AGBT talk: Obi Grifith

[Yay,we can blog Obi’s talk]

“Transcriptome and Exome Sequencing of Breast Cancer Cell Lines to Identigy Molecular Predictors of Response to Anti-Cancer Compounds.”

Using a panel of breast cancer cell lines.

Hypothesis: correlating drug response of the breast cancer cell lines with molecular characteristics will help identify drug treatments.

Combine drug response data with molecular data -> molecular signiture of drug response.

Panel: 72 cell lines. [I see a few normals in the cell line list, tho, so not all breast cancer cell lines… odd.]

after filtering : 82 drugs, 63 are publicly available, 19 are through private companies.

Methods: U133A, exon array, mehtylation, rppa, westerns, cnv .  Also: Cosmic, RNAseq (Alexa-seq), Exome-seq.  All data types can be combined and filtered (unsupervised) for variance.

Ideally, in the perfect world, all cell lines would be profiled with all of the techniques – but “we don’t live in that world”. [Speaks to a lot of issues in working with cell lines, IMHO.]

RNA-Seq: 55 cell lines.  See Alexa-seq poster [74, I think.]

Description of pipeline.  Looks relatively standard.

For each drug, cell lines separated into responders/non-responders.  Random Forests used for classification. (Internal cross validation step)

Exome-seq identifies known and novel cancer variants.  Concordant calls identify mutations. – Almost all discrepancies are false negatives in exome-seq. [interesting!]

RNA-seq recapitulates known subtypes with high accuracy.  (unsupervised clustering map shown, looks very pretty, and cell lines are nearly all in the correct clusters.)

Drugs with best predictors listed – 58/82 are better than random, (AUC > 0.5), [gets better the higher you go, obviously.]

Example given with Lapatinib – predicted by Her2 amplicon.  Over expression of her2 amplicon is visible across all of the methods, and thus, it’s possible to use that info for lapatinib response.

Example : BIBW2992 – drug response also associated with HER2, but also 3 other mutations.  classification requires use of multiple genes for response prediction.


  • Predictors can be found for many drugs
  • modst important predictors come from a wide range of data types.

Future work:

  • Experiment with sensitivity parameters/thresholds
  • control for subtype
  • compare performatnce of individual data types.

[I really didn’t do justice with the notes – much of this talk was visual, and data was hard to summarize in text.  Good talk, however, and nice to see that molecular classification is becoming more feasible.]

One thought on “AGBT talk: Obi Grifith

  1. Pingback: Tweets that mention AGBT talk: Obi Grifith | --

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.