Using Computer Vision To Resolve Proliferative Dynamics Within Therapeutic Responses in Large-Scale Screens of Patient-Derived Models

Authors Madhavi Kannan, Brian Larsen, Chi-Sing Ho, Jagadish Venkataraman, Martin Stumpe, Ameen Salahudeen

Background – IC50, Emax, and AUC values are commonly used to assess susceptibility for a given therapeutic candidate but are subject to the number of cell divisions for a given in vitro model. When comparing multiple in vitro models and especially patient-derived models such as patient-derived tumor organoids, the degree of intrinsic cell proliferation or doubling time substantially differs and therefore limits conventional metrics as above. Furthermore, assays that directly measure cellular viability can oftentimes overlook therapeutic agents that have cytostatic, versus cytotoxic mechanisms of action and require assays with similar sensitivity and precision to account for this phenotype. Finally, terminal endpoint assays can require tedious temporal optimization to capture the full dynamic range of a given therapeutic which makes
cross-therapeutic and cross-model comparisons difficult to interpret. An assay that incorporates these multiple features and in vitro specific doubling time would lead to highly accurate measurements of drug effects and allow the calculation of growth-adjusted IC50 values i.e. GI50 values.

Methods – We, therefore, sought to develop a computer vision model that utilizes label-free longitudinal light microscopy images as input to report the total number of nuclei present within a given experimental well, allowing for the monitoring of cell division and any cytostatic effects. To train the neural network, we use Hoechst-stained images as our ground truth. The model used is an extension of the Regularized Conditional Adversarial Network (PMID: 34320344). It takes as input Brightfield images along with the registered Hoescht stained wells. The label for each well is obtained using commercial image processing software to count the total number of nuclei present in each well from the Hoescht stained channel. The model was trained to predict the virtually stained Hoescht well for every Brightfield, as well as the aggregate nuclei count for all organoids present inside the well. The model was trained on 29000 individual sites across 9 distinct cancer types.

Results – Five experiments were conducted during training to test for bit depth and tonality effects of the input training data. Inference on 9240 images showed a pearsonR score of 0.81 on the best-performing model. Results show that the model performs well and can be applied in real-time using only experimental data to correct for differing cell division rates and measure cytostatic phenotypes all while using simple light microscopy.

Conclusions – The tool will enable cross-comparison of different therapeutic MoAs as well as enable cross-cancer type/indication comparison for a therapeutic in development to inform early development clinical strategy.