protocol

Nonlinear least-squares data fitting in Excel spreadsheets

Gerdi Kemmer & Sandro Keller
Leibniz Institute of Molecular Pharmacology FMP, Berlin, Germany. Correspondence should be addressed to S.K. ([email protected]). Published online 28 January 2010; doi:10.1038/nprot.2009.182

2010 Nature Publishing Group http://www.nature.com/natureprotocols

We describe an intuitive and rapid procedure for analyzing experimental data by nonlinear least-squares fitting (NLSF) in the most widely used spreadsheet program. Experimental data in x/y form and data calculated from a regression equation are inputted and plotted in a Microsoft Excel worksheet, and the sum of squared residuals is computed and minimized using the Solver add-in to obtain the set of parameter values that best describes the experimental data. The confidence of best-fit values is then visualized and assessed in a generally applicable and easily comprehensible way. Every user familiar with the most basic functions of Excel will be able to implement this protocol, without previous experience in data fitting or programming and without additional costs for specialist software. The application of this tool is exemplified using the well-known MichaelisMenten equation characterizing simple enzyme kinetics. Only slight modifications are required to adapt the protocol to virtually any other kind of dataset or regression equation. The entire protocol takes ~1 h.

INTRODUCTION
Modern measurement techniques allow researchers to gather ever more data in less time. In many cases, however, the primary or raw data have to be further analyzed, be it for the verification of a quantitative model (theory or hypothesis) thought to describe experimental data, quantitative comparison with other data, better visualization or simply data reduction. To this end, a wealth of information collected during a measurement or a series of measurements has to be reduced to a few characteristic parameters. This can be done by regression analysis, a statistical tool to find the set of parameter values that best describes the experimental data by assuming a certain relationship between two or more variables. Although many powerful and dedicated software packages have been developed for regression analysis, the most widely distributed regression tool is the Solver add-in bundled with Microsoft Excel. We first describe the following important terms and concepts used in regression analysis: independent variable; dependent variable; dataset; regression equation; coefficient/parameter; best fit/solution; constraint; sum of squared residuals; simple, multiple, linear, and nonlinear regression; and starting/initial values. Then, we list some examples in which Solver was used to fit or simulate data. Specific advantages and disadvantages of Solver with respect to other data fitting programs as well as general limitations and pitfalls inherent in nonlinear regression analysis are also addressed. The core part of the protocol lays out a general fitting strategy and explains the application of Solver to a typical nonlinear regression problem encountered in biochemical research. Finally, we discuss a more difficult example to highlight some of the most common problems and pitfalls encountered in fitting experimental data. Fundamentals of regression analysis Quantitative experiments aim at characterizing a relationship between an independent variable (x), which is varied throughout a measurement, and a dependent variable (yobs), which is observed/measured as a function of the former. The fitting method presented in this protocol requires that the independent variable can be measured with much greater precision than the dependent variable1. In other words, experimental errors (uncertainties) in the independent variable are small compared with errors in the dependent variable (see below). This is usually the case with experiments in which the value of the independent variable follows a predetermined trajectory and the experimental readout reports on the value of the dependent variable. The primary output of a measurement is a set of conjugated independent and dependent variables, which is called data or dataset. In addition to an experimental dataset, regression analysis requires a regression equation (also termed fitting function). This is a mathematical relationship describing the dependence of the dependent variable on the independent variable using one or more parameters. These parameters (also called adjustable parameters, fitting parameters or coefficients) are the same for every data point, i (i.e., every combination of xi and yi). In the simplest example of a proportionality (y = a x), the only parameter, a, is the slope of a straight line through zero. With these tools at hand and using a suitable computer program, one can employ regression analysis to find the combination of parameter values that best describes the experimental dataset. This combination of parameters is called the best fit or the solution. However, not every mathematically correct solution is physically possible or meaningful. For instance, temperature cannot fall below absolute zero. Mathematical relationships imposing such limits (e.g., T(K) > 0, where T is the absolute temperature) are called constraints and can, or in some cases must, be provided in order for regression analysis to find a reasonable solution. Furthermore, constraints may help reduce the parameter value space that the regression routine has to sample, which is particularly useful for regression equations containing many parameters. The decision as to which combination of parameters describes an experimental dataset best is most commonly done by leastsquares fitting (LSF), i.e., by minimizing the sum of squared residuals (SSR). In this context, a residual is defined as the difference () between an observed/measured data point (yobs) and its calculated counterpart (ycalc). The sum runs over all data points to be considered for regression analysis. Provided that the experimental errors in the dependent variable follow a Gaussian
nature protocols | VOL.5 NO.2 | 2010 | 267

protocol
(also known as normal) distribution with a mean of zero, the combination of parameter values having the lowest SSR has the highest probability of being correct. In a Gaussian distribution, small errors are more probable to occur than large ones. More precisely, for a very large number of independent replicate measurements, the frequency with which a certain error occurs follows a symmetrical bell-shaped function. A mean of zero implies that this distribution is centered at zero, such that errors are equally probable to be positive or negative. For many, but not all, experimental setups typically encountered in a biochemical or biophysical laboratory, errors (from pipetting, weighing, diluting, instrument noise, and so on) are or can be approximated to be distributed in a Gaussian way. For further discussion whether a particular dataset can be fitted reasonably using LSF, see reference 2. Regression analysis comes in different forms: simple, multiple, linear and nonlinear. In simple regression analysis, there is only one independent variable, whereas multiple regression analysis handles several independent variables. In general, there can be as many dependent and independent variables as experimentally possible and reasonable (e.g., one can simultaneously record absorbance, conductivity and light scattering data when performing size exclusion chromatography). Linear regression analysis or linear least-squares fitting (LLSF) refers to regression equations that are linear in their parameters (this, of course, includes but is not limi ted to equations that are linear themselves). By contrast, nonlinear regression analysis or nonlinear least-squares fitting (NLSF) refers to equations that are nonlinear in their parameters. For example, y = a exp(x) is said to be linear in a because y is linearly proportional to a if the values of all other fitting parameters and of the independent variables (here, x) are held constant. Mathematically, this is manifested in the first derivative of y with respect to a being independent of a: dy/da = exp(x). By contrast, y = exp(a x) is nonlinear in a because y is not linearly proportional to a if all other values are held constant. Here, differentiation of y with respect to a yields an expression dependent on a: dy/da = x exp(a x). The solution to a linear regression problem is exact because it can be calculated by analytical means. By contrast, nonlinear regression cannot offer an analytical solution, but has to rely on iterative procedures to find the best fit. For further reading on regression analysis, see one of the pertinent books36. This protocol focuses on the more general case of nonlinear regression analysis. The underlying fitting algorithms are based on a trial-and-error approach: beginning with starting values (also called initial values) for the parameters, these are changed during every regression step, and the similarity between the mea sured and calculated values is analyzed and compared with the preceding iteration. The algorithm according to which the para meter values are altered and analyzed during each step depends on the fitting program used and some advanced settings. Excel Solver employs the generalized reduced gradient algorithm7; a discussion of the mathematical formalism is beyond the scope of this protocol but can be found elsewhere8. Published applications of Solver Solver has been used in numerous and diverse applications, as illustrated by the following examples: modeling of detergent mixtures9,10, the hypothalamicpituitaryovarian axis11 or physio logical functions12; estimation of protein-binding capacity13 or magnetic relaxation times14; generation of complex pharmaco
268 | VOL.5 NO.2 | 2010 | nature protocols

