Professional Documents
Culture Documents
31:373-381 (2001)
1. Introduction After any fermentation process, two sub processes can be considered in the downstream processing, those are recovery and purification. The recovery stage includes the preliminary operations such as broth separation, cell rupture, debris removal and refolding. After the recovery subprocess the total protein concentration is 60-70 g/L (Asenjo, 1992). In the purification stage, all proteins are subject to a sequence of high-resolution operations in order to attain a specified level of purity in terms of the desired product (ideally the yield is 100%). Asenjo (1990) points to the need for the optimization of separation and downstream processes, since these represent the major manufacturing and investment costs in biochemical processes. The final quality of biotechnological products is determined at the purification level, which may be regarded as the most important stage in the whole fermentation process. This requirement is even more critical for therapeutic products such as vaccines and antibiotics that require an extremely high purity level. Among the diverse forms of separation and purification of a protein mixture, the most important group comprises chromatographic operations, also known as high-resolution steps. The specified purity is attained after several steps, and in each of them the mixture is basically split into two streams, one that contains the target protein and the other that is discarded. It is then fundamental to synthesize an
optimal sequence of purification steps such that their total number is as small as possible. Mao and Hearn (1996) studied several alternative approaches for the optimization of the operation in ion-exchange and affinity chromatographic methods for protein purification. The chromatographic operation has been considered as a four-stage process. Computer-aided process design combined with experimental planning was used for determining the optimal conditions in ion-exchange chromatography (Jungbauer and Kaltenbrunner, 1996). A review in computer techniques based on Artificial Intelligence that are applied to the analysis of chromatographic methods for protein separation has been recently reported in the literature by Bryant and Rowe (1998) with emphasis on commercially available tools. Eriksson et al. (1991) and Forslund (1995) describe the implementation of a deductive synthesizer for protein purification, which encompasses stages from extraction to high-resolution chromatography techniques. A more detailed planner that is restricted to liquid chromatography of high-resolution was proposed by Eriksson et al. (1991) and relies on a knowledge-based system. Leser and Asenjo (1994) developed a knowledgebased expert system for selection of protein recovery, separation and purification processes for multicomponent mixtures. Fundamental databases of characteristics of protein molecules, equations that describe heuristics and the behavior of proteins are elements of the expert system. The same system was further
373
Latin American Applied Research extended by Lienqueo et al. (1996) in order to include an algorithm to model the amount of each protein contaminant eliminated after each step. Lienqueo et al. (1998) developed two criteria to select the sequence of purification. One uses the selection separation coefficient and the other uses the final level of purity. Results for the synthesis of a model protein mixture as well as for the purification of -1,3-glucanase were validated experimentally by Lienqueo et al. (1999). The main objective of this paper is to determine the optimal selection and sequencing of purification steps that are required to separate contaminants from a protein mixture. For that purpose, mathematical models based on mixed integer linear optimization are developed. The main target is to obtain the minimum number of stages of purification, by applying highresolution chromatography in order to attain a purity level required to the final product and then maximizing the purity of the product. 2. Problem description Given a mixture of proteins in different concentration levels and a desired product specification in terms of a minimum purity, the problem is to synthesize a high-resolution purification process, which is usually carried out by chromatography. Selection of these purification operations is based on the efficiency of different chromatographic techniques that may be employed to separate the target protein from the contaminant ones. Therefore the objective is to select the chromatographic steps and their sequence in order to minimize the total number of operations and maximize the purity. Chromatographic techniques perform the separation of a protein mixture by exploiting a specific physicochemical property. For instance, ion exchange chromatography will separate proteins based