Interval cancers Intrinsic subtypes Interval cancers arise - - PowerPoint PPT Presentation

6/9/2017 1 Using a polygenic risk score and breast density to predict interval cancers

Yiwey Shieh, MD UCSF Division of General Internal Medicine

International Breast Densitometry & Cancer Risk Assessment Workshop June 9, 2017

I have no potential conflicts of interest to disclose.

Disclosures

I am a general internist.

Interval cancers

• Interval cancers arise

symptomatically between screening rounds, following a normal mammogram

• Present at more advanced stage,

worse survival in some studies

• Associated with dense breasts

(masking)

• “High-risk” biology: ER-negative,

proliferative

Kirsch JNCI 2011

Intrinsic subtypes

Subtype Receptor status Prevalence Luminal A ER+ or PR+ HER2- 30-70% Luminal B ER+ or PR+ HER2+ 10-20% Basal ER-/PR- HER2- 15-20% HER2 ER-/PR- HER2+ 5-10% Normal-like

low-grade ER+ ER- (high-grade) high-grade ER+

ER: estrogen receptor PR: progesterone receptor HER2: human epidermal growth factor receptor 2

6/9/2017 2

Genetic determinants of cancer

• Single nucleotide

polymorphisms (SNPs) → genetic variants responsible for differences in phenotype

• Individual SNP odds ratio =

0.8 to 1.3

• 157 SNPs associated with

breast cancer (p < 5 x 10-8), though ~90 published

Polygenic risk scores (PRS) represent cumulative effects of multiple SNPs

Risk stratification with a 77-SNP PRS

Mavaddat JNCI 2015

AUROC 0.62

Shieh BCRT 2016

Top vs bottom quartile: OR = 1.7 (95% CI 1.2-2.5) for BCSC model OR = 3.2 (95% CI 2.2-4.7) for BCSC-PRS model

p = 0.01

Our work: 83-SNP PRS improves

• n existing risk models

Why an interval cancer PRS?

• Many SNPs have differential associations with ER+ and ER-

cancers

• also survival, age of onset
• suggests some genetic determination of cancer phenotype
• Clinical setting:
• unfitted PRS → when to begin screening, consider prevention
• next steps (?)
• “interval cancer PRS” → how often to screen, modality of

screening

• risk-stratify women with dense breasts

6/9/2017 3

• 77-SNP PRS from Mavaddat et al tested in Swedish

cohort

• High PRS → lower risk of:
• interval cancers, OR 0.91 (95% CI 0.63-1.01) per

SD

• poor prognostic features: ER-negativity, high

grade

Li Annals of Oncology 2015 Holm JCO 2015

Can PRS predict interval cancers? PRS and tumor characteristics

Holm JCO 2015

High PRS protective against ER-negativity, high grade, and possibly larger tumor size and lymph node involvement

Alternate approach

Pick SNPs associated with “high risk” disease

Fit currently known SNPs to high-risk features of breast cancer Use high-risk SNPs to predict interval breast cancers in independent dataset

Methods

6/9/2017 4

Overview of methods

Fit currently known SNPs to high-risk features of breast cancer Use high-risk SNPs to predict interval breast cancers

Development set: TCGA Test set: nested study in screening cohort

The Cancer Genome Atlas

• The Cancer Genome Atlas (TCGA): publicly accessible

dataset with comprehensive molecular portraits of tumors

• 1,094 breast cancers with available data:
• mRNA expression (Agilent)
• array-based SNP genotypes (Affymetrix 6.0)
• whole-exome sequencing
• DNA methylation
• miRNA sequencing
• clinical outcomes

Identifying PC’s associated with “high-risk” features

• 1. Perform principal components (PC) analysis of

gene expression data in TCGA

• a. PC1-4, orthogonal (rotated)
• 2. Identify PCs associated with high-risk features like

proliferation or ER-negativity

• a. confirm ER-negative PCs using ESR1

expression

• b. confirm proliferation PCs using risk of

recurrence-proliferation (ROR-P) score

Risk of recurrence (ROR-P) score

• Based on PAM50, gene

expression array that classifies breast cancer into 4 intrinsic subtypes

• ROR-P score = model for risk
• f relapse based on tumor

subtype correlations + expression of a subset of 11 genes correlated with proliferation

ROR-P = -0.001*Basal + 0.7*Her2 - 0.95*LumA + 0.49*LumB + 0.34*Proliferation

Nielsen Clinical Cancer Research 2010

6/9/2017 5

Using PC’s to choose SNPs for polygenic risk score

• 1. Candidate SNPs = genome-wide significant

association vs breast cancer or phenotype (ER status, survival, etc)

• 2. Regress PC vs candidate SNPs
• a. adjusted for ancestry (principal components of

SNPs)

• 3. Select SNPs based on “direction” of association with

PC (pos or neg beta) & significance (p<0.2)

• 4. Use SNPs to modify existing 83-SNP polygenic risk

score

Calculating the polygenic risk score

• Step 1: for each SNP,

calculate probability of genotype given disease

• Step 2: calculate LR for SNP
• Step 3: multiply LRs for each
• f x SNPs to obtain final