2010 Nature Publishing Group http://www.nature.com/natureprotocols

kinetic models1517; microbial population counting18; application of game theory to medical management19; analysis of percentile growth curves20; optimization of feed formulation in poultry science21,22, operating-room allocation23 or production of proteins24, small molecules25 or cheese26; calibration of ionization chambers27; analysis of electrophysiological measurements28,29 or luminescence lifetime distributions30; deconvolution of chromatographic peaks3135 or thermoluminescence glow curves36; a number of applications in analytical chemistry3741; and high-throughput fitting of doseresponse curves42. We have employed Solver for deconvoluting absorbance and fluorescence spectra43,44, fitting chemical shift perturbations in NMR titration experiments (Vargas et al., unpublished data), simulating complex multicomponent/multiphase equilibria45 or cooperative ligand binding to proteins46, analyzing solubilization and reconstitution of lipid vesicles47,48 (for an experimental protocol, see ref. 49), and monitoring lipid membrane partitioning and translocation with the aid of uptake and release experiments44,5055 (see ref. 56 for a protocol and an Excel spreadsheet implementing Solver). The latter application is a particularly impressive demonstration of the power of Solver: the interactions of many charged proteins, peptides and small molecules with lipid membranes can be described by a simple surface partition equilibrium modulated by electrostatic effects 57,58. While the partition equilibrium yields the regression equation proper, membrane electrostatics have to be accounted for by Gouy Chapman theory (or another suitable model). In brief, this means that during fitting of the actual regression equation to experimental data, a nonlinear constraint has to be fulfilled for each data point. This is a challenging task that is beyond the capabilities of many dedicated fitting programs but is easily handled by Solver (see ref. 50 for a detailed description of regression and constraint equations). Solver can also easily be called up from a macro executable in Excel (see Box 1). Comparison with other fitting programs Data analysis using NLSF can be accomplished with the aid of many different programs. Examples include Origin (OriginLab, Northampton, MA, USA), Prism (GraphPad Software, La Jolla, CA, USA) and PSI-Plot (Poly Software International, Pearl River, NY, USA), but there is a plethora of other commercial or free programs developed for this purpose. Of course, sophisticated mathematical software packages like Mathematica (Wolfram Research, Champaign, IL, USA) and MathLab (MathWorks, Natick, MA, USA) can carry out nonlinear regression, as well. Moreover, alternative approaches like genetic algorithms 59 may have to be implemented for more demanding problems, especially those containing numerous adjustable parameters. Most of these programs have a wide array of additional features like built-in statistical tests to assess the confidence of the best-fit parameter values or advanced fitting procedures like automated global fits. However, a major problem with specialized fitting programs is that they tend to entice non-expert users to adopt a black-box approach, so that they can hardly judge on the meaningfulness of the results returned by the program. In light of this, the more basic Solver may be a better option for most non-expert users wishing to fit experimental data by NLSF. The procedure we describe here for Solver is simple, intuitive and fast, and can be implemented without previous knowledge of

protocol Box 1 | Macros

 Writing macros is beyond the scope of this protocol, but the following hints may aid users already familiar with this topic to incorporate Solver in Visual Basic for Applications (VBA) macros executable in Excel. Macros are particularly valuable for performing time-consuming and iterative tasks like confidence assessment (see Steps 1118). In order to start the Solver add-in within a procedure, a reference must be added. To do this, click: TOOLSREFERENCESSOLVER in the Visual Basic Editor.

? TROUBLESHOOTING

2010 Nature Publishing Group http://www.nature.com/natureprotocols

Solver can then be called using the syntax SolverOk SetCell:=B3, MaxMinVal:=3, ValueOf:=0, ByChange:=B1:B2 SolverSolve True  Here, SetCell:= is equal to the Set Target Cell input in Figure 4, MaxMinVal:=3 is identified with checking Value of: (the third option in the Equal To selection; analogously, 1 stands for Max and 2 for Min), ValueOf:= specifies the target value, and ByChange:= tells the program which cells to vary during optimization. The parameters given above correspond to those displayed in Figure 4 for our example. Defined names (see Box 3 for details on naming cells) can be used instead of cell addresses (for instance, ByChange:=v_max, K_m instead of the above). The term SolverSolve True is not necessary to run Solver but serves to replace manual confirmation (by clicking OK) upon completion of each fitting session. For further reading on advanced Excel solutions for scientific purposes, consult the excellent textbook of De Levie1.

or experience in data fitting or programming. The vast majority of scientists are familiar with the basic tools offered by Excel, such as data manipulation or graphing facilities; in fact, a great deal of experimental data will eventually be transferred into this or a similar spreadsheet program for analysis or visualization. Excel is available in most laboratories and can be run on most personal computers or laptops, so no additional costs are incurred for dedicated software packages or workstations. Limitations of regression analysis General pitfalls of the fitting process itself have to be accounted for regardless of the fitting equation used. NLSF requires that all of the experimental errors can be attributed to the dependent variable and that the values of the independent variable be known precisely, that the data points be independent of one another, and that a sufficient number of data points be measured (see ref. 2 for a discussion of these and other important prerequisites for NLSF). Furthermore, a minimum in the SSR corresponds to a maximum in the likelihood of having found the correct para meter values only if the experimental errors in the dependent

variable follow a Gaussian distribution with a mean of zero2. If this condition is not fulfilled, the minimal SSR does not represent the highest probability, and opinions diverge over the question if NLSF should still be used in such cases2,3. What is clear, however, is that systematic errors and improper data processing have to be avoided altogether. The latter includes nonlinear operations like taking a logarithm or smoothing, which diminishes the information content of the data. By contrast, linear operations like addition or subtraction (e.g., of blank values), as well as multiplication or division by a constant factor (e.g., normalization), are permitted. Moreover, one should keep in mind that NLSF is a trial-and-error approach and that there can never be an absolute certainty of having found the global SSR minimum unless a systematic sampling of all possible combinations of parameter values is carried out. Another matter of concern is weighting (see Box 2). These issues aside, the results of a fitting session have to be interpreted cautiously and in the light of the specific problem at hand. Not every parameter mathematically extractable through regression analysis makes sense physically (see Anticipated Results for an example), and not every regression equation allowing for a

Box 2 | Weighting
 The use of an unweighted SSR implies that the experimental errors of all data points included in the fit follow the same Gaussian distribution having the same mean ( = 0) and the same standard deviation (). There are, however, scenarios where varies considerably over the range of recorded data. This should be accounted for by giving different data points a different weight, i.e., a different impact on the SSR.  In its most general form, a weighted SSR is given by: SSR = wi ( yi , obs yi , calc )2 , where wi is the weighting factor for the ith data point. Thus, data points with higher weighting factors contribute more to the SSR than those given less weight. The special case of wi = 1 for all i corresponds to an unweighted SSR. The question then remains as to a meaningful and justified choice of weighting factors. Weighting by observed variability is a theoretically simple approach relying on the standard deviation of the ith data point, i: SSR = (( yi ,obs yi ,calc ) / s i )2 . This weighted SSR is also known as 2 value. From a practical viewpoint, however, this weighting method can be of limited utility because it might require that the standard deviation for each data point be determined from a fairly large number of replicate experiments (usually several dozen). Another important weighting scheme is relative weighting: SSR = (( yi ,obs yi ,calc ) / yi ,obs )2 . This approach is practically straightforward and is applicable whenever the error scales with the value of the measured variable. Discussions and comparisons of weighting methods can be found in the literature5,60. Once appropriate weighing factors have been determined, they can easily be included in Step 6 of the standard protocol by putting them into column H, changing the formula in column G2 to =H2*(E2F2)^2, and copying or dragging this formula down into the following cells in column G.
nature protocols | VOL.5 NO.2 | 2010 | 269
i i i

protocol
good fit is an appropriate description of the underlying relationship between independent and dependent variables. These precautions, however, pertain not to the fitting process as such but to the choice of regression equation. General fitting strategy Data fitting with the aid of Solver can be divided into the following stages: (i) an Excel worksheet is set up. Experimental data are pasted and plotted, and simulated data are calculated using a regression equation and plotted, too. If applicable and necessary, data processing like normalization and baseline subtraction should be completed first. The squared residuals between observed and calculated data are computed and summed up. Before moving on, the adjustable parameters should be varied to get a feel of the influence of each parameter on the calculated curve and to determine reasonable starting values. (ii) The Solver add-in is prepared. Cells containing the values to be changed (the fitting parameters) and the value to be minimized (the SSR) are specified. Solver options may be adapted and constraints can be added if necessary. (iii) The fitting procedure is carried out by running Solver. This is repeated several times using different starting values for the adjustable parameters. Although this approach cannot guarantee finding the global SSR minimum, it does decrease the probability of unwittingly getting stuck in a local minimum. (iv) The confidence of the best-fit parameter values is assessed by repeating the fitting procedure while fixing the parameter to be scrutinized at a value slightly different from the optimal one. After determining the best-fit parameter values, two important questions have to be answered: how good is the agreement between experimental data and fit, and how much confidence can be put in the best-fit parameter values returned by nonlinear regression? These two issues are sometimes confused, but it is crucial to appreciate the fundamental difference between them. The question of goodness of fit refers to the extent to which a dataset calculated assuming a certain fitting equation can approach the experimental dataset. The SSR and related (i.e., weighted or normalized) quantities, such as the 2 value (see Box 2) or the root-mean-square deviation (RMSD), are measures of the goodness of fit. For a given dataset and fitting equation, the goodness of fit generally increases with decreasing number of data points or increasing number of fitting parameters. Most importantly, the goodness of fit contains absolutely no information about the confidence of the fitted para meter values, which generally increases with increasing number of data points or decreasing number of fitting parameters. Thus, a different quantity is needed to describe confidence. To this end, virtually all commercial NLSF programs compute some kind of standard error or standard deviation which is displayed along with the corresponding best-fit value. Automatic generation of such confidence intervals might appear convenient but can be highly misleading. The underlying calculations are based on a series of linear approximations that are never fulfilled for NLSF. This invariably results in an underestimation of the real uncertainties60,61, which becomes particularly problematic for small datasets and strongly correlated parameters5. A robust solution to this problem that can easily be implemented in a spreadsheet program consists in perturbing one of the fitting parameters from its best-fit value and recording how the SSR is affected on fitting the remaining parameters2,5,62,63. Parameter values for which the SSR exceeds a certain threshold can then be used as measures of confidence for that parameter. Using this approach may be less convenient than relying on
270 | VOL.5 NO.2 | 2010 | nature protocols

