US 20060031024 A1
A method of optimizing parameter values in a process, which process is essentially controlled by a set of parameters affecting a set of properties characterizing an output of the process. The method may use an analytic hierarchy process (AHP) to associate a weight with each property according to its relative importance to obtain desired product characteristics. The method also uses parameter data and measured property data from a required number of experimental runs of the process, from which data property behavior relations between each property and the parameters are statistically established, which relations give estimated property values. Using the property weights, a process goal function is established, which is expressed in terms of weighted deviations between the estimated property values and the corresponding goal values for the properties. The process goal function is minimized in order to generate a set of optimal parameter values for the process.
1. A method of optimizing parameter values of a process, the process being essentially controlled by a set of n parameters X that affect a set of k properties Y characterizing an output of the process, said method comprising:
assigning values to a set of k property weights representing relative importance of said properties;
establishing property behavior mathematical relations giving an estimated property for each said property;
using said property weights to establish a goal function in terms of property weighted deviations between the estimated properties and corresponding specified goal values for said properties; and
minimizing the goal function to generate a set of n optimal parameter values for said parameters.
2. A method according to
3. A method according to
4. A method according to
5. An apparatus capable of optimizing parameter values of a process, the process being essentially controlled by a set of n parameters X that affect a set of k properties Y characterizing an output of the process, said apparatus comprising:
means for assigning values to a set of k property weights representing relative importance of said properties;
means for establishing property behavior mathematical relations giving an estimated property for each said property;
means for using said property weights to establish a goal function in terms of property weighted deviations between the estimated properties and corresponding specified goal values for said properties; and
means for minimizing the goal function to generate a set of n optimal parameter values for said parameters.
1. Field of the Invention
The invention relates to the process optimization field, and more particularly to a method of optimizing parameter values in a process of producing a product which is characterized by properties affected by the selected parameter values. This invention is applicable in different industries, such as the pharmaceutical, chemical, cosmetics, plastics, petrochemical, agriculture, metallurgy and food industries, as well as many other commercial and industrial applications.
2. Description of Prior Art
Processes for production of complex compositions such as those found in many pharmaceutical products generally require the mixing of many ingredients according to specific process parameters regarding formulation and production technology, to provide the product with properties at a level offering satisfactory performance according to predetermined specifications. In such complex production processes, it is not unusual that some process parameters involved exhibit interfering effects on the desired properties, further complicating the process design. Where possible, the designer may try to adapt the set of process parameters from known data derived from previous similar processes, and/or rely on conventional trial-and-error experimental schemes to optimize the set of process parameters values, in order to meet the product specifications. However, as the processes become more complex, optimization in such multidimensional space with high accuracy requirements turns out to be an extremely difficult task, even for the highly skilled designer. That limitation is particularly problematic in the design of pharmaceutical products, where one or more active substances mixed with a variety of excipients (e.g. carriers) must be produced in the form of a stable and highly effective standard delivery system such as a tablet, capsule, suspension, cream or injection, or even controlled release systems such as skin carriers and implants.
In the past years, many techniques have been developed to assist the process designer or formulator in optimizing values of parameters governing processes. These techniques aim at quantify existing relations between parameters and associated desired product performance characteristics. A conventional technique known as the Full Factorial Matrix (FFM) method consists of statistically deriving a behavior relations for the properties from a set of experimental runs of the process using selected initial values for the parameters. The established model being generally nonlinear, optimized parameter values are then derived using an optimization method such as the Multisimplex method described in “Practical Methods of Optimization” J, Wiley & Sons, Chichester, 2d, (1987), which essentially consists of linearizing the behavior functions related to the parameters according to straight lines or planes of different random directions. For any given property behavior relation of n parameters to be optimized in order to either minimize or maximize that behavior relation with or without constraints on the parameter values, a recursive estimation of the property is then performed using an initial set of parameter values according to a selected direction, until the obtained value for the property does not significantly vary in that direction. Then, a last unfavorable set of parameters is used as a new starting point for a following recursive estimation according to a different direction. Successive recursive estimation steps are performed until the resulting value for the property no longer significantly vary in any new direction. When applied to a model comprising a plurality of property behavior relations, the Multisimplex method allows a unique objective function to be created by proper transformation of the relations to adapt to different scales and/or units and by associating a relative importance weight to each property, either subjectively or through fuzzy logic algorithms.
