Applying Machine Vision To Empower Preclinical Development of Immunotherapies in Patient-Derived Organoid Models of Solid Tumors

Authors Chi-Sing Ho, Sonal Khare, Madhavi Kannan, Michael Streit, Timothy Lopez, Luca Lonini, Brian Larsen, Brandon Mapes, Jenna Shaxted, Martin Stumpe, Ameen Salahudeen and Jagadish Venkataraman

Background Paradigm shifting next-generation immuno-oncology therapeutics such as CAR-T cells and bispecific engagers are rapidly gaining interest as investigational therapies. However, the discovery and preclinical development of these therapies involve 2-D cell line based assays that require differential labeling, often with fluorescent dyes or transgenic fluorescent proteins. These conventional approaches are limited by cumbersome labeling, optimization for each tumor cell line, and dysregulation of effector cell function including primary immune cells. In this work, we build upon a patient-derived tumor organoid (TO) platform to measure TO-specific responses to next-gen immunotherapies using brightfield-only segmentation models. We utilize machine vision on time-lapse microscopy to obtain multiparameter kinetic readouts of tumor cell death, enabling dissection of mechanisms of both CAR-T cells and effector cell engaging bispecifics. The assay co-localizes target and effector cells spatiotemporally without the need for differential labeling, affording insights into therapeutic potency or resistance. Overall, the automated segmentation and co-localization pipeline enables a highly scalable and high-throughput imaging assay for rapid discovery and validation of candidate next-gen immunotherapies across all solid tumors.

Methods TOs were co-cultured with immune effector cells, optionally in the presence of a drug candidate, and imaged via a confocal microscope as a series of time-lapse images (figure 1A). We independently trained two fully-connected networks (FCNs) to segment the TOs and the immune cells directly from the brightfield channel. We then applied the TO masks to a registered vital dye channel to quantify TO cell death. Simultaneously, we used the TO masks, in conjunction with the immune cell masks from the second FCN, to quantify immune cells surrounding and infiltrating the TOs as a function of time (figure 1B).

Results We find that the rate of immune cell co-localization and infiltration is correlated with TO cell death over time, lending strong biological interpretability to the effectiveness of immunotherapies. We observe a differentiation of response between HER2-targeted and untargeted CAR-T lines, where targeted CAR-T lines exhibit higher infiltration rates with higher corresponding cell death rates. These correlations generate label-free insights into the pharmacokinetics and mechanisms for specific immune therapies.

Conclusions We present an effective solution for scalable, label-free quantification of TOs and immune cells from brightfield images to understand their dynamics in time-lapse imaging. Our machine vision platform enables high-throughput immune oncology preclinical studies to screen and mechanistically probe therapeutic candidates across dozens to hundreds of unique TO models, accelerating their evaluation as immuno-oncology therapeutic candidates in cancer patients.

Ethics Approval Human biospecimens and effector cells were obtained from third parties under IRB approved research protocols.