C1. Database layout.

An important aspect of GMP operation is a verifiable data management policy where data ownership is tightly controlled, data are stored in well-defined structured linkages where all data transactions can be audited and timestamped, a proper chain of data custody can be assured, and access by authorized individuals can be verified. Such GMP operation is greatly alleviated by the use of a relational database. Such databases enforce linkages between data tokens, and can specify allowed operations by using ‘stored procedures’. The UltraScan-III software uses its own relational database, MariaDB [25], an open source database which is derived from the MySQL database engine. In UltraScan, all relevant details from the experiment are stored in a relational database, using stored procedures. Stored procedures allow tight control of how data are used, deleted, modified, and accessed. Data access is tightly controlled by multiple user levels to fine-tune access permissions and safeguard data. Importantly, the database can enforce ‘foreign key constraints’, a mechanism that can be used to prevent inadvertent deletion of associated data, or used to enforce a cascading deletion operation, assuring removal of otherwise orphaned nodes in the database. Database functionality within UltraScan is provided through the desktop client software [13], as well as through the web-based UltraScan Laboratory Information Management System (UltraScan-LIMS) [26]. The new features described here address the use of UltraScan for GMP applications and is collectively referred to as UltraScan GMP.

The basic concept of the UltraScan-LIMS database is to start with a detailed experimental design, termed a ‘protocol’, which contains all relevant details about the experiment to be performed. This information is entered prior to the experiment at a time when all experimental details such as the loading pattern of cells, the rotor serial number, the types of windows and centerpieces used, etc. are readily available. Using hierarchical relationships in the database, each ‘protocol’ is connected to one or more ‘experiment’, which in turn is linked to a data owner (the ‘investigator’), a technician (the ‘operator’) who performs the experiment, a ‘rotor’, and a set of ‘cells’. Each experiment is performed at a specific temperature, and has a user-defined delay for temperature equilibration. An experiment is identified by its ‘experimentID’ and a verbose ‘label’. Each cell contains one or more ‘channels’ which can contain separate samples, identified by ‘solutions’. Solutions contain ‘buffers’ composed of one or more buffer components and one or more ‘analytes’ at specified molar ratios. ‘Solutions’, ‘buffers’ and ‘analytes’ have defined spectra recorded in the database in the form of intrinsic molar extinction coefficients, as well as automatically (in the case of protein sequences) or manually defined partial specific volume, density, viscosity, and pH values. Experiments are also linked to ‘projects’ which represent digital laboratory notebooks that track the history and ancillary details of an experimental process. Experiments are performed in ‘rotors’ which inherit one or more stretch ‘calibrations’ that contain the stretch profiles of the rotor at different loading weights and/or temperatures. Experimental data are also associated with an instrument and a laboratory where the experiment was performed. Experiments have one or more speed step associated with them (recorded in the timestate object, derived from data stored in the database), which specify the rotor speed of the step, as well as the scanning frequency, the stage delay and scan interval for each optical system. Experimental data are contained in a ‘triple’, which defines the cell, channel, and wavelength for each dataset. In the case of a multi-wavelength experiment, a channel can contain multiple triples, one for each wavelength. Cells are identified by their centerpiece type, sector geometry, pathlength and material, as well as the window type. Each channel can be scanned with either UV/visible absorbance optics or Rayleigh interference optics, and wavelength ranges can be specified for each cell.

Once experimental AUC data are collected, analysis starts with an ‘edit profile’, which defines the initial meniscus position, the bottom of the cell position, which is derived from the centerpiece geometry, the rotor stretch calibration, the right data limit, and the scans to be used for subsequent data fitting. Each triple can inherit one or more edit profiles. Fitted data result in ‘models’, which are linked to the edit profile, the analysis method, and the triple associated with the fitted data. These models represent spreadsheets that contain the uncorrected and 20,W corrected sedimentation and diffusion coefficients, frictional ratios, molar masses, and partial concentrations for each species fitted in the analysis. For reacting systems, models also contain equilibrium constants for each reaction, as well as kinetic rate constants, where applicable. All models can be passed between analysis methods and visualization and reporting modules to generate graphs and detailed reports. Reports are also linked to the associated triples, or a single experiment ID when global or combined analyses are performed that do not have a unique triple as a parent [13]. Together, these data tokens uniquely define an AUC experiment’s parameters, the analysis results, and reports.

