Regression plane concept for analysing continuous cellular processes with machine learning

Abel Szkalisity, Filippo Piccinini, Attila Beleon, Tamas Balassa, Istvan Gergely Varga, Ede Migh, Csaba Molnar, Lassi Paavolainen, Sanna Timonen, Indranil Banerjee, Elina Ikonen, Yohei Yamauchi, Istvan Ando, Jaakko Peltonen, Vilja Pietiäinen, Viktor Honti, Peter HorvathNature Communications 2021 · OpenReview

Abstract

Biological processes are inherently continuous, and the chance of phenotypic discovery is significantly restricted by discretising them. Using multi-parametric active regression we introduce the Regression Plane (RP), a user-friendly discovery tool enabling class-free phenotypic supervised machine learning, to describe and explore biological data in a continuous manner. First, we compare traditional classification with regression in a simulated experimental setup. Second, we use our framework to identify genes involved in regulating triglyceride levels in human cells. Subsequently, we analyse a time-lapse dataset on mitosis to demonstrate that the proposed methodology is capable of modelling complex processes at infinite resolution. Finally, we show that hemocyte differentiation in Drosophila melanogaster has continuous characteristics. High-content screening prompted the development of software enabling discrete phenotypic analysis of single cells. Here, the authors show that supervised continuous machine learning can drive novel discoveries in diverse imaging experiments and present the Regression Plane module of Advanced Cell Classifier.

Introduction: The Problem with Discrete Classification

In traditional bioimage analysis, researchers typically rely on classification to sort cells into discrete categories, such as specific stages of the cell cycle. However, biological transitions rarely occur in jumps. Forcing a cell that exists in a transitional state into a single category introduces "labeling noise" and obscures the very nuances that define biological progress. While unsupervised methods like t-SNE or UMAP can visualize clusters, they often lack the biological "ground truth" provided by expert knowledge. There is a critical need for a supervised, continuous tool that allows experts to define the biological landscape without the constraints of arbitrary class boundaries.

Methods: The Regression Plane and Active Learning

The authors propose the Regression Plane (RP), integrated into the open-source Advanced Cell Classifier (ACC) software. Instead of assigning a discrete class label, an expert "places" representative cells onto a 2D unit square (the plane). These coordinates represent biological progress or morphological shifts.

To learn the relationship between high-dimensional image features and these 2D coordinates, the system employs various models, including Random Forests and Gaussian Processes. To minimize the manual effort required for annotation, the framework utilizes Active Regression. The software identifies "uncertain" or "difficult" cells—such as those where models disagree or those falling into empty regions of the plane—and prompts the expert to label them, creating an efficient feedback loop.

The active learning loop for regression-based phenotypic discovery.

Experimental Validation and Key Insights

The RP framework was validated across four distinct scenarios, ranging from synthetic benchmarks to complex time-lapse imaging.

1. Synthetic Benchmarking

Using a synthetic dataset (SIMCEP) to simulate six continuous morphological processes, the authors compared RP against traditional classification. The RP approach significantly outperformed classification, improving precision by approximately 20% (0.88 vs 0.66). By treating phenotypes as a continuum, the model effectively handles "inter-class" cells that typically confuse standard classifiers.

Performance comparison and phenotypic mapping using the Regression Plane.

2. Lipid Droplet Regulation

In a study of human hepatocyte cells, the RP tool identified that knocking down the TM4SF5 gene promotes an increase in small lipid droplets. This finding supplements previous genome-wide association studies and demonstrates the tool's utility in gene discovery.

3. Mitosis and Differentiation

The RP was applied to mitosis time-lapse data, where it modeled the process at "infinite resolution." This allows researchers to zoom into specific sub-phases that are lost in discrete models. Furthermore, in a study of Drosophila hemocyte differentiation, the RP revealed that previously debated "Type I" and "Type II" cells are actually ends of a single size continuum, rather than distinct cell types.

Discussion and Limitations

The Regression Plane offers a powerful "expert-in-the-loop" approach, allowing biologists to steer the discovery process while machine learning handles high-dimensional data. However, the current 2D constraint is a trade-off for human interpretability; while biological processes can be higher-dimensional, 2D remains the limit for intuitive manual annotation. Additionally, the model's accuracy is dependent on the expert's initial placement, meaning it can reflect human bias. Future work may involve integrating deep learning backends to further enhance feature extraction and mapping accuracy for subtle phenotypes.

Reference: Szkalisity, A., Piccinini, F., Beleon, A. et al. Regression plane concept for analysing continuous cellular processes with machine learning. Nat Commun 12, 2532 (2021). https://doi.org/10.1038/s41467-021-22866-x

Summary updated

Abstract

Biological processes are inherently continuous, and the chance of phenotypic discovery is significantly restricted by discretising them. Using multi-parametric active regression we introduce the Regression Plane (RP), a user-friendly discovery tool enabling class-free phenotypic supervised machine learning, to describe and explore biological data in a continuous manner. First, we compare traditional classification with regression in a simulated experimental setup. Second, we use our framework to identify genes involved in regulating triglyceride levels in human cells. Subsequently, we analyse a time-lapse dataset on mitosis to demonstrate that the proposed methodology is capable of modelling complex processes at infinite resolution. Finally, we show that hemocyte differentiation in Drosophila melanogaster has continuous characteristics. High-content screening prompted the development of software enabling discrete phenotypic analysis of single cells. Here, the authors show that supervised continuous machine learning can drive novel discoveries in diverse imaging experiments and present the Regression Plane module of Advanced Cell Classifier.
BibTeX
@inproceedings{szkalisity2021regression,
  author = {Abel Szkalisity and Filippo Piccinini and Attila Beleon and Tamas Balassa and Istvan Gergely Varga and Ede Migh and Csaba Molnar and Lassi Paavolainen and Sanna Timonen and Indranil Banerjee and Elina Ikonen and Yohei Yamauchi and Istvan Ando and Jaakko Peltonen and Vilja Pietiäinen and Viktor Honti and Peter Horvath},
  title = {{Regression plane concept for analysing continuous cellular processes with machine learning}},
  booktitle = {Nature Communications},
  year = {2021},
  url = {https://openreview.net/forum?id=Y0MoZLhnV5},
}
Figure