statistical parameters reported by other fitting programs. However, it provides a more realistic picture and a much deeper understanding of the confidence with which the desired information can be extracted from the experimental data at hand (see Anticipated Results for an example in which commercial fitting programs suggest a misleadingly narrow confidence interval for an extremely poorly defined parameter). Example of nonlinear regression: enzyme kinetics A prominent example of a regression equation that is nonlinear in its parameters is the MichaelisMenten equation64 describing enzyme kinetics. The MichaelisMenten equation gives the initial velocity of an enzymatic reaction, v, in dependence on the concentration of substrate, [S]: v = v max [S]/(K m +[S]) . Here vmax is the maximal velocity (for [S] ), and Km is the MichaelisMenten constant (i.e., the substrate concentration at which the initial velocity is half the maxi mal velocity). In this example, [S] is the independent variable (x in generic terminology), v is the dependent variable (y), and vmax and Km are the fitting/adjustable parameters. Reasonable constraints for this case would be vmax 0 mM s1 and Km > 0 mM to avoid division by zero for the first data point (in the absence of substrate, [S] = 0 mM). In times when computers were not yet omnipresent, lineari zation approaches like the LineweaverBurk plot65 were frequently employed to circumvent nonlinear regression analysis. Although such plots may be of great didactic value, their principal drawback is that processing data in any nonlinear way (i.e., subjecting them to operations other than, e.g., subtraction of another dataset or multiplication by a constant factor) not only transforms the

2010 Nature Publishing Group http://www.nature.com/natureprotocols

Table 1 | Simulated data used for the example discussed in the protocol. [S] (mM) 0 2.5 5 10 15 20 30 40 50 60 70 80 90 100 v(mM s1) 2 153 231 342 396 438 467 505 523 523 539 548 555 554

Data points were simulated using the MichaelisMenten equation and vmax and Km values determined for the conversion of carbon dioxide to carbonic acid catalyzed by carbonic anhydrase at an enzyme concentration of 1 M (taken from ref. 66). A Gaussian random-error term was added to simulate experimental errors. [S], substrate concentration (independent variable, x); and v, initial reaction velocity (dependent variable, yobs).

protocol
measured values but also distorts the associated experimental errors (see ref. 62 for an example). In brief, linearization of data to render them amenable to linear regression can have adverse effects on the determination of the best-fit parameter values and, therefore, should be avoided. Nowadays, hard- and software necessary for nonlinear regression has become available to most scientists and has thus obviated the need for linearization. This protocol provides a step-by-step guide on how to extract vmax and Km from the dataset in Table 1 through nonlinear regression and how to assess the confidence of the best-fit values using Microsoft Excel Solver. The dataset was simulated on the basis of values determined for the conversion of carbon dioxide to carbonic acid catalyzed by carbonic anhydrase66; a Gaussian random-error term was added to simulate experimental noise.

MATERIALS
2010 Nature Publishing Group http://www.nature.com/natureprotocols
EQUIPMENT Computer equipped with Microsoft Excel 3.0 or newer. Excel Solver was developed by Frontline Systems and has been included in every distribution of Microsoft Excel since 1990. Solver is not included in the newest version of Excel for Mac (Excel 2008) but can be downloaded free of charge from http://www. solver.com/mac/dwnmacsolver.htm. This protocol was prepared using Office 2007; click paths may vary slightly depending on the Excel distribution used. Experimental data in x/y form. In the following, we exemplarily use the dataset in Table 1, which was simulated using the MichaelisMenten equation including a Gaussian random-error term. Regression equation describing the data; here, the MichaelisMenten equation given in the last section of the Introduction.

PROCEDURE Activating Solver TIMING 1 min 1| Open an Excel workbook. 2| Click OFFICE BUTTON EXCEL OPTIONS ADD-INS; in the MANAGE drop-down menu, choose EXCEL ADD-INS and hit the Go button (Excel 2003: TOOLS ADD-INS). In the ADD-INS window, check the Solver checkbox (see Fig. 1) and click OK. When asked to confirm, choose Yes. ? TROUBLESHOOTING Setting up worksheet and plotting data TIMING 10 min 3| Type vmax (the first adjustable parameter of the MichaelisMenten model) into cell A1 and define the name of cell B1 as v_max (see Box 3 for instructions on how to name cells). Type Km (the second adjustable parameter) into cell A2 and define the name of cell B2 as K_m. Set vmax to 400 mM s1 and Km to 2.0 mM by typing 400 and 2 into cells B1 and B2, respectively. See Figure 2 for clarification. 4| Use the first row of columns D, E, F and G to denote the corresponding columns as x (mM), yobs (mM s1), ycalc (mM s1) and 2 (mM2 s2), respectively. Plug the data listed in Table 1 into columns D and E (x in column D and yobs in column E). Graph the data as scatter plot and label the axes. In our example, x is the substrate concentration, [S], and yobs is the initial reaction velocity, v. 5| Type the regression equation, =v_max*D2/(K_m+D2) into cell F2. Copy or drag this formula down into the following rows of the column, which will be filled with the corresponding ycalc values. Add the calculated data to the graph as line plot using a different color for clearer discrimination between observed and calculated data. Figure 2 shows the resulting Excel worksheet. 6| Type =(E2F2)^2 into cell G2 and copy or drag this equation down into the rest of the column. This is the squared residual ( 2) between the observed data and data calculated using the regression equation. Type SSR into cell A3 and =SUM(G2:G15) (or the corresponding function name in your Excel language) into cell B3, which now displays the sum of the values listed in column G. The spreadsheet in FORMULA AUDITING MODE is shown in Figure 3 (FORMULA AUDITING MODE can be switched on/off in FORMULAS SHOW FORMULAS (Excel 2003: TOOLS FORMULA AUDITING FORMULA AUDITING MODE)). CRITICAL STEP As an alternative to Step 6, the SSR can be calculated directly by typing =SUMXMY2(E2:E15;F2: F15) into cell B3. It should, however, be realized that this function has rather different names in non-English versions of Excel.

Figure 1 | ADD-INS window. Make sure Solver is checked. To call this window: OFFICE BUTTON EXCEL OPTIONS ADD-INS; in the MANAGE menu, choose EXCEL ADD-INS and hit the Go button (Excel 2003: TOOLS ADD-INS). nature protocols | VOL.5 NO.2 | 2010 | 271

protocol Box 3 | Naming cells

 Naming cells is a particularly useful feature in carrying out spreadsheet calculations and fitting or simulating data. In contrast to referencing between cells, naming cells lacks the need for protection of absolute references (i.e., references that should not change on dragging or copying). Naming saves time and contributes to clarity and transparency, especially when typing lengthy formulas containing many variables and coefficients. Naming can be done either by using the menu command FORMULASDEFINE NAME (Excel 2003: INSERTNAMESCREATE) or by highlighting a cell or cell range and typing the desired name into the LABEL FIELD, located in the TOOLBAR directly above cell A1.  Names can be assigned to single cells or ranges of cells. The name of a single cell is displayed in the LABEL FIELD whenever the cell is activated (e.g., the LABEL FIELD in Fig. 2 displays v_max as cell B1 is activated). The name of a cell range appears in the LABEL FIELD only when the entire cell range is activated. For ranges spanning several cells in the same row (or column), the calculator will refer to the cell in the same column (or row) as the formula. Names are dragged and dropped with their respective cells. Names can be edited or deleted using FORMULASNAME MANAGER (Excel 2003: INSERTNAMES).  No two cells can have the same name, thus naming a cell D3 will not work because this name is already given to cell D3 in the third row of the fourth column. When trying to define a name that is already used within the same workbook, the cell to be named will remain unnamed, and the previously named cell will be highlighted instead. For further information, consult the Excel manual or one of the numerous pertinent books.