on their difference in charge. The charge of a protein changes with the pH following the titration curve. Ion exchange can use small differences in charge that give a very high resolution and hence is an extremely efficient operation to separate proteins. Hydrophobic interaction chromatography has been proposed only as a pretreatment step or as a first high-resolution purification step. Gel filtration for protein fractionation is normally not used as a large-scale highresolution operation owing to the low efficiency in exploiting differences in the molecular weight. Physicochemical property information must be available for the target protein and also for the major contaminants. Each chromatographic step is able to perform the separation of the protein mixture based on their physicochemical properties, which define the dimensionless retention times (Kdi,p). By comparing both the retention times of the desired protein and those of the contaminants, the deviation factors (DFi,p) are calculated. Finally, according to the deviation factor of each contaminant and the peak widths ( i ) of each step we are able to determine the concentration factor CFi,p for each protein, that is, the degree of
31:373-381 (2001) separation after making use of the chromatographic step. This information is then used to generate the best separation steps and their sequence within a set of candidate chromatographic units and operational conditions. For that purpose, a two-step procedure is proposed in which first an optimization model minimizes the total number of chromatographic steps for a given purity level. The result on the minimum number of total steps is then used to generate an MILP that maximizes the total purity. A particular case occurs when the selection of chromatographic techniques in the purification process is independent of their sequence. Alternative models can be derived from the previous models both for minimization of the total number of steps and for maximization of final purity. 3. Models for optimal selection and sequencing In this section, we describe optimization models for synthesizing a sequence of purification steps in order to attain specified protein purity. In the optimization models, the following are the indices: a dp i k k* p protein property desired protein chromatographic step (i = 1 , I) order in the sequence (k = 1, K) final order in the sequence (k* K) protein (product + contaminants) (p = 1, P)
Variables: mdp,k desired protein mass after chromatographic step in order k mp,k mass of contaminant p after chromatographic step in order k (k = 2, K) S function to be minimized T function to be minimized yi,k binary variable that denotes whether chromatographic step i is used in order k in the sequence. Zk binary variable that denotes whether order k is the last in the sequence and Parameters: Ai set of properties used in technique i p C dp purity specification of the desired protein dp mp ,1 initial mass of protein p CFi,p concentration factor of contaminant p after chromatographic step i (mass reduction of contaminant p after and before chromatographic step i) DFi,p deviation factor for protein p in chromatographic step i Kdi,p dimensionless retention time for protein p in chromatographic step i Pa,p value of property a for protein p U upper bound on mass i peak width of chromatographic step i
374
E. V. ALVAREZ, M. E. LIENQUEO and J. M. PINTO The first parameter to be calculated is the dimensionless retention time (Kdi,p), which is a function of the physicochemical properties and represents the behavior of the protein p in each chromatographic technique i. Mathematical relationships that predict values for this parameter for a number of chromatographic techniques are given in table 1. For instance, in table 2 it is shown that each protein p has a value for property a defined by Pa,p. Note that both anion and cation exchange steps are function of charge density (Qp/MWp). This charge was determined by eletrophoretic titration curves as indicated by Pharmacia (1991) and Watanabe et al. (1994) and presented a very good fit in order to predict the orders of protein elution. Table 1.Expressions and parameters for chromatographic steps
Chromatographic Technique Dimensionless retention times Kdi,p Peak Width
Then as a function of DFi,p calculated in (2), as well as i , (this parameter corresponds to the peak width value and is independent of protein mass) it is possible to calculate CFi,p from (3). Note that CFi,p represents the ratio between the mass of contaminant p after and before chromatographic step i. The mathematical relationships expressed in (3) represent graphical approximations of the chromatograms for two different proteins. As seen in Fig. 1. In the cases shown, one of the proteins is always the desired product. Therefore the deviation factor the difference between the two peak values.
