Development and evaluation of a novel digital pathology image analysis pipeline for prediction of clinical outcomes with the TROP2-directed antibody-drug conjugate (ADC) sacituzumab tirumotecan (sac-TMT) in triple-negative breast cancer (TNBC).

Abstract

1026 Background: Digital pathology (DP), including artificial intelligence (AI)-based image analysis methods, enables IHC biomarker assessment on a continuous scale, potentially improving precision and accuracy versus conventional IHC scoring. DP may also capture features relevant to ADC response that traditional methods miss. As proof of concept, we developed and evaluated an AI-assisted DP image analysis pipeline to explore the association between target antigen expression and clinical outcomes of sac-TMT, a TROP2-directed ADC with a unique bifunctional linker, in TNBC. Methods: We analyzed whole-slide images of IHC-stained tumor samples from participants with TNBC enrolled in the phase 1/2 MK-2870-001 study (NCT04152499) evaluating sac-TMT in pretreated advanced solid tumors. We established a set of prespecified human-interpretable features (HIFs), including IHC signal intensity, subcellular localization, and spatial patterns of cell and signal distribution. To mitigate overfitting, we prioritized a subset of 35 HIFs based on correlation structure and biological hypotheses, and in a blinded fashion, assessed their association with clinical outcomes (BOR; PFS) in a development cohort (DC) (n = 58). After unblinding clinical outcome data, additional HIFs and a multivariate model trained to predict clinical outcomes in the DC were prioritized for validation. Both sets of HIFs (identified from blinded and unblinded analyses) were assessed for their relationship to clinical outcomes in an independent validation cohort (IVC) from the same study (n = 34). In both cohorts, DP HIFs were compared with conventional TROP2 H-scores (on paired slides) for prediction of clinical outcome using the area under the receiver operating characteristic curve (AUROC) and Harrell C-index. Results: The blinded approach identified a HIF that was positively associated with BOR (multiplicity-adjusted P = 0.023), with an AUROC higher than TROP2 H-scores (0.76 vs 0.70). After unblinding, 5 additional HIFs and a multivariate model were selected for their associations with BOR (AUROC, 0.70-0.80). In the IVC, the 6 prioritized HIFs and 1 multivariate model were associated with response to sac-TMT (AUROC, 0.60-0.69); each outperformed TROP2 H-scores with respect to association with BOR (H-score AUROC, 0.57) and demonstrated incrementally better association with PFS than TROP2 H-scores (Harrell C-index, 0.64-0.66 vs 0.62). Conclusions: In this proof-of-concept study, DP-derived HIFs were associated with response to sac-TMT in TNBC and showed incrementally better nominal performance than conventional TROP2 H-scores. While the sample size was small, data from this study suggest that DP-based image analysis can identify novel biomarkers of response to sac-TMT in TNBC.