PRS

• Step 4: use Bayes theorem

to modify pretest probability as predicted by risk model

• California Pacific Medical Center Research Institute

(CPMCRI) cohort

• ~19,000 women undergoing screening 2004-2011 who

gave blood samples for research

• questionnaire data (SFMR), cancer outcomes
• density measured using BI-RADS
• Subset were genotyped (OncoArray):
• 481 cases: 102 interval, 369 screen-detected
• 496 controls matched by age, race/ethnicity

Testing the PRS versus interval cancers

• Case-control
• Outcome: interval cancer vs controls
• Predictors: PRS, density
• density adjusted for age, BMI, race/ethnicity
• Case-case
• Outcome: interval vs screen-detected cancers
• Same predictors, adjustment as above
• Evaluated discrimination with receiver operating characteristic

(ROC) curve analysis

Statistical methods

6/9/2017 6

Results

Principal components analysis SNP selection

Principal components

• 18,321 available transcripts
• dropped 5,662 genes with missing data
• 12,659 remaining genes

PC1 → ER status

positive PC1 → ER neg

6/9/2017 7

ρ = -0.80

PC1 and ESR1 expression

positive PC1 → ER neg

PC3 → proliferation/grade

negative PC3 → proliferative

PC3 and ROR-P

ρ = -0.64

SNP selection

SNPs in naive (unfitted) PRS hits vs ER+ or low-grade hits vs ER- or high-grade

83 4 13 +

• 74

SNP PRS fitted to “high risk” cancers

(subtract low-risk SNPs) (add high-risk SNPs) (retain “neutral” SNPs)

6/9/2017 8

Interval cancer prediction

Demographics

Controls n = 496 Interval cancers n = 102 Age, median (IQR) 55 (47-64) 53 (45-60) Race/Ethnicity, n(%) White Black Asian Hispanic Mixed 396 (80.1) 10 (2) 53 (10.1) 24 (4.9) 11 (2.2) 84 (82.3) 1 (1) 10 (9.8) 3 (2.9) 4 (3.9) BMI, median (IQR) 23.4 (21.2-26.4) 22.5 (20.9-25.8) Prior biopsy, n(%) 95 (19.2) 33 (32.4) Positive family history, n(%) 89 (17.9) 27 (26.5)

BIRADS density

interval cancers vs controls

OR 0.3

(0.1-1.7)

referent OR 3.0

(1.7-5.4)

OR 4.4

(2.1-9.2)

Adjusted for age, BMI, race/ethnicity

ROC curve, BIRADS density

interval cancer vs controls

Density* AUC = 0.66

(95% CI 0.60-0.71) *adjusted for age, BMI, race/ethnicity

6/9/2017 9

Histogram of 74-SNP PRS

interval cancer vs controls

PRS comparison

Case-control OR per SD (95% CI) Interval cancer vs control OR per SD (95% CI) PRS83 (unfitted) 1.39 (1.20-1.62) 1.32 (1.06-1.64) PRS74 (“high-risk”) 1.43 (1.23-1.69) 1.53 (1.21-1.92)

ROC curve, density + PRS74

interval cancer vs controls

Density AUC 0.66 (95% CI 0.60-0.71) PRS+density AUC 0.68 (95% CI 0.62-0.74) p = 0.14

Quartiles of Density/PRS model

Interval cancers vs controls

OR 1.9

(0.9-4.4)

OR 3.8

(1.8-8.0)

OR 7.3

(3.5-15.4)

6/9/2017 10

What about interval vs screen- detected cancers?

Density* AUC = 0.64

(95% CI 0.58-0.70) *adjusted for age, BMI, race/ethnicity

Histogram of 74-SNP PRS

interval vs screen-detected

PRS comparison

Case-control (OR, 95% CI) Interval cancer vs control (OR, 95% CI) Interval cancer vs screen- detected (OR, 95% CI) PRS83 (unfitted) 1.39 (1.20-1.62) 1.32 (1.06-1.64) 0.96 (0.78-1.19) PRS74 (“high-risk”) 1.43 (1.23-1.69) 1.53 (1.21-1.92) 1.07 (0.88-1.31)

Discussion

6/9/2017 11

Key findings

• Breast density is strongly associated with interval

cancers

• Able to identify SNPS associated with ER status &

proliferation (per gene expression) in TCGA

• Modifying existing PRS according to these SNPs →

minimal improvement over density in interval cancer prediction

Explanations (limitations)

• TCGA and/or CPMC datasets may be

underpowered

• ER-negativity and grade only modestly associated

with interval cancer status in CPMC dataset

• TCGA hits for ER-negative don’t replicate in CPMC

Next steps

• Incorporation of newly discovered ER-negative

SNPs from OncoArray in Breast Cancer Association Consortium

• Repeat SNP discovery in larger dataset
• Expand validation dataset

Acknowledgements

Donglei Hu Scott Huntsman Lin Ma Charlotte Gard Jessica Leung Celine Vachon Christopher Scott Jeff Tice Steve Cummings Celine Vachon Karla Kerlikowske Elad Ziv

6/9/2017 12

Questions?