur or not, and the number of repetitions may be retrieved from a look-up table or input by a user upon prompt. The number of repetitions may further be determined dynamically depending on the outcome from successive iterations of process simulation. The stochasticity flag may be stored in volatile or non-volatile memory.
The result of the fourth optimisation step is a utility value. The utility value is a numerical value associated to a certain recipe template composite, a certain set of input values, a certain set of likely values for the variable evolution parameter(s) and a certain simulation run. If simulation runs are affected by stochasticity, the effect of stochasticity may be mitigated or otherwise explicitly considered as a factor in determining the optimal recipe for the process by compounding a plurality of utility values, as explained below.
If the stochasticity is not introduced in the simulation (i.e. the flag is not set), the fourth optimisation step is performed only once and the utility tally is computed by merely setting it equal to the single utility value obtained from the fourth optimisation step.
If the stochasticity is introduced in the simulation (i.e. the flag is set), the fourth optimisation step is repeated at least once and, thus, a plurality of utility values is obtained. In this case, the utility tally is computed by compounding the plurality of utility values. “Compounding” means combining the utility values in a meaningful manner to address the issue of stochasticity.
Exemplarily, compounding may comprise computing statistics out of the utility values, e.g. location (average, median, mode...), spread (standard deviation, variance, mean absolute difference...), shape (skewness, kurtosis...). In one example, the utility tally may be the average of the utility values, so that the effect of stochasticity is mitigated. In another example, part of the optimisation in determining the recipe may be reducing the stochasticity, so that the utility tally may be the mean absolute difference of the utility values. In such a case, the utility score in the first optimisation step is optimised by finding the lowest value.
To summarize, the third optimisation step provides a utility tally for a given recipe template composite, a given set of input values for the variable recipe parameter(s) and a given set of likely values for the variable evolution parameter(s), which is then used to compute the utility score as described above. Each utility tally may exemplarily be stored, e.g. in a storage device such as a hard disk.
As mentioned above, the fourth optimisation step provides a utility value, which is obtained from the simulation trajectories and by potentially looping over different recipe templates. In particular, the fourth optimisation step comprises:
The simulation of the execution of the production process (or “simulation run”) is an imitation of the execution of the production process in the real world performed by means of a computer system.
A simulation is generated on the basis of a given recipe template in the recipe template composite and the set of input values for the variable recipe parameter(s), which together provide a recipe, and the process evolution information with the set of likely values for the variable evolution parameter(s). The recipe provides a description of the process to be simulated, while the process evolution information provides initial conditions for the process variables and models the evolution of the process with time.
Within the fourth optimisation step, the repetition of generating the simulation represents the innermost loop, nested within the fourth outermost loop for stochasticity. The innermost loop is a loop that generates a simulation for each recipe template within a recipe template composite. However, this loop involves only optionally a plurality of iterations, namely only if the recipe template composite comprises a plurality of recipe templates. If the recipe template composite comprises only one recipe template, the simulation is performed only once and, thus, the innermost loop “collapses” to a single iteration.
Optionally, the method may comprise retrieving a setup specification (also referred to as “machine specification”). A setup specification includes information about the setup of the process equipment used for executing the production process, such as the scale value of the equipment, e.g. the capacity of a bioreactor expressed in litres. Further, the setup specification may include at least one of the components of the setup and its characteristics, e.g. specifying that the equipment comprises an impeller and, optionally, which kind of impeller and so on. Also, other characteristics of the equipment may be indicated by referring to a model of a product, e.g. Sartorius ambr®. In other words the setup specification describes the equipment used for executing the production process.
Further, the setup specification refers to properties of the setup (machine) which mediate between the specification of the process (in the recipe) and the dynamics of the system (in the evolution parameters).
An example of a setup specification may be ambr® 250 bioreactor with standard sparger and mammalian impeller or Sartorius STR® 50 with ring sparger and two 3-blade impellers.
As explained above, the effect of the recipe parameters on the time evolution of the process variables, as governed by the process evolution information, may be mediated (or modulated) by the machine setup and in particular, machine parameters identified by that machine setup. This may introduce stochasticity in the simulation run. Indeed, the recipe with its recipe parameters dictates how the machine is set to operate. The effects on the process dynamics are mediated by the machine. The machine parameters therefore dictate how the machine desired behaviour (as prescribed by the recipe) maps onto its actual behaviour (as determined by the combination of the desired behaviour and the machine parameters). The actual behaviour of the setup then interacts with the process evolution information to determine how the process variables evolve over time.
Exemplarily, a setup specification may encapsulate information on the feedback loop for temperature control, in particular in terms of the manner in which the machine heats or cools the culture in response to deviation from set-point, and randomness in terms of deviation from set-point. By way of example, the temperature within the culture vessel may be modelled by a stochastic differential equation as shown below
where ε(f, dt) is a noise term, or e.g. equivalently:
where W is a Wiener process. In this example, f and a are both machine parameters and T* is the current temperature set-point, as dictated by the recipe. In this case, stochasticity then affects the trajectories by virtue of a dependency e.g. of growth rate on temperature. The effect can be modelled by the system of stochastic differential equations
where a, f are machine parameters as before, T* arises from the simulation of the recipe, as before, T is temperature as a function of time, W is a Wiener process, p is cell density, Topt, σ and rg are evolution parameters, respectively optimum temperature, temperature sensitivity, and maximum growth rate, N is the normal distribution with specified mean and standard deviation. The solution to the above system is stochastic in both T and p.
From the simulation(s) values of a process variable at the different times during the evolution of the process are derived in order to obtain a trajectory. Each trajectory may be understood to summarize and provide an overview of the associated process variable. Each trajectory may be expressed as a curve or graph that describes the time evolution of the process variable during the simulated execution of the production process. In particular, each trajectory may comprise a plurality of points representing values of a variable corresponding to different moments in time. For example, a time unit between successive points may be one hour. In alternative examples, the time unit between successive points may be ten seconds, or a day. The time units may also be unequally spaced within the trajectory.
If there is only one recipe template in the recipe template composite and only one process variable, then only one trajectory is determined. In all other cases, a plurality of trajectories is determined.