2010 Nature Publishing Group http://www.nature.com/natureprotocols

Setting up solver TIMING 10 min 7| Click DATA SOLVER (Excel 2003: TOOLS SOLVER). This opens the SOLVER PARAMETERS window, prompting for entries for Set Target Cell, Equal To, By Changing Cells and Subject to the Constraints. Figure 4 shows the entries appropriate for our example. Set Target Cell determines which value is to be optimized. When using the least-squares procedure, this is the SSR in cell B3, which has to be minimized. This is accomplished by opting for Value of: 0 in the Equal To selection (see below). By Changing Cells defines the adjustable parameters, which in our case are vmax and Km contained in cells B1 and B2, respectively. Subject to the Constraints offers the relational operators >=, <=, =, int and bin to impose constraints on the adjustable parameters or any other variables in the spreadsheet. int allows only integer values for the specified variable, and bin allows only binary values. Constraints can be added, changed or deleted by clicking the corresponding buttons. A useful constraint in the present example is Km >= 0.000001 to prevent division by zero for the first data point recorded in the absence of substrate ([S] = 0 mM). CRITICAL STEP Minimization of the target-cell value can, in principle, be carried out by checking the Min checkbox in the Equal To selection. However, we have noticed in numerous cases that Solver is more likely to get stuck in a local minimum when using this option. Upon checking Value of: 0, the algorithm tries harder to reduce the SSR to zero. As this is not perfectly feasible with experimental data, Solver will open a window stating that it was unable to find a solution. This message can be ignored. CRITICAL STEP More sophisticated Solver settings are available in the SOLVER OPTIONS window, which is accessible by clicking the OPTIONS button in the SOLVER PARAMETERS window. Figure 5 shows the default values, which work well in most cases. The meanings of these parameters are explained in Box 4. ? TROUBLESHOOTING Fitting TIMING 10 min 8| In the SOLVER PARAMETERS window (see Fig. 4), click the Solve button to initialize Solver. At the end of the fitting procedure, the SOLVER RESULTS window opens, reporting that no feasible solution could be found. As explained in Step 7, this is because of our selection in the SOLVER PARAMETERS window and can be ignored. One can also choose to view additional information on the fitting procedure, which in most cases is little enlightening, and therefore can be skipped. Hit the OK button in the SOLVER RESULTS window. ? TROUBLESHOOTING

Figure 2 | Worksheet after the completion of Step 5. Highlighted cell B2 is named v_max, as can be seen in the LABEL FIELD in the upper left. Columns D and E are filled with, respectively, x and yobs data from Table 1, and yobs is plotted against x (blue scatter plot). Column F contains the values calculated using the parameters in cells B1 and B2 and the MichaelisMenten equation, ycalc. Calculated values are plotted against x as solid red line. 272 | VOL.5 NO.2 | 2010 | nature protocols

protocol
Figure 3 | Worksheet in FORMULA AUDITING MODE after the completion of Step 6. Cell B3 contains the sum of squared residuals (SSR), column F the values calculated using the MichaelisMenten equation (ycalc) and column G the squared residuals (2). yobs and ycalc are plotted against x (blue scatter plot and solid red line, respectively). FORMULA AUDITING MODE can be switched on/off by clicking FORMULAS SHOW FORMULAS (Excel 2003: TOOLS FORMULA AUDITING FORMULA AUDITING MODE).

2010 Nature Publishing Group http://www.nature.com/natureprotocols

9| Repeat the fitting procedure by clicking DATA SOLVER (Excel 2003: TOOLS SOLVER) and hitting the Solve button until the SSR in cell B3 no longer decreases. ? TROUBLESHOOTING 10| Inspect the x/y plot containing the experimental data and the data calculated using the best-fit parameter values returned by Solver: vmax = 598 mM s1 and Km = 7.6 mM in cells B1 and B2, respectively. Experimental and fitted data should resemble one another reasonably well, as illustrated in Figure 6. This is reflected in a small SSR value of ~298 mM2 s2. CRITICAL STEP Once the best fit for a given set of initial parameter values has been found, it is highly advisable to repeat the fitting procedure using different starting values for the adjustable parameters to reduce the risk of getting stuck in a local rather than the global SSR minimum. CRITICAL STEP Visual inspection of the goodness of fit is often facilitated by plotting the residuals (not the squared residuals) versus the independent parameter. This helps in judging the quality of the fit and the applicability of the chosen regression equation (see ref. 67 for an example of analysis of residual plots). CRITICAL STEP The steps outlined thus far yield the vmax and Km values that best fit the experimental data in Table 1 in terms of the MichaelisMenten model. Yet, this procedure affords no information on the confidence to put in the obtained values. In contrast to linear regression, simple statistical measures like the standard error or the standard deviation are not readily applicable to nonlinear regression analysis. In the following, we describe a straightforward and general way of assessing confidence by variation of the SSR near a minimum2,5,62,63. ? TROUBLESHOOTING Confidence assessment TIMING 30 min (15 min per parameter) 11| The basic idea of variation of the SSR near a minimum is to fix one of the adjustable parameters at various values close to but different from the optimal solution and to monitor the impact on the SSR on optimizing the other parameters. To this end, fix the value of Km at 7.4 mM by typing 7.4 into cell B2 while leaving the value of vmax in cell B1 unchanged. Click DATA SOLVER (Excel 2003: TOOLS SOLVER) and select only cell B1 in the By Changing Cells option in the SOLVER PARAMETERS window. Then, hit the Solve button and repeat this one-parameter fit until the SSR in cell B3 no longer decreases. In this example, the SSR will amount to 330 mM2 s2. Copy this value as well as the fixed Km and fitted vmax values into another worksheet using PASTE SPECIAL: VALUES.

Figure 4 | SOLVER PARAMETERS window with the appropriate values for Set Target Cell, Equal To, By Changing Cells and Subject to the Constraints. An equally appropriate entry for By Changing Cells would be K_m, v_max. To call this window: DATA SOLVER (Excel 2003: TOOLS SOLVER).

Figure 5 | SOLVER OPTIONS window with default selection of values appropriate for most fitting problems. See Box 4 for further information on the settings. nature protocols | VOL.5 NO.2 | 2010 | 273

protocol Box 4 | Solver options

 The default values shown in the SOLVER OPTIONS window in Figure 5 are a good starting point for most fitting tasks. Max Time and Iterations define the time and number of iterations, respectively, after which Solver interrupts the fitting procedure and asks whether it should continue. Irrespective of these two parameters, Solver can be halted any time during the fitting procedure by pressing the ESC key. The value given in the Precision box determines how strictly constraints have to be fulfilled, whereas Tolerance specifies how strictly the integer criterion will be applied (if applicable at all). Convergence determines when optimization is considered to be achieved: this is the case as soon as five consecutive iterations do not yield a relative change in the target-cell value greater than the value specified in this box. If Assume Linear Model is checked, Solver uses linear regression and cannot solve nonlinear problems. Assume Non-Negative can be checked whenever none of the adjustable parameters is allowed to become negative. If Use Automatic Scaling is enabled, Solver changes the parameter values in relation to the magnitude of the initial values, which is recommended for all applications. Checking Show Iteration Results causes Solver to pause after every iteration and display the current values. The Estimates, Derivatives and Search options determine the details of the regression process. For difficult problems, it may be advantageous to choose Central in the Derivatives option. See the literature7,28 for more information on these advanced settings.

2010 Nature Publishing Group http://www.nature.com/natureprotocols

12| Repeat Step 11 by decreasing the fixed Km value in 0.2-mM increments until the SSR is more than four times greater than the best-fit SSR, i.e., >1200 mM2 s2. This should be the case for Km = 6.4 mM, for which the SSR amounts to 1547 mM2 s2. 13| Starting from a Km value of 7.8 mM, repeat Step 11 by increasing the fixed Km value in 0.2-mM increments until the SSR is more than four times greater than the best-fit SSR, i.e., >1200 mM2 s2. This should be the case for Km = 8.8 mM, for which the SSR amounts to 1225 mM2 s2. 14| Plot the SSR values calculated in Steps 9 and 1113 against the corresponding Km values. The resulting diagram should show a well-defined minimum at Km = 7.6 mM, as depicted in Figure 7a. The steepness of the SSR curve on both sides of the minimum is a measure of the precision with which the best-fit value can be determined from the given experimental data. It is obvious from this plot that the SSR rises sharply (i.e., the goodness of fit decreases dramatically) upon small deviations from the optimal Km, indicating that Km can be determined from the given dataset with high confidence. CRITICAL STEP Plots of the SSR against one or several adjustable parameters are sometimes rather shallow or extremely asymmetric (see Anticipated Results for an example). In such cases, the following Steps (15, 16 and 18) employed to derive confidence intervals are not applicable (see Box 5). This emphasizes the importance of visual inspection and judicious interpretation of SSR plots. CRITICAL STEP Instead of plotting the SSR, the value of the second adjustable parameter (vmax, which is varied in Steps 1113) can be plotted against that of the frozen parameter. This gives an idea about the correlation (interdependence) of the two parameters, i.e., if one parameter changes vigorously upon alteration of the other, then the two parameters are highly correlated (interdependent). This can be visualized in a contour plot or a three dimensional graph (see refs. 2,5,62,63 for examples). ? TROUBLESHOOTING 15| In cases where the SSR minimum is defined as clearly as in the present example, true confidence intervals can be approximated by the following procedure: to calculate a 95% confidence interval for Km, type =B3*(1 + 2/12*FINV(1-95/ 100; 2; 12)) into an empty cell in the Excel workbook. This should return a so-called threshold value of 491 mM2 s2. In the SSR plot, draw a horizontal line at SSR = 491 mM2 s2 and determine the two Km values at which this straight

