A team of researchers have created a convolutional neural network (CNN) that is able to accurately classify cell death and potentially benefit drug screening trials.
Deep learning (DL) models are a class of machine learning whereby artificial neural networks, inspired by the human brain, learn from large amounts of data. A convolution neural network (CNN) is a DL algorithm that can analyse an input image, assigning importance to objects and differentiating various aspects from each other. This allows them to classify large sets of complex information and act as one of the most prominent methods for image recognition. CNN techniques are growing in popularity within the biomedical field, particularly for the detection of morphological changes.
Cell death plays an important role in many human diseases. Therefore, several strategies aimed at modulating cell death and its associated pathways have been applied to treat various disorders, such as cancer and neurodegenerative diseases. It is unsurprising that cell death experiments are key for the discovery and development of numerous drugs. Typically, each study analyses features of a dying cell, such as DNA fragmentation, membrane protein flipping or protein modifications. However, performing the individual aspects of these assays requires time and money.
Training CNNs to classify cells undergoing cell death
Researchers have recently developed a CNN for fast and accurate classification of cell death in cultures using transmitted light microscopy (TLM) images. TLM is a technique employed to distinguish the morphological characteristics of a sample.
The team incubated pluripotent stem cell lines and cancer cell lines with camptothecin (CPT), a topoisomerase inhibitor that induces very rapid cell death in human stems cells. TLM images were then taken hourly at different CPT concentrations. After one hour, it was found that the trained CNN was able to correctly classify around 98% of the images based on cell death features. This validation accuracy increased further after exposure to CPT for two and three hours, probably due to the drug-associated effect becoming more pronounced.
It was found that the CNN predications largely outperformed human ability. After one hour of cell death induction, experienced scientists were unable to correctly identify images showing slight morphological changes in cell lines caused by CPT. Even after the researchers were ‘trained’ by studying 500 labelled images of cell death, their answers were no more accurate than before. This provides evidence for the benefit of using CNNs instead of human eyes when identifying subtle features of dying cells.
Furthermore, the CNN was partially able to classify images from other unseen cell lines with an accuracy of 75%. Although this feature may be useful for classification during drug validation in cell lines that have not been seen before, it reinforces the fact that training the model on certain cell lines is preferrable.
Cell death experiments for drug development
This research showed that CNNs can be trained to recognise extremely early features of cell death. It has shown that using DL methods to recognise features in TLM images can be highly accurate, which may be a significant benefit in everyday laboratory practise – it was clarified that any researcher could obtain similar results using the CNN on their own images, even with a limited understanding of DL.
Therefore, the novel tool has the potential to be used in the majority of cell death experiments. Perhaps, it will be particularly useful in those that require huge and repetitive experimental settings, such as drug screening in cancer research. Nevertheless, the use of these data-driven techniques alongside automation will certainly increase reproducibility and reduce costs of the drug development process in the future.
Image credit: FreePik starline