As explained above, the utility function depends at least on the performance indicator(s). The value(s) of the performance indicator(s) can be obtained from the value(s) of the process variable(s) by means of the retrieved mapping and the value(s) of the process variable(s) are obtained from the simulation(s). Accordingly, the utility function can be evaluated based on the trajectories via the mapping. As already explained above, in some cases a process variable may be itself a performance indicator and the mapping is simply the identity function. Generally, if U is the utility function, I is the tuple of the performance indicators, M is the mapping and V is the tuple of the process variables, U(I) = U{M[V(t)]}. The utility value computed within the fourth optimisation step may be, in general, a function of the output of the utility function, i.e. UV=f[U(I)]. Exemplarily, UV=U(1).
In some examples, the utility function may be cumulatively evaluated over time in order to compute the utility value. In other words, if the trajectory points (i.e. the different values of a process variable over time) correspond to times t1, t2 ... tn, the utility value may be computed as
or some other composite of utility values at selected time points. In other examples, the utility value may correspond to the utility function evaluated at a single time, e.g. tn.
In some cases, each of the one or more variable recipe parameters may be a timepoint. For example, there may be only one variable recipe parameter, the harvest time. The purpose of the loop over the sets of input values in the first optimisation step is to obtain utility scores related to different sets of input values, which, in these specific cases, means obtaining utility scores related to different timepoints. In such cases, a trajectory already provides the values of the process variables as a function of time, so that it is possible to obtain a plurality of utility scores (via the steps in between) directly from the trajectories. Accordingly, in these specific cases, the second outermost loop may be dispensed with.
If there is more than one recipe template in the recipe template composite, i.e. if parallel or sequential runs are to be performed, the overall utility shall be considered. In one example, computing the utility value may comprise combining partial utility values, each partial utility value associated to a recipe template. The utility value may e.g. be given by the arithmetic mean, geometric mean, sum or product of the partial utility values.
In another example, as mentioned above, the mapping to the performance indicator may already take into consideration a plurality of trajectories for a given process variable, each trajectory corresponding to a different recipe template in the recipe template composite. In this case, for m recipe templates in a recipe template composite, the mapping M may be M(V1(t), V2(t), ... Vm(t)), wherein Vj(t) is the trajectory of the process variable V associated with the j-th recipe template.
In yet another example, as mentioned above, the utility function itself may take as input values of performance indicators corresponding to simulation runs with different recipe templates within a recipe template composite, namely U{M[V1(t)], M[V2(t)], ... M[Vm(t)]}.
To summarize, the fourth optimisation step provides a utility value for a given recipe template composite, a given set of input values for the variable recipe parameter(s) and a given set of likely values for the variable evolution parameter(s), which is then used to compute the utility tally as described above. Each utility value may exemplarily be stored, e.g. in a storage device such as a hard disk.
In conclusion, the path from the utility function to its form that is optimized, i.e. the score function, may - only exemplarily - be described as following (assuming a recipe template composite comprising only one recipe template, for the sake of clarity):
The steps 1-3 illustrated above provide a general mathematical formalism on how to arrive at the score function from the utility function; however, they are not meant to cover all possible cases. As already explained in detail previously, for example, in some cases only one utility value is computed, or higher moments of the utility tally function are considered.
In a particular example, the method may further comprise providing the at least one recipe to a control system and executing, by the control system, the production process based on the at least one recipe.
In other words, the recipe template composite with relative input value(s) as selected by the above-described method may be fed directly to a control system configured to control the setup equipment to carry out the process. In the case of a recipe template composite comprising multiple recipe templates, the process is run multiple times in parallel or sequentially, e.g. on multiple bioreactors.
When the production process is executed in the real world, the results may be used to provide feedback for the ongoing run and/or future runs. Accordingly, the method may further comprise storing data from the executed production process as historical data and/or storing data from the executing production process as current data.
The method described herein, or at least aspects thereof, may be applied in the context of scaling the production process. Processes for the production of chemical, pharmaceutical and/or biotechnological products are scale dependent; in other words, a process behaves at least partly differently on a small scale (e.g., in a laboratory) in comparison to a large scale (e.g., in production). Usually the process is first performed at small scales and then at successively larger scales. In the following, two exemplary scaling scenarios will be considered, namely scaling from a scale for process development to a scale for manufacturing and scaling from a scale for clone selection to a scale for process development.
In this case, the process development is at source scale A and the manufacturing is at target scale B. There may be two recipe spaces α and β that contain recipe template composites for scales A and B, respectively. The goal may be to find the recipe (recipe template composite plus values for variable recipe parameters) at scale A that, when scaled, will optimize the execution of the production process at scale B.
Similarly to what is described above, one or more recipe template composites (from space α) may be retrieved, as well as process evolution information and a likelihood function. A recipe template composite is selected, a set of input values is assigned to the variable recipe parameters and a set of likely values is assigned to the variable evolution parameters on the basis of the likelihood function. Using all these inputs, the production process at scale A may be simulated, thereby obtaining one or more trajectories for the process variables, e.g. trajectories corresponding to different recipe templates within the composite.
Similarly to what described above about akin data, current data and historical data, also the simulated data, i.e. the trajectories, may be used as evidence to create a posterior distribution. In other words, an updated likelihood function may be computed, based on the retrieved likelihood function as a prior, and the simulated data as evidence.
The above-described method for determining at least one optimal recipe may then be performed for scale B, wherein the recipe template composites are taken from space β and the retrieved likelihood function is the updated likelihood function previously obtained. Accordingly, a recipe at scale B may be obtained.
Alternatively, the recipe at scale B may be obtained via: (a) general scaling procedures (b) DoE approaches and linear / quadratic curve fitting (c) the methods described in European patent application 18 000 132 (d) some combination of (a) to (c).
Irrespectively of how the recipe at scale B is obtained, it may be used, together with the process evolution information and values for the evolution parameters drawn from the updated likelihood function, to perform a simulation at scale B. From the trajectories obtained in the simulation, a utility function may be evaluated at scale B, wherein the output of the utility function quantifies the overall worth or benefit of the performance of the production process at scale B.