Figure 6 | Worksheet after the completion of Step 10. Cells B1 and B2 contain the best-fit values for vmax and Km, respectively. The solid red line represents the best fit (ycalc) to the measured data (yobs; blue scatter). 274 | VOL.5 NO.2 | 2010 | nature protocols

protocol
line intersects the SSR curve (light gray lines in Fig. 7a). The lower and upper values amount to 7.1 and 8.1 mM, respectively. Thus, the 95% confidence interval of Km can be approximated as 7.18.1 mM. See Box 5 for further details. 16| Repeat Step 15 for a 99% confidence interval by using the formula =B3*(1 + 2/12*FINV(1-99/100; 2; 12)). The threshold value now is 642 mM2 s2, which intersects the SSR curve at Km values of 7.0 and 8.3 mM (dark gray lines in Fig. 7a). Thus, the 99% confidence interval of Km can be approximated as 7.08.3 mM. See also Box 5. 17| Repeat Steps 1114 with the second adjustable para meter to obtain a plot of the SSR versus vmax. In order to span the same range of SSR values as above (Steps 12 and 13), vmax should be fixed at values ranging from 580 to 620 mM s1 using increments of 5 mM s1. The resulting diagram is presented in Figure 7b. Again, the steep slopes on both sides of the minimum indicate that vmax is determined with high confidence. ? TROUBLESHOOTING

a
SSR (mM2 s2)

1,200 1,000 800 600 400 200 0

6.5

b
SSR (mM2 s2)

7.5 Km (mM)

8.5

1,200 1,000 800 600 400 200 0 570 580 590 600 610 620

2010 Nature Publishing Group http://www.nature.com/natureprotocols

vmax (mM s1)

18| Repeat Steps 15 and 16 with the second adjustable parameter to derive 95% and 99% confidence intervals of vmax. To this end, use the threshold SSR values calculated above (491 and 642 mM2 s2, for 95% and 99% confidence intervals, respectively) and determine the vmax values at which the corresponding horizontal lines intersect the SSR curve (gray lines in Fig. 7b). The confidence intervals thus obtained should be 590606 mM s1 at 95% and 587609 mM s1 at 99% confidence. See also Box 5. TIMING Steps 12, Activating Solver: 1 min (provided Solver is installed) Steps 36, Setting up worksheet and plotting data: 10 min Step 7, Setting up Solver: 10 min Steps 810, Fitting: 10 min (may vary depending on the performance of the computer being used) Steps 1118, Confidence assessment: 30 min (15 min per parameter)

Figure 7 | Confidence assessment of best-fit parameter values. Plots of sum of squared residuals (SSR) against (a) Km and (b) vmax. Light and dark gray lines mark parameter values at which the SSR amounts to, respectively, 491 and 642 mM2 s2. These parameter ranges correspond to the approximate 95% and 99% confidence intervals, respectively. Vertical black lines indicate best-fit values.

Box 5 | Calculation of confidence intervals using Fishers F distribution

 The connection between an SSR plot (see Fig. 7) and lower and upper confidence intervals at a desired confidence level (P, expressed in percent) is established by a threshold SSR value (SSRth). Under certain conditions (as explained below), the latter can be calculated from the best-fit SSR value (SSRbf) on the basis of Fishers F distribution2,5,62,63: upper (1P/100) quantile of Fishers F distribution with M being the number of adjustable parameters and N the number of data points included in the fit. The difference NM is also referred to as the number of degrees of freedom. F can easily be calculated in Excel using the formula =FINV(1P/100; M; NM), as shown in Steps 15 and 16 for M = 2 adjustable parameters (Km and vmax) and N = 14 independent data points at confidence levels of 95 and 99%, respectively.  Importantly, application of Fishers F distribution is strictly valid only for linear fitting equations2,63. However, if an SSR plot reveals a clear minimum with steep slopes on both sides (as in Fig. 7), even a nonlinear fitting equation can be assumed to be approximately linear for small deviations from the best-fit value. Then, confidence intervals at a certain confidence level P can be approximated according to the above procedure (see Steps 1118). By contrast, if the minimum in SSR is not that well defined (as in Fig. 9), this approximation no longer holds. Nevertheless, lower and upper parameter values can be derived from an arbitrarily chosen threshold SSR, but there are no simple means of ascribing a confidence level to parameter ranges thus obtained (see Anticipated Results).
nature protocols | VOL.5 NO.2 | 2010 | 275
M SSR th = SSR bf 1 + F (1 P /100; M ; N M ) . N M

F is the so-called

protocol
? TROUBLESHOOTING Troubleshooting advice can be found in Table 2. Further information on specific Solver messages can be found online at http://www.solver.com/suppstdmessages.htm.
Table 2 | Troubleshooting table. Step 2 Problem Solver is not listed in the ADD-INS window Solver cannot be found at the specified location Possible reason Solver is not installed Solution Install the Solver add-in from the Microsoft Office CD/DVD Activate the Solver add-in as described in Step 2 Activate any spreadsheet cell. Solver can be run only from a spreadsheet No problem. Provided that the fitted curve resembles the experimental data, ignore the error message. See Step 7 for details Use commercial Solver version (Premium Solver)

2010 Nature Publishing Group http://www.nature.com/natureprotocols

Solver is not activated as an add-in A graph window is activated

8,9

Error message: Solver cannot find a solution to the problem Error message: The problem is too large for Solver to handle

In the SOLVER PARAMETERS window, Value of: 0 is checked in the Equal To selection The Solver version shipped with Microsoft Excel can handle up to 200 adjustable parameters and 100 constraints on not adjustable variables as well as upper and lower bounds on all adjustable parameters

Divide dataset into smaller subsets and fit these independently (usually not recommended) Try different starting values. Ideally, they should be as close as possible to the final values. Qualitatively, the calculated curve should resemble the experimental data already at the beginning of the fitting session Use different regression algorithm. In the SOLVER OPTIONS window (see Fig. 5), try the Central option in the Derivatives selection. See Box 4 for details In the SOLVER OPTIONS window (see Fig. 5), make sure Use Automatic Scaling is checked. See Box 4 for details Use appropriate model, keeping in mind that simply adding more adjustable parameters will always result in a better fit. Avoid overfitting (see Anticipated Results for an example) Use starting values close to the expected global minimum

10

The calculated curve does not approach the experimental data points

Poor starting values for the adjustable parameters

The regression algorithm is not suitable for the problem at hand

The model is poorly scaled

The model used does not describe the experimental data appropriately

The fitted values and the SSR returned by Solver depend on the starting values

Solver gets stuck in local minima

(continued)
276 | VOL.5 NO.2 | 2010 | nature protocols

protocol
Table 2 | Troubleshooting table (continued). Step Problem Possible reason Solution Apply a more stringent convergence criterion by reducing the value in the Convergence box in the SOLVER OPTIONS window (see Fig. 5; lowest possible value is 10300). See Box 4 for details Use different regression algorithm. In the SOLVER OPTIONS window (see Fig. 5), try the Central option in the Derivatives selection. See Box 4 for details The model is poorly scaled In the SOLVER OPTIONS window (see Fig. 5), make sure Use Automatic Scaling is checked. See Box 4 for details Carry out confidence analysis, as detailed in Steps 1118. If the fitted parameters are not defined sufficiently well, fit several datasets simultaneously, apply constraints, reduce the number of adjustable parameters, optimize experimental conditions or use different experimental method (see Anticipated Results) Check and fix regression equation. Pay particular attention to units, as these do not show up in spreadsheet formulas Plot the residuals ( = yobsycalc) against x as scatter plot. If the residuals are not distributed randomly around zero (i.e., if there are systematic deviations), use a different model. Keep in mind that simply adding more adjustable parameters will always result in a better fit. Avoid overfitting (see Anticipated Results) Fit several datasets simultaneously, apply constraints, reduce the number of adjustable parameters, optimize experimental conditions or use different experimental method (see ANTICIPATED RESULTS). Do not carry out a detailed calculation of confidence intervals, as done in Steps 15, 16 and 18. See Box 5 for details Go to TOOLSREFERENCESBROWSE and open Solver in \OFFICE12\LIBRARY\SOLVER (Excel 2003: OFFICE11\LIBRARY\SOLVER)

