Abstract 5233: Towards patient-centric drug discovery: Analyzing drug action in malignant pleural effusions and ascites using high content imaging and deep learning

Abstract

Abstract Introduction: A vast majority of investigational anticancer drugs found to be active in preclinical development later do not show the desired effect in the clinic (Wong et al., Biostatistics 2019). This suggests that currently used preclinical models do not fully recapitulate the complexity of the human disease. The study of drug activity in primary human tissue samples, by contrast, could provide a more immediate picture of the activity of a molecule's effect in a patient. Factors that have so far hampered the use of primary tissue samples for drug discovery and development include access in sufficient quantity as well as robust analytical methods. We hypothesised that malignant pleural effusions and ascites (MPAs) of solid tumour patients are a promising model system to study drug activity in a preclinical setting. They are easily accessible in large quantities and contain cancer cells as well as major recruited immune populations. The latter could render them interesting model systems for studying I/O drug activity. Following previous successes in studying drug action in primary human tissues of haematological cancer patients with automated microscopy (Snijder et al 2017, Lancet Haem, NCT03096821) we here describe advances in using high content imaging and deep learning-based image analysis to study drug action in MPAs of solid tumour patients. Methods: MPAs from patients with metastatic breast, pancreatic, lung and ovarian cancer (at least n=10 of each) were collected under appropriate ethics approval. The response of EpCam+/CD45− and CD45+ cells against small molecule drugs was evaluated using high content microscopy. Drug response was quantified with single cell resolution using regional convolutional neural networks (R-CNNs) comprising an object detection and a single cell classification stage. EpCam+ and live/dead cell classification accuracies on the validation set were >94%. Results: MPAs contain both cancer cells (range: 1-37%) and recruited myeloid and lymphoid immune populations with varying activation states (e.g. CD69+/PD-1+ CD8+ T-cells). Ex vivo drug responses from each patient sample were measured and combined to form an overall map of drug susceptibility across the patient population, pan-indication. Sensitivity mirrors drug approvals and also reveals drugs with potential off label use. Conclusions: Single-cell phenotypic analysis of MPAs enables the study of anticancer drug action in a setting that is one step closer to the clinic than cell line or outgrown organoid models of solid tumor. While initial response patterns can be observed that mirror current approvals, further biological and clinical validation must occur to understand in how far these data can be used for drug discovery and translational research purposes. Application as a functional tool for selecting salvage therapies for late-stage solid tumor patients (“functional prevision medicine”) can also be envisioned. Citation Format: Robert Sehlke, Christina Taubert, Isabella Alt, Florian Rohrer, Elisabeth Fuchs, Bojan Vilagos, Nikolaus Krall, Minichsdorfer Christoph, Michael Schumacher, Dominik Maurer, Klaus Hackner, Judith Lafleur, Lukas Hefler, Thorsten Fureder, Gregory Ian Vladimer. Towards patient-centric drug discovery: Analyzing drug action in malignant pleural effusions and ascites using high content imaging and deep learning [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 5233.