Abstract PO2-07-06: Multimodal learning predictor of HER2-positive breast cancer therapy response in the randomized PREDIX HER2 trial

Abstract Background: The PREDIX HER2 trial, compared six courses of docetaxel, trastuzumab, and pertuzumab (DTP) vs. trastuzumab emtansine (T-DM1) as neoadjuvant treatment for HER2-positive breast cancer (BC). Similar rates of pathologic complete response (pCR) were seen. Methods: Clinicopathological, shallow whole-genome sequencing (CUTseq, n=176), whole exome sequencing (WES, n=192), and RNA-sequencing (RNA-seq, n=187) data were generated using fresh-frozen baseline core biopsies. Potential tumor intrinsic resistance factors and microenvironment components were quantified by multi-omics analysis, including BC-specific somatic mutations and copy number alterations (CNA), COSMIC mutational signatures, CNA-based chromosomal instability signatures (CIN), subclone percentage, PAM50 subtype, GGI/PIK3CA score, HER2DX score, immune profiles (Danaher signature score, TIDE score and immune repertoires). We assessed the association of biomarkers with pCR in each treatment arm using logistic regression adjusting for hormone receptor (HR) status, and evaluated their predictive value by adding the interaction term (biomarker x treatment arm). In addition, a machine learning (ML) analysis was conducted from different classifiers, comprising unimodal ML-based models from clinical, RNA and DNA information, respectively. Model performance was assessed using the mean and standard deviation (mean ± std) of the area under receiver operator characteristic curve (AUC), positive predictive value (PPV) and negative predictive value (NPV) using a nested stratified cross-validation (CV) schema of 200 outer shuffle splits and 100 inner 5-fold CV to mitigate potential risk of overfitting. Results: In DTP arm, patients with higher ERBB2 copy ratio (ORadj=1.98, p=0.004) or mRNA (ORadj=3.08, p< 0.001) or HER2-enriched subtype (PAM50) (ORadj=1.78, p=0.02) had higher pCR rates, while ESR1 gene expression (ORadj=0.59, p=0.07) predicted treatment resistance despite adjustment for HR status. Conversely, response to T-DM1 was less likely to depend on ERBB2 profiles and only PAM50 HER2 enriched subtype (ORadj=1.53, p=0.1) showed higher pCR rate (52% vs. 25%) than other subtypes. Both ESR1 (ORadj=0.4, p=0.008) and PGR (ORadj=0.5, p=0.03) gene expression were independent predictors of T-DM1 resistance. Pre-treatment immune exclusion metrics could predict resistance to DTP (endothelial cell, ORadj=0.67, p=0.07) and T-DM1 (neutrophils, ORadj=0.54, p=0.02; mast cells, ORadj=0.57, p=0.02; cancer-associated fibroblasts, ORadj=0.67, p=0.09)), respectively. Predefined metrics such as PIK3CA signature score (ORadj =1.67, p=0.04) and Taxane response score (ORadj =1.64, p=0.03) were positively related to pCR in DTP arm. Genome instability, involving CIN CX2 signature (impaired homologous recombination) (ORadj=1.71, p=0.05), COSMIC signature6 (ORadj =1.53, p=0.07) and signature13 (ORadj =1.57, p=0.05), predicted benefit from DTP. The biomarker-treatment interaction tests were significant for HER2DX (pinteraction=0.004) and COSMIC signature15 (defective DNA mismatch repair) (pinteraction=0.007): lower HERDX score (ORadj =0.73, p=0.14) or higher COSMIC signature15 score (ORadj =1.51, p=0.1) could identify patients benefiting from T-DM1, while being resistant to DTP (HERDX: ORadj =1.46, p=0.13; signature15: ORadj =0.70, p=0.1). In the ML models, clinical information yielded an AUC=0.62±0.07, PPV=0.64±0.12 and NPV=0.64±0.06; for DNA data, AUC was equal to 0.70±0.08, PPV=0.72±0.09 and NPV=0.71±0.07; an adaptive boosting ensemble learning on RNA reported slightly increased pCR prediction performance (AUC=0.72±0.06, PPV=0.64±0.06, NPV=0.80±0.09). Conclusion: This study demonstrates that antibody–drug conjugates and standard treatment harbor strikingly distinctive biomarkers across tumor ecosystem. Further external validation and integrated ML model comprising all unimodal models are ongoing. Citation Format: Kang Wang, Yajing Zhu, Ioannis Zerdes, Emmanouil Sifakis, Georgios Manikis, Dimitrios Salgkamis, Nikolaos Tsiknakis, Luuk Harbers, Nicola Crosetto, Jonas Bergh, Alexios Matikas, Thomas Hatschek, Theodoros Foukakis. Multimodal learning predictor of HER2-positive breast cancer therapy response in the randomized PREDIX HER2 trial [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO2-07-06.