2010 Nature Publishing Group http://www.nature.com/natureprotocols

The fitted values returned by Solver depend on the starting values, but the SSR is about the same

Overfitting: experimental data are described equally well by different solutions

The fitted curve resembles the There is a mistake in the experimental data, but the regression equation best-fit values are unrealistic The model used does not describe the experimental data appropriately

14,17

Plot of the SSR versus an adjustable parameter reveals a broad minimum or several minima

Overfitting: experimental data are described equally well by different solutions

Box 1 In VBA mode: Solver is not There is no reference between the listed in the VBA PROJECT Solver add-in and the VBA code REFERENCES window

ANTICIPATED RESULTS The protocol outlined above is, in principle, applicable to any other dataset or regression equation, provided that the basic prerequisites of NLSF outlined in the Introduction are met. Of course, the adjustable variables in Steps 3 and 7, the regression equation in Step 5, as well as the value ranges and increments used during confidence assessment in Steps 1118
nature protocols | VOL.5 NO.2 | 2010 | 277

protocol
6 will have to be adapted. In any case, it is wise to repeat the 8 fitting procedure several times using different initial values 10 for the adjustable parameters and to subject the best-fit 12 values to careful confidence analysis. 14 The way replicate measurements (repeated measurements 16 of the dependent variable at the same value of the inde18 pendent variable) should be considered in a fit depends on 20 30 40 50 60 70 80 90 100 T (C) whether the replicates are dependent on or independent of one another. For example, two readings on the same solution Figure 8 | Circular dichroism (CD) melting curve of 20 M RNAse A (SigmaAldrich, Steinheim, Germany) in 15 mM potassium acetate buffer in a spectrometer are not considered independent, whereas (pH 5.5). Experimental data (blue scatter) and best fit (red line) determined readings on two separately prepared solutions are. In the with the aid of Excel Solver using seven adjustable parameters. The best case of independent measurements, each data point should fit requiring only six fitting parameters is indistinguishable (not shown). be included in the fit, whereas dependent measurements , ellipticity at 220 nm (given in millidegrees); T, temperature (shown in should be averaged and only the mean should be included C for convenience). in the fit. Finally, it cannot be overemphasized that a good fit in terms of a low SSR need not necessarily imply that the model invoked to derive a regression equation and, consequently, the parameter values retrieved from a fitting procedure are correct or make sense. The goodness of fit as expressed by the SSR (see Step 9) and confidence intervals of the adjustable parameters (see Steps 1418) only make statements about, respectively, the agreement between experimental and fitted data and the reliability with which the best-fit values can be determined from a given dataset. It is of special importance here to note the difference between precision and accuracy. Quoting from a textbook on data reduction and error analysis4: The accuracy of an experiment is a measure of how close the result of the experiment is to the true value; the precision is a measure of how well the result has been determined, without reference to its agreement with the true value.

2010 Nature Publishing Group http://www.nature.com/natureprotocols

An example of overfitting: thermal unfolding of a protein Overfitting is probably the single most important pitfall encountered in data fitting. Overfitting refers to any attempt to extract more parameter values from an experimental dataset than the latter can actually afford. Thus, the fit is said to be underconstrained, or the parameters are underdetermined or redundant. This does not necessarily imply that the experimental data are of poor quality; in most cases, it simply means that the chosen experimental method or setup cannot unambiguously supply the desired number of parameter values. In any case, overfitting causes the values returned for some or all of the adjustable parameters to be imprecise or inaccurate. This shall be exemplified using the thermal protein unfolding data depicted in Figure 8 (blue circles). In this experiment, the small globular protein RNAse A was unfolded by raising the temperature from 24 to 94 C, and loss of secondary structure was monitored by recording the circular dichroism (CD) signal in the far-UV range (see ref. 68 for an experimental protocol). Under equilibrium conditions and making certain assumptions, thermal unfolding of RNAse A and many other proteins can be described by the following regression equation:
e = 1 T 1 1 1 + exp C p ln (Hm TmC p ) R T T T m m
o eu e fo + (mu mf )T

+ e fo + mf T

Here the independent variable is T, the absolute temperature and the dependent variable is , the ellipticity at 220 nm (given in millidegrees). R is the universal gas constant. The adjustable parameters are: Tm, the midpoint temperature of
Table 3 | Best-fit parameter values returned by Solver on fitting the thermal unfolding curve depicted in Figure 8 using either six or seven adjustable parameters. #P 6 7 Tm (C) 61.6 61.6 Hm (kJ mol )
1

Cp (J (mol K) )
1

f (mdeg) 29.5 28.9

(mdeg)

u 0.24 0.95

mf (mdeg C )
1

mu 0.02 0.02

SSR (mdeg2) 43.234 43.045

427 427

5 16

0.05 0.04

Cp, molar isobaric heat capacity change; f and u, ellipticities of, respectively, folded and unfolded protein extrapolated to T = 0 K; Hm, molar enthalpy change on unfolding at Tm; mf and mu, temperature dependencies of the ellipticities of, respectively, folded and unfolded protein; #P, number of adjustable parameters; SSR, sum of squared residuals; Tm, midpoint temperature of thermal unfolding.

278 | VOL.5 NO.2 | 2010 | nature protocols

protocol
thermal unfolding (i.e., the temperature at which half of the a 5 protein is unfolded); Hm, the molar enthalpy change on 4 unfolding at Tm; Cp, the (constant) molar isobaric 3 heat capacity change on unfolding; f and u, the ellipticities of, respectively, folded and unfolded protein 2 extrapolated to T = 0 K; as well as mf and mu, the (constant) 1 temperature dependencies of the ellipticities of folded and 0 unfolded protein, respectively. Thus, the regression equation 56 58 60 62 64 66 68 T m (C) contains seven fitting parameters, three of which (Tm, Hm and Cp) are thermodynamic parameters, whereas b 5 the other four (f, u, mf and mu) are required to represent 4 the dependencies of pre- and post-transition baselines on 3 temperature. A discussion of the assumptions made in 2 deriving the above regression equation is beyond the 68 scope of this protocol but can be found elsewhere . 1 Irrespective of this, it is obvious that fitting a simple 0 0 200 400 600 800 1,000 1,200 1,400 1,600 1,800 2,000 sigmoidal curve with sloping pre- and post-transition Hm (J (mol K)1) baselines using seven adjustable parameters is prone 5 c to serious overfitting. 4 As illustrated in Figure 8, the best fit (red line) based on seven adjustable parameters is excellent and returns the 3 values summarized in Table 3. However, Figure 9 reveals 2 that the confidence of the best-fit values of some parameters 1 is, at best, mediocre (orange lines and symbols). Tm displays 0 the highest confidence among the three parameters (Fig. 9a), but a plot of the SSR against Hm is strongly Cp (J (mol K)1) asymmetric (Fig. 9b), revealing that the SSR is extremely Figure 9 | Confidence assessment of best-fit parameter values. Plots of sum insensitive to unrealistically positive deviations from the of squared residuals (SSR) against (a) midpoint temperature (Tm), (b) molar optimal Hm value. Finally, a diagram depicting the dependenthalpy change (Hm) and (c) molar isobaric heat capacity change (Cp). Orange lines and symbols refer to a seven-parameter fit (with adjustable ence of the SSR on Cp (Fig. 9c) is characterized by a broad 1 1 Cp), whereas green lines and symbols refer to a six-parameter fit (with Cp trough ranging from 20 J (mol K) to +20 J (mol K) in fixed at 5 J (mol K)1). Dashed lines mark parameter values at which the SSR which the SSR is almost indifferent to Cp. Negative values amounts to 125% of its minimal value. Vertical black lines indicate best-fit are particularly suspect because unfolding of globular provalues. teins is usually accompanied by an increase in heat capacity. In the specific case of RNAse A, direct determination of the heat capacity change by differential scanning calorimetry has yielded a value of Cp = 5 J (mol K)1 (ref. 69). The considerable deviation between this and the best-fit value of Cp = 16 J (mol K)1 might be surprising at first glance, but inspection of Figure 9c reveals that the difference in the SSR is negligible and that the tremendously poor confidence of this parameter is because of overfitting. It is obvious that the complex, asymmetric shapes obtained on plotting the SSR versus Hm or Cp cannot be captured by a single parameter, as might be implied by fitting programs reporting a best-fit value a standard error or standard deviation. To illustrate this point, we analyzed the same dataset using the commercial fitting program Origin. Like most other commercially available software packages for NLSF, this program calculates the so-called asymptotic standard errors. Confidence intervals can then be assigned by multiplying these standard errors by an appropriate constant taken from statistical tables5. Although Origin determined the same best-fit values as found by Solver, the 95% confidence intervals calculated by Origin were only 0.1 C for Tm, 24 kJ mol1 for Hm and 5 J (mol K)1 for Cp, which clearly are gross underestimations of the real confidence intervals. Also, derivation of asymmetric confidence intervals as outlined above in Steps 15, 16 and 18 is not applicable (see Box 5 for details). In the absence of such well-established tools for confidence assessment, lower and upper parameter estimates can be obtained by arbitrarily defining a threshold SSR value. In the present example, the dashed lines in Figure 9 delimit the parameter ranges within which the SSR does not exceed 125% of the minimal value (as suggested in ref. 63). Lower and upper parameter estimates retrieved from such a procedure may help in quantifying the qualitative statements of the previous paragraph, but it should be kept in mind that the SSR threshold was chosen arbitrarily and that no single quantity can account for the complex shape of the SSR curves. The problem of overfitting can be minimally alleviated by fixing the ill-behaved parameter at Cp = 5 J (mol K)1, thus decreasing the number of fitting parameters to six and slightly improving the confidence of the other adjustable parameters (green lines and symbols in Fig. 9a,b). However, the goodness of the best fit in terms of the SSR is virtually unaffected
SSR (mdeg2) SSR (mdeg2) SSR (mdeg2)

