Testing Precision Limits of Neural Network-Based Quality Control Metrics in High-Throughput Digital Microscopy.

OBJECTIVE: Digital microscopy is used to monitor particulates such as protein aggregates within biopharmaceutical products. The images that result encode a wealth of information that is underutilized in pharmaceutical process monitoring. For example, images of particles in protein drug products typically are analyzed only to obtain particle counts and size distributions, even though the images also reflect particle characteristics such as shape and refractive index. Multiple groups have demonstrated that convolutional neural networks (CNNs) can extract information from images of protein aggregates allowing assignment of the likely stress at the "root-cause" of aggregation. A practical limitation of previous CNN-based approaches is that the potential aggregation-inducing stresses must be known a priori, disallowing identification of particles produced by unknown stresses. METHODS: We demonstrate an expanded CNN analysis of flow imaging microscopy (FIM) images incorporating judiciously chosen particle standards within a recently proposed "fingerprinting algorithm" (Biotechnol. & Bioeng. (2020) 117:3322) that allows detection of particles formed by unknown root-causes. We focus on ethylene tetrafluoroethylene (ETFE) microparticles as standard surrogates for protein aggregates. We quantify the sensitivity of the new algorithm to experimental parameters such as microscope focus and solution refractive index changes, and explore how FIM sample noise affects statistical testing procedures. RESULTS & CONCLUSIONS: Applied to real-world microscopy images of protein aggregates, the algorithm reproducibly detects complex, distinguishing "textural features" of particles that are not easily described by standard morphological measurements. This offers promise for quality control applications and for detecting shifts in protein aggregate populations due to stresses resulting from unknown process upsets.

An Interlaboratory Study to Identify Potential Visible Protein-Like Particle Standards.

Visible protein-like particle standards may improve visual inspection and/or appearance testing practices used in the biotechnology industry. They may improve assay performance resulting in better alignment and more standardized training among different companies. The National Institute of Standards and Technology (NIST) has conducted an interlaboratory study to test whether the standards under development mimic typical proteinaceous particles found in biotherapeutics and if they can be implemented during the visual inspection process. Fourteen organizations from industry and government have participated. A total of 20 labs from these 14 organizations participated with analysts from 6 formulation, 7 analytical, 4 quality control, and 3 manufacturing labs. The circulated samples consisted of abraded ethylene tetrafluoroethylene (ETFE) particles or photolithographic particles. The results consist of qualitative ratings, which varied substantially among organizations and within labs. Polydisperse ETFE particle suspensions, containing particles enriched in greater than 150 µm in size, were rated more favorably than the photolithographic particles by formulation and analytical scientists. The largest monodisperse photolithographic particles (approximately 300 µm in size) were favored equally compared to ETFE by all scientists. Solution modifications to decrease the settling rate or to alter optical properties of the ETFE solutions yielded lower ratings by the analysts. Both particle types received mixed ratings for their usability and for their application for visual inspection and for training purposes. Industry feedback will assist NIST in developing reference material(s) for visible protein-like particles.

Evaluating changes to Ralstonia pickettii in high-purity water to guide selection of potential calibration materials for online water bioburden analyzers.

Online water bioburden analyzers (OWBAs) can provide real-time feedback on viable bacteria in high-purity water (HPW) systems for pharmaceutical manufacturers. To calibrate and validate OWBAs, which detect bacteria using scattered light and bacterial autofluorescence, standards are needed that mimic the characteristics of bacteria in HPW. To guide selection of potential standards, e.g., fluorescent microspheres, a relevant bacterial contaminant, Ralstonia pickettii, was characterized for size, count, viability, and autofluorescence after exposure for 24 h to HPW or a nutrient environment. The cells exposed to HPW showed smaller sizes, with lower counts and autofluorescence intensities, but similar spectral features. The cell characteristics are discussed in comparison with a set of fluorescent microspheres, considering factors relevant to OWBAs. These studies suggest that fluorescent microspheres should be relatively small (< 1 µm diameter) and dim, while covering a broad emission range from ≈ (420 to 600) nm to best mimic the representative R. pickettii.

Primary Determination of Particle Number Concentration with Light Obscuration and Dynamic Imaging Particle Counters.

