Abstract 729: Applying novel transcriptomics-based markers to predict response to durvalumab and olaparib in mTNBC from histopathological slides

Abstract Background: Transcriptomics-based predictors show great promise in identifying responders to immune checkpoint inhibitors (ICI) who are rejected by standard biomarkers like PD-L1. However, the reliance on gene-expression data makes these biomarkers less applicable in the clinic. On the other hand, H&E-stained slides are commonly available, so predicting response from slide scans could greatly improve accessibility. We have previously described DeepPT - a deep learning model to infer gene expression from H&E slides. Such a model could bridge that gap and allow clinicians to use transcriptomics-based biomarkers directly with H&E slides. Methods: We tested two transcriptomics-based biomarkers of response to PD-1/PD-L1 inhibitors on a dataset of 21 mTNBC patients treated with olaparib and durvalumab. The first biomarker was derived using our published ENLIGHT platform, and the second is an XGBoost model on a curated set of 839 genes associated with ICI response, trained on 24 datasets with measured gene-expression and response data. We blindly applied the two models to the pre-treated tumor RNAseq data of the patient tumors, as well as to expression profiles inferred from their H&E slides using DeepPT. We also examined whether combining the scores from multiple slides per sample improves prediction. We then unblinded the outcomes and analyzed the model's ability to predict objective response (RECIST 1.1). Results: Both models showed high predictivity in distinguishing responders (CR/PR) from non responders (SD/PD) when applied to RNAseq gene-expression profiles: AUC of 0.73 for the ENLIGHT model and 0.85 for the XGB model (Odds ratio for classification of 2.1 and 5.6, respectively). The performance on the gene-expression profiles that were inferred from a single H&E slide was also high - AUC of 0.82 for ENLIGHT and 0.83 for XGB (OR of 44 and 12). The performance of the XGB model improves when we average over all slides that correspond to a single patient (AUC of 0.85, OR of 16.7) while we observe no such improvement for ENLIGHT (AUC of 0.78, OR of 6.7). Interestingly, both models also distinguish stable from progressing disease within the nonresponders group. Conclusion: We introduce a method to tackle the challenge of applying gene-expression based biomarkers for response to treatment by PD1/PDL1 inhibitors in the clinic. We do this by combining transcriptomics-based models and a deep-learning model for predicting gene-expression from H&E slides. We demonstrate that this method correctly predicts response on an external dataset, favorably comparing with direct application of the models to RNAseq data. This approach may be extended to additional drugs where transcriptomics-based biomarkers exist, and can be further improved by taking into consideration the differential predictivity of genes from H&E slides. Citation Format: Yaron Kinar, Gal Dinstag, Omer Tirosh, Doreen S. Ben-Zvi, SMMART Clinical Trials Program, Tuvik Beker, Aharonov Ranit, Gordon B. Mills. Applying novel transcriptomics-based markers to predict response to durvalumab and olaparib in mTNBC from histopathological slides [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 729.