To aid biological assessments, new technology called Deepcell provides a mechanism to look at the morphology of cells. The platform is an artificial intelligence foundation model trained on millions of cell images. The technology uses a combination of artificial intelligence, microfluidics and high-resolution optics.
The resultant patterns enable scientists for the first time to explore a novel multidimensional analyte.
To support the innovation, Deepcell has issued some new data sets. These data sets enable an exploration of novel high-dimensional morphology data. In particular, the technology allows scientists to easily produce high-dimensional readouts of known and novel morphology features from unlabelled cells in an ‘unbounded hypothesis approach’ (this means regional diversity depends only on time and diversification rate and increases without limit).
These sets were presented at the Advances in Genome Biology and Technology (AGBT) General Meeting in Hollywood, Florida, U.S. in February 2023.
The AI model is termed the Human Foundation Model (HFM), and it has been trained using millions of cell images. The AI trains through a form of self-supervised learning. Self-supervised learning refers to a machine learning paradigm for processing unlabelled data to obtain useful representations that can help with downstream learning tasks.
The software suite facilitates the creation of custom cell classifications and identification of morphologically similar cell groups for sorting of viable cells to enable downstream molecular or functional analysis.
This insight should help to inform researchers about discoveries across a broad variety of sample types, such as cell lines, primary body fluids, and dissociated tissue samples. Other areas of application include the characterization of complex samples, cell atlasing (the process of charts the cell types in the healthy body), cell and gene therapy development, functional screening, cancer biology, and stem cell research.
The new data sets demonstrate how the technology can characterize different cell types in a heterogeneous sample in a label-free manner. Through this, three human cancer data sets are available for exploration.
In the first data set, the Deepcell platform assessed a mixture of human melanoma cell lines and primary tumour samples to identify tumour, immune, and stromal cell populations in a label-free manner, using only morphology.
Following this, the melanoma tumour cell population data from this data set was then selected in the Deepcell software suite and re-projected to gain additional resolution into this morphologically distinct subpopulation to create a second data set. This process revealed the heterogeneity within these cells based on subtle morphological distinctions, including pigmentation, which can be difficult to identify using conventional methods.
In the final data set, the Deepcell label-free technology was used to explore the morphological diversity of immune cell populations in the lung tumour microenvironment from a variety of human dissociated tumour cell (DTC) samples.
