# Data-Driven Analysis of Fluorescence Lifetime Imaging Experiments: Unraveling the Signal/Stress Relationship of Polluted Microalgae Cells with Machine Learning

## TL;DR

This paper uses machine learning to analyze how microalgae cells respond to copper pollution by studying their fluorescence signals, revealing key features linked to cell stress.

## Contribution

The novel contribution is applying data-driven machine learning to map fluorescence signal features to stress levels in microalgae exposed to copper.

## Key findings

- Random forest and ridge regressors achieved strong predictive performance in mapping fluorescence signals to copper stress.
- Feature importance analysis identified statistical signal features and decay rates as most relevant to cell health.
- Simpler ML models with engineered features performed as well as complex neural networks but with faster training times.

## Abstract

Chlorophyll a fluorescence decay profiles of biological
cells may be used as indicators of the ability of a plant to tolerate
environmental stress and the extent of the associated damage to its
photosynthetic apparatus. However, the interpretation of data remains
often complex and sometimes controversial. Based on previously recorded
experimental data from fluorescence lifetime imaging microscopy (FLIM)
on the freshwater microalga Coccomyxa cimbrica exposed to the Cu­(II) toxic agent, in this work, we set out to investigate
the relationship between FLIM measurements and cell stress conditions
based on a data-driven approach. In particular, we analyze the changes
induced by Cu­(II) in the photosynthetic cycle of the microalga by
monitoring the decay profiles of single cells exposed to different
concentrations of Cu­(II) (0, 30, 100, 300, 500, and 700 μg mL–1) as a function of time (0, 24, 48, 72, and 96 h)
and use Machine Learning to train predictive models mapping the signal
shape to Cu­(II) dosage (defined here as the product of the concentration
of Cu­(II) and the time of exposure to it) and to gain insights into
the signal features that are more deeply connected with the cell health
status. Results show that a good tabularization of the data can lead
to acceptable predictions with several standard models, with random
forest and ridge regressors showing the best performances. Feature-importance
analysis of the forest model reveals that a few statistical features
of the fluorescence signal, in combination with its decay rate, are
the most relevant descriptors. A final analysis of the predictive
performances of more sophisticated models, including fully connected
and convolutional neural networks, confirms that careful feature engineering
coupled with simpler ML models can lead to equally good performances
in shorter training times.