Principal Investigator/Program Director (Last, First, Middle):
Response to Critiques from the Summary Statement.
Include both response and overview of changes:
Resume and Summary of Discussion:
While the importance of such predictive tools in bioseparations was acknowledged, the review panel considered it unlikely that the proposed approach would yield much useful fundamental information: The proposed approach will yield powerful tools for predicting protein binding affinity and preparative chromatographic behavior in ion exchange systems. The SPR and chromatographic experiments will produce a large amount of important and useful data and will provide the foundation for the modeling work. The molecular descriptors that will be selected during generation of a wide variety of QSPR models (e.g. models of free energies and isotherm parameters for various pH, salts, and stationary phase conditions) will yield significant information about the relative importance of key physico-chemical phenomena involved in protein binding affinity under various conditions.Something about Shekhar’s contribution here.
Concerns mentioned include the lack of calibration and validations schemes for the computational modeling systems:
Ambitious nature of the project’s goals: While we agree that the project is ambitious, we have included recent results in the revised application that indicate that we can indeed carry out a-priori prediction of both isotherm parameters and preparative chromatographic behavior. Further, the MD simulations section in the proposal has been significantly modified to better the connect the goals of this section with the other parts of the proposed work.
Questionable utility of a biochemical tool that requires knowing the crystal structure of a protein:As stated in the revised application, a key component of this work will be to examine ways to predict protein binding affinity from sequence information using threading or homology modeling approaches to generate approximate structures. Since ion exchange is an “averaged phenomenon” it may indeed be possible to take this approach, particularly for antibodies where homology modeling may be useful. Alternative methods such as RECON 2D (Breneman 1997) that do not rely on structure will also be examined and modified for use in predicting protein binding affinity in ion exchange.
Uncertainty of relating the experimental and computational data: Shekhar and Ravi?
Critique 1.
The practical significance of the proposed work is in establishing conditions for routine purification of well-characterized proteins: The practical significance of the proposed work goes beyond routine purification of well-characterized proteins. The ability to predict isotherm parameters for a range of operating conditions enables these isotherms to then be employed for the simulation and optimization of large scale preparative protein purifications using various modes of operation (e.g. linear gradient, step gradient, displacement). Further, if we are successful in predicting protein binding affinity from sequence information used in threading or homology modeling, we will be able to predict the behavior of proteins that are not well-characterized. In addition to using a wide spectrum of model proteins, we will also examine the utility of this technique for predicting the affinity of proteins of industrial significance. As indicated in the attached supporting letter from Amgen, relevant industrial biological mixtures (e.g. fusion proteins, antibodies) will be included in the protein data set for proof of concept as we use the QSPR models to predict the chromatographic behavior. The QSPR models will be used for both the main bioproduct (e.g. fusion protein) as well as key impurities (e.g. clipped variants) to identify optimal conditions for their separation.
The PI proposes not only to be able to predict retention of proteins, a tall order, but to predict the nonlinear behavior as well:While we realize that this is a tall order, as shown in the revised application, we have recently results which indicate that the use of QSPR with appropriate molecular descriptors and machine learning techniques can indeed be used for predicting the isotherm parameters which can then be used to predict nonlinear chromatographic behavior.
The PI has not proposed any methods to determine the nature of the modified surface. It is particularly important to know whether or not there is 2-D phase separation: As pointed out in the proposal, the surfaces will be characterized by a variety of techniques including infra-red spectroscopy, ellipsometry, and contact angle goniometry. The reviewer has brought up the important issue of 2-D phase separation. 2-D phase separation is often observed in mixed SAMs formed by immersing a gold substrate in a solution containing a mixture of two or more different alkanethiols. We have therefore proposed a different method for generating mixed surfaces (i.e. surfaces presenting different densities of a desired ligand). We will first form a homogeneous self-assembled monolayer using a carboxylic acid-terminated alkanethiol. Reaction of the free carboxylic acid groups with a mixture of suitable amines will allow us to control the density of the desired ligand. As pointed out in papers by the Whitesides group (e.g. Yan et al., Langmuir, 1997, 13, 6704, and Lahiri et al. Anal. Chem. 1999, 71, 777), mixed surfaces generated in this manner are unlikely to contain phase-separated domains. We will confirm the absence of 2-D phase separation by using atomic force microscopy.
Protein adsorption will be determined by SPR. This has the virtue of simplicity and reproducibility. It is not clear, however, how the adsorption will be calibrated:We agree with the reviewer that SPR provides a simple and reproducible technique for determining the amount of adsorbed protein. We will use SPR to quantitate the amount of adsorbed protein by utilizing the relationship between SPR response units and the mass of adsorbed protein per unit area as described by Sigal, Mrksich, and Whitesides, J. Am. Chem. Soc. 1998, 120, 3469). The resulting large set of data for protein adsorption on well-defined surfaces under a variety of experimental conditions will form the basis for the chromatographic modeling work described in specific aims three and four.
There is no description of which proteins will be studied theoretically. Nor is there a description of how the theory will be related to experiment… What is the method going to be used for after confirmation?Shekhar..
A serious missing element acknowledged by the coinvestigator is that the TAE’s have never been calculated in water. Curt: The electron density-derived descriptors associated with TAE reconstructions represent molecular properties prior to their interaction with any other molecular species - including solvent molecules or resin surfaces. During the process of model construction, the weights of various descriptors are altered in response to actual training case responses, adding empirical knowledge to the model. The resulting descriptor weight patterns implicitly incorporate the effects of the electronic environment perturbations associated with solvation and intermolecular interactions because they stem from actual experimental data. Solvent effects will be investigatedin the revised applicationto examine changes in electronic environments during solvation, and will also include new descriptors that explicitly include water distribution information.