C2. UltraScan GMP graphical user interface.

The UltraScan GMP software provides a multi-platform graphical user interface for user interaction that is available for Windows, Macintosh, and Linux computers. The goal for this interface is to allow a technician with minimal training in biophysics to successfully operate the instrument in a GMP environment and monitor data acquisition as well as the analysis workflow. A separate interface for a senior scientist provides easy access to all functions needed for the development of an experimental design and analysis workflow. The type of interface presented is dependent on the user level of the individual logged into this system. Permissions are set based on the user level when the user logs in during the initial opening of the program. The user level then controls permissions for all subsequent actions, including preventing the modification of read-only elements in the database.

The program is organized into 6 distinct steps, shown in a vertical tab row on the left (see S2 Fig). The program will automatically cycle from one tab to the next and prompt the user only when any interaction is needed. At the top, a horizontal tab may offer sub-tasks, depending on each step that needs to be completed sequentially. If the program is opened during an ongoing experiment, the program will detect the status of the experiment running on any of one or more instruments connected to this system, and present the option to either re-attach to a currently running experiment, or to design and submit an experiment to a machine that is currently unoccupied.

C3. Experimental analysis profile definition.

In the first step, the experimental analysis profile is either loaded by a technician in read-only mode, allowing only the experiment name to be changed, or a new design is created (Fig 1) by a user with elevated privileges. This way, GMP procedures are exactly replicated in a repeat experiment, without any possibility of critical parameters being accidentally changed. The only field that can be changed for a stored profile is its run name, which needs to be adapted for a new lot or different date of the experiment. The UltraScan software prevents the user from using the same name twice for different experiments. In this screen, the run temperature and the delay for rotor temperature equilibration are also set. In the next screen, the laboratory in which this experiment is carried out is identified, and the instrument, rotor and rotor stretch calibration, operator, and the type of experiment are selected. The program will check the current connection status of the instrument (see S3 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. UltraScan GMP opening screen for the experimental analysis profile definition.

https://doi.org/10.1371/journal.pcbi.1007942.g001

Single- or multi-speed experiments are defined on the next screen. For each speed step, initial delay, scanning time, and scan interval can be set. The total number of scans are automatically calculated based on the length of active scanning time for a single scan, which is speed dependent. The user can readily adapt the duration of a particular speed step to achieve a desired number of scans with user-selected scan delays. A minimum initial scan delay up to the first scan is a function of rotor acceleration time as well as the internal calibration steps for the science modules of the Optima, which can be increased by the user. Minimum initial scan delays, as well as scan times, differ for the Rayleigh interference and the UV/visible absorbance optical systems, hence separate controls are available for each system, in case both optical systems are used in parallel for an experiment. A screenshot of the speed setting screen is shown in S4 Fig.

In the next screen, cells to be scanned are identified by the centerpiece and windows used. If one cell in the rotor is configured with quartz or sapphire windows, the same window and centerpiece configuration is automatically chosen for the opposite cell to assure weight balance. UltraScan is able to utilize all rotor positions for data acquisition in UV/visible detection mode, and a counterbalance does not need to be located in hole 4 (An60Ti rotor) or 8 (An50Ti rotor), as long as a 3,000 rpm radial calibration has already been performed with the counterbalance in place in a previous calibration run (see S5 Fig). For GMP experiments, the calibration centerpiece [12] can be used instead of a counterbalance to validate radial calibration, and both a 2-channel centerpiece or the calibration centerpiece can be used for the delay calibration.