2010 Nature Publishing Group http://www.nature.com/natureprotocols

0 6 0 4 0 2 0

20

40

80

20 1 00

nature protocols | VOL.5 NO.2 | 2010 | 279

10

0 12 0

40

60

protocol
because both the best-fit and the independently determined Cp values lie in the SSR trough in Figure 9c. In this particular case, further enhancement could be achieved by a number of remedies: (i) the temperature could be scanned over a greater range (e.g., 0100 C rather than 2494 C). This would give greater confidence to the four parameters defining the pre- and post-transition baselines, such that the latter could compensate less readily for variations in the thermodynamic parameters. (ii) Several experiments carried out under identical conditions might be fitted simultaneously (globally). (iii) Experiments could be carried out under various conditions (e.g., different pH values or in the presence of stabilizing or destabilizing additives like urea or trimethylamine N-oxide, respectively), and the data could be fitted simultaneously. This would require an additional term in the regression equation accounting for the effect of the varied parameter (i.e., a linkage between the influence of temperature and the influence of pH or additive). However, as the baselines would be the same for all datasets, the number of adjustable parameters per dataset would be reduced (A. Sieber and S. Keller, unpublished data). (iv) In addition to Cp, either Tm or Hm might be determined using an independent method, and the remaining thermodynamic parameter could be extracted from CD unfolding curves with greater confidence.
Acknowledgments We thank Professor Heiko Heerklotz (University of Toronto, Canada), Dr. Alekos Tsamaloukas (Rensselaer Polytechnic Institute, USA), Natalie Bordag (FMP) and Martin Kemmer (23karat GmbH, Berlin, Germany) for helpful discussions and constructive comments on the manuscript. This work was supported by grant KE 1478/1-1 from the Deutsche Forschungsgemeinschaft (DFG) to S.K. AUTHOR CONTRIBUTIONS G.K. designed and performed experiments, analyzed and fitted data, and wrote the manuscript. S.K. designed experiments, analyzed data and wrote the manuscript. Published online at http://www.natureprotocols.com. Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions. 1. De Levie, R. Advanced Excel for Scientific Data Analysis 2nd edn. (Oxford University Press, New York, 2008). 2. Johnson, M.L. Why, when, and how biochemists should use least squares. Anal. Biochem. 206, 215225 (1992). 3. Press, W.H., Teukolsky, A.S., Vetterling, W.T. & Flannery, B.P. Modeling of data. In Numerical Recipes in C: The Art of Scientific Computing 2nd edn. 656706 (Cambridge University Press, New York, 1992). 4. Bevington, P.R. & Robinson, D.K. Least-squares fit to an arbitrary function. In Data Reduction and Error Analysis for the Physical Sciences 3rd edn. 142167 (McGraw-Hill Higher Education, New York, 2009). 5. Motulsky, H. & Christopoulos, A. Fitting Models to Biological Data Using Linear and Nonlinear Regression: A Practical Guide to Curve Fitting 2nd edn. (GraphPad Software, San Diego, 2003). 6. Orvis, W.J. Excel for Scientists and Engineers 1st edn. (SYBEX, Alameda, 1995). 7. Fylstra, D., Lasdon, L., Watson, J. & Waren, A. Design and use of the Microsoft Excel Solver. Interfaces 28, 2955 (1998). 8. Lasdon, L.S., Waren, A.D., Jain, A. & Ratner, M. Design and testing of a generalized reduced gradient code for nonlinear programming. ACM T. Math. Software 4, 3450 (1987). 9. Beck, A., Tsamaloukas, A.D., Jurcevic, P. & Heerklotz, H. Additive action of two or more solutes on lipid membranes. Langmuir 24, 88338840 (2008). 10. Tsamaloukas, A.D., Beck, A. & Heerklotz, H. Modeling the micellization behavior of mixed and pure n-alkyl-maltosides. Langmuir 25, 43934401 (2009). 11. Plouffe, L. Jr. & Luxenberg, S.N. Biological modeling on a microcomputer using standard spreadsheet and equation solver programs: the hypothalamic-pituitary-ovarian axis as an example. Comput. Biomed. Res. 25, 117130 (1992). 12. Hargrove, J.L., Heinz, G. & Heinz, O. Modeling transitions in body composition: the approach to steady state for anthropometric measures and physiological functions in the Minnesota human starvation study. Dyn. Med. 7, 16 (2008). 13. Stevens, P.W. & Kelso, D.M. Estimation of the protein-binding capacity of microplate wells using sequential ELISAs. J. Immunol. Methods 178, 5970 (1995). 14. Kawamata, W. & Toyoshima, H. Estimation of T1 and T2 using generalpurpose spreadsheet software. Nippon Hoshasen Gijutsu Gakkai Zasshi 65, 306311 (2009). 15. Rohatagi, S., Hochhaus, G., Mllmann, H., Barth, J. & Derendorf, H. Pharmacokinetic interaction between endogenous cortisol and exogenous corticosteroids. Pharmazie 50, 610613 (1995). 16. Dansirikul, C., Choi, M. & Duffull, S.B. Estimation of pharmacokinetic parameters from non-compartmental variables using Microsoft Excel. Comput. Biol. Med. 35, 389403 (2005). 17. Meineke, I. & Brockmller, J. Simulation of complex pharmacokinetic models in Microsoft Excel. Comput. Methods Programs Biomed. 88, 239245 (2007). 18. Briones, A.M. Jr. & Reichardt, W. Estimating microbial population counts by most probable number using Microsoft Excel. J. Microbiol. Methods 35, 157161 (1999). 19. Sonnenberg, A. Special review: game theory to analyse management options in gastro-oesophageal reflux disease. Aliment. Pharmacol. Ther. 14, 14111417 (2000). 20. Ward, R., Schlenker, J. & Anderson, G.S. Simple method for developing percentile growth curves for height and weight. Am. J. Phys. Anthropol. 116, 246250 (2001). 21. Zhang, F. & Roush, W.B. Multiple-objective (goal) programming model for feed formulation: an example for reducing nutrient variation. Poult. Sci. 81, 182192 (2002). 22. Guevara, V.R. Use of nonlinear programming to optimize performance response to energy density in broiler feed formulation. Poult. Sci. 83, 147151 (2004). 23. Kuo, P.C., Schroeder, R.A., Mahaffey, S. & Bollinger, R.R. Optimization of operating room allocation using linear programming techniques. J. Am. Coll. Surg. 197, 889895 (2003). 24. Maurer, M., Khleitner, M., Gasser, B. & Mattanovich, D. Versatile modeling and optimization of fed batch processes for the production of secreted heterologous proteins with Pichia pastoris. Microb. Cell Fact. 5, 37 (2006). 25. Abdel-Fattah, Y.R. et al. Application of factorial experimental designs for optimization of cyclosporin A production by Tolypocladium inflatum in submerged culture. J. Mol. Microbiol. Biotechnol. 17, 19301936 (2007). 26. Burke, J.A. Two mathematical programming models of cheese manufacture. J. Dairy Sci. 89, 799809 (2006). 27. Schrader, H. & Svec, A. Comparison of ionization chamber efficiencies for activity measurements. Appl. Radiat. Isot. 60, 369378 (2004). 28. Brown, A.M. A step-by-step guide to non-linear regression analysis of experimental data using a Microsoft Excel spreadsheet. Comput. Methods Programs Biomed. 65, 191200 (2001). 29. Brown, A.M. A non-linear regression analysis program for describing electrophysiological data with multiple functions using Microsoft Excel. Comput. Methods Programs Biomed. 82, 5157 (2006). 30. Branco, T.J., Botelho do Rego, A.M., Ferreira, M.I. & Vieira Ferreira, L.F. Luminescence lifetime distributions analysis in heterogeneous systems by the use of Excels Solver. J. Phys. Chem. B 109, 1595815967 (2005). 31. Li, J. Comparison of the capability of peak functions in describing real chromatographic peaks. J. Chromatogr. A 952, 6370 (2002). 32. Nikitas, P., Pappa-Louisi, A. & Papageorgiou, A. On the equations describing chromatographic peaks and the problem of the deconvolution of overlapped peaks. J. Chromatogr. A 912, 1329 (2001). 33. Nikitas, P., Pappa-Louisi, A., Papageorgiou, A. & Zitrou, A. On the use of genetic algorithms for response surface modeling in high-performance liquid chromatography and their combination with the Microsoft Solver. J. Chromatogr. A 942, 93105 (2002). 34. Karmarkar, S., Garber, R., Kluza, J. & Koberda, M. Gel permeation chromatography of dextrans in parenteral solutions: calibration procedure development and method validation. J. Pharm. Biomed. Anal. 41, 12601267 (2006).

