Quirks and details

There are additional details that are not commonly asked but it is important to retain on record. This is a compendium of those.

  • Source 1 and 9 use higher-density plates (1536 vs the standard 384)
  • Source 9’s nine TARGET2 plates lack expected positive-control activity and should be excluded from cross-site alignment; its 99 COMPOUND plates retain that activity. Public profiles show positive-control activity near chance for TARGET2 (median normalized AP: -0.013) but strong for COMPOUND (0.974); MorphEM on the original images gives -0.016 and 0.978, respectively. Evidence and code.
  • Genetic replicates were usually not randomized across well positions. Approximately 99.4% of ORF IDs and 94.9% of CRISPR IDs occur at a single well across all plates on which they appear.
  • Source 7 and 13 are the same laboratory
  • In JUMP-Target there is an InChIKey that maps to 2 different perturbations: ‘LOUPRKONTZGTKE-UHFFFAOYSA-N’ maps to both quinidine and quinine.
  • The definition of controls, especially positive controls, can be tricky: Some are hard-coded in broad_babel, based on internal knowledge that was not recorded at the time of assembling the datasets. In certain datasets, such as JUMP-ORF, there are additional types of positive controls: poscon_orf, poscon_cp (compound probe), and poscon_diverse.
  • The treatment compounds were assayed at 10 uM for all sources, except for source_7 where the compounds were assayed at 0.625 uM (the goal being to assay some of the compounds at a low concentration in addition to the higher concentration used for most of data production). The positive control compounds in compound, ORF and CRISPR plates were assayed at 5 uM. JUMP-Target-1-Compound and JUMP-Target-2-Compound plates were also assayed at 5 uM
  • Due to some plates having letters and numbers and others only numbers, be careful when loading multiple load_data_csvs. We treat all columns as strings to avoid any potential casting issue.