Whole Genome Approaches to Cancer 1. What other tumor is a given rare tumor most like? 2. Is tumor X likely to respond to drug Y?
Jan 06, 2016
Whole Genome Approaches to Cancer
1. What other tumor is a given rare tumor most like?
2. Is tumor X likely to respond to drug Y?
Oligonucleotide Arrays
• 300,000 25-mer probes• in situ photolithographic synthesis• single color hybridizations• chips available for 40,000 human genes and 25,000 murine genes
1.28cm
1.28 cm
1 2 3
n
106 oligos
24 micronsGenes
1 2 ... ..20
matchmismatch
3’ UTRcodingGene n
Estimating Message Abundance
perfect match (PM)
mismatch (MM)
Message abundance = trimmed mean (PM1-MM1 . . .PM20-MM20)
1 2 3. . . . . .20
Confidence measure:
‘A’ low confidence‘P’ high confidence
Oligonucleotide Arrays: Sample Preparation
AAA TTT-T7TTT-T7AAA-T7
TTTB B
BBSA
computer
10 g total RNA cDNA ds cDNA
cRNA
RT
IVTbio-NTPs
hybSAPEscan
ion argon laser
chip
+2X
-2X
100 1000 10000 100000100
1000
10000
100000
‘P’ calls (2301) ‘A’ calls (4830)
Reproducibility ExperimentsSame Target on 2 Arrays
+2X
-2X
Cancer Classification
Identify previouslyunrecognized classes
Assign new tumors to known classes
Class PredictionClass Discovery
Type 1
Type 2
Type 3Type 1 Type 2 Type 3
Proof of Concept: Acute Leukemia Diagnosis
ALL AML
Molecularly distinct tumors are morphologically similar
ALL
genes
low high
normalizedexpression
AML
Gene Expression Correlates of LeukemiaGenes sorted according to correlation with ALL/AML distinction
Permutation Test
1000 genes more highlycorrelated than expected
Terminal transferase
Myelo-peroxidase
38 pre-treatment marrows (ALL or AML)No leukemia cell purification
3-10 g total RNA per sample
Proof of Principle: ALL vs. AML Distinction
Sort genes by degree of correlation with ALL vs. AML
Randomly withhold one sample
6800 gene arrays
Biotin label RNA
Choose most highly correlated genes
Predict class of withheld sample Error Rate
Removeeach sample in turn
Results
Initial set (n=38) 36 predictions 2 uncertain
100% correct
Independent set (n=34) 29 predictions 5 uncertain
100% correct
AML T-ALLB-ALL
Class DiscoveryWhat if ALL/AML distinction was not previously known?
Could we discover it by expression alone?
38 samples Cluster by SOM
Golub et al., Science, 1999
Can a gene expression-based model ‘learn’how to predict treatment response?
p = 0.0003
Lymphoma Outcome Prediction: All patients (n=58)M. Shipp, J. Aster
predicted ‘good’
predicted ‘bad’
Chemosensitivity Prediction: NCI-60J. Staunton, J. Weinstein
• panel of 60 human cancer cell lines• known sensitivity to 000’s of compounds• we measured expression of 6800 genes in untreated cells
Are gene expression patterns sufficient to predict sensitivity?
Choose pair of sens/resistant within each tissue type
Build best model
Test on remaining samples
0
10
20
30
40
50
10 20 30 40 50 60 70 80 90 100
% accuracy
num
ber
of d
rugs
random prediction gene-based prediction
Kolmogorov-Smirnov p = 10-24
Expression-Based Prediction
50
40
30
20
10
0
Num
ber
of D
rugs
% Accuracy
BR PR LU CR LY M LEU R P
First Generation Global Cancer MapS. Ramaswamy
BR PR CO CNS LY ME LEU RE PA
genes
300 tumors and normals27 tumor classes
13,000 genes/ESTs