As mentioned with reference to the features of the utility function above, the output of the utility function depends indirectly on the likelihood function. In this case, the updated likelihood function obtained from the simulated trajectories at scale A is used for the simulation at scale B. Hence, the output of the utility function will indirectly depend on the inputs of the simulation at scale A, in particular the selected recipe template composite and the set of input values assigned to the variable recipe parameters.
Therefore, numerical optimisation methods may be used to find the recipe template composite and the set of input values for the variable recipe parameters at scale A that optimise the output of the utility function at scale B. In this way the recipe (recipe template composite plus values for variable recipe parameters) at scale A that, when scaled, will optimize the execution of the process at scale B can be determined.
The computation of an updated likelihood function as illustrated above may be performed as a stand-alone preparatory step for determining the optimal recipe at the process development scale in view of future manufacturing. Alternatively, the updated likelihood function may be computed and stored while performing the method according to the invention simply for determining a recipe at scale A. In this case, if scaling to B is later to be considered, the updated likelihood function may be already available. Accordingly, in a particular example, the method may further comprise updating the likelihood function based on the trajectory(ies) and storing the updated likelihood function.
In this case, the clone selection is at source scale A and the process development is at target scale B. There may be two recipe spaces α and β that contain recipe template composites for scales A and B, respectively. The goal may be to find the clone at scale A that will optimize the execution when the production process is executed at scale B.
Similarly to what described above, one or more recipe template composites (from space α) may be retrieved, as well as process evolution information and a likelihood function. Each recipe template composite in the recipe space α contains, in this case, a plurality of recipe templates, wherein each recipe template is associated with a clone. In the typical case, the recipe templates here are identical, although in an alternative realisation, a multiplicity of different recipe templates may be associated with each clone, e.g. a set of templates (D, E, F) may be used for clones (1, 2, 3, 4) in all combinations, i.e. D1, D2, D3, D4, E1, E2, E3, E4, F1, F2, F3, F4.
A recipe template composite is selected, a set of input values is assigned to the variable recipe parameters and a set of likely values is assigned to the variable evolution parameters on the basis of the likelihood function. In particular, the set of likely values G comprises a plurality of subsets G1, ..., Gn corresponding to the plurality of clones. Accordingly, each subset is assigned to a recipe template in the selected recipe template composite.
Using all these inputs, the production process at scale A may be simulated, thereby obtaining a plurality of trajectories for the process variables, wherein, for a given process variable, a plurality of trajectories T1, ...,Tn corresponding to the plurality of clones is obtained.
Based on the trajectories, the best clone may be identified, e.g. by evaluating some metric using the values of the trajectories. Performance indicators and/or a utility function may be used, similarly to the way in which they are used in the above-described method for determining the optimal recipe.
Once the best clone is selected, the associated subset Gb of values for the variable evolution parameters may be used, together with the process evolution information and a recipe, to perform a simulation at scale B. The recipe (i.e. a recipe template composite from space β and values for its variable recipe parameters) may be determined using the method according to this invention.
From the trajectories obtained in the simulation, a utility function may be evaluated at scale B, wherein the output of the utility function quantifies the overall worth or benefit of the performance of the production process at scale B.
As mentioned with reference to the features of the utility function above, the output of the utility function depends indirectly on the values for the evolution parameters. In this case, the values for the evolution parameters are associated with a given clone. Hence, the output of the utility function will indirectly depend on the clone chosen at scale A.
Therefore, numerical optimisation methods may be used to find the clone at scale A that optimises the output of the utility function at scale B.
According to another aspect of the present invention, a computer program product is provided. The computer program product comprises computer readable instructions, which, when loaded and executed on a computer system, cause the computer system to perform operations as described above. The computer program product may be tangibly embodied in a computer readable medium.
Yet another aspect of the present invention relates to a computer system operable to determine at least one recipe for a production process to produce a chemical, pharmaceutical and/or biotechnological product, wherein the production process is defined by a plurality of steps specified by recipe parameter(s) controlling an execution of the production process and a recipe comprises the plurality of steps defining the production process. The computer system comprises:
In other words, the computer system is operable to perform a method according to the aspect discussed above.
The computer system may particularly comprise a memory and a processor to operate the modules. The retrieving module and the computing module may be separate entities or may at least partly overlap with each other.
Further, the computer system may be configured to interface with a control system via a network, shared disk, or database system, wherein the control system controls the execution of the production process in the real world. The interface may particularly allow transfer of data and/or commands. In some examples, the computer system may coincide with the control system.
To summarize, the method, the product and the system discussed above enable the determination of an optimal recipe for a production process, wherein the recipe can be determined prior to execution or during the process execution, thereby supporting dynamic recipe adaptation.
Examples in which the recipe is determined prior to a run of the production process include the following scenarios:
Examples in which the recipe is determined (i.e. adapted) during a run of the production process include the following scenarios:
Further, according to the invention, knowledge about the process can be fully and accurately taken into consideration, in particular minimizing the role of assumptions for the production process during its whole course. In particular, ignorance can be accounted for by maintaining probability landscapes throughout the optimization procedure thanks to the use of the likelihood function for the variable evolution parameters and by considering multiple recipe template composites. In addition, transfer of process knowledge between different organisms, cell types, media etc. is made possible by the use of relatedness information.
Accordingly, the above-described method, product and system can be successfully employed in many different contexts, some of which are reported below, together with an explanation of the issues of conventional approaches for each case:
At the cell line selection stage, the use of a recipe that is simply a “scale down” model of the recipe of a known large-scale target process fails to acknowledge that the best “cell line and recipe” combination may comprise a different cell line than that which would be selected against some (essentially arbitrary) recipe. For example, given two recipes A and B, and two cell lines 1 and 2, where recipe A is the standard “platform” recipe, performance at scale and during cell line (clone) selection may rank B1 > A2 > B2 > A1. However, just using the platform recipe for clone selection, cell line 2 would be selected, resulting in ultimately lower-than-optimal performance.
At the point in time when an organisation is conducting cell line selection, it will typically have access to data from previous molecule development processes. This gives an opportunity to determine prior estimates of variation expected within a candidate set of clones. Such estimates have the potential to be used to drive recipe design during clone selection (with a view to subsequent scale-up for manufacture). This invention provides a mechanism to do so thanks to the plurality of recipe template composites and the degrees of freedom offered by the variable evolution parameters.
