Deep Neural Networks for Analysis of Microscopy Images—Synthetic Data Generation and Adaptive Sampling

Applications of Machine Learning to the Study of Crystalline Materials

This Special Issue, "Applications of Machine Learning to the Study of Crystalline Materials", is a collection […]

Machine learning-based protein crystal detection for monitoring of crystallization processes enabled with large-scale synthetic data sets of photorealistic images

Since preparative chromatography is a sustainability challenge due to large amounts of consumables used in downstream processing of biomolecules, protein crystallization offers a promising alternative as a purification method. While the limited crystallizability of proteins often restricts a broad application of crystallization as a purification method, advances in molecular biology, as well as computational methods are pushing the applicability towards integration in biotechnological downstream processes. However, in industrial and academic settings, monitoring protein crystallization processes non-invasively by microscopic photography and automated image evaluation remains a challenging problem. Recently, the identification of single crystal objects using deep learning has been the subject of increased attention for various model systems. However, the advancement of crystal detection using deep learning for biotechnological applications is limited: robust models obtained through supervised machine learning tasks require large-scale and high-quality data sets usually obtained in large projects through extensive manual labeling, an approach that is highly error-prone for dense systems of transparent crystals. For the first time, recent trends involving the use of synthetic data sets for supervised learning are transferred, thus generating photorealistic images of virtual protein crystals in suspension (PCS) through the use of ray tracing algorithms, accompanied by specialized data augmentations modelling experimental noise. Further, it is demonstrated that state-of-the-art models trained with the large-scale synthetic PCS data set outperform similar fine-tuned models based on the average precision metric on a validation data set, followed by experimental validation using high-resolution photomicrographs from stirred tank protein crystallization processes.