Published: March 10th, 2023
This protocol details high-throughput crystallization screening, ranging from the 1,536 microassay plate preparation to the end of a 6 week experimental time window. Details are included about the sample setup, the imaging obtained, and how users can perform analyses using an artificial intelligence-enabled graphical user interface to quickly and efficiently identify macromolecular crystallization conditions.
X-ray crystallography is the most commonly employed technique to discern macromolecular structures, but the crucial step of crystallizing a protein into an ordered lattice amenable to diffraction remains challenging. The crystallization of biomolecules is largely experimentally defined, and this process can be labor-intensive and prohibitive to researchers at resource-limited institutions. At the National High-Throughput Crystallization (HTX) Center, highly reproducible methods have been implemented to facilitate crystal growth, including an automated high-throughput 1,536-well microbatch-under-oil plate setup designed to sample a wide breadth of crystallization parameters. Plates are monitored using state-of-the-art imaging modalities over the course of 6 weeks to provide insight into crystal growth, as well as to accurately distinguish valuable crystal hits. Furthermore, the implementation of a trained artificial intelligence scoring algorithm for identifying crystal hits, coupled with an open-source, user-friendly interface for viewing experimental images, streamlines the process of analyzing crystal growth images. Here, the key procedures and instrumentation are described for the preparation of the cocktails and crystallization plates, imaging the plates, and identifying hits in a way that ensures reproducibility and increases the likelihood of successful crystallization.
Even in an age of tremendous progress in structural biology methods, X-ray crystallography continues to be a dependable and popular method for generating high-quality structural models of macromolecules. Over 85% of all three-dimensional structural models deposited to the Protein Data Bank (PDB) are from crystal-based structural methods (as of January, 2023).1 Furthermore, X-ray crystallography remains indispensable for solving protein-ligand structures, a crucial component of the drug discovery and development process2. Despite protein crystallization having remained the dominant structural biology technique for over half a century, methods to predict crystallization likelihood based on physical properties3 or sequence4,5 are still in their infancy.
The prediction of crystallization conditions is even more obscure; limited progress has been made to predict likely crystallization conditions even for model proteins6,7. Other studies have attempted to identify crystallization conditions based on protein homology and conditions mined from the PDB8,9,10. The predictive power to be found in the PDB is limited, however, as only the final, successful crystallization conditions are deposited, which, by necessity, misses the often extensive optimization experiments required to fine-tune crystal growth. Further, many PDB entries lack metadata containing these details, including the cocktail formulas, crystallization format, temperature, and time to crystallize11,12. Therefore, for many proteins of interest, the most accessible way to determine the crystallization conditions is experimentally, using as many conditions as possible across a wide range of chemical possibilities.
Several approaches to make crystallization screening as fruitful and thorough as possible have been explored to great effect, including sparse matrices13, incomplete factorial screening14, additives15,16, seeding17, and nucleating agents18. The National HTX Center at Hauptman-Woodward Medical Research Institute (HWI) has developed an efficient pipeline for crystallization screening using the microbatch-under-oil approach19, which utilizes automated liquid handling and imaging modalities to streamline the identification of initial crystallization conditions using comparatively minimal sample and cocktail volumes (Figure 1). The set of 1,536 unique cocktails are based on conditions previously determined to be conducive to protein crystal growth and are designed to be chemically diverse in order to sample a large range of possible crystallization conditions20,21,22. The broad sampling of crystallization conditions increases the likelihood of observing one or more crystallization leads.
Few formal analyses of how many conditions are needed for screening have appeared in the literature. One study focused on the sampling layout of different screens and found that the random sampling of components (similar to an incomplete factorial) represented the most thorough and efficient sampling method23. Another study of screening noted that there have been numerous instances when the very thorough 1,536 screen has yielded only a single crystal hit24, and a very recent study highlighted that most commercial screens undersample the crystallization space known to be associated with screening hits25. Not all crystallization leads will yield a diffraction quality crystal suitable for data collection due to inherent disorder within the crystal, diffraction limitations, or crystal flaws; therefore, casting a wider net for conditions has the additional benefit of providing alternative crystal forms for optimization.