At the process development stage, the range of potentially beneficial recipe variations that can be evaluated is limited by available time and resources from an effectively infinite number of possibilities resulting from the interaction of recipe parameters (e.g. pH setpoint) with time (e.g. when to perform a change in pH setpoint). As a result, process development is normally limited to a very small subspace of the range of possible recipe parameters values, and within this subspace, conclusions are drawn on the basis of linear behaviour of the biological systems.
Currently it is impossible to establish upfront whether the subspace being explored is itself close to optimal or far off, and whether linearity assumptions within this subspace are justifiable. For example, a series of experiments may be conducted to determine whether a switch in pH from 7 to 6.6, 6.7 or 6.8 is optimal after 20 hrs of culture, in conjunction with DO % setpoints of 20, 25 and 30, using a DoE. The result may be a conclusion, via statistical modelling (e.g. fitting of a low-degree polynomial and identification of an optimum based upon the fitted surface) of the pH and DO response surface, that pH 6.65 and DO 23% are optimal. However, this may be ill-informed if, for example, the response to pH is highly nonlinear, with 6.6 being much worse than 6.7, which itself is only marginally worse than 6.8. Equally, it may be ill-informed because the true optimum may be a change of pH after 15 hrs of culture rather than 20.
This invention provides a mechanism by which an appropriate subspace of recipe parameters can be explored and their optimal values for a given process determined.
Previous approaches to scale-up and transfer to manufacture consider only the most likely parameters for simulation models which drive the choice of scale-up recipes. These approaches have a deficiency that this invention resolves, namely the use of a single model with fixed evolution parameters for simulation of the target scale for manufacture. There is indeed the risk of weighting too highly the choice of model (the scale-up optimal therefore responding too strongly to the qualitative understanding of the person using the system). Further, the determination of the optimal recipe may: (a) lack robustness to deviations in process conditions, (b) be sub-optimal due to the inherent non-linearities of the process in question.
During process execution, deviations from expected conditions may occur due to stochastic variation in the biological, chemical or physical systems or failure or compromise of part of the equipment. It is well established that “Model Based Control” is a powerful approach to respond to such a deviation. Traditional Model Based Control involves e.g. adjusting set-points to restore equilibrium or the original trajectory of the process. It takes a model of the process, and, working from the deviation, determines the optimum adjusted setpoints by simulating the process subject to the model.
Such Model Based Control has two important deficiencies:
The invention presented here resolves the above two deficiencies.
It is worth noting that the same principles of the method discussed above could be applied to solve a different problem than finding the optimal recipe, namely the definition of the utility function. Indeed, it may be useful to get an insight in how to systematically evaluate the performance of such a complex process as the production process in a quantifiable, objective manner. Accordingly, there may be one or more loops pertaining to properties of the utility function (e.g. parameters in it, dependencies on which performance indicators... ) instead of the loops on the recipe template composite, the variable recipe parameters and the recipe templates.
Details of exemplary embodiments are set forth below with reference to the exemplary drawings. Other features will be apparent from the description, the drawings, and from the claims. It should be understood, however, that even though embodiments are separately described, single features of different embodiments may be combined to further embodiments.
In the following, a detailed description of examples will be given with reference to the drawings. It should be understood that various modifications to the examples may be made. Unless explicitly indicated otherwise, elements of one example may be combined and used in other examples to form new examples.
The computer system 10 may include a processing unit, a system memory, and a system bus. The system bus couples various system components including the system memory to the processing unit. The processing unit may perform arithmetic, logic and/or control operations by accessing the system memory. The system memory may store information and/or instructions for use in combination with the processing unit. The system memory may include volatile and non-volatile memory, such as a random access memory (RAM) and a read only memory (ROM).
The computer system 10 may further include a hard disk drive for reading from and writing to a hard disk (not shown), and an external unit drive for reading from or writing to a removable unit. The drives and their associated computer-readable media provide nonvolatile storage of computer readable instructions, data structures, program modules and other data for the personal computer 920. The data structures may include relevant data for the implementation of the method as described above.
A number of program modules may be stored on the hard disk, external disk, ROM or RAM, including an operating system (not shown), one or more application programs, other program modules (not shown), and program data. The application programs may include at least a part of the functionality as described below.
A user may enter commands and information into the computer system 10 through input devices such as keyboard and mouse. A monitor or other type of display device is also connected to the system bus via an interface, such as a video input/output.
The computer system 10 may communicate with other electronic devices. To communicate, the computer system 10 may operate in a networked environment using connections to one or more electronic devices.
In particular, the computer system may interface and communicate with a control system 20. The computer system 10 may be operable, possibly in conjunction with other devices, to determine a recipe for a production process.
The control system 20 may be connected to a bioreactor 30 constituting the equipment for performing the production process. The control system 20 and the computer system 10 may be located in different rooms of the same facility or in any two informatically connected (e.g. via a network) locations. The computer system 10 may be a separate entity from the control systems 20. In other examples (not shown) the computer system 10 may coincide with the control system 20.
In some examples, a database 40 may be provided. The database may be connected to a network, such that the database is accessible by multiple devices/users. The database may be implemented as a cloud database, i.e. a database that runs on a cloud computing platform. In other words, the database may be accessible over the Internet via a provider that makes shared processing resources and data available to computers and other devices on demand. The database may be implemented using a virtual machine image or a database service. The database may use an SQL based or NoSQL data model.
The database 40 may store any of: sets of values for recipe parameters, recipe template composites, process evolution information, setup specification, likelihood functions, utility functions, mappings. Communications between the database 40 and the computer system 10 may be secured, e.g. via Internet protocol security (IPSEC) or other security protocols. A virtual private network (VPN) may also be used.
The database 40 may be hosted by a service provider, possibly on a virtual machine, and may be accessible by various users from multiple organizations, possibly located in a variety of different geographic locations around the world. Alternatively, the database 40 may be hosted locally, e.g. in the computer system 10. Accordingly, the computer system 10 and the database 40 may or may not be located in physical proximity. In particular, the database 40 may be located in a location that is geographically distant (e.g. on another continent) from the computer system.
