Using Time Courses To Enrich the Information Obtained from Images of Crystallization Trials

Abstract

Visual identification of crystals from experimental setups (droplets) underpins the field of structural biology, and there is an increasing use of automation to capture records of crystallization trials  snapshots of the experiments over time  which are then used to identify crystals for subsequent analysis. Here we describe an algorithm that locates differences between images within an image time-course of a crystallization experiment. A user-accessible "front end" to this functionality is provided by tagging images, which have been identified as changed, and using an existing (commercial) viewing software package to display the tagged images. The existence of change appears to be a powerful tool to filter out images that do not need further human examination, as the rate of false negatives is low and the method is general. We identify significant change in a time sequence of crystallization images by using a process of image alignment, drop recognition, masking, filtering, and comparison. We propose that this process might be a powerful way of picking up changes in crystallization experiments which happen after weeks or months.

Exploring the in meso crystallization mechanism by characterizing the lipid mesophase microenvironment during the growth of single transmembrane α-helical peptide crystals

“…None of the lysozyme crystals co-localized with a local lamellar phase, as expected for proteins that crystallize in the aqueous phase rather than from the lipid bilayer under conditions of in meso crystallization [49,50]. Images shown here are taken with the Minstrel HTUV imaging system [38]. These crystals were used for SAXS analysis, as shown in figures 3 and 4.…”

Using textons to rank crystallization droplets by the likely presence of crystals

“…Here, we target a more realistic goal, namely to prioritize droplets for viewing: we judge this to be more useful because it does not pre-empt decisions but helps them to be made more accurately as well as more rapidly. One version of this approach is filtering, as employed, for example, in Rigaku Automation's viewing software for images captured on a UVenabled instrument, or more recently by Mele et al (2014), who filter out images from further examination based on the lack of change (differences) in a droplet over time, on the basis that such changes may indicate the formation, growth or disappearance of crystals. However, filtering is merely an extended case of classification, where instead of a single classification cutoff a tuneable cutoff or criterion is still required to hide a subset of data.…”

“…An aspect that has not been much explored in the existing literature is how to present most effectively the computed classifications, or for that matter the features obtained from alternative imaging techniques (SONICC or UV). A common method in vendor software is to label the images with tags and different colour schemes (Mele et al, 2014), and sophistication is introduced to allow users to sort by labels or to hide subsets of data.…”

Time-Dependent Multi-Light-Source Image Classification Combined With Automated Multidimensional Protein Phase Diagram Construction for Protein Phase Behavior Analysis

“…However, incorporation of information obtained during protein phase behavior experiments (time-dependent information) has shown to aid crystallization screenings. [29][30][31] To explore improvements of protein phase behavior classification, this study aims to combine time-dependent and end point features obtained from multi-light-source images and assess the impact by means of a random forest classification algorithm. A random forest classification algorithm was selected because the optimization of protein phase behavior classification via more complex algorithms was considered outside the scope of this work.…”