The format of protein crystallization experiments also has an impact on the success of the screen. Vapor diffusion is the most commonly used setup for high-throughput crystallization applications and is utilized at state-of-the art crystallization centers, including the EMBL Hamburg and Institut Pasteur high-throughput screening centers26,27,28. The HTX Center uses the microbatch-under-oil method; while less commonly used, it is a robust method that minimizes the consumption of sample and crystallization cocktails20,21,22. One advantage of the microbatch-under-oil method, particularly when using a high-viscosity paraffin oil, is that only slight evaporation occurs within the drop during the experiment, meaning that the equilibrium concentration is achieved upon drop mixing. If positive crystallization results are observed in the microbatch-under-oil method, the reproduction of these conditions is typically more straightforward than in vapor diffusion setups, in which crystallization occurs at some undefined point during the equilibration between the crystallization drop and the reservoir. The reproducibility of hits is desirable for high-throughput crystallization approaches, which produce prohibitively tiny protein crystals that typically need to be optimized for single-crystal X-ray experiments.
The high-throughput crystallization screen for soluble proteins is made up of cocktails that are prepared in-house, ready-made commercial screens, and in-house-modified commercial screens22. The cocktails were initially developed using the incomplete factorial strategy using previously successful crystallization cocktails20. The reagents in the screen that are commercially available include arrays of polymers, crystallization salts, PEG, and ion combinations and screens that utilize sparse matrix and incomplete factorial approaches. There are also reagents that are modified before inclusion in the screen: an additive screen, a pH and buffer screen, an ionic liquid additive screen, and a polymer screen.
The power of known crystallization conditions and strategies has been leveraged in the 1,536 crystallization cocktails, along with the benefits of the microbatch-under-oil system to generate a pipeline that employs automated liquid handling, automated brightfield imaging, and second order nonlinear imaging of chiral crystals (SONICC). The automation of both the liquid handling and imaging provides the benefits of fewer wet lab hours and higher reproducibility. The high-throughput nature of automated crystallization screening necessitates the automation of the process of monitoring for crystal growth. These advances are achieved with state-of-the-art imaging technologies to assist in the identification of positive crystal hits. Both standard brightfield imaging of plates, as well as multi-photon methods for enhanced detection, are used via a crystal imaging system with SONICC (Figure 2). SONICC combines second harmonic generation (SHG)29 microscopy and ultraviolet two-photon excited fluorescence (UV-TPEF)30 microscopy to detect very small crystals, as well as those obscured by precipitate. The SONICC imaging informs on whether the wells contain protein (via UV-TPEF) and crystals (via SHG). Beyond the positive identification of protein crystals, additional information can also be obtained using state-of-the-art imaging methods. Cocktail-only imaging prior to sample addition serves as a negative control; these images can identify the well appearance prior to sample addition, including in terms of salt crystals and debris. Additionally, SHG and UV-TPEF imaging help differentiate protein crystals from salt crystals and can be used for visualizing protein-nucleic acid complexed material31.
High-throughput crystallization experiments undergoing repeated monitoring via imaging result in a very large volume of images needing examination. Automated crystal scoring methods have been developed to reduce the burden on the user and increase the probability of identifying positive crystal hits. The HTX Center partcipated in the development of the MAchine Recognition of Crystallization Outcomes (MARCO) scoring algorithm, a trained deep convolutional neural network architecture developed by a consortium of academic, non-profit, government, and industry partners to classify brightfield well images32. The algorithm was trained on nearly half a million brightfield images from crystallization experiments from multiple institutions using different crystallization methods and different imagers. The algorithm outputs a probabilistic score indicating whether a given image falls into four possible image classes: "crystal", "clear", "precipitate", and "other". MARCO has a reported classification accuracy of 94.5%. Crystal detection is further enhanced with software that implements the algorithm and provides a graphical user interface (GUI) for accessible and simple image viewing, enabled with the AI-enabled scoring capabilities32,33. The MARCO Polo GUI is designed to work seamlessly with the setup of the imaging and data management system in the HTX Center to identify hits in the 1,536-well screen, with human engagement to examine the output of sorted lists. Additionally, as open-source software available on GitHub, the GUI is readily available for modification to reflect the specific needs of other laboratory groups.