An example of a production process for which a recipe is determined by the computer system 10 may be a fed-batch process comprising the following phases:
Starting with
Further, at 310, process evolution information 210 is retrieved, wherein the process evolution information describes the time evolution of one or more process variables that describe a state of the production process. The process evolution information 210 characterises how the process variables change with time, exemplarily including initial conditions for the process variables and relations among the process variables.
In particular, the process evolution information 210 may comprise relations empirically derived from previous executions of the production process and/or equations derived by theoretical models about the evolution of the production process. These relations and equations may be characterized by various parameters, denoted as evolution parameters. At least one of these evolution parameters may be a free parameter.
Further, at 310, a likelihood function 220, a mapping 230 and a utility function 240 are retrieved. The likelihood function 220 is a probability distribution for the values of the evolution parameters. In other words, the likelihood function 220 indicates how likely it is that the process evolution information 210, given a certain value for an evolution parameter, models the production process in an accurate manner. An exemplary plot of a likelihood function 220 for two evolution parameters, growth rate and death rate, is shown in
The mapping 230 translates the values of the process variables into values of performance indicators. In some cases, it is not straightforward to understand the link between the process variables and the relevant outputs of the production process, and the mapping 230 provides exactly this link.
The utility function 240 adds a further refinement to assessing the performance of the production process by associating a desirability value to the values of the performance indicators. In other words, the utility function 240 provides a quantitative characterization of the process performance in terms of the values of its performance indicators. An exemplary plot of a utility function 240 for two performance indicators, titre and quality of the product, is shown in
The recipe template composites 200, the process evolution information 210, the likelihood function 220, the mapping 230 and the utility function 240 may be retrieved by means of a file system. Exemplarily, these input data may be stored in a data storage device.
At 320, a recipe template composite is selected and a first optimisation step is performed. The first optimisation step is illustrated in
The second optimisation step is illustrated in
The third optimisation step is illustrated in
The fourth optimisation step is illustrated in
Once the loop 540 has completed, at 465 the utility value is computed by evaluating the utility function 240 using the values of the process variables from the trajectories 250 and the mapping 230.
Once one or more utility values have been obtained for the same recipe template composite, set of input values and set of likely values, depending on whether stochasticity is accounted for, the method proceeds at 450 by computing the utility tally from the one or more utility values.
Once one or more utility tallies have been obtained corresponding, respectively, to one or more sets of likely values, and for the same recipe template composite and set of input values, at 440 the utility score is computed by weighting the utility tally(ies) with the likelihood function 220.
From the plurality of utility scores obtained at 420, at 425 the optimal utility score is identified. In other words, the optimal set of input values that optimizes the utility score is identified. If there is only one recipe template composite, that recipe template composite 270 combined with the optimal set of input values 280 is determined to be the at least one recipe for the production process.
If there is more than one recipe template composite, each of the recipe template composites is selected in turn and at 330 the first optimisation step is repeated for all remaining recipe template composites. The repetition of the first optimisation step corresponds to the outermost loop 500 shown in
Accordingly, a plurality of optimal utility scores corresponding to the plurality of recipe template composites has been identified and, then, at 340 the best optimal utility score 260 is identified among those, i.e. the maximum or minimum of the plurality of optimal utility scores. The combination of the recipe template composite 270 and the set of input values 280 associated with (i.e. that result in) the best optimal utility score 260 is determined at 350 to be the at least one recipe for the production process.
The general method illustrated above with reference to
At 310, the process evolution information 210 is retrieved. The model for this production process may assume that viable cells reproduce at a maximum rate, r, and that growth is slowed down as the cell population approaches maximum capacity. The process evolution information 210 in this case comprises the following two differential equations:
wherein V is the viable cell density (VCD) [cells/ml], r is the cell growth rate [hr-1], kv is half of a capacity constant [cells.hr/ml] and lv is the cumulative cell count [cells.hr/ml]. The process evolution information 210 in this case comprises only one variable evolution parameter, namely the cell growth rate.
Further, a single recipe template composite 200 comprising one recipe template is retrieved, and the recipe template has one variable recipe parameter, the harvest time. A mapping 230 is also retrieved, wherein the mapping 230 is a function of the VCD and is the identity function, i.e. the VCD is itself considered a performance indicator. Additionally, a utility function 240 is retrieved, wherein the utility function 240 is only a function of the VCD, U(V)= a* log(V), wherein a is a constant.
In this example retrieving the likelihood function 220 comprises retrieving akin data, relatedness information and historical data and computing the likelihood function 220 based on all those, as explained in the following.
The cell type that will be used for the production process is cell type B. The prior knowledge available includes historical data for cell type B, in particular trajectories for the process variables, Db and a prior on the growth rate for cell type B, rb, derived e.g. from literature or from experience, as well as akin data related to an akin production process that used cell type A together with relatedness information. The akin data include trajectories for the process variables, Da, and a prior for the growth rate of cell type A, ra. The relatedness information between process B (using cell type B) and process A (using cell type A) may be based on experimental/literature knowledge of cell line behaviours and/or derived from a comparison between the historical data and the akin data.
The prior on rb is a uniform distribution in the interval [0, 0.15) and the prior on ra is a uniform distribution in the interval [0.02, 0.12). For each cell type x, Bayesian inference may be used to obtain a posterior distribution p(rx | Dx) ~ p(Dx | rx) p(rx), wherein p(rx) is the prior and p(Dx | rx) is the conditional distribution of the data for a given value of the growth rate. The value of the conditional distribution may be calculated at each data point e.g. by assuming a Gaussian distribution centred on the mean rx. The posterior distributions obtained in this manner are shown in
The relatedness information in this case is the probability of a value for rb given a value for ra. The relatedness is R(rb, ra) = N(rb- ra, σab), wherein N is the normal distribution.
The available prior knowledge discussed above is merged in the joint probability distribution for ra and rb, p(ra, rb | Da, Db) ~ p(ra | Da) p(rb | Db) R (rb, ra). It should be noted that, if the relatedness is not taken into consideration, the joint probability reduces to p(ra | Da) p(rb | Db).