2010 Nature Publishing Group http://www.nature.com/natureprotocols

280 | VOL.5 NO.2 | 2010 | nature protocols

protocol
35. Dasgupta, P.K. Chromatographic peak resolution using Microsoft Excel Solver. The merit of time shifting input arrays. J. Chromatogr. A 1213, 5055 (2008). 36. van Dijk, J.W. Thermoluminescence glow curve deconvolution and its statistical analysis using the flexibility of spreadsheet programs. Radiat. Prot. Dosimetry 119, 332338 (2006). 37. Walsh, S. & Diamond, D. Non-linear curve fitting using Microsoft Excel Solver. Talanta 42, 561572 (1995). 38. Kane, P. & Diamond, D. Determination of ion-selective electrode characteristics by non-linear curve fitting. Talanta 44, 18471858 (1997). 39. Luther, G.W. III, Theberge, S.M. & Rickard, D. Determination of stability constants for metal-ligand complexes using the voltammetric oxidation wave of the anion/ligand and the DeFord and Hume formalism. Talanta 51, 1120 (2000). 40. Comuzzi, C., Polese, P., Melchior, A., Portanova, R. & Tolazzi, M. SOLVERSTAT: a new utility for multipurpose analysis. An application to the investigation of dioxygenated Co(II) complex formation in dimethylsulfoxide solution. Talanta 59, 6780 (2003). 41. Safavi, A., Maleki, N., Rostamzadeh, A. & Maesum, S. CCD camera full range pH sensor array. Talanta 71, 498501 (2007). 42. Parsons, J.D. A high-throughput method for fitting doseresponse curves using Microsoft Excel. Anal. Biochem. 360, 309311 (2007). 43. Brny-Wallje, E. et al. A critical reassessment of penetratin translocation across lipid membranes. Biophys. J. 89, 25132521 (2005). 44. Keller, S., Bthe, M., Bienert, M., Dathe, M. & Blume, A. A simple fluorescence-spectroscopic membrane translocation assay. ChemBioChem 8, 546552 (2007). 45. Keller, S., Tsamaloukas, A. & Heerklotz, H. A quantitative model describing the selective solubilization of membrane domains. J. Am. Chem. Soc. 127, 1146911476 (2005). 46. Schmidt, M.F., El-Dahshan, A., Keller, S. & Rademann, J. Selective identification of cooperatively binding fragments in a high-throughput ligation assay enables the evolution of a picomolar caspase-3 inhibitor. Angew. Chem. Int. Ed. 48, 63466349 (2009). 47. Keller, S. et al. Membrane-mimetic nanocarriers formed by a dipalmitoylated cell-penetrating peptide. Angew. Chem. Int. Ed. 44, 52525255 (2005). 48. Keller, S., Heerklotz, H., Jahnke, N. & Blume, A. Thermodynamics of lipid membrane solubilization by sodium dodecyl sulfate. Biophys. J. 90, 45094521 (2006). 49. Heerklotz, H., Tsamaloukas, A.D. & Keller, S. Monitoring detergentmediated solubilization and reconstitution of lipid membranes by isothermal titration calorimetry. Nat. Protoc. 4, 686697 (2009). 50. Keller, S., Heerklotz, H. & Blume, A. Monitoring lipid membrane translocation of sodium dodecyl sulfate by isothermal titration calorimetry. J. Am. Chem. Soc. 128, 12791286 (2006). 51. Geissler, D. et al. (Coumarin-4-yl)methyl esters as highly efficient, ultrafast phototriggers for protons and their application to acidifying membrane surfaces. Angew. Chem. Int. Ed. 44, 11951198 (2005). 52. Hagen, V. et al. Coumarinylmethyl esters for ultrafast release of high concentrations of cyclic nucleotides upon one- and two-photon photolysis. Angew. Chem. Int. Ed. 44, 78877891 (2005). 53. Cambridge, S.B., Geissler, D., Keller, S. & Crten, B. A caged doxycycline analogue for photoactivated gene expression. Angew. Chem. Int. Ed. 45, 22292231 (2006). 54. Gilbert, D. et al. Caged capsaicins: new tools for the examination of TRPV1 channels in somatosensory neurons. ChemBioChem 8, 8997 (2007). 55. Sauer, I. et al. Dipalmitoylation of a cellular uptake-mediating apolipoprotein E-derived peptide as a promising modification for stable anchorage in liposomal drug carriers. Biochim. Biophys. Acta. 1758, 552561 (2006). 56. Tsamaloukas, A.D., Keller, S. & Heerklotz, H. Uptake and release protocol for assessing membrane binding and permeation by way of isothermal titration calorimetry. Nat. Protoc. 2, 695704 (2007). 57. Seelig, J. Titration calorimetry of lipidpeptide interactions. Biochim. Biophys. Acta. 1331, 103116 (1997). 58. Seelig, J. Thermodynamics of lipidpeptide interactions. Biochim. Biophys. Acta. 1666, 4050 (2004). 59. Forrest, S. Genetic algorithms: principles of natural selection applied to computation. Science 261, 872878 (1993). 60. Motulsky, H.J. & Ransnas, L.A. Fitting curves to data using nonlinear regression: a practical and nonmathematical review. FASEB J. 1, 365374 (1987). 61. Johnson, M.L. Evaluation and propagation of confidence intervals in nonlinear, asymmetrical variance spaces. Analysis of ligand-binding data. Biophys. J. 44, 101106 (1983). 62. Johnson, M.L. & Frasier, S.G. Nonlinear least-squares analysis. Methods Enzymol. 117, 301342 (1985). 63. Johnson, K.A., Simpson, Z.B. & Blom, T. FitSpace Explorer: an algorithm to evaluate multidimensional parameter space in fitting kinetic data. Anal. Biochem. 387, 3041 (2009). 64. Michaelis, L. & Menten, M.L. Die Kinetik der Invertinwirkung. Biochem. Z. 49, 333369 (1913). 65. Lineweaver, H. & Burk, D. The determination of enzyme dissociation constants. J. Am. Chem. Soc. 56, 658666 (1934). 66. Berg, J.M., Tymoczko, J.L. & Stryer, L. Biochemistry 5th edn. (W.H. Freeman & Company, New York, 2002). 67. Wisniak, J. & Polishuk, A. Analysis of residualsa useful tool for phase equilibrium data analysis. Fluid Phase Equilib. 164, 6182 (1999). 68. Greenfield, N.J. Using circular dichroism collected as a function of temperature to determine the thermodynamics of protein unfolding and binding interactions. Nat. Protoc. 1, 25272535 (2006). 69. Pace, C.N., Grimsley, G.R., Thomas, S.T. & Makhatadze, G.I. Heat capacity change for ribonuclease A folding. Protein Sci. 8, 15001504 (1999).

2010 Nature Publishing Group http://www.nature.com/natureprotocols

nature protocols | VOL.5 NO.2 | 2010 | 281