W0348

A high-throughput crystallization facility at LLNL is being developed in support of the TB structural genomics center. Current capacity for setup of crystallization experiments is ~3600 experiments/day. With the facility working at full capacity observation of crystallization experiments quickly becomes the most labor-intensive process step. Crystal detection is also tedious for human observers and prone to subjectivity. In an effort to address these issues, crystal recognition software, named ‘CrysFind’ has been developed in our group. Crystal detection is known to be a difficult task because of the diversity of shapes that can show in an experiment. In addition, crystals are usually low-contrast features compared to high-contrast ones like heavy precipitates. Noisy imaging system, unpredictable crystal shapes and complex background also increase the difficulty of this task. Moreover, crystallization events are rare, so that an automatic detection has to be highly discriminating. CrysFind appears to be reliable and robust in overcoming the complexities. The software was able to pick up both large and small crystals, very faint crystals hidden in partial shadow, and at the same time discards precipitates, spots, scratches and other defects in pictures taken by a economical imaging system. For a relatively large set of experiments (~700 images) CrysFind gives both false-positive false-negative rates under 4% compared to a human observer. Both the quality of the edge detection algorithm used by CrysFind and the careful analysis of shape geometry are CrysFind strengths. CrysFind will soon be incorporated into a fully automated HT crystallization system, greatly reducing time, effort, and cost of viewing crystallization experiments.