The likelihood function 220, L(rb), is computed as the integral over the joint probability distribution:
After having retrieved all the necessary inputs, the first optimisation step is performed at 320, which comprises providing 410 an input value (set of one element) for the harvest time and then performing 415 the second optimisation step. The second optimisation step comprises providing 430 a likely value (set of one element) for the growth rate based on the likelihood function 220. In this example, the likely value for rb is chosen as the most likely value, i.e. the value that maximises the likelihood function
argmax L(rb). Then at 435 the third optimisation step is performed. No other values are considered for the growth rate, this means there is no loop within the second optimisation step and the utility score computed at 440 will coincide with the only utility tally. Further, if stochasticity is not considered, the utility tally computed at 450 will coincide with the utility value computed at 465 by performing the fourth optimisation step at 445.
In the fourth optimisation step, at 455 a simulation is generated using the process evolution information 210 of equations (5) with the most likely value for the growth rate and given initial conditions, the retrieved recipe template composite 200 comprising only one recipe template, and the input value for the harvest time, denoted e.g. with ht1. From the simulation, the values of the VCD at different times are obtained at 460, i.e. the trajectory for the VCD. In particular, in order to evaluate the utility function 240, the value of the VCD at the time coinciding with the harvest time may be considered. At 465 the utility value is then computed as U[V(ht1)].
At 420 the above steps are repeated and a plurality of utility scores U[V(ht1)], ... U[V(htn)] is obtained, with n greater than or equal to 2. The plurality of input values for the harvest time may be retrieved together with the recipe template composite, or input by a user, or drawn from a distribution according to certain criteria.
Alternatively, in this very specific case, instead of performing the loop to obtain the plurality of utility scores, the trajectory for the VCD obtained from a single simulation may be used to compute U[V(ht1)], ... U[V(htn)], since the trajectory is indeed V(t). This “shortcut” can be generally applied to cases in which the variable recipe parameter is a timepoint e.g. for performing an action, because the trajectories already provide the values of the process variables at different times. The same holds if there are more variable recipe parameters and each one of them is a timepoint.
At 425 the score function US(ht) = U[V(ht)] is optimised e.g. using numerical methods in order to find the optimal utility score and the associated htopt. The recipe template with harvest time htopt is provided to the control system of the bioreactor, so that harvest is carried out automatically at htopt, thereby optimizing the utility score and, thus, the production process.
At 310, the process evolution information 210 is retrieved. The model for this production process may assume that viable cells reproduce at a maximum rate, r, and that growth is slowed down is slowed by an inhibitory toxin which is produced as a byproduct of cell growth. Further, cells have a finite lifetime and become apoptotic at a constant rate δv, and apoptotic cells quickly lyse at rate δA, after which they can no longer be detected. The process evolution information 210 in this case comprises the following three differential equations:
wherein V is the viable cell density [cells/ml], r is the cell growth rate [hr-1], s is the sensitivity to inhibitor [mM], c is the sigmoid centre for inhibitor response [mM], δv is the apoptosis rate [hr-1], A is the apoptotic cell density [cells/ml], δA is the lysis rate [hr-1], τ is the inhibitor concentration [mM] and θ is the inhibitor production [µmol/cell]. The process evolution information 210 in this case comprises two variable evolution parameters, namely the cell growth rate and the inhibitor production.
Further, a single recipe template composite 200 comprising one recipe template is retrieved, and the recipe template has one variable recipe parameter, the harvest time. A mapping 230 is also retrieved, wherein the mapping 230 is a function of the VCD and the inhibitor concentration and is the identity function, i.e. V and τ are both process variables and performance indicators. Additionally, a utility function 240 is retrieved, wherein the utility function 240 depends not only on reaching sufficient cell density but also on the amount of inhibitor in the media, namely
In this example retrieving the likelihood function 220 comprises retrieving akin data related to two different akin processes, A and B, and corresponding relatedness information and computing the likelihood function 220 based on all those, as explained in the following.
The production process C for which an optimal recipe should be determined uses a certain cell type at a large scale. The prior knowledge available includes akin data for akin production process A, Da, which uses the same cell type as C but at a smaller scale, and akin data for production process B, Db, which uses a different cell type than C but at the same scale as C. For example, the datasets Da and Db may include trajectory data from a large number of runs.
For each process (A and B), Bayesian inference may be used to obtain a posterior distribution p(rx, θX | Dx) ~ p(Dx | rx, θx), wherein p(Dx | rx, θx) is the conditional distribution of the data for a given value of the growth rate and the inhibitor production, while the prior is uniform. The conditional distribution may be obtained by computing the probability that, for each run in Dx, the run data could have been produced if the “true” values for the evolution parameters had been rx, θx. For example, the run data in Dx may comprise a time series of cell densities, a(t), and simulated data for the times series, b(t), may be obtained based on the evolution parameters rx, θx. Exemplarily, the probability is estimated as the integral of N(a(t)-b(t), σ), where σ is an estimate of the standard error of the measurement system for the cell density. The posterior distributions obtained in this manner are shown in
The relatedness information may be based on experimental/literature knowledge about the production processes. The relatedness information in this case is the probability of value for the evolution parameters for process C given values for evolution parameters for process A/B. It may be assumed that R(A, C) = R(rc, ra)* R(θc, θa ), and similarly for B.
As mentioned, the cell type for processes A and C is the same. According to the scientific literature, this cell type slows down its processes at the scale of C due to the limits on the dissolved oxygen. Even if the dissolved oxygen is not an evolution parameter in the process evolution information, this information can be taken into consideration in the relatedness between A and C. In particular, the relatedness is chosen as a distribution with a longer tail towards the lower end, reflecting the probability that both the growth rate and the inhibitor production will be lower for process C than for process A.
The cell types for processes B and C are different but similar, and it is assumed that they behave similarly in terms of growth rate and inhibitor production, although there is a larger uncertainty about the latter, which is reflected in the choice of a larger variance.
