Weakly Supervised Learning of Single-Cell Feature Embeddings.

Area of Science:

• Computational biology
• Biomedical imaging
• Machine learning

Background:

• Accurate cell morphology representation is crucial for drug discovery and functional genomics.
• Deep convolutional neural networks (CNNs) offer powerful visual representation learning but typically require extensive labeled data, which is scarce in biomedical profiling.
• Existing methods often struggle with the nuances of cellular morphology in complex experimental conditions.

Purpose of the Study:

• To develop a weakly supervised method for learning single-cell representations from microscopy images.
• To enable the discovery of biological relationships based on cellular morphology without precise ground truth labels.
• To improve the comprehensive capture of individual cell morphology for advanced biological applications.

Main Methods:

• Training CNNs using a weakly supervised approach that leverages the assumption that cells from the same experimental condition exhibit similar morphologies.
• Employing regularization techniques such as mixup or recurrent neural networks (RNNs) to mitigate unwanted variations during training.
• Conducting experiments on two distinct biological datasets to validate the proposed methodology.

Main Results:

• The proposed weakly supervised approach generated single-cell embeddings that demonstrated higher accuracy compared to classical feature extraction methods.
• The learned embeddings were found to be competitive with established transfer learning techniques in the field.
• The method effectively captured relevant morphological features for biological interpretation.

Conclusions:

• Weakly supervised learning is a viable and effective strategy for cell representation learning in microscopy images, especially when ground truth labels are unavailable.
• The developed method enhances the utility of CNNs for biomedical profiling, facilitating deeper insights into experimental conditions.
• This approach offers a promising direction for advancing drug discovery and functional genomics through improved analysis of cellular morphology.