Accurate number concentrations of particles in liquid media are needed to assess the quality of water, pharmaceuticals, and other liquids, yet there are limited reference materials or calibration services available with clear traceability to the International System of Units. We describe two methods, based on very simple modifications of commercial particle counter instruments, that can provide traceable number concentration measurements. One method used a light obscuration counter. Fitting a model to the data enabled correction for timing and coincidence errors, and gravimetric calibration of the syringe pump gave a traceable determination of measured volume. Other potential biases were diagnosed by analysis of the particle size distribution. The other method used a dynamic imaging particle counter (a flow imaging microscope). The instrument was intentionally configured so that each particle passing through the flow cell was imaged multiple times. Following the particle image acquisition runs, runs with a rinse solution released and counted microspheres adsorbed to tubing or flow-cell walls. Software assembled the redundant particle images into tracks, and the total number of tracks was assigned as the number of particles counted. Both light obscuration and dynamic imaging methods, when applied to polystyrene microspheres of approximately 4 µm diameter, achieved expanded uncertainties (k = 2) of approximately 2 % of number concentration and agreed to within a difference of 1.1 %.

Correcting the Relative Bias of Light Obscuration and Flow Imaging Particle Counters.

PURPOSE: Industry and regulatory bodies desire more accurate methods for counting and characterizing particles. Measurements of proteinaceous-particle concentrations by light obscuration and flow imaging can differ by factors of ten or more. METHODS: We propose methods to correct the diameters reported by light obscuration and flow imaging instruments. For light obscuration, diameters were rescaled based on characterization of the refractive index of typical particles and a light scattering model for the extinction efficiency factor. The light obscuration models are applicable for either homogeneous materials (e.g., silicone oil) or for chemically homogeneous, but spatially non-uniform aggregates (e.g., protein aggregates). For flow imaging, the method relied on calibration of the instrument with silica beads suspended in water-glycerol mixtures. RESULTS: These methods were applied to a silicone-oil droplet suspension and four particle suspensions containing particles produced from heat stressed and agitated human serum albumin, agitated polyclonal immunoglobulin, and abraded ethylene tetrafluoroethylene polymer. All suspensions were measured by two flow imaging and one light obscuration apparatus. Prior to correction, results from the three instruments disagreed by a factor ranging from 3.1 to 48 in particle concentration over the size range from 2 to 20 µm. Bias corrections reduced the disagreement from an average factor of 14 down to an average factor of 1.5. CONCLUSIONS: The methods presented show promise in reducing the relative bias between light obscuration and flow imaging.

The Use of Index-Matched Beads in Optical Particle Counters.

In this paper, we demonstrate the use of 2-pyridinemethanol (2P) aqueous solutions as a refractive index matching liquid. The high refractive index and low viscosity of 2P-water mixtures enables refractive index matching of beads that cannot be index matched with glycerol-water or sucrose-water solutions, such as silica beads that have the refractive index of bulk fused silica or of polymethylmethacrylate beads. Suspensions of beads in a nearly index-matching liquid are a useful tool to understand the response of particle counting instruments to particles of low optical contrast, such as aggregated protein particles. Data from flow imaging and light obscuration instruments are presented for bead diameters ranging from 6 µm to 69 µm, in a matrix liquid spanning the point of matched refractive index.

Morphologically-Directed Raman Spectroscopy as an Analytical Method for Subvisible Particle Characterization in Therapeutic Protein Product Quality.

Subvisible particles (SVPs) are a critical quality attribute of injectable therapeutic proteins (TPs) that needs to be controlled due to potential risks associated with drug product quality. The current compendial methods routinely used to analyze SVPs for lot release provide information on particle size and count. However, chemical identification of individual particles is also important to address root-cause analysis. Herein, we introduce Morphologically-Directed Raman Spectroscopy (MDRS) for SVP characterization of TPs. The following particles were used for method development: (1) polystyrene microspheres, a traditional standard used in industry; (2) photolithographic (SU-8); and (3) ethylene tetrafluoroethylene (ETFE) particles, candidate reference materials developed by NIST. In our study, MDRS rendered high-resolution images for the ETFE particles (> 90%) ranging from 19 to 100 µm in size, covering most of SVP range, and generated comparable morphology data to flow imaging microscopy. Our method was applied to characterize particles formed in stressed TPs and was able to chemically identify individual particles using Raman spectroscopy. MDRS was able to compare morphology and transparency properties of proteinaceous particles with reference materials. The data suggests MDRS may complement the current TPs SVP analysis system and product quality characterization workflow throughout development and commercial lifecycle.