i
Kd AE , p = KdAE,p = 0 7383 | ( Q p / MWp ) | if Q p < 0 25 + 15844. | Q p / MWp ) |
Anion Exchange
1. 10
0.15
if QP > 0 5972. | ( Q p / MW p ) | 1. 10
25
Cation Exchange
Kd CE , p = KdCE,p = 0
+ 17065 . | ( Q p / MWp ) |
if Q p > 0
0.15
Kd HI , p = 1
[(NH 4 )2 SO 4 ]p
1. 5
if QP < 0
0.22 0.46
Figure 1. Peak representation in a chromatogram 3.1 Model for minimizing total number of steps Model (M1a) for protein purification using chromatographic steps of high resolution that must be selected to attain a final purity specification of the desired product is composed of the following constraints: a) assignment constraints k yi ,k 1
i
Then in general every chromatographic technique can be expressed as: Kdi,p = fi (Pa,p) i, p,a Ai (1)
For every protein and chromatographic step i the deviation factor can be calculated with respect to the desired protein dp (see Fig. 1), according to eqn. (2): DFi,p = |Kdi,dp - Kdi,p| i, p (2)
(4a) (4b)
yi ,k
k
For any protein p and chromatographic step i: 0 DFi,p < i 10 i 2 i 10 i 2 then then CFi,p = 1 CFi,p= 1.02.(
2 i
b) ordering constraints yi ,k +1 yi ,k
i i 2 i, p
(3a) 2 DF
2 i
Zk yi ,k yi ,k +1 i i ) yi ,k Z k
i i
DFi,p <
yi ,k + Z k 1 Zk = 1
k
DFi,p i
(6)
375
Latin American Applied Research d) purity constraints p mdp ,k+1 Cdp . mp ',k+1 U *(1 Z k )
31:373-381 (2001)
defined to satisfy the required purity level. Nevertheless, several sequences may be selected (7) k<K p which attain this specified purity. Note that for any given selection, any sequence would satisfy final Equation (4a) states that at most one purity for the objective function defined in (10). chromatographic process i is chosen in order k, while (4b) enforces that each chromatographic step is 3.2 Model for maximizing purity selected at most once in the sequence. A second model (M1b) is proposed for maximizing In order to avoid that empty steps are chosen as the purity level for the number of steps obtained by well as to reduce the combinatorial search, Eqn. (5a) model (M1a). By denoting k* the optimal number of states that steps k are assigned in increasing order, that steps found in the first model, the objective function is, step k+1 can only be chosen if step k is also for maximizing the purity of the mixture is defined as selected. Constraints (5b)-(5d) are used to define the follows: last step of the sequence, denoted by Zk. According to (5b), if step k is chosen (yi,k =1 for some i) but not k+1 (yi,k+1 =0 for all i) then k is the last step in the max T = mdp ,k*+1 (11) m p , k* +1 sequence. On the other hand, if k is the last step in the p sequence (Zk =1) then (5c) enforces that all previous steps are selected (k' k) while (5d) states that all the where m dp,k*+1 represents the mass of the desired steps that succeed k are empty. Equation (5e) imposes protein after step k* purifies the mixture. that there is exactly one final step in the sequence. Note that (11) is in nonlinear form. However, Equation (6) relates the mass levels of each protein since the mass of the desired product remains the same in subsequent steps. Note, however that they are in along the steps, function (11) is equivalent to: nonlinear form, since the summation term contains the product of a continuous and a discrete variable. m p , k*+ 1 Equation (6) can be linearized as: min T = p min T = mp ,k*+1 (12) p dp mdp ,k* +1 p (8a) mp ,2 = CFi ,p . yi ,1.mp ,1 i In (12), the second objective (T) is equivalent to T (8b) mp ,k CFi, p .mp ,k 1 U .(1 yi,k 1 ) i,p, k 3 since the mass of the desired protein is constant and i,p, k 3 (8c) therefore can be removed. mp ,k CFi, p .mp ,k 1 U .(1 yi ,k 1 ) The constraints defined for model (M1b) are as It is important to note that Eqn. (6) may be used follows: directly for the first chromatographic step, since the initial mass of all proteins mp,1 is known a priori, as e) assignment constraints expressed in (8a). Constraints (8b) and (8c) on yi ,k = 1 (13) k k* i subsequent steps are enforced only if any of them is k* selected. U is a valid upper bound on protein mass. i (14) yi ,k 1 Constraint (7) enforces the purity specification for k =1 the target protein. Note from (7) that if Zk* is set to one, (k* is the last chromatographic step), then the c) Contaminant constraints (linearized) following condition is imposed: (8a) mp ,2 = CFi ,p . yi ,1.mp ,1 p
i
mdp ,k*+1
m p , k* +1
p
mp ,k
p Cdp
(9)
Equation (9) imposes that the concentration of the desired protein after the last step achieves its specified value. If Zk = 0, then constraint (7) is relaxed. Besides these restrictions, model (M1a) has as constraints yi,k { 0,1} and mp,k , Zk 0 . The objective function is the minimization of the total number of chromatographic steps, given as:
yi ,k = k * Z k Min S = i k k
CFi, p .mp ,k 1 U .(1 yi,k 1 ) , i,p,3 k (k*+1) mp ,k CFi, p .mp ,k 1 U .(1 yi ,k 1 ) i,p,3 k (k*+1)
(15a) (15b)
(10)
The solution of model (M1a) provides a sequence, which contains the minimum number of steps that are
Equation (13) indicates that only one chromatographic process i is chosen in order k, while (14) enforces that each chromatographic step is selected at most once in the sequence. Note that (13) is an equation as opposed to (4a) which is an inequality, due to the fact that the minimum number of steps is known for model (M1b). The linearized contaminant constraints (15a) to (15b) are also used in the range of k from 1 to k*. In summary, (M1b) is an MILP composed of objective function (12) and constraints (8a) and (13) to (15) with yi,k { 0,1}; mp,k , Zk 0 .