Here, the process of setting up a high-throughput microbatch-under-oil experiment using robotic liquid handling to deliver both the cocktail and protein is described. The HTX Center has a unique array of instrumentation and resources that are not found at other institutions, with the goal of providing screening services and educational resources to interested users. Demonstrating the methods and capabilities of robotics-enabled high-throughput techniques will enable the community to have knowledge of available technologies and make decisions for their own structure determination efforts.
1. Preparation or purchase of cocktails for sixteen 96-well deep well blocks
2. Dispensing the cocktails to 384-well plates
3. Preparing the 1,536-well plates with oil and crystallization cocktails
4. Sample submission
5. Sample setup in the prepared 1,536-well plates
6. Monitor 1,536-well plates for crystal formation
7. Image analysis
The outcomes of the 1,536-well crystal screening experiment consist of seven complete brightfield image sets collected at day 0 (negative control), day 1, week 1, week 2, week 3, week 4, and week 6 (Figure 4). SONICC images are collected at the 4 week time point for plates incubated at 23 °C and at the 6 week time point for plates incubated at 4 °C or 14 °C. Altogether, once a sample has been shipped, users can anticipate having their plates set up within 1 day of arrival. The images will be uploaded as they are collected. The crystallization screening experiment concludes after 6 weeks.
The 1,536-well plate setup allows all the screening experiments to be conducted within the same plate, thus limiting sample consumption and facilitating imaging and direct comparison between imaging modalities. Representative results for the time course of crystal growth for a single cocktail condition are shown in Figure 4. Automated plate imaging throughout the course of the experiment allows the identification of both rapidly and slowly growing crystals by brightfield imaging. The UV-TPEF and SHG imaging allow cross-validation of the hits observed by brightfield imaging and indicate that the crystals observed are proteinaceous and crystalline, respectively (Figure 5A,B). Furthermore, SONICC imaging enables the identification of crystals that are visually obscured by precipitate or films (Figure 5C) or microcrystals that may otherwise be mistaken for precipitate (Figure 5D). For some crystals, a lack of SHG signal is not disqualifying, as some point groups do not produce an SHG signal35,36, as exemplified by the tetragonal thaumatin crystal in Figure 5C. Conversely, a lack of UV-TPEF signal for proteins lacking tryptophan residues should be anticipated. The observation of UV-TPEF and SHG signals also facilitates the identification of non-protein salt crystals, which will appear in brightfield and exhibit a strong positive SHG signal but will lack a UV-TPEF signal (Figure 5E).
Image analysis for the plate setup is streamlined with the MARCO Polo GUI, which also bundles the ftp data transfer from the HWI servers (as an alternative to transferring files with FileZilla). The MARCO Polo GUI allows for easily navigable plate and image viewing and performs computational image scoring using the MARCO algorithm so that the image results can be rapidly downloaded, viewed, and analyzed from the HTX Center. The MARCO scoring algorithm, as implemented in the MARCO Polo GUI, is capable of scoring images from the entire 1,536-well plate in less than 5 min. Images flagged as crystalline by the MARCO algorithm can be subsequently sorted by the Polo GUI for display. Since the MARCO algorithm was optimized for crystal identification and minimizing false negatives so as not to miss any positive hits, the scoring can result in false positive flags. Nevertheless, the ability of MARCO to limit the set of images needing to be examined by focusing attention on the wells with a high probability of containing crystals results in a substantial reduction in data processing burden for users. The convenient implementation of the algorithm in the user-friendly MARCO Polo viewing platform, with its ability to sort images based on MARCO scores, greatly improves the user's ability to analyze the dataset quickly and to accurately determine crystal hits.
Figure 1: Schematic of a high-throughput 1,536-well crystallization screening experiment performed at the HTX Center. (1) In this step, 5 µL of paraffin oil and 200 nL of cocktail are added to each well (protocol step 3.1 and step 3.5). A cartoon illustration of one well containing only oil and cocktail and a representative image are shown to the right. (2) Samples arrive at the HTX Center (protocol step 5.1). 3) In this step, 200 nL of sample is added to each well (protocol step 5.4). (4) All 1,536 wells are monitored over time using brightfield imaging, 5) as well as the UV-TPEF and SHG modalities (protocol step 6). 6) The AI-enabled open-source GUI is used to view, score, and analyze the crystallization images (protocol step 7). Abbreviations: HTX = high-throughput crystallization; UV-TPEF = UV-two-photon excited fluorescence; SHG = second harmonic generation; AI = artificial intelligence; GUI = graphical user interface. Please click here to view a larger version of this figure.
