Sparse mode analysis

Analyzing data generated with sparse mode is nearly identical to the dense mode analysis steps, with a few key differences:

  1. The machine learning model file must be provided

  2. The prediction step is performed before calculating synergy

Gather files

Gather the required metadata file and raw file(s), we recommend putting them into separate folders for simplicity:

  • meta 📂

    • Contains the single metadata .csv file

  • raw 📂

    • Contains the raw file(s) in 1536-well format (and nothing else)

Nothing else should be in these folders at the start of analysis.

Deployable ensemble model

Combocat’s ensemble machine learning model is packaged (serialized) and ready to be deployed. As new iterations of models are released, they will be available in the repository’s models folder.

You can read the model file directly into R from the repository:

model_link <- "https://raw.githubusercontent.com/wcwr/combocat/main/models/m03_Feb_2026_02.23PM_Train500_Test167_TuneT.RDS"

my_model <- 
  readRDS(url(model_link))
Note

Check to make sure this is the latest/desired model before loading

Model nomenclature:

The model is a single file and is named according to specific nomenclature which can tell us some useful information. For example, the model m10_Jun_2024_05.00PM_Train552_Test184_TuneT.RDS, can tell us the following:

  • “m” : prepended to all models so files don’t start with numbers

  • <date> the model was created

  • <time> the model was created

  • Train<number> : Number of samples that were used in the training set

  • Test<number>: Number of samples that were used in the test set

  • Tune<T/F>: Indicates if hyperparameter tuning of each model was used

The number of training/test set samples also indicates their proportion and the total sample size. 552 + 184 = 736 total samples, indicating a 75/25 split (the default in cc_buildModel)

Quick start

The entire sparse mode analysis can be run in the following single block.

Note

You can try this code yourself by downloading the sparse mode example from GitHub.

#Map, normalize, predict
normData <- 
  cc_map(meta_sparse, getwd(), save_raw_plate_heatmaps = TRUE) %>%
  cc_norm(.) %>%
  cc_predict(my_model, .) #Use ensemble model to predict non-measured values

#Fit DR models
drData <- 
  cc_getDR(normData)

#Calculate synergy, QC
synData <- 
  cc_getSyn(normData, drData) %>%
  cc_getQC(., drData)

#Return/save results
cc_report(
  synData,   
  drData,   
  cd_plots    = cc_plotMat(synData, "perc_cell_death", color_midpoint = 50),
  syn_plots   = cc_plotMat(synData, "bliss_synergy",   color_midpoint = 20),
  extra_plots = cc_plotExtras(synData,                 color_midpoint = 20))
Caution

For larger datasets, these steps should be broken up to avoid running out of resources