Deepcell, a Silicon Valley startup spun out of Stanford University in 2017 with a novel way to identify cells using AI, today raised $20 million. The company says the funds will be put toward bolstering development of its technology and enabling it to drive a “hypothesis-free” approach to cell classification and sorting.
Cells have distinctive characteristics that an AI system can identify, which has applications in translational research, diagnostic testing, and therapeutics. But some AI-based cell identification techniques fail to isolate and collect label-free cells of any type, making it impossible to keep any target cells intact for subsequent classification.
That’s in contrast to Deepcell’s approach, which targets whole cells instead of cell-free DNA. Using microfluidics, high-resolution imaging sensors, and pretrained computer vision models, the platform sorts cells based on visual features, delivering insights through a granular view of biology.
Deepcell’s technology gives researchers access to cell-specific information, including RNA, epigenetics, and protein contents. The company claims it can be used to isolate virtually any type of cell, even those occurring at frequencies as low as one in a billion. Moreover, Deepcell’s machine learning algorithms add as many as 1 million cells per day to a “cell atlas” of some 400 million cells, which serves as a ground truth for cell classification.
“With its AI-powered approach, Deepcell’s technology is able to differentiate among cell types with greater accuracy than traditional cell isolation techniques that rely on antibody staining or similar methods,” cofounder and CEO Maddison Masaeli told VentureBeat via email. “The company’s AI identifies cells based on infinitesimal morphological differences that may not be visible to the human eye and continually improves through a closed-loop process in which results from each analysis are fed back into the AI to hone its performance.”
Masaeli claims Deepcell’s supervised and unsupervised algorithms lack “inherent bias” of any kind. (Supervised algorithms require labeled datasets for training, while unsupervised algorithms infer labels from unlabeled data.) Considering the body of research demonstrating computer vision algorithms’ susceptibility to bias, this seems unlikely. But Masaeli says Deepcell takes steps to detect and mitigate imbalances in all of its training datasets.
The company plans to make its platform available as a service, potentially with tools for characterizing tissue composition and enriching cell populations for further study. There is particular promise in cell enrichment, a fundamental process for research into tumor microenvironments and other groups of cells that have to be filtered out of broader populations for analysis. Deepcell’s technology could allow researchers to pull out cells based on morphological differences for a sample enriched with the desired cell type, laying the groundwork for new drugs and therapies.
“Cell morphology is a phenotype with a long history in clinical application that has to date been based on the eyes of a human expert,” Masaeli said. “Deepcell is bringing this phenotype into modern use by adding scale, interpretability, and actionability, thanks to our innovations in AI, microfluidics, and multiomics.”
Bow Capital and Andreessen Horowitz led Deepcell’s series A round announced today, with participation from 50Y, DCVC, Stanford University, and angel investors that include Google AI head Jeff Dean. This brings the company’s total raised to over $25 million.