Figure 2: Single 1,536-well plates containing screening experiments, imaged using brightfield, UV-TPEF, and SHG imaging. The 1,536-well plates are shown with an American penny for scale (top). Each screening experiment is imaged once prior to setup and six times after sample addition with brightfield imaging (seven total brightfield image sets, left). The plates undergo UV-TPEF (center) and SHG (right) imaging at 4 weeks or 6 weeks. Abbreviations: UV-TPEF = UV-two-photon excited fluorescence; SHG = second harmonic generation. Please click here to view a larger version of this figure.
Figure 3: Schematic showing how the 1,536-well plates are generated. Sixteen 96-well DW blocks are used to stamp out four 384-well plates, with each quadrant of each 384-well plate filled by dispensing crystallization cocktails. Four 96-well DW blocks fill one 384-well plate (middle). Four 384-well plates are used to stamp out the single 1,536-well plate (right). Abbreviation: DW = deep well. Please click here to view a larger version of this figure.
Figure 4: Representative time course of a single well in a 1,536-well screening experiment. Plates are imaged prior to sample setup (day 0), as well as with brightfield imaging on day 1, week 1, week 2, week 3, week 4, and week 6. The plates incubated at 23 °C are imaged with SONICC at week 4. Scale bars = 80 µm (brightfield), 200 µm (SHG, UV-TPEF). Abbreviations: SONICC = second order nonlinear imaging of chiral crystals; UV-TPEF = UV-two-photon excited fluorescence; SHG = second harmonic generation. Please click here to view a larger version of this figure.
Figure 5: Representative imaging results for the HT 1,536 crystal screening experiments. Brightfield, UV-TPEF, and SHG imaging results are shown for five example wells. (A,B) Protein crystals observed by brightfield, UV-TPEF, and SHG imaging are clearly apparent in all three imaging modalities. (C) A protein crystal obscured by film in brightfield imaging is visible by UV-TPEF imaging; the crystal is not observed by SHG imaging due to point group incompatibility. (D) Example of microcrystals verified by UV-TPEF and SHG imaging that may otherwise be considered precipitate. (E) Example of salt crystals that appear crystalline by brightfield and SHG imaging but do not exhibit a UV-TPEF signal. Scale bars = 200 µm. Well diameter = 0.9 mm. Abbreviations: UV-TPEF = UV-two-photon excited fluorescence; SHG = second harmonic generation. Please click here to view a larger version of this figure.
Supplementary Figure S1: Opening image files in MARCO Polo. Image files can be opened within the MARCO Polo GUI by navigating to the Import | Images tab at the top (a). Note that files can also be transferred via the From FTP tool directly in MARCO Polo (a) or can be transferred via FileZilla as described in protocol step 7.2. To import files that have already been downloaded, select Images | From Rar Archive/Directory. In the popup window that appears, select Browse for Folder (b), and navigate to the file directory where the plate image files are saved. Once the files are in the Selected Paths window (c), highlight a file, and click on Import Runs (d). The MARCO Polo GUI will identify the correct Cocktail File metadata to import with the images. Please click here to download this File.
The method describes a high-throughput pipeline for protein crystallization screening that requires as little as 500 µL of sample for 1,536 individual crystallization experiments in the microbatch-under-oil format. The pipeline relies on liquid-handling robotics to rapidly and reproducibly aid the experimental setup, as well as the computational image analysis resource MARCO Polo, which is customized to analyze 1,536-well plate images using the MARCO algorithm to identify and isolate crystal hits.
The small volume of individual screening drops (400 nL total with a 1:1 ratio of sample:cocktail) means that extremely small sample volumes are required to identify positive crystallization conditions. These small drop sizes necessarily produce small crystals that cannot be fished by traditional looping. Methods have been developed to harvest from the 1,536 plates37; additionally, the plates with crystals have been used directly at synchrotron sources for in situ data collection38. If a robust method for harvesting these crystals were developed, advances in synchrotron technology and micro-focused beams would further enable useful datasets to be obtained. Additionally, the crystals obtained could potentially be used as seeds for optimization efforts.
