Article Text

Download PDFPDF

1277 Self-supervised deep learning enables label-free high-dimensional morphology profiling of immune cell types
  1. Stephane Boutet,
  2. Kiran Saini,
  3. Senzeyu Zhang,
  4. Ryan Carelli,
  5. Kevin B Jacobs,
  6. Amy Wong-Thai,
  7. Vivian Lu,
  8. Andreja Jovic,
  9. Maddison Masaeli and
  10. Anastasia Mavropoulos
  1. Deepcell, Inc, Menlo Park, CA, USA
  • Journal for ImmunoTherapy of Cancer (JITC) preprint. The copyright holder for this preprint are the authors/funders, who have granted JITC permission to display the preprint. All rights reserved. No reuse allowed without permission.


Background Morphology is a fundamental cell property associated with cell identity, state, function, and disease. Current methods to quantify morphology lack high discriminative power and are generally destructive to cells, limiting data that can be extracted from samples. Morphology as a quantitative, high-dimensional analyte (‘morpholome’) can provide a more comprehensive understanding of cell biology, particularly when integrated with other multi-omics approaches. We applied deep learning and computer vision to characterize and sort label-free cells based on real-time interpretation of high-content morphology data.

Methods The REM-I platform combines microfluidics, high-speed imaging, deep learning models, and data analysis/visualization software. Data is generated by loading single cell suspensions of human cells, such as patient-derived immune cells and dissociated tissues, into the system. As the label-free cells flow through the microfluidic channel, high-resolution (0.16 µm/pixel) brightfield images are captured and sent to deep learning models for characterization and identification in real-time. The current model iteration, termed ‘Human Foundation Model’ (HFM) combines deep learning with computer vision trained and validated on a wide range of human cells. Cell types that can be classified and characterized by the HFM include immune populations such as peripheral blood mononuclear cells (PBMCs), quiescent/activated T cells, monocyte-derived macrophages, hematopoietic stem cells (HSCs), and erythroid progenitors for broad applicability.

Results The hybrid self-supervised deep learning model produces embeddings which are reproducible, quantitative, and high-dimensional descriptions that distinguish individual cells from one another. Once cell images are processed by the HFM, deep learning embeddings can be analyzed through interactive UMAPs, population analytics, and image visualizations. Using these features, investigators can define cell populations of interest and reproducibly use select features for subsequent experiments or sort them into six wells for further multi-omic and functional analysis. High-dimensional profiling of different immune cells identified distinct morphotypes associated with subpopulation of PBMCs, HSCs, and other immune cells. Remarkably, results showed clearly distinguished profiles between several comparison experiments, including quiescent and activated T cells, monocytes and macrophages, as well as bone marrow derived CD34+ HSCs, cord blood CD34+ HSCs, and CD36+ erythroid progenitors.

Conclusions These proof-of-concept experiments, including profiling tumor heterogeneity, enriching carcinoma cells in malignant body fluids, and analyzing immune cell activation/differentiation states, revealed associations between morphological fingerprints and cell state, function, and identity. Thus, the REM-I platform can help usher in quantitative morphology as a new biological readout by revealing morphology-based information critical for biological discoveries through the use of AI.

This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.