376
4. Mathematical models for optimal selection A particular case may occur, in which only the selection of steps is performed. This is especially important whenever the sequence does not affect the optimal solution, as is the case when the total number of steps must be minimized in order to attain a final purity level. A key issue in this formulation is that the decision variables must simply select the subset of chromatographic steps from the candidate steps. A binary variable wi that is activated when step i is selected is defined in place of yi,k . In this formulation, steps are selected independently of their sequence. Fig. 2 illustrates the difference between this representation and the previous one.
The first term in disjunction (16) models the case of selection of steps i and the second models no selection of i. Thus, in (16) the mass of protein p after going through step i is modified if this is selected, otherwise it remains the same. Disjunction (16) generates the following constraints in mixed-integer representation: U .(1 w1 ) mo1,p CF1,p .m p ,1 U .(1 w1 ) p U .w1 mo1, p C p ,1 U .w1 p U .(1 wi ) moi ,p CFi ,p .moi 1, p U .(1 wi ) p, i 2 U .wi moi1, p moi, p U .wi p, i 2
(a)
As defined in the previous section, mp,1 corresponds to the initial mass of protein p in the mixture. Constraints (17) enforce that if the first chromatographic step is chosen (w1 = 1) then the resulting mass is mo1,p = CF1,p.mp,1. Otherwise it remains the same (mo1,p = mp,1). The same approach is taken in (18) for the remaining steps i = 2, I. Finally, there is the purity constraint, which is written as: moI ,dp Cdp moI , p '
p p'
(19)
(b) Figure 2. Selection of techniques for models (a) M1a and (b) M2a Consider a simple example in which there are 6 candidate steps. Suppose that among these steps 2, 4 and 5 are chosen in the following sequence: 2-5-4. In model (M1a) assignment constraints would enforce that y2,1 = y5,2 = y4,3 =1. In this model we would simply have w2 = w4 = w5 =1. The optimization model, denoted by (M2a), contains the following new variables: moi,p mass of protein p after chromatographic step i R function to be minimized wi binary variable that denotes whether chromatographic step i is selected 4.1 Model for minimizing total number of steps In this case, contaminant constraints can be represented by the following disjunction:
Note that moi,p corresponds to the final purity of the mixture, regardless the selection of techniques, as shown in disjunction (16). The model is also restricted to wi { 0 ,1} and moi,p 0. For the minimization of the total number of selected steps the objective is written as follows: Min R = wi
i
(20)
Note that there is a significant reduction in the model size with respect to (M1a). The number of binary variables is I as opposed to I*K variables in (M1a), in which K is of the same order of I. 4.2 Maximizing the purity for optimal selection In the previous section, model (M1b) was proposed for the maximization of a given purity level for a fixed number of steps. In a similar way, an alternative model (M2b) may be formulated for the present case. For a given number of steps k* (which may be given or determined by the solution of M1a or M2a) we define: wi = k *
i
(21)
377
31:373-381 (2001)
implementation of OSL used several options in order with binary variables wi { 0 ,1}and moi,p 0. The objective function for the maximization of the to achieve better efficiency: dual simplex for LP purity of the desired protein may be defined as follows: solution, branch and bound preprocessing and model scaling. The models described in the previous sections are mo I ,dp max R = min R = moI , p (22) solved for 2 different examples of increasing size in p dp moI , p terms of protein types and candidate techniques. The p results are compared to solutions generated by expert where I corresponds to the last chromatographic step systems (Lienqueo et al., 1998; 1999). in the set of candidate techniques. Eqn. (22) follows the same procedure of generating (12). Models (M2a) and (M2b) may be either solved in a two-step procedure or individually. 5. Computational results The software General Algebraic Modeling System (GAMS), version 2.25 (Brooke et al., 1992) was used to implement the model and its solution method. This software is used for large-scale optimization and has a number of solvers for linear, non-linear and mixedinteger programming applications. In this paper the proposed models were solved with the Optimization Subroutine Library OSL (IBM, 1991) which is a high performance integrated solver for Linear Programming (LP) and Mixed Integer Linear Programming (MIP). The LP based branch and bound 5.1 Example 1 In this example, taken from Lienqueo et al. (1998), we consider the purification of a mixture containing four proteins: Serum from bovine albumin (p1), Ovalbumin (p2), Soybean trypsin inhibitor (p3) and Thaumatin (p4). Their physicochemical properties as well as the initial protein mass of the mixture are shown in table 2. The purity level required for protein p1 is 98% and may be attained through the appropriate selection from multimodal chromatographic techniques, which exploit differences in molecular charge, size and hydrophobicity. Overall, there are twelve chromatographic techniques that are ion exchange at 4.0, 5.0, 6.0, 7.0, and 8.0 pH levels, gel filtration and hydrophobic interaction (see table 2).
