Not getting in too deep: A practical deep learning approach to routine crystallisation image classification.

Area of Science:

• Structural biology
• Computational chemistry
• Machine learning

Background:

• Macromolecular crystallization is crucial for drug discovery.
• Accurate classification of crystallization outcomes aids experimental optimization.
• Deep learning offers potential for automating this classification.

Purpose of the Study:

• To compare the performance of four convolutional deep-learning network architectures for classifying crystallization images.
• To develop an ensemble classifier by combining individual models.
• To assess the feasibility of using these methods with limited computational resources.

Main Methods:

• Trained ~16,000 macromolecular crystallization images using four common convolutional neural network architectures.
• Evaluated individual model performance and combined them into an ensemble classifier.
• Utilized eight distinct classes to categorize experimental outcomes.

Main Results:

• Individual deep learning models demonstrated varying classification strengths.
• The ensemble classifier achieved classification accuracy comparable to a large consortium initiative.
• The method effectively ranks experimental outcomes, providing detailed insights.

Conclusions:

• Deep learning, particularly ensemble methods, can accurately classify macromolecular crystallization outcomes with a small dataset.
• This approach facilitates automated identification of crystal formation in routine crystallography.
• The findings support further exploration of crystal formation-condition relationships for drug discovery.