The posterior distributions obtain from the akin data and the relatedness information are combined to compute the likelihood function 220 for process C. Specifically,
After having retrieved all the necessary inputs, the first optimisation step is performed at 320, which comprises providing 410 an input value (set of one element) for the harvest time and then performing 415 the second optimisation step. The second optimisation step comprises providing 430 a set of likely values for the variable evolution parameters based on the likelihood function 220 of equation (9). In this example, the likely values for the growth rate and the inhibitor production are chosen as the most likely values, i.e. the values that maximise the likelihood function
= argmax L(rc, θc). Then at 435 the third optimisation step is performed. No other values are considered for the variable evolution parameters, this means there is no loop within the second optimisation step and the utility score computed at 440 will coincide with the only utility tally. Further, if stochasticity is not considered, the utility tally computed at 450 will coincide with the utility value computed at 465 by performing the fourth optimisation step at 445.
In the fourth optimisation step, at 455 a simulation is generated using the process evolution information 210 of equations (7) with the most likely values for (rc, θc) and given initial conditions, the retrieved recipe template composite 200 comprising only one recipe template, and the input value for the harvest time, denoted e.g. with ht1. From the simulation, the values of the process variables (in particular VCD and inhibitor concentration) at different times are obtained at 460, i.e. the trajectories. The left-side plot in
In this example, similarly to the first scenario, the utility score function can be directly obtained from the trajectories, which already provide V and τ as a function of time, without looping over the second optimisation step. The values of the process variables are fed to the utility function 240 of equation (8), a plot of which is shown on the right side of
At 425 the score function US(ht) = U[V(ht), τ (ht)] is optimised e.g. using numerical methods in order to find the optimal utility score and the associated htopt. In this example the optimal utility score equals 13.13, while htopt = 108 hours. These values correspond to V = 23.65*105 cells/ml and τ= 2.54 mM. The recipe template with harvest time htopt is provided to the control system of the bioreactor, so that harvest is carried out automatically at htopt, thereby optimizing the utility score and, thus, the production process C.
At 310, the process evolution information 210 is retrieved. The model for this production process may assume that viable cells reproduce at a maximum rate, r, and that growth is slowed down is slowed by an inhibitory toxin which is produced as a byproduct of cell growth. Further, cells have a finite lifetime and become apoptotic at a constant rate δv Viable cells produce protein with activity r and protein is degraded by the build-up of the inhibitor at a rate δτ. The process evolution information 210 in this case comprises the following four differential equations:
wherein V is the viable cell density [cells/ml], r is the cell growth rate [hr-1], k is the inhibitor concentration capacity [mM*hr/ml], δv is the apoptosis rate [hr-1], lτ is the cumulative inhibitor concentration [inhibitor*hr/ml], τ is the inhibitor concentration [mM], θ is the inhibitor production [µmol/cell], δτ is the protein degradation [mM-1*hr-1], p is the viable protein concentration [units/ml] and ρ is the protein production rate [units/cell*hr]. The process evolution information 210 in this case comprises two variable evolution parameters, namely the cell growth rate and the inhibitor concentration capacity.
Further, a mapping 230 is retrieved, wherein the mapping 230 is a function of the VCD and the viable protein concentration and is the identity function, i.e. V and p are both process variables and performance indicators. Additionally, a utility function U(V,p) 240 is retrieved.
In this scenario, the production process has already started and the recipe is updated and optimized on the fly. Accordingly, the recipe template composite 200 retrieved corresponds to the recipe currently being used, in which the harvest time is the variable recipe parameter whose value has to be chosen in order to optimize the utility function.
In this example, retrieving the likelihood function 220 comprises retrieving current data related to an ongoing run of the production process B, Db, akin data related to an akin process A, Da, as well as corresponding relatedness information and computing the likelihood function 220 based on all those, as explained in the following. Process B uses cell type B and process A uses cell type A.
In particular, the dataset Da may include trajectory data from a large number of runs. For process A, Bayesian inference may be used to obtain a probability or posterior distribution p(ra, ka | Da) ~ p(Da | ra, ka), wherein p(Da | ra, ka) is the conditional distribution of the data given values for the growth rate r, and the inhibitor concentration saturation constant k, while the prior is again uniform. The conditional distribution may be obtained by computing the probability that, for each run in Da, the run data could have been produced if the “true” values for the evolution parameters had been ra, ka.
Indeed, if data relative to more than a single run are available, the probability surfaces can be combined additively to give an increasingly more informative description of the probability landscape. The larger the dataset, the closer is the probability landscape to predicting the population mean for each of the evolution parameters of interest.
The relatedness information may be based on experimental/literature knowledge about the production processes and may be a joint distribution over the two variable evolution parameters, R(rb, ra, kb , ka), e.g. a multivariate normal distribution having standard deviation equal to 0.012 for the growth rate and standard deviation equal to 0.008 for the inhibitor saturation constant (the off-diagonal terms of the covariance matrix are set to zero).
Based on the partial data Db from the ongoing process, it is also possible to obtain a posterior or probability distribution p(rb, kb | Db) similarly to what explained above for A. Since it is based on a single, partial run, this distribution may not be particularly informative, but it can be combined with p(ra, ka | Da) and the relatedness information R(rb, ra, kb , ka) to obtain a likelihood function 220 that is more informative, namely
After having retrieved all the necessary inputs, the first optimisation step is performed at 320, which comprises providing 410 an input value (set of one element) for the harvest time and then performing 415 the second optimisation step. The second optimisation step comprises providing 430 a set of likely values for the variable evolution parameters based on the likelihood function 220 of equation (11). In this example, the likely values for the growth rate and the inhibitor concentration capacity are chosen as the most likely values, i.e. the values that maximise the likelihood function (rb,
= argmax L(rb,kb ). Then at 435 the third optimisation step is performed. No other values are considered for the variable evolution parameters, this means there is no loop within the second optimisation step and the utility score computed at 440 will coincide with the only utility tally. Further, if stochasticity is not considered, the utility tally computed at 450 will coincide with the utility value computed at 465 by performing the fourth optimisation step at 445.
In the fourth optimisation step, at 455 a simulation is generated using the process evolution information 210 of equations (10) with the most likely values for (rb, kb) and given initial conditions, the recipe being used in the ongoing run, and the input value for the harvest time, denoted e.g. with ht1. From the simulation, the values of the process variables at different times are obtained at 460, i.e. the trajectories.