The known optimization processes based on Full Factorial Matrix-Multisimplex methods suffer from several drawbacks. As a general rule, the number of experimental runs required to obtain a model of sufficient reliability is proportional to the total number of significant parameters involved. Therefore, the cost and time frame of the experimental work will therefore be essentially proportional to the number of runs required. Although a variant of the method known as the Fractional Factorial Matrix has been proposed in order to reduce the number of runs to be performed, the provided reduction of experimental runs may not significantly reduce the total cost and time frame of the work required to complete the design of a complex product involving many production technologies. While adequate formulations complying with constraints imposed on the parameter values can nevertheless be obtained, these formulations generally cannot be qualified as optimal when comparing actual property performance with desired property values set forth in the product specifications.
A technique which attempts to improve parameter optimization in process design is disclosed in European Patent Office laid-open patent application publication number 0,430,753 dated Jun. 5, 1991 and in U.S. Pat. No. 5,218,526 issued on Jun. 8, 1993 to Mozzo. According to the technique in Mozzo, from a set of property relations expressed in terms of parameters which is obtained by standard statistical methods using the results of a number of experimental runs of the process, a corresponding set of property relations expressed in terms of weighted parameters is derived. For each actual value of a parameter, a first weighting is expressed as the ratio of: (a) the deviation of the actual value from the mean value of the parameter over the experimental range, on (b) the range between extreme values for that parameter over the experimental range. Then, a goal function is established in term of deviations between weighted values of property values as estimated by the property relations and corresponding weighted values of specified goal values for the properties. For each goal value of a property, a second weighting is expressed as the ratio of: (a) the deviation of the actual value from the mean value of the property over the experimental range, on (b) the range between extreme values for that property over the experimental range. Then, according to a recursive geometric algorithm aimed at successively minimizing the established goal function, a set of optimal parameter values is generated. While being an improvement over the conventional Full/Fractional Factorial Matrix—Multisimplex methods regarding the capability to consider specified goal values for the properties, the weightings as taught by Mozzo do not reflect the relative importance of the properties involved, and that limitation may therefore affect the convergence of the algorithm toward an optimal solution.
A review of modern techniques and software systems for the design of pharmaceutical product formulations is given in “Intelligent Software System For Pharmaceutical Product Formulation” R. C. Rowe, Pharmaceutical Technology, March 1997. In that paper, expert systems, rule induction algorithms, case-based reasoning algorithms, neural networks and genetic networks are presented as modern tools for supporting formulation design, and a number of available software systems using some of these tools are summarized. As indicated in the Rowe paper, although a knowledgeable expert system could be a powerful tool to assist the process designer in the formulation task, its development is generally a high risk, time consuming and expensive process. Rule induction is a knowledge-based algorithm which allows hierarchical classification of objects, using statistical methods which are found generally effective only if the input data is continuous, which is often not the case in practice. Moreover, since rule induction is limited to establishing whether or not a given object is close to another, it generally cannot provide an optimal solution. Case-based reasoning is a knowledge-based iterative technique which can be used to design formulations, which consists of matching the desired specifications for the product with the specifications of the most relevant known formulation(s), and adapting the selected formulation(s) as necessary, followed by an evaluation. Although effective for optimizing the parameters of a variant process from a family of similar processes and corresponding formulations, case-based reasoning generally cannot be used where the design of a significantly different formulation is contemplated. As to neural networks, in which each neuron input is modified by a weight associated with that neuron, they appear to be effective tools for assisting formulation design only in cases where no constraint applies on either the parameter or property values, such cases being rarely found in practice. Finally, regarding the genetics algorithms, they are cyclic methods based on Markov chains for predicting from a starting point a solution likely to result from a sequence of operations, in order to allow making changes to obtain a desired solution. Since these changes are generally made arbitrarily, in most cases, the resulting solution cannot be considered as optimal.
It is therefore an object of the present invention to provide a systematic method of optimizing parameter values in a process for producing a product which minimizes the number of experimental runs required to obtain an optimal solution complying with the product specifications.
According to the above object, from a broad aspect of the present invention, there is provided a method of optimizing parameter values in a process of producing a product, the process being essentially controlled by a set of n parameters Xi affecting a set of k properties Yj characterizing the product. The method comprises the steps of: i) assigning values to a set of k property weights wj representing relative importance of the properties Yj for the characterization of the product; ii) establishing property behavior mathematical relations giving an estimated property Yej for each property Yj in terms of the parameters Xi from given parameter data and associated property data; iii) using the property weights wj to establish a goal function in terms of property weighted deviations between the estimated properties Yej and corresponding specified goal values for the properties Yj; and iv) optimizing the goal function to generate a set of n optimal parameter values for the parameters Xi.