Cp,1
MWp (Da) a1
[mg/ml]
-25
p1 p2 p3 p4
The retention times (Kd) for each protein in every chromatographic step are calculated according to the
relations and data shown in tables 1 and 2. These values, the deviation factor (DF) and the concentration
378
E. V. ALVAREZ, M. E. LIENQUEO and J. M. PINTO factor (CF) are calculated with Eqns. (1) to (3). These parameters are used in the models for determining the optimal synthesis of chromatographic techniques. The purity obtained is 99.8%, which is attained in 3 steps and can be seen in table 3 for all models. When the optimization is performed with model (M1a) the final purity and selection of steps are the same as those from model (M1b) but their sequence is different. In the case of model (M1a) the result shown is one example of several possible sequences that this model generates, while model (M1b) not only calculates the maximum purity but also gives the optimal selection. Models (M2a,) and (M2b) generate the same solution, which selects the same operations as those chosen by the previous models. Note from table 2 that the hidrophobicity of p2 is much lower than of the remaining proteins. Therefore by choosing hidrophobic interaction as the first the chromatographic technique, p2 is almost completely eliminated, which is the step choice for (M1b). The same techniques are also selected for the other models, in a different sequence. Proteins p3 and p4 are separated by choosing techniques for which their charge is lowest that are anion exchange at pH 7.0 and 8.0 respectively. The optimal results can be compared to the solution obtained by Expert System (ES) (Lienqueo et al., 1998) for 94.5% purity. We obtain the same sequence and final purity both in model (M1a) and the ES, with a small discrepancy in purity for anion exchange at pH 7.0 and hydrophobic interaction. However, since the required purity for the MILP models is 98% an additional step must be added, which is anion exchange at pH 8.0 as seen in table 3. It is important to note that if model (M1b) is used for maximizing purity in two steps a value of 96.3% is obtained in exactly the same sequence as that from the ES. 5.2 Example 2 In this case we consider the purification of -1,3 glucanase produced from recombinant bacillus subtilis culture, in which the product must be separated from eight contaminants, denoted by cont_1 to cont_8. Their physicochemical properties are shown in table 4 and were provided by Lienqueo et al. (1999).
379
Latin American Applied Research The purification requires 94% -1,3 glucanase and the steps must be synthesized from 22 candidate chromatographic techniques. Note that besides the chromatographic techniques from example 1 there are 10 additional candidate techniques that are anion and cation exchange at 4.5, 5.5, 6.5, 7.5 and 8.5 pH levels. However the same mathematical relationships shown in table 1 are used, since these are valid for ion exchange despite the pH level. The resulting techniques and their sequence are shown in Fig. 3 for model (M1b). Note that this mixture is very difficult to purify. Overall, six steps are required and a final maximum purity of 94.8% is attained for models (M1b) and (M2b). In the particular case when maximizing purity with (M1b) the global minimum is not obtained as shown in table 5. Models (M1a) and (M2a) generate solutions that minimize the number of steps and still satisfy the required purity. The solution provided by the Expert System (Lienqueo et al., 1999) is limited to a final purity of 70%. The suggested sequence is hydrophobic interaction (reaching 32.7% purity) followed by anion exchange at 6.5 pH level (70.3%). Interestingly, we obtained the same steps with purity levels of 32.5% and 69.25% respectively by model (M1a). Moreover, we attained final higher purity levels of 94.4% for model (M1a) and 94.8% with (M1b) that are similar results with little discrepancy (see Fig. 3).
