Recapitulating Spatial Tumor Morphology Using Automated Classifier In Triple Negative Breast Cancer.

Abstract Introduction Triple negative breast cancer (TNBC) is a heterogeneous disease. Spatial transcriptomics (ST) interrogates gene expression in spatially defined spots. Here, we assessed whether ST expression data could identify specific histomorphologic features from TNBC samples, the ultimate goal being to develop a tool allowing to automatically annotate ST slides from gene expression data. Methods ST expression profiles were obtained from 41 TNBC samples. T and B cell immune phenotyping was performed on a consecutive slide using double CD3/CD20 IHC. All samples were manually annotated in 11 histomorphological categories (see Table). ST spots with artifacts or less than 500 reads were removed from the analysis. Each spot was assigned to its most common annotation. A linear booster classifier was used to classify the spots into 11 histomorphological categories based on gene expression. Classification accuracy was summarized by area under the ROC curve (AUC), assessed using leave-a-patient-out (LPO) or in-patient 10-fold cross-validation (xVal). Results Most ST spots were annotated as invasive tumor and stroma with low TILs as well stroma with high TILs (see Table). Overall, the performance of the classifier was very high. However, xVal method obtained higher AUC than the LPO method, reflecting higher interpatient than intrapatient variations. Necrosis, in situ carcinoma and B cells were very variable between patients, but not intrapatient, as shown by low LPO and high xVal AUCs. Of note, misclassification of B cells could be due to poor distinction of B vs T cells on HE slides despite IHC support. Spots with multiple annotations were also more difficult to classify. Conclusion Here, we developed a high-performance classifier based on the biggest series of ST data allowing to identify several histomorphological features from TNBC tissues, representing a unique tool for further ST research. SupercategoriesCategoriesAbbreviationsN spotsN samplesAUC LPO (%)AUC Xval (%)StromaStroma220484180.292.5Stroma with low TILslow sTILs140924178.291.8Stroma with high TILshigh sTILs75543984.394.8Blood vesselsBV1912774.283.6Lymphoid nodulesLN185148994.9High lymphocytes B areasBcell26564.495.9Invasive tumorIT126364088.894.7Invasive tumor with low or no intratumoral TILslow iTILs101773987.594.6Invasive tumor with high intratumoral TILshigh iTILs24592484.897.1OthersFat tissuesFT27212292.897.4NecrosisN10621975.295.4In situ carcinomaIS289749.399.1Normal glandsNG1741686.794.4 Citation Format: Xiaoxiao Wang, David Venet, Floriane Dupont, Ghizlane Rouas, Linnea Stenberg, Annelie Mollbrink, Denis Larsimont, Joakim Lundeberg, Françoise Rothé, Christos Sotiriou. Recapitulating spatial tumor morphology using automated classifier in triple negative breast cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 182.