In the next screen, the solution for each channel needs to be chosen (S6 Fig). These are selected from pre-defined solutions in the database that contain a buffer solution and one or more analytes (proteins, nucleic acids, carbohydrates, or others). An analyte in the database provides a partial specific volume, extinction coefficients and absorbance spectra to a solution, the buffer components of the solution specify the density and viscosity of the solution. These values are automatically used for hydrodynamic corrections to the analysis results.

In the “Optics” screen (S7 Fig), individual cells are associated with the available optical systems in the instrument. Use of a sample cell in holes 4 (for An60Ti rotors) or 8 (for An50Ti rotors) prevents usage of the interference optical system, since Rayleigh interference measurements require a counterbalance to perform the delay calibration. By default, the operation of the Optima acquires intensity data in UV/visible detection mode. Avoiding absorbance detection is preferred, and intensity mode is enforced in UltraScan to double the capacity of the instrument, by allowing for samples to be run in both channels of the cell, and to decrease stochastic noise contributions [27]. The definition of radial measurement ranges and wavelength ranges in UV/visible detection mode is performed in the next screen (S8 Fig). To facilitate wavelength selections, UltraScan GMP offers a simple scripting language, as well as a wizard to select desired wavelengths or wavelength ranges (S9 Fig).

An important innovation in UltraScan GMP is the ability to define analysis profiles. These profiles become part of the experimental design stored in the database, and must be defined before the experiment is performed. For GMP operation, it is critical that all aspects of a GMP analysis are reproducible and that all conditions of a replicate experiment or comparison experiment are identical to the extent that this can be achieved. The variables that can be controlled here do not only extend to the rotor speed, temperature, and experimental duration, but include many other conditions, such as loading concentration, loading volume, solution properties, data owner, optical systems, window type, scan frequency, scanning wavelength, centerpiece geometry, rotor stretch properties, as well as the instrument on which the data collection is performed. A second requirement is that the analysis parameters with which the experiments are analyzed are replicated identically for each sample. For example, comparisons between a market brand drug and a generic biosimilar formulation will require identical analysis workflows of the brand name drug and the biosimilar formulation in order for the comparison to be valid. Our approach is to not only define specific analysis parameters that need to be used during fitting of experimental data, but also to define analysis tolerances that reflect intrinsic confidence intervals of the technique, given the available instrumentation.

These tolerances need to be defined for each experimental system, since they are dependent on rotor speed, optical properties of the sample, optical systems and settings, and other factors that are instrument related. As is shown in S10 Fig, each channel is assigned a concentration loading ratio. By default, this ratio is unity, indicating the same concentration as was used in a reference experiment is applied here. Maintenance of identical concentration in biosimilar comparison experiments is important, since some samples change association state as a function of mass action, which changes sample composition. If mass action is to be monitored, multiple concentration ratios should be examined. Due to experimental error and micro-variations in the centerpiece pathlength, the exact same concentration may not be reproducible, therefore a tolerance value within which the experiment is considered to be repeatable is assigned. This value will guide the reporting module, and deviations from this tolerance will be flagged in the report. The loading volume of a sample in a velocity experiment is proportional to the signal-to-noise ratio and the resolution that can be obtained. The longer the solution column, the higher the resolution with which different sedimenting species can be resolved. Therefore, it is important that the loading volume is maintained across comparison experiments. To account for pipetting error, a loading volume tolerance value can be specified. Like the solution column length, the data range to be evaluated is also proportional to the resolution observed, hence the right data limit must be defined as well. Rotor-stretch dependent boundary conditions for solutions of the Lamm equation are calculated from the rotor stretch calibration profile, and independent from the right data limit [12].

