saezlab / footprints Goto Github PK

preliminary results looked like the expected RPPA phosph. do not correlate with mutations

could be interesting to look at this, but TCGA data is doubtful to work

is outdated

• make sure speed score scaling is correct and does not suffer from different magnitudes because of different arrays having different numbers of genes on them we scale per (pathway x method), no issue
• bootstrap of signature creation + scores
• use NES instead of kernel density estimator for GSEA scores using NES + KD
• scale scores tissue-wise? no
• ...
• add here

prerequesites

• map all scores to specific set of pathways
• use same pathway set input for glmnet

kinds of models

• pathways
• mutations and pathways as one input (to compare mut to pathways)
• mutations + pathways (as in, formula syntax - to compare which pathways add most on top)

generation of null models

• null models: 100 repetitions, shuffled labels
• null models: 1000 repetitions, shuffled labels
• real model: one repetition

calc

• empirical p-values for models
• bar plots for best models (as defined with p-val)

• Fix ArrayExpress package to work for enough arrays
  • handle duplicate rows in .sdrf better
  • A-AFFY-141
  • A-AFFY-33
  • A-AFFY-44
  • A-AFFY-37
  • A-GEOD-9419
  • A-AFFY-143
  • A-GEOD-16239
  • A-GEOD-10520
  • A-GEOD-10200
  • A-GEOD-8300
  • A-AFFY-1
  • A-GEOD-13667
  • ...any more...?
  • Agilent 1-color arrays
  • Illumina arrays
• Map expression to genes
• Assemble Z-score matrix + index for all pathways

• just take the SPEED2 exps where too few controls + rank them with precision-recall curve
• LINCS data

should do volcano with covariate, maybe highlight in addition VHL that is only mutated in KIRC so we have another angle as well

Fit using pathway factors:

• That's how we do it now for speed_linear (w/o intercept)

Fit using perturbed vs basal for each pathway:

• That's how we do it now for speed_matrix; VEGF surv increase?!
• MAPK += EGFR - MAPK drug assocs > EGFR for MEKi

Fit using one matrix with 1/-1 coefficients:

• MAPK += EGFR EGFR MEKi resistance (because non-MEK EGFR targets, esp. PI3K)
• MAPK, PI3K += EGFR - mut assocs good (TP53, BRAF); EGFR survival increase (why?)
• intercept but no xtalk - drug assocs good (MAPK>EGFR; good pvals: Trametinib<1e-15)
• no intercept, no xtalk - ??

Right now, GO and Reactome associations are conditioned on the best SPEED association to show whether or not they can be explained by the response-genes alone.

We could go one step further: make the statement that SPEED explains most of the GO/Reactome associations, but not the other way around. This would require a conditioning of SPEED scores on either/both of GO/Reactome association scores.

Also, it might make sense to condition on all pathways/all significant pathways - not sure, this might just get too many random correlations with 10 conditioning vars.

Overall, this would make the statement that SPEED outperforms the others stronger.

but around 5, 9, etc.

blocking #3

Cell line scores for different signatures

• GDSC baseline expression
  • new SPEED
  • old SPEED
  • other genesets
• TCGA
  • new SPEED
  • old SPEED
  • other genesets

Associations with

• Tissue
• Drugs
• Mutations
• Survival

when accessing the TCGA data directly, they provide references to which batch patients belong to; in addition, barcodes provide information about the collection and analysis centres

this info could be used to remove batch effects, if the TCGA didn't already do so - should talk to someone who worked with the data for longer + see if adjustment changes the survival associations

• RPPA data
• Gene sets
  • BioCarta
  • Reactome
  • Gatza
  • GO
  • old SPEED
  • Gatza signatures

need to impute them using sig-only neighbourhood

does that change the assocs?

also, speed in general should work better with mutation-heavy tumors and CNA-heavy - can we find something interesting there?

For now it's almost only a recap of the results; add a bit more about when pathway expression, individual signatures are useful

Discuss a bit about the limitations

Should probably still add:

• Curated exps are a resource for further study
• Pathway enrichment is still useful & when it is
• Limitation that when a specific signature is available, it may still be better

Z-scores: min 2 arrays in basal condition, average for perturbed

Format: yaml

---
id        : <pathway.accession.#>
accession : <arrayexpress accession>
platform  : <arrayexpress platform id>
pathway   : <out of set below>
cells     : <human-readable cell description>
treatment : <human-readable treatment description>
effect    : activating|inhibiting
hours     : <int> number of hours treatment

control   :
    - <list of control arrays>
perturbed :
    - <list of perturbed arrays>
...

Pathways included:

• EGFR
• MAPK
• PI3K
• H2O2
• Hypoxia
• Trail
• p53
• TNFa
• NFkB
• VEGF
• TGFb
• Insulin
• IL1
• RAR
• notch
• PPAR
• Estrogen
• JAK-STAT
• Wnt
• acidosis ?
• senescence ?
• starvataion ?

2-4 features, glmnet

• Drug response
• Survival

Model creation

• Linear model
• SEM lavaan not feasible for number of genes

QC

• scores of model on original perturbation files
• GO enrichment of "gold standard" categories

• also PARADIGM because of NA or NULL
• also errors for unfiltered continuous for gatza, biocarta, reactome

see what's going on here

na.omit() for zscores leaves 5k genes, top assoc is p53 ~ 1e-19, some EGFR ~ 1e-14.

for 20k genes is is the same, that's a bit surprising

Right now, I scale the resampled scores per experiment.

• don't scale in scores, recompute LOOCV model scores
• column-scale in heatmap
• scale for pathway ROC curves
• don't scale for t-SNE plots (?)

some survival times are 0 w/ both status alive/deceased, this should not be

run clustering

• batch correct tcga+gdsc
• run NMF clustering

clustering analysis

• for different NMF clusters, do some sort of silhouette + discard weak points
• use this to also select with k (in case cophenetic implies more than one)

clustering plots

• tSNE for vis, but color NMF clusters + shape tumor/cell lines

quite hard to find a specific script right now

maybe some crosstalk to explicitly model?

to compare against, at least

• SPIA
• PARADIGM
• Pathifier

• relationship between signalling footprints and pathway expression
  • overlap
  • which signal activates transcription of which pathway?
  • which pathways are more posttranslationally controlled, which more by expression? (would need phospho for this, but could probably do drug/survival response)
  • is the effect the same for cell lines and primary tumors?
  • for each cancer type, where do pathway expression/footprints correlate more? less?
    • should do a pancan and tissue-specific correlation plot here (resp<->resp and resp<->expr)
• TCGA pathway activation + enrichment of variants that activate those
  • e.g., we find PI3K_variantTP53_R282W and enrichment of this variant in some cancers (BRCA)
    • does the same cancer need to have PI3K activated?
    • does p53 bind more PI3K-responsive genes with modified motif?
• tcga_pathway_per_mutation: use GISTIC scores as regression coeffs? (= include +/-1)