@ArranTurnbull Arran Turnbull Accurate and robust prediction of clinical response to aromatase inhibitors by two weeks of neoadjuvant breast cancer treatment AK Turnbull 1 , LM Arthur 1 , V Webber 1 , AA Larionov 1 , L Renshaw 1 , C Kay 1 , A Dunbier 2 , M Dowsett 2 , AH Sims 1 , JM Dixon 1 1 Breakthrough Breast Research Team, Division of Pathology Laboratories, IGMM, The University of Edinburgh, UK 2 Breakthrough Breast Research Unit, Institute of Cancer Research, The Royal Marsden, Chester Beatty Laboratories, London, UK SABCS 2012 Printed by Background Methods Results Summary of Results Conclusions • Training Dataset: Pre- and on-treatment (at 14 days and 3 months) biopsies obtained from 89 post-menopausal women with ER+ breast cancer treated with neoadjuvant Letrozole. Illumina Beadarray gene expression data (n=34 patients) combined with Affymetrix GeneChip data (n=55 patients) and integrated [3] (figure 1). • Validation Dataset: Illumina gene expression data generated from 44 patients with ER+ cancers biopsied before, and after 2 weeks of anastrozole in anastrozole- only arm of a neoadjuvant clinical trial [4]. • Response: Based on periodic ultrasound (US) 3D volume measurements (figure 1): 16 patients with incomplete response data excluded. • Classification of Response: • ‘Quick stable response’: reduction of at least 50% by 45 days and 70% by 3 months. • ‘Slow response’ : reduction of between 0-50% by 45 days and at least 70% by 3 months. • ‘Non-response’: increase or partial reduction that never exceeds 50%. • Data Analysis: Pairwise and grouped differential gene expression analyses performed using rank product (RP). RandomForest analysis used to identify the 200 most informative features (consisting of 131 genes) differentiating between responsive and non-responsive patients. These were then used as input for regression tree (CART) analysis. Functional analysis performed in David Bioinformatics Resources 6.7. • Genes associated with response to Letrozole enriched for proliferation, glycolysis/oxidative phosphorylation, immune/inflammatory and extra-cellular matrix (ECM) remodelling (figure 2A). • Slow responders had similar gene expression profile to quick stable responders in respect of the major functional groups. (figure 2A). These were subsequently all classified as ‘responsive’. • RandomForest analysis used to identify the 200 most informative features differentiating between responsive and non-responsive tumours, comprising 82 pre-treatment (PT), 82 day 14 (D14) and 35 delta variables, amounting to 131 distinct genes. Each gene was assigned to one function category (figure 2B). • A CART model approach predicting response to neoadjuvant AIs was developed. The model comprises PT expression of an immune signalling (Gene A) an apoptosis related (Gene D) gene and D14 expression of two proliferation (Genes B & C) genes (figures 2C & E). • Majority of responsive tumours characterised by high expression of gene A and low expression of gene C. Most non-responsive tumours characterised by low expression of gene A and high expression of gene B (figure 2C). • The model had 96% accuracy in training set (n=73) and 91% accuracy in independent validation dataset (n=44) (figure 2D). • Patients corresponding to different terminal nodes in the tree (figure 2C) had significantly poorer recurrence free survival (RFS). Patients with low initial expression of gene A had significantly poorer RFS. Those with high initial expression of gene A, could be divided into significantly different high and low risk RFS groups based on D14 expression of the proliferation gene C (P<0.0001) (figure F). Immune Signalling Gene A • Expression of this gene alone predicted response with 85% and 82% accuracy in training and validation datasets respectively. • Higher levels at diagnosis were associated with significantly better 1 year PFS (P=0.0001). • High expression was also associated with a significantly improved 10 year RFS in the adjuvant setting in two other patient groups, a Letrozole treated dataset (n=129) (P=0.035) and a tamoxifen treated cohort (n=212) (P=0.033). • A 4 gene classifier with huge potential clinical value has been developed and validated to accurately predict response to neoadjuvant aromatase inhibitors and this classifier predicts for long-term benefit from endocrine therapy. References A B C D Pre-treatment Expression Genes Day 14 Expression Genes E High Gene A – Low Gene C High Gene A – High Gene C Low Gene A F Figure 1: Top: Study design: tumour biopsies and US measurements taken periodically throughout treatment. Bottom: Graph of relative USS measurements over time. Patients divided into different response groups. Recurrence Free Survival Figure 2: Heatmap of expression of major functional groups in different clinical response cohorts. Samples in numerical order of patient ID for each time point shown. Colours represent relative differences in log2 gene expression. Red denotes higher expression and green lower expression. B: pie chart showing functional groups of 131 genes identified from RandomForest analysis. C: (left) CART model, right and left branches from each node denote higher and lower expression respectively, as determined by a model defined cut-off value, blue numbers denote node numbers, (centre): pie charts showing proportion of responsive (green) and non-responsive (red) patients at each terminal node in the model, (right) PFS survival, progression defined as any increase in tumour volume greater than 20% above the previous measurement. D: table of training and validation model statistics. E: Log2 mean expression of the 4 genes over time in respect of the 3 clinical response cohorts: quick stable response (green), slow response (light green) and non-response (red), error bars denote standard error. F: Kaplan-meier plot of 10 year RFS for the 73 patients in training model. [1] Smith IE, Dowsett M (2003) N Engl J Med, 348(24):2431-2442. [2] Miller WR et al (2009) J Clin Oncol, 27(9):1382-1387. [3] Turnbull AK et al (2012) BMC Medical Genomics, 2012:1. [4] Smith IE et al (2007) J Clin Oncol (25):3816-3822. PD3-2 Printed by San Antonio Breast Cancer Symposium - Cancer Therapy and Research Centre at UT Health Science Centre - December 10-14, 2013 This presentation is the intellectual property of the author/presenter. Contact a.turnbull@ed.ac.uk for permission to reprint and/or distribute. • Aromatase inhibitors (AIs) are used in the treatment of estrogen receptor positive (ER+) post- menopausal breast cancer and work by reducing estrogen production [1]. • Response rates to AIs are only 50-70% in the neoadjuvant setting and lower in advanced disease [2]. • Predicting response based on clinicopathological features is not possible. • There is no current method of predicting early on-treatment whether the cancer will respond. • Need to have validated biomarkers to predict response to AIs which out-perform currently used clinical and pathological factors. Relative Volume (%) Days from Treatment -0.8 0.8 0