In the next screen (S11 Fig), the sedimentation velocity analysis workflow with the 2-Dimensional Spectrum Analysis (2DSA) [28, 29] or Custom Grid (CG) analysis [30] methods can be pre-defined for each channel. The user can specify grid options, including predefined custom grids, grid resolution, subgrid size, and two-dimensional grid parameters, including exchange of frictional ratio for partial specific volume as one of the grid dimensions. The analysis flowchart provides default options for fitting of time- and radially-invariant noise as well as boundary conditions, refinements, and Monte Carlo iterations. Meniscus fits can be manually inspected before proceeding, or automatically processed. All default options follow the analysis workflow recommended in [27]. An optional parametrically constrained spectrum analysis [18] with all options available in the non-GMP UltraScan version can be added for each channel (see S12 Fig).

After defining the analysis workflow for each channel, a complete AUC analysis profile has been created, which is saved using a global unique identifier (GUID) in the database. This profile can be indefinitely reused, but any changes will force a name change and a new GUID to be assigned. This prevents any changes to a GMP protocol and assures the same method can be reused for future repeat experiments. In the next screen (S13 Fig), a check is performed to assure that all relevant run parameters have been defined, and can be submitted to the Optima instrument and to the UltraScan LIMS database where it will be stored in read-only mode. Only instrument specific details will be sent to the Optima. In the Optima, the GMP analysis profile will appear as a “Method Scan” entry in the PostgreSQL database. It contains all of the instructions to perform the entire data acquisition process, and cannot be edited from the Beckman-Coulter Tomcat web interface. Immediately after submission, the user is instructed to load the rotor, select the newly submitted method scan and start the experiment from the instrument. Further, the software will assure that any GMP repeat experiments derived from this profile are launched from the UltraScan software, and are executed under the controlled workflow orchestrated by UltraScan by checking the associated PostgreSQL ID issued by the Optima whenever a protocol has been submitted to the Optima.

C4. Data acquisition and monitoring.

After the experimental analysis profile has been submitted, the UltraScan GMP software automatically switches to the “Live Update” panel, monitoring any data acquisition activity occurring on the Optima AUC instrument. As soon as the run is started, temperature and rotor speed monitoring commences (S14 Fig). During data acquisition, experimental data are displayed on the monitor screen and updated in real time. A scan selection panel allows the user to switch between optical systems if both absorbance and interference optics are used, and to switch between cells, channels, and wavelengths (see S15 Fig). The rotor speed and chamber temperature are continuously updated in the GUI to show the current status of the instrument.

C5. Data import into the UltraScan LIMS database.

Once all expected data have been retrieved from the Optima AUC instrument, the UltraScan GMP program switches to the “LIMS Import” stage, and uploads the data into the LIMS database (S16 Fig). If intensity data were acquired, the user is able to manually select the air absorbance region above the meniscus, which defines the maximum transmitted light intensity reference for all scans, to be used in the conversion of intensity data to pseudo-absorbance data, using a different reference intensity for each wavelength. Upon conversion to pseudo-absorbance (S17 Fig), the data are uploaded into the UltraScan LIMS database and associated with all relevant metrics from the experiment and the experimental analysis profile definition. It should be stressed that GMP operation benefits from the direct and automated import of experimental data from the instrument into the UltraScan LIMS database in important ways, and has several key advantages compared to the default data import method performed with the Beckman-supplied software: 1. Experimental data stored in the UltraScan LIMS remain in their native double-precision format for all downstream operations, and do not suffer the information loss incurred when data are converted into ASCII formatted data by the Beckman program; 2. The default method for downloading data is tedious, requires special software to uncompress and rename data files before they can be consumed by third-party analysis software; 3. The default ASCII format consumes significantly more disk space and takes more time to read and write; 4. At no time are data manually stored by the user in an intermediate disk storage environment which introduces issues for GMP operation because of the possibility of unauthorized access and alteration of experimental data.

C6. Data editing.

