Artificial intelligence (AI)/machine-learning (ML) algorithms have been heavily used in data processing in various biological and clinical applications. Sensitive biological signaling can be monitored using two-photon fluorescence lifetime imaging microscopy (FLIM). Lifetime fitting, processing, and analyzing FLIM data of biological specimens can be a challenging and time-consuming affair. The recently developed Fluorescence Lifetime Redox Ratio (FLIRR) focuses on tracking metabolic changes ‘before-and-after-treatment’ in live cells using only two lifetime parameters. FLIM data produces many data parameters which are all associated with drug response in living cells. To predict drug cellular responses, we have chosen the Becker & Hickl SPCImage software to fit the lifetime images and the resultant data was used in ML analysis. With the objective of achieving even more robust statistical power predicting earliest drug effects, we developed Python software and autoencoder (AE) models to analyze the multiple biophysical FLIM parameters acquired in 2p-FLIM images of drug response in cervical cancer cells. The use of systematic hyperparameter (HP) tuning shows the variation in performance across the different models, enabling the selection of the highest performing model and HPs for repetition. Our results show that our optimized multi-parameter trained AE models can statistically outperform single FLIRR time-course analysis in discriminating earliest metabolic changes following drug treatment.
|