Oral Presentation 47th Lorne Genome Conference 2026

EmbryoRadar – a machine learning model to uncover the impact of developmental transcriptional program reawakening in cancer (133068)

Tongtong Wang 1 2 , Benjamin Hernandez-Rodriguez 1 2 , Janith A Seneviratne 1 2 , Alicia Oshlack 1 2 3 , Melanie Eckersley-Maslin 1 2 4
  1. Peter MacCallum Cancer Centre, Melbourne, Victoria, Australia
  2. Sir Peter MacCallum Department of Oncology, University of Melbourne, Melbourne, Victoria, Australia
  3. School of Mathematics & Statistics, The University of Melbourne, Melbourne, Victoria, Australia
  4. Department of Anatomy and Physiology, University of Melbourne, Melbourne, Victoria, Australia

Embryos and cancers share many similarities, such as proliferation and phenotypic plasticity. The preimplantation embryo is associated with the highest developmental potential. Whilst there have been many studies reporting that re-expression of certain embryonic genes in cancer is associated with poorer prognosis, these have been largely limited to comparisons with the later stages of development or embryonic stem cells, which have adapted to in vitro culture conditions. To address this gap, we developed EmbryoRadar, a suite of machine-learning models that scan transcriptomic profiles to detect the re-emergence of different preimplantation embryonic states.

Surprisingly, when applied to TCGA and other bulk cancer cohorts, EmbryoRadar revealed a broadly protective pattern. This favourable association was evident in cancers including uveal melanoma (UVM), lower-grade glioma (LGG), acute myeloid leukaemia (LAML), kidney chromophobe/papillary subtypes (KIRP, KIRC), and bladder carcinoma (BLCA), where tumours with higher embryonic scores conferred longer survival. However, this effect was not universal. We identified three tumour types, cutaneous melanoma (SKCM), liver hepatocellular carcinoma (LIHC), and paediatric B-cell acute lymphoblastic leukemia (B-ALL), where higher embryonic reactivation scores instead tracked with worse outcomes. Adapting EmbryoRadar to score single-cell transcriptomic and spatial datasets enabled us to deconvolve the cell types that contribute to these embryonic scores and uncover tumour-type–specific transcriptional and microenvironmental contexts. This provides a plausible basis for why the same developmental signal can be beneficial in some cancers but detrimental in others.

EmbryoRadar provides a direct, generalisable, and scalable way to quantify the reactivation of developmental transcriptional programs in cancer. These findings challenge the prevailing assumption that embryonic program reactivation is uniformly adverse and instead suggest that its prognostic impact is tumour-context dependent.