In this example, similarly to the first and second scenarios, the utility score function can be directly obtained from the trajectories, which already provide V and p as a function of time, without looping within the first optimisation step. At 425 the score function US(ht) = U[V(ht), p (ht)] is optimised e.g. using numerical methods in order to find the optimal utility score and the associated htopt. The value for harvest time htopt is provided to the control system of the bioreactor, so that the recipe is updated on the fly and harvest is carried out automatically at htopt, thereby optimizing the utility score and, thus, the production process B.
Once the run is completed, the trajectories data are stored so that they can be part of the prior knowledge space for subsequent runs.
At 310, the process evolution information 210 is retrieved. The model for this production process may assume that viable cells reproduce at a maximum rate, r, and that growth is slowed down is slowed by an inhibitory toxin which is produced as a byproduct of cell growth. Further, cells have a finite lifetime and become apoptotic at a constant rate δv. Viable cells produce protein with activity ρ and protein is degraded by the build-up of the inhibitor at a rate δτ. The cell growth rate is dependent on the pH being maintained around an optimal level pHgro, while deviations on either side of this optimum will inhibit growth. The protein is stable for high pH but is denatured at low pH at a rate δpH and with centre point of the response c.
The process evolution information 210 in this case comprises the following four differential equations:
wherein V is the viable cell density [cells/ml], r is the cell growth rate [hr-1], k is the inhibitor concentration capacity [mM*hr/ml], δv is the apoptosis rate [hr-1], lτ is the cumulative inhibitor concentration [inhibitor*hr/ml], τ is the inhibitor concentration [mM], θ is the inhibitor production [µmol/cell], δτ is the protein degradation [mM-1*hr-1], p is the viable protein concentration [units/ml], ρ is the protein production rate [units/cell*hr], pHgro is the optimal pH for growth [pH], σ is the tolerance to deviation from pHgro [pH], δpH is the rate of pH-mediated protein degradation [hr-1], s is the sensitivity to pH change [pH-1] and c is the centre point of response [pH]. The process evolution information 210 in this case comprises three variable evolution parameters, namely the cell growth rate, the inhibitor production and the protein production rate.
Besides the process evolution information, two recipe template composites 200 are retrieved, each comprising a single recipe template. The first recipe template involves a shift of the pH during the execution and comprises three variable recipe templates, the starting value for the pH, the magnitude of the pH shift, ΔpH, and the time of the shift, tpH. The second recipe template comprises one variable recipe template, the value at which the pH is set, and according to this recipe template the pH is kept constant.
Further, a mapping 230 is retrieved, wherein the mapping 230 is a function M of the viable protein concentration and associates it to the titre T as performance indicator. Additionally, a utility function 240 is retrieved. The goal of the optimisation is not simply to produce the most protein but to consider also the costs of the process, in particular due to the running time ft. The utility function 240 is given by U(T,ft) = 10T
(13) and is plotted in
In this example, retrieving the likelihood function 220 comprises retrieving historical data related to one or more previous runs of the production process for which the recipe has to be optimized and computing the likelihood function 220 based on those, as explained in the following. In particular, the historical data may include trajectory data for at least some of the process variables.
Based on the historical data, Bayesian inference may be used to obtain the likelihood function L(r, ρ, θ) as the probability or posterior distribution p(r, ρ, θ| D) ~ p(D | r, ρ, θ), wherein p(D | r, ρ, θ) is the conditional distribution of the data for a given value of the growth rate and the inhibitor concentration capacity, while the priors are uniform. The conditional distribution may be obtained by computing the probability that, for each run in D, the run data could have been produced if the “true” values for the evolution parameters had been r, ρ, θ.
After having retrieved all the necessary inputs, the first recipe template is selected and the first optimisation step is performed at 320, which comprises providing 410 a set of input values for pH, ΔpH, and tpH, and then performing 415 the second optimisation step. The second optimisation step comprises providing 430 a set of likely values for the variable evolution parameters r, θ, ρ based on the likelihood function 220. In this example, the likely values may be drawn randomly with the probability of selection dependant on the likelihood function 220.
Then at 435 the third optimisation step is performed. If stochasticity is not considered, the third optimisation step coincides with the fourth optimisation step and the utility tally computed at 450 will coincide with the utility value computed at 465 by performing the fourth optimisation step at 445. In the fourth optimisation step, at 455 a simulation is generated using the process evolution information 210 of equations (12) with the provided set of likely values, given initial conditions, the first recipe template, and the set of input values for the variable recipe parameters.
From the simulation, the values of the process variables at different times are obtained at 460, i.e. the trajectories, in particular for the protein concentration p. Via the mapping, the values for the titre T can be obtained and fed to the utility function 240 of equation (13), in particular T at the final time, which is also the running time, T(ft).The utility value is then UV=U(T(ft), ft] and is associated to the given recipe template composite, the given set of input values and the set of likely values.
The third/fourth optimisation step is repeated by providing different sets of likely values, thereby obtaining a plurality of utility values/utility tallies UV1, UV2, ... UVn. At 440 the utility score for a specific set of input values
is computed by weighting the utility values by the likelihood function 220:
At 420 the second optimisation step is repeated by iterating over different values of pH, ΔpH, tpH, thereby obtaining a plurality of utility scores, from which a score function US (pH,tpH,ΔpH)|〈R1〉 is defined, and at 425 the optimal utility score is identified, namely by optimizing the score function
The optimal utility score
= US (pHopt,
, ΔpHopt)|〈R1〉 associated with the first recipe template R1 equals 78.9.
The whole procedure is repeated again at 330 for the second recipe template R2, obtaining an optimal utility score associate thereto,
= 37.55. At 340 the best optimal utility score is identified, which in this case is
The first recipe template with the values starting pH = 7.1, ΔpH = 0.24 and tpH = 106 hours is provided to the control system of the bioreactor, so that the production process is optimised, i.e. executed in a manner that maximises utility.
| Number | Date | Country | Kind |
|---|---|---|---|
| 20171236.1 | Apr 2020 | EP | regional |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/EP2021/060289 | 4/21/2021 | WO |
