Koufariotis, L. T. | Search bioRxiv

bioRxiv2022

Generalising uncertainty improves accuracy and safety of deep learning analytics applied to oncology

Trust and transparency are critical for deploying deep learning (DL) models into the clinic. DL application poses generalisation obstacles since training/development datasets often have different data distributions to clinical/production datasets that can lead to incorrect predictions with underestimated uncertainty. To investigate this pitfall, we benchmarked one pointwise and three approximate Bayesian DL models used to predict cancer of unknown primary with three independent RNA-seq datasets covering 10,968 samples across 57 primary cancer types. Our results highlight simple and scalable Bayesian DL significantly improves the generalisation of uncertainty estimation (e.g., p-value = 0.0013 for calibration). Moreover, we demonstrate Bayesian DL substantially improves accuracy under data distributional shifts when utilising uncertainty thresholding by designing a prototypical metric that evaluates the expected (accuracy) loss when deploying models from development to production, which we call the Area between Development and Production curve (ADP). In summary, Bayesian DL is a hopeful avenue of research for generalising uncertainty, which improves performance, transparency, and therefore safety of DL models for deployment in real-world.

bioinformatics↗

bioRxiv2021

Genomic analysis of patient-derived xenograft models reveals intra-tumor heterogeneity in endometrial cancer and can predict tumor growth inhibition with talazoparib

BackgroundEndometrial cancer (EC) is a major gynecological cancer with increasing incidence. It comprised of four molecular subtypes with differing etiology, prognoses, and response to chemotherapy. In the future, clinical trials testing new single agents or combination therapies will be targeted to the molecular subtype most likely to respond. Pre-clinical models that faithfully represent the molecular subtypes of EC are urgently needed, we sought to develop and characterize a panel of novel EC patient-derived xenograft (PDX) models. MethodsHere, we report whole exome or whole genome sequencing of 11 PDX models and the matched primary tumor. Analysis of multiple PDX lineages and passages was performed to study tumor heterogeneity across lineages and/or passages. Based on recent reports of frequent defects in the homologous recombination (HR) pathway in EC, we assessed mutational signatures and HR deficiency scores and correlated these with in vivo responses to the PARP inhibitor (PARPi) talazoparib in six PDXs representing the different molecular subtypes of EC. ResultsPDX models were successfully generated from all four molecular subtypes of EC and uterine carcinosarcomas, and they recapitulated morphology and the molecular landscape of primary tumors without major genomic drift. We also observed a wide range of inter-tumor and intra-tumor heterogeneity, well captured by different PDX lineages, which could lead to different treatment responses. An in vivo response to talazoparib was detected in two p53mut models consistent with stable disease, however both lacked the HR deficiency genomic signature. ConclusionsEC PDX models represent the four molecular subtypes of disease and can capture intra-tumoral heterogeneity of the original primary tumor. PDXs of the p53mut molecular subtype showed sensitivity to PARPi, however, deeper and more durable responses will likely require combination of PARPi with other agents.

genomics↗