emanuelsoda / fucci_ml

Automated workflow for the cell cycle analysis of (non-)adherent cells using a machine learning approach

Kourosh Hayatigolkhatmi, Chiara Soriani, Emanuel Soda, Elena Ceccacci, Oualid El Menna, Sebastiano Peri, Ivan Negrelli, Giacomo Bertolini, Gian Martino Franchi, Roberta Carbone, Saverio Minucci, Simona Rodighiero

Pipeline overview

0. Model Creation

The pipeline is created using the tidymodels framework. The random forest model trained and stored as docker container can be found here .

The model stored as .rds file can be found here.

NB: If the provided model does not yield satisfactory results for your data, we recommend provided pipeline to train a new model tailored to your specific dataset. While this process involves some data annotation, tidymodels offers a variety of models that are not data-intensive and perform well on a low amount of data.

1. Inference

To extract the time series feature the function timetk_feature_extraction(). We suggest to have a look to the official time_tk page.

After training the model, making predictions is straightforward. Simply load the trained model and utilize either the parsnip::predict() function or, alternatively, parsnip::augment(). The latter not only provides predictions but also includes the associated predicted probabilities.

rf_model_quality <- 
  readr::read_rds("models/random_forest_model.rds")
  
parsnip::augment(rf_model_quality, tsfeature_tbl)

2. Phase assignment

The normalized HUE intensity value for cycling cells serves as a basis for assigning the cell cycle phase according to the FUCCI construct. Various threshold values can be employed, but based on empirical testing, we recommend considering the following threshold:

Phase	Lower threshold	Upper threshold
G1	0	0.65
G2/M	0.65	0.85
S	0.85	1

Assigning the cell cycle phase to the tracks that have passed the filter is straightforward, involving the addition of a new column to the table. This can be achieved, for example, using dplyr::mutate() and dplyr::case_when() as follows:

data_table <- data_table %>%  
  dplyr::mutate(phase = dplyr::case_when(
    HUE >= 0 & HUE < 0.65 ~ "G1",
    HUE >= 0.65 & HUE <= 0.85 ~ "G2/M",
    HUE >= 0.85 ~ "S"
  ))

3. Result

The filtered tracks offer a valuable avenue for assessing differences among treatments. Specifically, by enumerating the number of frames each cell persists in a given phase and then dividing by the number of acquisitions per hour, we can compute the time spent (in hours) in each phase for each experimental condition. The subsequent plots depict the time spent in each phase for individual cells across various conditions