According to a further broad aspect of the present invention, there is provided a method of producing a pharmaceutical product using optimized process parameter values, the process being essentially controlled by a set of n parameters Xi characterizing a formulation for the product, the parameters Xi affecting a set of k properties Yj characterizing the product. The method comprises the steps of: a) conducting a number of l of experimental runs of the process each using a selected distinct set of values for the parameters Xi covering substantially all extreme values within a chosen range of values for each one of the parameters Xi, wherein l is at least equal to n+1 and is substantially less than a number used in the Fractional Factorial Matrix method; b) measuring values for the properties Yj characterizing the product in each of the l experimental runs, whereby parameter data and associated property data are obtained from the selected distinct set of values for the parameters Xi and the measured values for the properties Yj, respectively; c) determining an importance of the properties Yj for the characterization of the product, comparing the importance of the properties Yj relative to one another, and assigning values to a set of k property weights Wj representing a relative importance of the properties Yj for the characterization of the product; d)calculating a set of optimal parameter values for the parameters Xi using the measured values for the properties Yj and the assigned values of the set of k property weights wj; and e) producing the pharmaceutical product using the optimized process parameter values Xi calculated in the previous step.
The invention will be better understood by way of the following detailed description of a preferred embodiment with reference to the appended drawings, in which:
In the following description, a preferred embodiment of the present invention applied to product formulation design will be described. However, it is to be understood that the present invention can be also be used to optimize parameter values of processes related to the production of many types of products which cannot be associated with a formulation, while being characterized by a number of properties affected by process parameters, such as biotechnological products, electronic components, etc.
Referring now to
Modules 16, 20 and 22 are linked to a property behavior models module 24 that uses experimental data, parameter interaction data and remaining significant parameters for determining an optimal mathematical model for each property which is likely to better estimate that property. The model data as generated at module 24 is fed to a property behavior relation module 26 that also receives experimental data from module 16 to statistically estimate polynomial coefficients to be incorporated within the established property behavior models, thereby generating a behavior relation for each property. The S-Plus statistical software from MathSoft may be used to program module 26 to apply the appropriate regression methods to the data. System 10 is further provided with a goal function module 28 linked to property weighting module 14 and property behavior relation module 26 to generate, from specified goal values for the properties, a goal function in term of property weighted deviations between properties as estimated by the behavior relations and the corresponding specified goal values for these properties.
An optimization module 30 is provided to optimize the goal function as established by module 28 through successive iterations and according to the type of each variable (discrete or continuous) and according to one or more ranges specified as constraints imposed on one or more optimal parameter values. Module 30 can be programmed using Matlab™ software supplied by The Math Works Inc to implement network optimization methods. Optimization module 30 is linked to the experimental data entry module 16 to transfer thereto the generated set of optimal parameter values, which module 16 also stores the actual property values obtained from an experimental run based on the set of optimal parameter values. All experimental data is then transferred to the evaluation module 18 as mentioned before.
A preferred embodiment of an optimization method according to the present invention will now be described with reference to
The AHP method consists of building a hierarchical tree from all properties, with one or more hierarchical levels depending on existing relations between the properties. For each level, a pair-wise comparison matrix is built between the properties of this level and presented at an input of the parameters weighting module 14 shown in
In a parallel direction, each pair-wise comparison is associated with a consistency index reflecting the transitivity relation between all comparison by pairs given by the formulator. Multi-criteria analysis software which is commercially available, such as Expertchoice™, Criterium™ or Ergo™, may be used to program module 14. For example, to one or more m main properties classified at a first (higher) level, may correspond one or more groups of properties classified at a second (lower) level, the latter properties being therefore identified as sub-properties. For each main property associated with a group of p sub-properties, a matrix of dimension [p+1×p+1] is built and filled, as a result of a pair-wise comparison between each property and sub-property, using relative importance values selected from a standard AHP scale. Next, a suitable algorithm performed by parameter weighting module 14 consists of first calculating the higher eigenvalue of the resulting numerical matrix, and then deriving a normalized relative importance vector of dimension [p+1] by an estimation of the left principal eigenvector of that matrix associated with the calculated main eigenvalue of the input matrix. The above algorithm is then applied to compare the m main properties of the higher level, from a pair-wise comparison matrix of dimension [m×m] from which a normalized relative importance vector of dimension [m] is derived. Finally, the above normalized vectors are combined according to the hierarchical relations to generate a global relative importance weight vector for the k properties of dimension [m+Σp] or [k]. In practice, it is generally appropriate to retain only each group of sub-properties without the corresponding main property, the sum of the weights related to the retained k properties/sub-properties being always equal to unity.
