MOA Multi-label Classification Predictions

[AdeboyeML]

MOA Multi-label Classification Predictions

@gway @shntnu

• A compound profile has multiple MOAs i.e. more than one target label (multi-label classification)
• All MOAs found in just one compound are given the same MOA label called unknown moa
• Train / Test split == ~78% / ~22%
• 271 MOAs labels
• Level 4 profiles

- Above figure interpretation

• ~ 187 MOAs are found in 2 - 5 compounds
• ~ 67 MOAs are found in 6 - 15 compounds
• ~ 10 MOAs are found in 16 - 25 compounds
• ~ 5 MOAs are found in 26 - 42 compounds
• 1 MOA (unknown moa) is found in 248 compounds

Overall Model Prediction results

- Model predictions with respect to MOA distribution among the 1,398 distinct compounds

Top predicted MOAs based on Precision-Recall AUC score

- Cell painting

- L1000

- Cell painting & L1000

30 different MOAs with higher ROC-AUC score in one profiling assay than in another one

- 76 MOAs have higher ROC AUC score in Cell painting than in L1000 and Integrated Cell painting & L1000 level 4 data

- 49 MOAs have higher ROC AUC score in Integrated Cell painting & L1000 than in L1000 and Cell painting level 4 data

- 53 MOA have higher ROC-AUC score in L1000 than in Cell painting and Integrated Cell painting & L1000 level 4 data

[shntnu]

Thanks @AdeboyeML

• Please clarify how you did the train test split – was it 80/20 per MOA? (which then works out to be ~78/22) What do you do with MOAs with few compounds? How do you handle compounds with multiple MOAs
• Please plot a graph similar to the first one, reporting the "polypharmacology" i.e. number of MOAs per compound (X = number of MOAs of that a compound is annotated with; Y = number of compounds)
• How do you assign a label? Presumably not a max because that would give you only one label.

We can discuss over a call if that's easier

[AdeboyeML]

@shntnu Yes we can clarify some of the above bullet points over a zoom call.

[AdeboyeML]

@shntnu @gwaygenomics

- Polypharmacology

Number of MOAs	Number of compounds
1	1176
2	157
3	39
4	14
5	4
7	3
6	3
10	1
11	1