Number Concentration Measurements of Polystyrene Submicrometer Particles.

The number of techniques to measure number concentrations and size distributions of submicrometer particles has recently increased. Submicrometer particle standards are needed to improve the accuracy and reproducibility of these techniques. The number concentrations of fluorescently labeled polystyrene submicrometer sphere suspensions with nominal 100 nm, 200 nm and 500 nm diameters were measured using seven different techniques. Diameter values were also measured where possible. The diameter values were found to agree within 20%, but the number concentration values differed by as much as a factor of two. Accuracy and reproducibility related with the different techniques are discussed with the goal of using number concentration standards for instrument calibration. Three of the techniques were used to determine SI-traceable number concentration values, and the three independent values were averaged to give consensus values. This consensus approach is proposed as a protocol for certifying SI-traceable number concentration standards.

An Interlaboratory Comparison on the Characterization of a Sub-micrometer Polydisperse Particle Dispersion.

The measurement of polydisperse protein aggregates and particles in biotherapeutics remains a challenge, especially for particles with diameters of ≈ 1 µm and below (sub-micrometer). This paper describes an interlaboratory comparison with the goal of assessing the measurement variability for the characterization of a sub-micrometer polydisperse particle dispersion composed of five sub-populations of poly(methyl methacrylate) (PMMA) and silica beads. The study included 20 participating laboratories from industry, academia, and government, and a variety of state-of-the-art particle-counting instruments. The received datasets were organized by instrument class to enable comparison of intralaboratory and interlaboratory performance. The main findings included high variability between datasets from different laboratories, with coefficients of variation from 13 % to 189 %. Intralaboratory variability was, on average, 37 % of the interlaboratory variability for an instrument class and particle sub-population. Drop-offs at either end of the size range and poor agreement on maximum counts of particle sub-populations were noted. The mean distributions from an instrument class, however, showed the size-coverage range for that class. The study shows that a polydisperse sample can be used to assess performance capabilities of an instrument set-up (including hardware, software, and user settings) and provides guidance for the development of polydisperse reference materials.

Characterization of gold nanoparticles modified with single-stranded DNA using analytical ultracentrifugation and dynamic light scattering.

We report the characterization of gold nanoparticles modified with thiol-terminated single stranded DNA (ssDNA) using analytical ultracentrifugation. Dynamic light scattering was used to measure the diameter of bare and ssDNA modified gold nanoparticles to corroborate the predictions of our models. Sedimentation coefficients of nominally 10 and 20 nm diameter gold nanoparticles modified with thiol-terminated thymidine homo-oligonucleotides, 5-30 bases in length, were determined with analytical ultracentrifugation. The sedimentation coefficients of gold nanoparticles modified with ssDNA were found to decrease with increasing coverage of ssDNA and increasing length of ssDNA. The sedimentation coefficients of ssDNA modified gold particles were most closely predicted when the strands were modeled as fully extended chains (FEC). Apparent particle densities of bare gold nanoparticles calculated from measured sedimentation coefficients decreased significantly below that of bulk gold with decreasing size of nanoparticles. This finding suggests that hydration layer effects are an important factor in the sedimentation behavior for both bare and short ssDNA chain modified gold particles.

Improving Diameter Accuracy for Dynamic Imaging Microscopy for Different Particle Types.

Dynamic imaging analysis instruments are used for sizing particles of different types that might appear in a biopharmaceutical. These instruments are calibrated using polystyrene latex microspheres in water, which is a significantly different system than the typical particles imaged in a formulation. We show how the instruments, when reporting an equivalent diameter, set a threshold for image processing and then apply a built-in correction to account for fuzzy boundary effects. We investigate the degree to which the threshold value and built-in correction influences the size, and ultimately particle size distribution, that the instrument reports on other particle types. Size corrections for a dynamic imaging system in a typical optical configuration were determined by comparison of equivalent image diameters with diameters from Brownian motion tracking of particles. A variety of particles were characterized: aggregates made from a monoclonal antibody available as reference material RM 8671 from the National Institute of Standards and Technology, bovine serum albumin aggregates, silicone oil droplets, polystyrene microspheres, and ethylene tetrafluoroethylene particles, a protein aggregate simulant (National Institute of Standards and Technology reference material RM 8634). The results show that the protein aggregates and ethylene tetrafluoroethylene are very similar to one another but quite different from the polystyrene calibration spheres. This points the way to developing new correction factors and calibration procedures based on particle type.

