Integrative and deep learning-based prediction of therapy response in ovarian cancer

Ovarian cancer comprises a highly complex ecosystem of malignant cells and their surrounding tumor microenvironment (TME), where intricate interactions shape therapeutic responses. Most current predictive models fail to capture the full extent of these interactions. Here, we performed a comprehensive multi-omic analysis of pre-treatment ovarian tumor tissues, integrating clinical, genomic, transcriptomic, and immune features to correlate with pathological therapy response. Our results show that integrating genetic and immune parameters—particularly the interplay between NK cells and TP53 status in high grade serous ovarian cancer (HGSOC), and diverse genetic alterations in non-HGSOC—markedly improves therapy response prediction. We demonstrate that tumor TP53 status governs the persistence of early NK cells in HGSOC, and this persistent NK phenotype is associated with favorable clinical outcomes. Machine learning models harnessing these multi-omic features significantly outperform those based on any single information type alone. These findings highlight the central role of the baseline tumor ecosystem and support a precision oncology framework leveraging integrated multi-omic profiling and advanced analytics to improve prediction and guide treatment strategies. • Clinical features are insufficient predictors of therapy response. • Genetic and immune features are linked to therapy response. • Early NK cells exhibit a persistent phenotype associated with favorable clinical outcomes. • Machine learning models integrating genomic and immune features significantly improve therapy response prediction. • In HGSOC, therapy response depends on NK cell—TP53 interplay, while in non-HGSOC, it is driven by diverse genetic mutations. • Tumor TP53 status regulates persistence of NK cells in HGSOC.