31:373-381 (2001) solved to global optimality, in a personal workstation Pentium 266 with GAMS /OSL. Note from table 5 that as K increases (in model M1a) there is a significant increase in model size as well as in computational effort. 6. Conclusions This paper presented the development of mixed integer linear models and optimization strategies for tackling the problem of synthesis of chromatographic steps for the purification of protein mixtures. Results indicate that appropriate chromatographic sequences have been obtained, which can be used as a starting point for purification processes. Acknowledgements The authors would like to acknowledge financial support received from PADCT under grant 62.0239/97-QEQ and from VITAE (Cooperation Program Argentina-Brasil-Chile) under grant B11487/GB004. References Asenjo J.A., Separation processes in biotechnology, Marcel Dekker, New York, NY (1990). Asenjo J.A., Profiles on Biotechnology, Santiago de Compostela, Spain, 505-522 (1992). Brooke A., D. Kendrick and A. Meeraus, GAMS A User's Guide (release 2.25), The Scientific Press, San Francisco, CA. (1992). Bryant C.H. and R.C. Rowe, "Knowledge discovery in databases: application to chromatography," Tractrends in analytical chemistry 17, 18-24 (1998). Eriksson H., K. Sandahl, G. Forslund and B. sterlund, "Knowledge-based planning for protein purification," Chemometrics and intelligent laboratory systems: Laboratory information management 13, 173-184 (1991). Eriksson H., K. Sandahl, G. Forslund and B. sterlund, "Reactive planning for chromatography," Chemometrics and intelligent laboratory systems: Laboratory information management 13, 185-194 (1991). Forslund G., "Designing for flexibility: a case study," Expert systems 12, (1) 27-37 (1995). IBM, OSL (Optimization Subroutine Library) Guide and reference (release 2), IBM, Kingston, NY (1991). Jungbauer A. and O. Kaltenbrunner, "Fundamental questions in optimizing ion-exchange chromatography of proteins using computer-aided process design," Biotechnology and Bioengineering 52, 223-236 (1996). Leser E.W. and J.A. Asenjo, "Building an expert system to assist the rational selection of large scale protein purification processes," In D. L. Pyle (ed.). Separations for Biotechnology III, Royal Society of Chemistry, London, 260-266 (1994). Lienqueo M.E.; E.W. Leser and J.A. Asenjo, "An expert system for the selection and synthesis of
Figure 3. Optimal results for (M1b) in example 2 5.3 Model statistics Tables 5 present statistical data for all models. These results correspond to cases for which the specified purity is 98% for example 1 and 94% for example 2. In these tables model sizes in terms of binary variables, continuous variables and constraints are given. In this table there is an additional entry concerning the size of the set K, which defines the total number of possible steps that can be assigned to the sequence. As expected, for models (M2a) and (M2b) there is a significant reduction in the number of binary variables and constraints. However, the number of continuous variables in fact increases. Also shown in table 5 are overall CPU time, relaxed objective value and nodes enumerated. Except for the cases that were explicitly mentioned, all models were
380
E. V. ALVAREZ, M. E. LIENQUEO and J. M. PINTO multistep protein separation processes," Computers chem. Engng. 20 (suppl) S189-S194 (1996). Lienqueo M.E., J.C. Salgado and J.A. Asenjo "Design of an expert system for selection of protein purification processes: comparison between different selection criteria," Computer applications in biotechnology CAB7, Japan, 321-324 (1998). Lienqueo M.E., J.C. Salgado and J.A. Asenjo, "An expert system for selection of protein purification processes: experimental validation," Chemical Technology and Biotechnology 74, 293-299 (1999). Mao Q.M. and M.T.W. Hearn, "Optimization of affinity and ion-exchange chromatographic processes for the purification of proteins," Biotechnology and Bioengineering 52, 204-222 (1996). Pharmacia, "IEF and electrophoretic titration curve analysis: Separation Technique File 100, PhastSystem, Uppsala (1991). Watanabe, E.; S. Tsoka. and J.A. Asenjo "Selection of chromatographic protein purification operations based on physicochemical properties," Annals of the New York Academy of Sciences, 721, 348-364 (1994).
Received August 30, 1999; accepted for publication May 15, 2001 Recomme nded by T. Pinto and Secchi
381