Measurement of Average Aggregate Density by Sedimentation and Brownian Motion Analysis.

The spatially averaged density of protein aggregates is an important parameter that can be used to relate size distributions measured by orthogonal methods, to characterize protein particles, and perhaps to estimate the amount of protein in aggregate form in a sample. We obtained a series of images of protein aggregates exhibiting Brownian diffusion while settling under the influence of gravity in a sealed capillary. The aggregates were formed by stir-stressing a monoclonal antibody (NISTmAb). Image processing yielded particle tracks, which were then examined to determine settling velocity and hydrodynamic diameter down to 1 µm based on mean square displacement analysis. Measurements on polystyrene calibration microspheres ranging in size from 1 to 5 µm showed that the mean square displacement diameter had improved accuracy over the diameter derived from imaged particle area, suggesting a future method for correcting size distributions based on imaging. Stokes' law was used to estimate the density of each particle. It was found that the aggregates were highly porous with density decreasing from 1.080 to 1.028 g/cm3 as the size increased from 1.37 to 4.9 µm.

Using Image Attributes to Assure Accurate Particle Size and Count Using Nanoparticle Tracking Analysis.

Nanoparticle tracking analysis (NTA) obtains particle size by analysis of particle diffusion through a time series of micrographs and particle count by a count of imaged particles. The number of observed particles imaged is controlled by the scattering cross-section of the particles and by camera settings such as sensitivity and shutter speed. Appropriate camera settings are defined as those that image, track, and analyze a sufficient number of particles for statistical repeatability. Here, we test if image attributes, features captured within the image itself, can provide measurable guidelines to assess the accuracy for particle size and count measurements using NTA. The results show that particle sizing is a robust process independent of image attributes for model systems. However, particle count is sensitive to camera settings. Using open-source software analysis, it was found that a median pixel area, 4 pixels2, results in a particle concentration within 20% of the expected value. The distribution of these illuminated pixel areas can also provide clues about the polydispersity of particle solutions prior to using a particle tracking analysis. Using the median pixel area serves as an operator-independent means to assess the quality of the NTA measurement for count.

Variable Threshold Method for Determining the Boundaries of Imaged Subvisible Particles.

An accurate assessment of particle characteristics and concentrations in pharmaceutical products by flow imaging requires accurate particle sizing and morphological analysis. Analysis of images begins with the definition of particle boundaries. Commonly a single threshold defines the level for a pixel in the image to be included in the detection of particles, but depending on the threshold level, this results in either missing translucent particles or oversizing of less transparent particles due to the halos and gradients in intensity near the particle boundaries. We have developed an imaging analysis algorithm that sets the threshold for a particle based on the maximum gray value of the particle. We show that this results in tighter boundaries for particles with high contrast, while conserving the number of highly translucent particles detected. The method is implemented as a plugin for FIJI, an open-source image analysis software. The method is tested for calibration beads in water and glycerol/water solutions, a suspension of microfabricated rods, and stir-stressed aggregates made from IgG. The result is that appropriate thresholds are automatically set for solutions with a range of particle properties, and that improved boundaries will allow for more accurate sizing results and potentially improved particle classification studies.

An interlaboratory comparison of sizing and counting of subvisible particles mimicking protein aggregates.

Accurate counting and sizing of protein particles has been limited by discrepancies of counts obtained by different methods. To understand the bias and repeatability of techniques in common use in the biopharmaceutical community, the National Institute of Standards and Technology has conducted an interlaboratory comparison for sizing and counting subvisible particles from 1 to 25 µm. Twenty-three laboratories from industry, government, and academic institutions participated. The circulated samples consisted of a polydisperse suspension of abraded ethylene tetrafluoroethylene particles, which closely mimic the optical contrast and morphology of protein particles. For restricted data sets, agreement between data sets was reasonably good: relative standard deviations (RSDs) of approximately 25% for light obscuration counts with lower diameter limits from 1 to 5 µm, and approximately 30% for flow imaging with specified manufacturer and instrument setting. RSDs of the reported counts for unrestricted data sets were approximately 50% for both light obscuration and flow imaging. Differences between instrument manufacturers were not statistically significant for light obscuration but were significant for flow imaging. We also report a method for accounting for differences in the reported diameter for flow imaging and electrical sensing zone techniques; the method worked well for diameters greater than 15 µm.