SONICC imaging is clearly advantageous in identifying both small protein crystals and protein crystals hidden beneath precipitate. Despite these advantages, not all sample types are amenable to SHG and UV-TPEF imaging. For example, proteins with few or no aromatic tryptophan residues will show an ambiguous UV-TPEF signal. Furthermore, crystals in specific space groups, including centrosymmetric groups or point group 432, will be undetected by SHG imaging. Samples with fluorophores sometimes interfere with the SHG signal, resulting in the cancellation of the signal or increased intensity, meaning careful interpretation of SHG signals is required for metal-containing proteins and proteins containing fluorescent moieties. However, in many cases, it is possible to rationalize the absence of an SHG or UV-TPEF signal, and the lack of these signals should not necessarily rule out the presence of a protein crystal.
The microbatch-under-oil format provides an alternative to the more common vapor diffusion method used for high-throughput crystallography. Importantly, the crystallization format impacts hit identification39, which provides a rationale for the use of different crystallization formats for high-throughput screening efforts. Automated imaging and SONICC-enabled modalities aid in the rapid identification of protein crystals throughout the 6 week experimental time course. Finally, the MARCO Polo GUI enables users to rapidly analyze images from 1,536 conditions to identify promising hit wells for optimization. The capabilities at the HTX Center, including the robotics-enabled high-throughput experimental setup, coupled with the state-of the-art imaging and computational tools for analyses, provide a major contribution to the structural biology community by empowering researchers to effectively address a primary bottleneck in crystal-based structural work: finding crystallization conditions.
The authors have no competing financial interests or other conflicts of interest.
We would like to extend our gratitude to our users for entrusting their precious samples to us for crystal screening, as well as for providing critical feedback and requests that have helped us refine and develop our resources to better serve the structural biology community. We would also like to acknowledge Ethan Holleman, Dr. Lisa J Keefe, and Dr. Erica Duguid, who drove the development of the MARCO Polo GUI. We would like to thank the HWI colleagues for their support and suggestions, especially Dr. Diana CF Monteiro. We acknowledge funding support from the National Institutes of Health, R24GM141256.
|1536 Well Imp@ct LBR LoBase
|AlumaSeal II Sealing Film
|Ammonium phosphate dibasic
|Ammonium phosphate monobasic
|Bicine pH 9.0
|Bis-tris propane pH 7.0
|Calcium chloride dihydrate
|CAPS pH 10.0
|Cobalt sulfate heptahydrate
|Crystal Screen HT screen
|Gryphon liquid handling robot
|Art Robbins Instruments
|HEPES pH 7.0
|HEPES pH 7.5
|HWI HTX Center sample submission form
|Index HT screen
|Ionic Liquid screen
|Lithium sulfate monohydrate
|Magnesium acetate tetrahydrate
|Magnesium chloride hexahydrate
|Magnesium nitrate hexahydrate
|Magnesium sulfate heptahydrate
|Manganese chloride tetrahydrate
|Manganese sulfate monohydrate
|MARCO Polo GUI download
|Matrix Platemate 2 x 3 liquid handling robot
|MES pH 6.0
|Mosquito liquid handling robot
|Paraffin Oil/White Mineral Oil Saybolt Viscosity 340-365 at 100 °F
|PEG/Ion HT screen
|PEGRx HT screen
|Potassium phosphate dibasic
|Rock Imager 1000 with SONICC
|Rock Imager 54
|SaltRx HT screen
|Silver Bullets screen
|Slice pH screen
|Sodium acetate pH 5.0
|Sodium citrate pH 4.2
|Sodium citrate pH 5.6
|Sodium molybdate dihydrate
|Sodium phosphate monobasic
|Sodium thiosulfate pentahydrate
|StockOptions Polymer screen
|Tacsimate pH 7
|TAPS pH 9.0
|Tris pH 8
|Tris pH 8.5
|ViaFLO 384 384 channel pipettor head (0.5-12.5µL)
|ViaFLO 384 96 channel pipettor head (300µL)
|Zinc acetate dihydrate
Copyright © 2024 MyJoVE Corporation. All rights reserved