High-Throughput, Label-Free, Unsupervised Image Analysis of CAR-T Cell Data

Project Summary Over the last few decades, there has been a paradigm shift in the pharmaceutical industry with compa- nies transitioning from developing small molecule drugs to biologics. Biologic drugs have several advantages, the primary one being that biologic drugs are better suited to high target specificity (traditional small molecule drugs often exhibit non-target side effects). Some biologics show great promise as a personalized medicine, e.g. chimeric antigen receptor (CAR)- T cell therapies. CAR-T cell therapies are a type of biologic falling under the cell based medicinal product (CBMP) category, where a patient's own T cells can be modified and used as a drug. In 2017, the first CAR-T cell based therapy was approved by the U.S. FDA; so far CAR-T therapies have demonstrated success in treating lymphomas, leukemias, and multiple myelomas, but active research will expand the scope of CAR-T therapies greatly within in the next few years. Despite their many benefits and common use, biologic drugs can still elicit serious adverse side-effects. Many of these side-effects are believed to be caused by small particle aggregates inherent in multiple biologic drugs. For CBMPs, the regulation issue is even more complex since the drugs (often human cells) are comparable in size to particle impurities historically monitored to mitigate adverse side-effects. Furthermore, modern CBMP therapies are currently very expensive, 500K-1M USD per treatment, and face nontrivial quality control challenges in both autologous and allogeneic formulations. Pharmaceutical companies are currently required to record and catalog vast volumes of particle data of biolog- ics, but are only mandated under FDA regulations (i.e., USP 788 ) to control the number of particles exceeding 10 and 25m in delivered products. Historically, particle size and counts have been used to analyze subvisible particles (those 100m in size) in biologics, but the many images of the particles inherent in biologics contain a rich amount of “morphological information” which can be extracted using convolutional neural networks (CNNs) in combination with computational statistics as we have shown in our previous work in collaboration with academia, industry, and government agencies (FDA and NIST). This proposal will expand our previous supervised CNN algorithms to include unsupervised machine learn- ing algorithms for analyzing particle images measured in biologics using label-free high-throughput microscopy. Unsupervised machine learning algorithms show promise in situations where multiple biologic drug formulations exhibit highly heterogeneous particle populations that may change over time. This particular situation is com- monly encountered in batch biologic drug production. In Phase I, we will develop new algorithms and apply these algorithms to characterize T cells modified by lentivirus transduction under good manufacturing practice (GMP) conditions with the intent to optimize process parameters improving yield of viable CAR-T cells. Finding process parameters producing a large quantity of consistent viable CAR-T cells will aid in driving down production cost.