Using light scattering to evaluate the separation of polydisperse nanoparticles.

The analysis of natural and otherwise complex samples is challenging and yields uncertainty about the accuracy and precision of measurements. Here we present a practical tool to assess relative accuracy among separation protocols for techniques using light scattering detection. Due to the highly non-linear relationship between particle size and the intensity of scattered light, a few large particles may obfuscate greater numbers of small particles. Therefore, insufficiently separated mixtures may result in an overestimate of the average measured particle size. Complete separation of complex samples is needed to mitigate this challenge. A separation protocol can be considered improved if the average measured size is smaller than a previous separation protocol. Further, the protocol resulting in the smallest average measured particle size yields the best separation among those explored. If the differential in average measured size between protocols is less than the measurement uncertainty, then the selected protocols are of equivalent precision. As a demonstration, this assessment metric is applied to optimization of cross flow (V(x)) protocols in asymmetric flow field flow fractionation (AF(4)) separation interfaced with online quasi-elastic light scattering (QELS) detection using mixtures of polystyrene beads spanning a large size range. Using this assessment metric, the V(x) parameter was modulated to improve separation until the average measured size of the mixture was in statistical agreement with the calculated average size of particles in the mixture. While we demonstrate this metric by improving AF(4) V(x) protocols, it can be applied to any given separation parameters for separation techniques that employ dynamic light scattering detectors.

Particle shape effects on subvisible particle sizing measurements.

Particle analysis tools for the subvisible (<100 µm) size range, such as light obscuration, flow imaging (FI), and electrical sensing zone (ESZ), often produce results that do not agree with one another, despite their general agreement when characterizing polystyrene latex spheres of different sizes. To include the effect of shape in comparison studies, we have used the methods of photolithography to create rods and disks. Although the rods are highly monodisperse, the instruments produce broadened peaks and report mean size parameters that are different for different instruments. We have fabricated a microfluidic device that simultaneously performs ESZ and FI measurements on each particle to elucidate the causes of discrepancies and broadening. Alignment of the rods with flow causes an oversizing by FI and undersizing by ESZ. FI also oversizes rods because of the incorrect edge definition that results from diffraction and imperfect focus. We present an improved correction algorithm for this effect that reduces discrepancies for rod-shaped particles. Tumbling of particles is observed in the microfluidic ESZ/FI and results in particle oversizing and breadth of size distribution for the monodisperse rods.

Subvisible (2-100 µm) Particle Analysis During Biotherapeutic Drug Product Development: Part 1, Considerations and Strategy.

Measurement and characterization of subvisible particles (defined here as those ranging in size from 2 to 100 µm), including proteinaceous and nonproteinaceous particles, is an important part of every stage of protein therapeutic development. The tools used and the ways in which the information generated is applied depends on the particular product development stage, the amount of material, and the time available for the analysis. In order to compare results across laboratories and products, it is important to harmonize nomenclature, experimental protocols, data analysis, and interpretation. In this manuscript on perspectives on subvisible particles in protein therapeutic drug products, we focus on the tools available for detection, characterization, and quantification of these species and the strategy around their application.

Protein particles: what we know and what we do not know.

All therapeutic protein products contain intrinsic particles formed by the aggregation of protein monomers. There is growing interest in understanding particles in biopharmaceutical products, fostered on one hand by significant advancements in particle analysis and on the other hand by concerns about potential impact of particles on product quality and safety. With currently available methods, particles in therapeutic proteins can be counted, sized, and characterized in a rudimentary way over a broad size range (from 10s of nanometers to 100 s of micrometers). Here, we review the known attributes of common protein particles, and then discuss the gaps in our current knowledge. The capabilities, limitations, and opportunities for improvement of common particle counting and characterization methods are listed. We conclude that further analytical progress is needed to better classify and characterize the diversity of particles encountered in therapeutic proteins, which may vary in the degree of protein unfolding, the inclusion of nonprotein nucleation centers, and aggregate morphology. Very little is known about the potential correlation between specific particle attributes and increased immunogenicity. In this environment of uncertainty, a deeper understanding about specific particle attributes and potentially increased immunogenicity is greatly needed and will likely be an area of future intensive research.