Cell growth and division are two significant areas of research in the field of cell physiology. Of late, there have been several developments in microfluidic devices that enable more precise and effective tracking of microbial cells and various other physiological parameters such as cell sizes.
Simple mathematical models have been developed and tested based on their capabilities to recreate the observed information to elucidate laws of cell growth and characteristics such as cell size. However, models have failed to record specific data characteristics allegedly due to presumed assumptions or simplification, especially when the data are multidimensional.
Furthermore, bio-data has considerable variability, which can only be explained in part by a basic model. A simple distribution usually represents the unexplained part as noise. But, again, the assumed distribution form may not always be correct when unclarified components are represented. This shows that developing a more versatile way to clarify the laws of the data is needed. Such an approach should also process and integrate multidimensional noisy data and have a flexible representation power for both deterministic and noise distribution relationships.
Now researchers from the University of Tokyo’s Institute of Industrial Science have developed a machine learning system to estimate cell size as it grows and divides. The computer algorithm creates more detailed and accurate forecasts by employing an artificial neural network that avoids making assumptions that are frequently used in biology.
ML for cell size estimation
In the last decade, machine learning (ML) methods have been prodigiously developed and applied to a wide variety of problems[18]. ML methods, especially deep learning (DL), have demonstrated that they can semi-automatically extract complicated patterns underlying large amounts of noisy data sets by removing irrelevant dimensions and unexplainable noise factors. Even though ML, in reality, is not at all the magic wand that can solve problems without any human help or preconceived assumptions, we can use it to support us for searching a defined but huge model space without relying on our insights and intuition.
Many experiments in the past have shown that machine learning, especially deep learning methods can semi-automatically extract complicated patterns underlying a large number of data sets. These methods generally achieve this by removing irrelevant dimensions and other noise factors.
With this underlying principle, the authors of the study employed a neural network to the problem of cell size regulation. Following the formulation used to model cell size control as a stochastic process, the researchers represented the cell size dynamics (interrupted by division events) using the temporal point process.
This research introduces a neural network (NN) method to represent the process’s intensity function that depends on the history of cell size over retrospective lineage. The team was able to obtain conditional probability density functions (PDFs) of cell sizes at both birth and division by training these NNs. The trained PDFs reproduce and confirm the previous results and presumed assumptions of size control.
To evaluate when cells divide, biologists typically utilise a “sizer” model based on the cell’s absolute size and an“adder” model based on the cell size increase since birth. The NN model here supports the “adder” principle, but only as part of complicated metabolic events and communications networks. The results demonstrate that the inferred NN model is more advantageous than standard descriptive statistics for data representation for model searching. In short, the NN approach can be an effective tool for uncovering latent dynamic laws from noisy data using NN methods.
Wrapping up
This research could lead to discoveries in quantitative biology and could also positively impact the manufacturing of pharmaceuticals or fermented foods.
In future, the team wants to extend the NN model architecture to incorporate cell size, division interval and other variables like gene expression as a multidimensional history.
Access the complete research here