After experimental import, an automated editing step is included that defines the meniscus position and the start and end data range for each experimental dataset. The end data range is based on the value entered in the general analysis profile screen (S10 Fig), while the meniscus position is determined with a peak finding algorithm near the column position predicted by the loading volume entered on the same screen. The left experimental data range is automatically identified 0.03 cm to the right of the determined meniscus position to insure boundary conditions will not be set inside the data range during a meniscus fit. Meniscus positions, baseline and plateau values (needed for the van Holde–Weischet [31] and the dC/dt [32] analysis) for each triple are determined automatically during this stage. First, the meniscus position is estimated based on a simple functional analysis of the data scans corresponding to the region in the vicinity of the cell top which, in turn, is estimated based on the parameters defining the centerpiece geometry discussed above (all derived from the LIMS database), and the loading volume for the channel (defined in the analysis_profile). For example, the initial estimate of the meniscus position, r_m, for a sector shaped centerpiece is inferred from the loading volume, v (in μl), the sector angle θ (in degrees), the bottom position of the cell, r_b (calculated in cm from the rotor stretch profile [12]), and the pathlength h (in cm) of the centerpiece as shown in Eq 1: Eq 1

Each scan in the triple dataset is then examined for the existence of a prominent peak in the immediate vicinity of the initial estimate for the r_m value, which is accepted as the meniscus value for each scan. The actual meniscus position is defined as the average over the meniscus positions from all scans belonging to the same triple. To improve the reliability of the method, only the last ~20% of the scans in the triple are used for this analysis, since later scans in an experiment are expected to have absorbances closer to the baseline near the meniscus and are therefore more easily to recognize as meniscus positions. Next, the baseline of the dataset is determined by taking the average absorbance value of 20 points, centered at the radial position in the last scan of the experiment 0.03 cm to the right of the meniscus position. Similarly, the plateau value for each scan is defined as the absorbance average of 20 points at the right limit of the data range, shifted to the left by a predefined offset value (also 0.03 cm). All computed parameters (meniscus, data range, plateau, and baseline) are recorded in the ‘Edit Profile' associated with the corresponding raw dataset (a triple). In the case of a multi-wavelength AUC experiment, these parameters are applied to all triples from all wavelengths associated with the same cell and channel. Optionally, a manual editing window is available (S18 Fig) to fine-tune other parameters of the editing process. At this time, only intensity data are supported, since special cases encountered during interference data editing prevent reliable automation. We hope to arrive at robust algorithms in the future to include interference experiments as well.

C7. Data analysis.

The data analysis phase is an automated process that sequentially applies the fitting refinement steps defined in the analysis profile above on a remote supercomputer or local cluster. All communications with the cluster use secure sockets layer (SSL) encryption to prevent unauthorized man-in-the-middle access of experimental data or results. The sequence of the data analysis process is defined by the analysis process record and managed by the ‘Autoflow’ supervisor, a process control system that monitors each step of the GMP data collection and analysis workflow, which is illustrated in S19 Fig. The analysis workflow logic was previously discussed in [27]. Briefly, the default analysis workflow starts with a 2DSA analysis with time-invariant noise removal, followed by a meniscus and bottom fit performed with time- and radially invariant noise removal. Additional refinement is achieved with the iterative 2DSA step, which again includes time- and radially-invariant noise removal, followed either by genetic algorithm analysis [33] for pauci-disperse samples, by 2DSA-Monte Carlo analysis [34], or by PCSA analysis when heterogeneous samples are involved. All models generated by sequential analysis steps are deposited in the UltraScan LIMS database for further meta data analysis or analysis reporting. During analysis, all datasets are submitted to a high-performance computer for parallel analysis, using batch processing [35]. In the Analysis tab, the user can follow the progress of each dataset’s analysis and receive real-time job status updates (see S20 Fig, S21 Fig).

C8. Reporting.

