Reconstructing cell cycle and disease progression using deep learning

Philipp Eulenberg, Niklas Köhler, Thomas Blasi, Andrew Filby, Anne E. Carpenter, Paul Rees, Fabian J. Theis, F. Alexander WolfNature Communications 2017 · OpenReview

Abstract

We show that deep convolutional neural networks combined with nonlinear dimension reduction enable reconstructing biological processes based on raw image data. We demonstrate this by reconstructing the cell cycle of Jurkat cells and disease progression in diabetic retinopathy. In further analysis of Jurkat cells, we detect and separate a subpopulation of dead cells in an unsupervised manner and, in classifying discrete cell cycle stages, we reach a sixfold reduction in error rate compared to a recent approach based on boosting on image features. In contrast to previous methods, deep learning based predictions are fast enough for on-the-fly analysis in an imaging flow cytometer. The interpretation of information-rich, high-throughput single-cell data is a challenge requiring sophisticated computational tools. Here the authors demonstrate a deep convolutional neural network that can classify cell cycle status on-the-fly.

Introduction

Biological processes, such as the cell cycle or the progression of a disease, are inherently continuous. However, experimental datasets often consist of discrete, coarse-grained labels. Traditionally, reconstructing these continuous trajectories—often referred to as "pseudotime"—has relied on specific molecular markers or manual feature engineering. These methods are frequently invasive, expensive, or limited by human bias, potentially missing subtle morphological cues essential for accurate mapping. This study explores whether deep learning can learn these continuous biological manifolds directly from raw images using only discrete labels.

Methods: The DeepFlow Architecture

The researchers developed DeepFlow, a 42-layer convolutional neural network (CNN) based on the Inception architecture. It features 13 "dual-path" modules designed to process visual information at multiple scales.

Schematic of the DeepFlow workflow, from raw image input to feature extraction and final classification/visualization.

The workflow involves feeding raw images into the CNN, which is trained to classify them into discrete categories. The critical innovation lies in extracting activations from the network's final layer and applying t-SNE (t-distributed Stochastic Neighbor Embedding). This nonlinear dimension reduction projects high-dimensional features into a lower-dimensional space, revealing the continuous relationships the network perceives between cells.

Datasets

The approach was validated on two distinct datasets:

  1. Jurkat Cells: 32,266 human T-lymphocyte images from imaging flow cytometry. While ground truth labels for seven cell cycle stages (G1, S, G2, Prophase, Metaphase, Anaphase, Telophase) were established via fluorescent stains, the model was trained primarily on raw Brightfield and Darkfield images.
  2. Diabetic Retinopathy (DR): 30,000 color fundus photographs classified into four states (healthy to severe), demonstrating the model's applicability to medical diagnostics.

Key Insights

  • Continuity from Discreteness: A network trained on discrete "buckets" naturally learns the continuum between them. Cells at the transition between phases share morphological similarities that the network's feature space organizes chronologically.
  • Unsupervised Discovery: The t-SNE visualization of Jurkat cells revealed a distinct, separate cluster. These were identified as damaged or dead cells, which the network isolated without being explicitly trained on a "dead cell" label.
  • Implicit Segmentation: Intermediate layers showed that the network autonomously learned to perform image segmentation to improve classification accuracy.
  • Real-time Processing: The CNN's forward pass is efficient enough (~1000 cells/second) for on-the-fly sorting in imaging flow cytometers.

Results

The DeepFlow model achieved a 98.73% accuracy in Jurkat cell cycle classification, representing a sixfold reduction in error compared to previous state-of-the-art boosting methods (92.35%).

Reconstructed cell cycle trajectory (a), unsupervised discovery of dead cells (b, arrow), and representative images of healthy (c) vs. damaged (d) cells.

The reconstructed trajectories in the latent space showed a clear chronological ordering. In the diabetic retinopathy data, the model correctly ordered disease states and identified a "density gap" between healthy and mild stages, suggesting a significant morphological shift at the onset of the disease.

Conclusion and Future Work

While highly effective, the model requires large labeled datasets and continuous sampling to avoid artificial gaps in the latent space. Future research could apply this framework to higher-resolution microscopy or multi-spectral data. Despite these challenges, DeepFlow represents a significant step toward automated, unbiased, and real-time analysis of complex biological processes.

Summary updated

Abstract

We show that deep convolutional neural networks combined with nonlinear dimension reduction enable reconstructing biological processes based on raw image data. We demonstrate this by reconstructing the cell cycle of Jurkat cells and disease progression in diabetic retinopathy. In further analysis of Jurkat cells, we detect and separate a subpopulation of dead cells in an unsupervised manner and, in classifying discrete cell cycle stages, we reach a sixfold reduction in error rate compared to a recent approach based on boosting on image features. In contrast to previous methods, deep learning based predictions are fast enough for on-the-fly analysis in an imaging flow cytometer. The interpretation of information-rich, high-throughput single-cell data is a challenge requiring sophisticated computational tools. Here the authors demonstrate a deep convolutional neural network that can classify cell cycle status on-the-fly.
BibTeX
@inproceedings{eulenberg2017reconstructing,
  author = {Philipp Eulenberg and Niklas Köhler and Thomas Blasi and Andrew Filby and Anne E. Carpenter and Paul Rees and Fabian J. Theis and F. Alexander Wolf},
  title = {{Reconstructing cell cycle and disease progression using deep learning}},
  booktitle = {Nature Communications},
  year = {2017},
  url = {https://openreview.net/forum?id=BYj5Vry5mj},
}
Figure