According to the next step, namely step 42, parameter data and property data values are provided, which data is obtained from experimental runs using different sets of parameter values for the process, the various values for each parameter being preferably selected according to an expected operation range within which an optimal parameter value is likely to be found. The parameters Xi used in the experimental runs should cover the extremes of the expected operational range for each parameter. Generally, the number of formulation combinations required to determine an optimal formulation depends on many factors among which the more important ones are: 1) the formulation designer experience; 2) complexity of the formulation; 3) the availability of literature and experimental data available on the desired product; and 4) the analytical laboratory workload and throughput. According to the method of the present invention, the minimal number of experimental runs l to perform has been found to be equal to n+1, wherein n is the number of relevant parameters involved. A greater number of runs is certainly possible. Step 42 is performed by experimental data entry module 16 shown in
The method then comprises a step 44 of establishing property behavior mathematical relations linking the properties with the parameters and interactions thereof, in polynomial form. These property behavior relations provide an estimated property Yej for each of the k properties Yj in terms of a number n of parameters Xi from the parameter data and associated property data provided at step 42. Step 44 is typically comprised of four sub-steps, namely 1) a parameters reduction step performed by module 22, 2) a parameters interaction analysis step performed by module 20, 3) a property behavior modeling step performed by module 24, and 4) a property behavior relations generating step performed by module 26, as shown in
The parameters associated with the retained correlation factors form the reduced set of n parameters.
It can be also shown that a minimum number l of runs at least equal to n+1 is required to obtain reliable parameters estimation. Then, parameter interactions, that are in the form XiXj with i≠j and which are significant, can be identified using the above relations (1), with the suggested specific ranges given in (2). The values for Xi from the l experimental runs are combined with the retained correlation factors ρij to form a final matrix W, with each element of the first column being equal to unity for the purpose of following sub-step 4). As to sub-step 3), it consists of establishing, for each property Yj, a best model in terms of retained parameters and parameter interactions. A standard variance analysis is carried out to confirm relevancy of all parameter coefficients and parameter interaction coefficients, and to select by successive variance analysis operations through the use of modules 24, 20 and 22, a suitable model amongst different predetermined models of upgraded degrees, whenever difference in performance between a given model of degree r and a following model of degree r+1 is found to be not significant. The resulting best model is taken along with matrix W and property experimental data in matrix Y, as inputs for following sub-step 4) aimed at generating property behavior relations for each property Yj. A matrix C of coefficient values is given by the matrix:
A following step 46 as shown in
The goal function to be minimized may be expressed as follows:
The “G” goal function is determinated by experimentation. The optimization of the “G” function is a step by step procedure. The first step is to obtain the behavior laws with the best fit between the experimental data and their corresponding ideal value factor.
The second step, the optimization is based on a initial point.
H=the Hessian of G
We observe a perfect overlap between the two goal functions and on the stationary point the goal function will be;
These equations supply the maxima and minima of the goal
This mathematical approach induces a reduction of the dimension of the variables, consequently we pass from “n” variables to “n−1” variables. In the actual case, we start with the most important variables from the behavior laws with the highest weight values of the factor.
This approach is known under the name of network optimization, in this case the network nodes are built by the optimal values of the variable by decreasing order of the factor's rank.
After the iterative optimization step 48 is completed, although the set of optimal parameter values X0 i obtained can generally be considered as the solution to recommend, that solution is preferably evaluated amongst other alternative solutions by following steps 50 and 52 as shown in
An example illustrating an application of the method according to the present invention in the pharmaceutical field will now be described.
Formulation and production process for enalapril maleate tablets were optimized in order to provide a drug product with satisfactory biological performance as well as stability when packaged and stored under ICH (International Conference on Harmonization) conditions. Three (3) independent formulation and process parameters (n=3) were identified as having an impact on the stability of the drug product: 1) the degree of drug neutralization during granulation (X1); 2) the manufacturing technology (X2); and 3) The drug-to-excipient ratio in the formulation, i.e. dose strength (X3).
