ESTRO 2021 Abstract Book

S903

ESTRO 2021

Pathology, London, United Kingdom; 5 Institute of Cancer Research, Centre for Evolution and Cancer, London, United Kingdom; 6 Royal Marsden NHS Foundation Trust, Head and Neck Unit, London, United Kingdom

Purpose or Objective The prognostic significance of tumour infiltrating lymphocytes (TILs) in breast cancer (BC) is increasingly recognised. However, there remains a paucity of data describing the effect of radiotherapy on the BC tumour immune microenvironment in vivo. The PRADA trial (NCT 02771938) tested the feasibility and safety of neoadjuvant radiotherapy (NRT) in BC. The linked translational component (Trans-PRADA) provided an opportunity to collect novel radiobiological data with direct clinical relevance. We investigated changes in stromal TIL density occurring with neoadjuvant treatment in serial tumour samples and stratified results according to pathological outcome. Results will inform rational design of clinical trials of RT-IO combinations in high-risk disease. Materials and Methods Patients with high-risk, non-metastatic BC completed sequential neoadjuvant systemic therapy (NST) and locoregional NRT (40.05-42.72Gy in 15-16 fractions) prior to mastectomy and immediate reconstruction (mean RT-to-surgery interval = 22 days, range:12-53 days). Formalin-fixed, paraffin-embedded tissue was obtained from the primary tumour: before treatment (T0), after fraction 1 (T1) and final fraction (T2) of RT and at mastectomy (T3). TILs scores were generated from H&E sections at each time point, in accordance with international consensus guidelines. Peripheral lymphocyte counts (PLCs) were evaluated in a subset of cases treated at one centre. Machine learning algorithms for automated TILs quantification were developed in parallel. Results Material was analysed for 30 NST-NRT patients (5/30 triple-negative, 7/30 HER2+, 18/30 ER+HER2 neg). Pre- treatment samples showed highest TILs scores and greatest inter-patient variation (range:1-80%, median=9%). TILs fell significantly by T1 with striking convergence (range:1-7%, median=2%) and did not recover by the time of surgery (range:0-10%, median=2%). Higher baseline TILs correlated with favourable pathological outcomes (Fig 1). PLCs fell between completion of NST (mean=1.66 x10 9 n=9) and completion of NRT (0.91 x10 9 , n=14), and this was statistically significant (P=0.0049, paired t test). An AI superpixel-based model applied to low power magnification (5x) whole-slide images generated maps of stromal regions 1 . A second neural network performed single cell classification based on features of small patches centred on nuclei (Fig 2).

Conclusion This study represents the first application of current manual TILs scoring methodology to the assessment of BC in the NRT setting. Findings indicate a rapid locally lymphodepletive effect of NRT, and support the utility of baseline TILs as a biomarker predicting response to sequential NST-NRT. Machine learning algorithms will