All reporting is performed automatically and is based on the fitted models deposited in the UltraScan LIMS database. Optionally, analysis results can be compared to a validated reference data set. A reference data set has been processed with the same experimental analysis profile, and detailed results are available for comparison. For example, a reference data set may stem from the analysis of a brand name pharmaceutical, or a DNA standard previously analyzed under GMP conditions. In the report, a list of values obtained from tests performed on the primary data and the fitted models, are compared to the values obtained from the same tests performed on the reference dataset. Alternatively, the values can be reported as-is, if a reference dataset is not available. The criteria tested include properties of the experimental data, such as loading volume, loading concentration, duration of the experiment, rotor speed, number of scans collected, wavelength ranges used for measurement, and the solution column length, as well as the root mean square deviations for the fitted models. In addition, the results from a metadata analysis, where partial concentrations of individual analytes are compared, or percentages of signal observed in subsections of the 2DSA grid, will be reported. Reports for comparisons with reference data sets will provide overlay plots and percent deviations for the tested values. If values exceed the pre-configured tolerance values, the report will be flagged. Thus, the software is able to provide the user with an entirely unbiased analysis report. For custom reports and additional analysis not considered in UltraScan GMP, the same data can also be retrieved (in read-only mode) from the academic UltraScan software and further processed with the full set of functionality available in UltraScan. In the reporting screen (S22 Fig), a report for each dataset can be viewed or downloaded for each dataset, and reports on all parameters discussed above, and their deviations from the set tolerances. A Pass/Fail grade will be assigned based on whether a data analysis met all previously set tolerance values (see Section C3). Instrument specific performance data, such as RMSD values expected from a well-tuned instrument, as well as expected agreement with reference standards will be determined in future experimental work.

C9. Autoflow Supervisor.

The data acquisition program has a flexible yet simple and intuitive interface to manage experiments submitted to the Optima AUC. It features screens for designing and/or submitting a new experimental analysis profile, to monitor the experimental progress, and to import and post-process acquired data. In addition, it allows the user to re-attach to any stage of this process, and to delete abandoned experiments from the list it manages. The management is performed by the autoflow supervisor, which keeps track of each experiment by linking the details of the designed experimental protocol to the details on the associated experiment that was launched from a particular Optima instrument. The protocol details are represented by an experimental description, a profile name and a unique ID in the Optima and the LIMS database. Experimental profiles include information required for the Optima to execute the profile, such as the duration of the experiment, details on the wavelength range(s), scan frequency, and optical systems used. For laboratories with multiple instruments, the supervisor must also keep track of multiple Optima instruments. This information represents a small subset of the protocol record kept in the LIMS database and is used by the autoflow supervisor whenever protocol details need to be retrieved at different stages of the program execution. To keep track of the stage for each experiment, an autoflow database table has been added to the UltraScan LIMS database, which keeps the state information for real time access, and can be queried by any software instance, at any time for the up-to-date system state of the UltraScan GMP module.

For each experiment, the Optima AUC stores the associated runID (generated upon launching the protocol from the Optima panel), the run status, and the time stamp information. Like the experimental protocol details, the Optima AUC’s runID can be used to access a wide range of information about the instrument status or acquired data, as soon as it becomes available. Together, these concise sets of parameters provide a unique signature for any experiment. These run signatures are stored in the autoflow LIMS database table. In addition to the experimental analysis profile’s ID and the Optima status information, these table records include a field identifying the state of each experiment, which is used to switch UltraScan GMP to the correct processing stage. For example, when the experiment is submitted to the Optima instrument, the program switches to the “Live Update” stage, an autoflow table record is generated which includes protocol and run details, and the program status (“LIVE_UPDATE”). When the Optima run is completed, the program proceeds to the LIMS Import stage, and the autoflow table record for this run is updated with status “IMPORT” (and other data relevant parameters). The autoflow table records are used to both characterize Optima runs, and identify and reattach to the appropriate program’s stage, be it monitoring during data acquisition (2. Live Update), importing of the retrieved data into the relational LIMS database (3. LIMS Import), data post-processing during data editing (4. Editing) or data analysis (5. Analysis). A report is generated at the end, concluding the autoflow record. Together, these features allow multiple users to monitor ongoing experiments with random access. To prevent corruption from multiple accesses, a global unique identifier (GUID) is used as a session variable that prevents multiple users from simultaneously importing the same experimental data into the LIMS database, or submitting different experiments to the same instrument. The autoflow process is illustrated in S19 Fig.