Rapid video-based deep learning of cognate versus non-cognate T cell-dendritic cell interactions

Priya N. Anandakumaran,Remi J. Creusot,Samuel K. Sia,Pawel Muranski,Abigail G. Ayers

doi:10.1038/s41598-021-04286-5

Priya N. Anandakumaran, Remi J. Creusot + Show 3 more

Open Access

https://doi.org/10.1038/s41598-021-04286-5

Copy DOI

Abstract

Identification of cognate interactions between antigen-specific T cells and dendritic cells (DCs) is essential to understanding immunity and tolerance, and for developing therapies for cancer and autoimmune diseases. Conventional techniques for selecting antigen-specific T cells are time-consuming and limited to pre-defined antigenic peptide sequences. Here, we demonstrate the ability to use deep learning to rapidly classify videos of antigen-specific CD8+ T cells. The trained model distinguishes distinct interaction dynamics (in motility and morphology) between cognate and non-cognate T cells and DCs over 20 to 80 min. The model classified high affinity antigen-specific CD8+ T cells from OT-I mice with an area under the curve (AUC) of 0.91, and generalized well to other types of high and low affinity CD8+ T cells. The classification accuracy achieved by the model was consistently higher than simple image analysis techniques, and conventional metrics used to differentiate between cognate and non-cognate T cells, such as speed. Also, we demonstrated that experimental addition of anti-CD40 antibodies improved model prediction. Overall, this method demonstrates the potential of video-based deep learning to rapidly classify cognate T cell-DC interactions, which may also be potentially integrated into high-throughput methods for selecting antigen-specific T cells in the future.

Highlights

Interactions between cognate T cells and antigen-presenting dendritic cells (DCs) are a critical component of the cell-mediated adaptive immune response and peripheral tolerance[1,2]
We hypothesize that a deep learning model, aided by anti-CD40 antibodies, can classify C D8+ T cells interacting with cognate DCs, and CD8+ T cells interacting with non-cognate DCs, based on their distinct interaction dynamics in terms of morphology and motility
Consistent with previous observations, we found that T cells interacting with Q4-DCs had a mean speed in between that of T cells interacting with N4 peptide-presenting DCs (N4-DCs), and T cells interacting with non-cognate DCs (Supplementary Fig. S4)

Summary

Introduction

Interactions between cognate T cells and antigen-presenting dendritic cells (DCs) are a critical component of the cell-mediated adaptive immune response and peripheral tolerance[1,2]. Machine learning-based approaches have been used for T cell classification, in which quantitative variables such as autofluorescence lifetime[11], or morphological variables[12] are extracted from images and used to classify quiescent and activated T cells, or cognate and non-cognate T cell-DC contacts, respectively. These models are both limited by the need for long T cell activation times, while the latter omits critical temporal information that can be used to improve classification accuracy (Supplementary Table S1). We hypothesize that a deep learning model, aided by anti-CD40 antibodies (aCD40), can classify C D8+ T cells interacting with cognate DCs, and CD8+ T cells interacting with non-cognate DCs, based on their distinct interaction dynamics in terms of morphology and motility

Methods

Results

Conclusion