As to the degree of drug neutralization during granulation (X1), it was classified as either complete, partial or no neutralization. In the case of complete neutralization, the drug and the alkaline agent were both added to the granulation fluid, i.e. water. Therefore, the alkaline agent neutralized the drug prior to its addition to the powder blend for the granulation procedure. In partial neutralization, both the drug and the alkaline agent were added to the powder blend, blended and water added as the granulation fluid for the granulation procedure. When water and/or alkaline agent were not added to the formulation, the drug was not neutralized. The level of water added as well as the drug-to-alkaline agent ratio were kept constant for all of the formulations. The level of the alkaline agent was determined by the stoichiometry of the reaction.
The manufacturing technology (X2 ) was either wet granulation (X2=0) or direct compression (X2=1). These two technologies are used worldwide for the manufacturing of probably more than 90% of all of the solid oral dosage forms. In the wet granulation technology, the drug and other functional materials added to impart good processing attributes to the drug, often called excipients, are first blended together and agglomerated into larger particles by the addition of a granulating fluid. The role of the granulating fluid is to promote the development of adhesive forces between the materials required for the agglomeration process. After granulation, the granulating fluid is removed by drying. When a direct compression approach is selected as a manufacturing method, the drug is first blended with the excipients and tablets produced without the use of a granulating fluid.
As to the dose strength (X3), four doses of the product were developed, which were obtained by using two formulations with different drug-to-excipient ratios (continuous parameter values) compressed at different tablet weights.
A total of nine (9) experimental runs involving different formulations based on a combination of the three parameters were prepared, as shown in Table 1.
The nine formulations covered all of the six (6) possible combinations for the wet granulation technology and three (3) combinations of direction compression. Tablets were manufactured by using enalapril maleate with USP/NF and EP excipients. In the direct compression technology, there is not a sufficient amount of moisture to dissolve all the drug and alkaline agent and provide for any significant neutralization reaction. However, excipients do contain a certain level of adsorbed free moisture capable of creating a microenvironment where small quantities of the drug and alkaline agent can be dissolved and become available for the neutralization reaction. These phenomena could be responsible of the appearance of physical as well as chemical stability problems and where taken into account by evaluating three (3) formulation combinations. The nine (9) formulation combinations where prepared and the tablets were stored in opened containers at 25° C./60% RH and 40° C./75% RH for a 2-week period. These open container studies are typically conducted during the early formulation development phases of a product to purposely accelerate physical and chemical changes in formulations in order to select the lead candidate, i.e., the formulation with the best stability profile. After the 2-week time period, the tablets were removed from the environmental chambers and sent to the analytical department for their performance evaluation. The performance of the formulations was determined by measuring ten (k=10) properties as a function of time and temperature, which properties were selected as follows, according to a hierarchical tree comprising properties and sub-properties:
Applying the AHP process with the standard scale for these properties, the decision matrixes given in Table 2 for the properties and in Tables 3, 4 and 5 for the sub-properties were built.
From the decision matrixes, the following weight values for the k=10 properties/sub-properties are given in Table 6, the sum of the weights being equal to unity.
Experimental property data that were obtained from nine (9) runs of the process using the selected nine (9) combinations of parameter values of Table 1, are given in Table 7.
Since n=3<8, the parameter reduction step is not required for the purpose of the instant case. As to the statistical analysis of parameters interaction, since a correlation factor ρ13=0.7013 for the X1X3 interaction was calculated, that interaction can be considered as significant since the condition 0.5<ρ13<0.95 is satisfied. The following property behavior relations were established:
The specified goal values for the properties as given in Table 8 were used to establish the goal function that was minimized to generate the following set of optimal parameters:
The associated experimental property values are given in Table 9.
Applying the method for the particular case where only the minimum four (n+1=3+1=4) experimental runs required were used, runs 1, 3, 6 and 9 were selected to provide the parameter and property data as given in Table 7. As to the statistical analysis of parameters interaction, since a correlation factor ρ13=0.332 for the X1X3 interaction was calculated, that interaction cannot be considered as significant since 0.5<ρ13<0.95 is not satisfied. The following property behavior relations were established:
The same specified goal values for the properties as given in Table 8 were used to establish the goal function that was minimized to generate the following set of optimal parameters:
The associated experimental property values are given in Table 10.
Comparing the set of parameter values given at (20) with the former set obtained from all nine (9) experimental runs given at (19), it can be noted that both sets are very similar. Actually, from a pharmaceutical standpoint, they could almost be considered as identical.