Level-4 Exploratory analysis results - L1000 vs Cell painting Comparison

[AdeboyeML]

L1000 vs Cell Painting Comparison based on median correlation values from compound replicates per dose

@gwaygenomics @shntnu

• This is a recap of what I presented about 2 weeks ago during the profiling check-in meeting

- Median score scatter plot

- Median score distribution across doses

- Compounds with reproducible median correlation values (i.e. p_values below 0.05)

- Reproducible median scores scatterplot per dose

[gwaybio]

• We should also plot the null distribution plots for the second figure above
• Add a dashed line at the 95% of the null distribution for the scatter plots for both x and y axes
• Add alpha channel to scatterplots

[gwaybio]

☝️ @AdeboyeML and I discussed this in real time during our checkin :)

[AdeboyeML]

@gwaygenomics @shntnu

Updated figures

- L1000 vs Cell painting median score scatter plots (includes 95th percentile of the null distribution per dose)

- Reproducible median scores scatter plot per dose (includes 95th percentile of null distribution per dose)

- Cell painting -- Replicate versus Non-replicate (Null) correlation values distribution across all doses (1-6)

- L1000 -- Replicate versus Non-replicate (Null) correlation values distribution across all doses (1-6)

[AdeboyeML]

Analysis based on Signature Strength and Transcriptional/Morphological activity scores of L1000 and Cell painting

@gwaygenomics @shntnu

Definitions from clue.io

• Signature strength - Signature strength is a measure of the magnitude of the response elicited by a given treatment and is computed as the number of landmark genes (out of 978) with absolute z-score greater than or equal to 2. SS helps to further discriminate signatures that were consistent (high CC) from those that did or did not impact many genes.

• Transcriptional Activity Score (TAS) - is an aggregate measure of signature strength (SS) and replicate correlation (CC) that is intended to represent a perturbagen's transcriptional activity. The more transcriptionally active a perturbagen, the higher its TAS.

• Morphological activity score (MAS) is the equivalent of TAS for Cell Painting.

Signature strength VS Replicate correlation (median scores) - L1000

- (includes 95th percentile null distribution median score)

TAS VS Replicate correlation (median scores) - L1000

• Morphological activity score is the equivalent of TAS for Cell Painting.

Signature strength VS Replicate correlation (median scores) - Cell painting

MAS VS Replicate correlation (median scores) - Cell painting

Signature Strength comparison between L1000 landmark genes and Cell painting morphological features

L1000 TAS vs Cell painting MAS

[gwaybio]

I deleted all the notes from the 1.5 hour meeting @shntnu, @AdeboyeML and I had just now....and there were a lot!

I'll try to remember the key pieces, but please add to this list:

• Ted et al. have done similar work (and went down some time-consuming albeit useful rabbit holes, e.g. null distribution from same well position) - we should consider adding some of these refined analyses to the manuscript to supplement Adeniyi's next step work
• Some notes on the figures
  • we need to double-check how the compounds were subset
  • Add DMSO 95% thresholds to the TAS/MAS plots
    • I also want to see the genes that are part of the compounds with high MAS but low TAS
  • Show the "Compounds with reproducible median correlation values" figure split across dose
• Next steps:
  • Some diversity analysis of compounds - how diverse is the phenotypic space in L1000 vs. Cell Painting
  • Concatenate Cell painting, L1000, and PRISM and predict MOA
    • Look at feature importance scores
  • We also need to consider across batch reproducibility of each modality
    • We may consider using existing analysis by Ted et al for this

[AdeboyeML]

Updated Results and Figures

@gwaygenomics @shntnu

- Calculated the null distribution per dose from 5 randomly selected plates per dose (as suggested from the last meeting)

- Median score scatter plot

- Reproducible compound median scores above 4 doses (Corrected figure)

- Number of compounds with reproducible median scores

- Cell painting -- Replicate versus Non-replicate (Null) correlation values distribution across all doses (1-6)

- L1000 -- Replicate versus Non-replicate (Null) correlation values distribution across all doses (1-6)

Signature Strength, Morphological Activity Score (MAS) and Transcriptional Activity Score (TAS) Viz..

• Only showing L1000 vs Cell painting comparison plots
• Includes 95th percentile DMSO score

- TAS vs MAS

- Signature Strength Comparison

- - Compounds with high MAS but low TAS (i.e. TAS below 0.3 and MAS > 0.7)

- Plots for compounds with reproducible median scores

- TAS vs MAS

- Signature Strength Comparison

[gwaybio]

This analysis is painting quite a nice picture @AdeboyeML - great work. Three thoughts:

1. For this figure: "Compounds with high MAS but low TAS (i.e. TAS below 0.3 and MAS > 0.7)"
   • Can you provide me with the underlying data for this? I'd like to perform one follow-up exploratory analysis.
   • I specifically want the transcriptionally active genes for each compound. Think of a tidy, long data table with the following format:

Compound	MOA	Transcriptionally active genes	MAS	TAS
X	Y	Gene A	0.25	0.7
X	Y	Gene B	0.25	0.7
X	Y	Gene C	0.25	0.7

If it's easier, the data can be output in a different tidy format, but this one will work nicely (but it will be large!).

2. For the figures where we subset to only looking at compounds with 6 reproducible doses, can we bump this down to having 4, 5, or 6 reproducible doses? I don't want to drop a compound if there was a single dose anomaly, or if the compounds were used at too low a dose.
3. What do you see as the next steps?

This is exciting progress! Looking forward to discussing this further 💯

[AdeboyeML]

Assessing quality of clustering between Cell painting and L1000

- PCA and UMAP

-- Cell painting

- Explained variance by PCs -- ~77% variance explained by 25 Principal Components (PCs)

-- L1000

- Explained variance by PCs -- ~75% variance explained by 200 Principal Components (PCs)

- UMAP -- Cell Painting

- UMAP -- L1000

- Silhouette Score

• Silhouette score takes into consideration the intra-cluster distance between the sample and other data points within the same cluster and inter-cluster distance between the sample and the next nearest cluster.
• The silhouette score falls within the range [-1, 1].

- The silhouette score of 1 means that the clusters are very dense and nicely separated. The score of 0 means that clusters are overlapping. The score of less than 0 means that data belonging to clusters may be wrong/incorrect.

- Davies-Bouldin Index

• Davies-Bouldin (DB) Index evaluates intra-cluster similarity and inter-cluster differences.

• The DB index captures the intuition that clusters that are (1) well-spaced from each other and (2) are very dense

• Low Davies-Bouldin Index score indicates likely a ‘good’ clustering.

- Gausian Mixture Model (GMM) - Evaluating based on BIC (Dose 3 - 6)

-- Cell painting

-- L1000

[AdeboyeML]

[AdeboyeML]