The theoretical approaches will result in scalar descriptors, some of which are moments. These are inherently averaged properties. There is nothing proposed to come from theory that could possible be predictive of the sort of adsorption shown on p. 35: The SMA isotherm is also an “averaged” phenomenon since the adsorption data represents an ensemble average of all of the interactions of a given protein with the resin during the experiment. There is no attempt to predict the binding orientation of proteins using the QSPR approach, thus there is no disconnect here between the QSPR approach which employs scalar descriptors and the SMA isotherm predictions. Inherent in many QSPR and QSAR approaches is a presumption that the energetics of a distribution of interaction modes between interacting molecules is represented by the resulting averaged models – in this case, the interactions between proteins and resins.
There is no doubt that having an approximate method to estimate the conditions required to elute a particular protein (…) will same some laboratory time. The real problems, however, are in dealing with minor and major impurities about which little may be known. The main focus of this project will be on the development of novel methods for the a-priori prediction of protein chromatographic behavior. However, as described above, we will also examine the utility of the QSPR technique for predicting the affinity of proteins of industrial significance. The QSPR models will be used for both the main bioproduct (e.g. fusion protein) as well as key impurities (e.g. clipped variants) to identify optimal conditions for their separation. We will work with Amgen to identify some of the key minor and major impurities and will examine ways to predict the binding affinity of these impurities as well as the main bioproduct from sequence information using threading or homology modeling approaches to generate approximate structures for the QSPR analysis. Obviously, this will require that sequence information is available, however, this is often the case for clipped and other variants. In addition, we will use the models to predict appropriate conditions for high dynamic capacities and conditions for efficient capture and elution of the bioproduct.
Critique 2
Some concerns arise regarding the proposed research with respect to both the significance of, and the ability to meaningfully perform some of the computational studies. In particular, a) there does not appear to be a way to determine if the model polyelectrolyte surfaces are indeed representative of those of IEC columns. Shekhar and Curt need to answer. From Ravi, We will be studying protein adsorption on surfaces presenting different polyelectrolytes such as polystyrene sulfonate, polymethacrylic acid, poly(dimethylaminoethyl methacrylate), poly(methacryloyloxyethyltrimethyl ammonium bromide), and dextran, diethylaminoether. These surfaces present functional groups commonly used in ion exchange chromatography, such as sulfonate, carboxylate, and quaternary ammonium groups, as well as commonly used polymer backbones such as polystyrene, methacrylate polymers, and dextran. We believe that these will be excellent model surfaces, and will provide an intermediate degree of control between SAMs and commercial chromatographic stationary phases. The use of model surfaces in the development of descriptor-based correlative machine learning approaches is desirable since they allow computational descriptor patterns to be associated quantitatively with well known protein-surface interaction modes. Later, this information can be used to interpret the descriptor patterns found in models developed for real chromatographic systems, and allow insights about important interactions to be realized.
MDS force field representation… Shekhar
Selection of descriptors/properties: Curt A wide variety of molecular property descriptors is required in order to capture enough chemical information to build generalizable models of intermolecular interactions. Important descriptor types include those related to protein shape, size, charge distribution and local solvent environments. In most cases, making an a priori choice of which descriptors to utilize in a particular modeling situation would be difficult, due to the non-linear (and non-intuitive) relationship between many descriptor types and the magnitude of an observed binding interaction. Consequently, a heuristic “feature selection” process is employed to remove descriptors from consideration if they would not contribute significant chemical information to the resulting model. Care must be taken to reduce the likelihood of fortuitous correlations and overdetermined models, and to address this, our group has developed techniques to minimize this risk. As part of the revised application, we have added a description of the modern feature selection and model-building processes that will be employed in proposed work.
Listing of protein models to be used in the research nor reasons as to how models will be selected and why choices are meaningful.. Shekhar and rest
Free energy often used, what about entropy component of free energy? Shekhar
Little test set experimental data, to both calibrate and validate many of the simulation results and QSPR models is provided. Steve and Ravi
Manner and strategy of employing chemometric tools is not always clear: Curt Thegoal of developing and utilizing chemometric tools for protein ion-exchange chromatographic systems is two-fold: First, the development of cross-validated machine-learning models of protein-resin interactions in ion-exchange scenarios will permit predictions of the behavior of untested systems to be made prior to the availability of experimental results. The models developed as part of this study would become part of computational tools that would be made available to the scientific community. Secondly, during the development of such models, descriptor patterns are developed and can be analyzedto interpret the relative magnitude of specific chemical effects that are controlling the observed protein-resin interactions. Throughout the proposed work, correlative models of this type will be constructed and evaluated as a means of developing new chemical insights related to selectivity, specificity and overall binding affinity. New protein property descriptor types will be developed, and fruitful descriptor families will be identified.
Software does not report a major MDS package. Shekhar
Validation Schemes for proposed computational model: Shekhar
Self-consistent overall strategy to bring all the IEC pieces together within the framework of chemometric based QSPR models (Steve and Curt).
Add some thoughts about Useful information that wil result from this work
A powerful predictive tool, relative contributions of backbone chemistry . ligands, ligand density, ideas for how to create new phases with unique selecitivty,..
Map out connectivity,
Breneman Changes - notes
TAE/RECON section – cut down and revised.
Table 2 removed.
Figure 14 deleted.
Revisions made throughout the text in order to clarify and integrate the use of chemometric tools with the MD work.
PHS 298/2590 (Rev. 05/01)Page Continuation Page