Senior Application Scientist Metafora Biosystems Inc. Paris, Ile-de-France, France
Abstract: Recent successes of a multitude of immunotherapies ranging from chimeric antigen receptors, cancer vaccines, and immunomodulators has resulted in a mountain of flow cytometry (FCM) data. Unlike many other scientific outputs, FCM data are inherently complex and often highly variable, requiring resident cytometrists with significant expertise to plod through the piles of data. These constraints significantly reduce the speed of interpretation of results; worse, they can disguise underlying technical problems, lead to false positives or negatives. Taken together, the slow processing of FCM data and the perils associated with misinterpretation, pose costly risks to any high throughput application requiring flow cytometry.
To increase the speed of FCM data analysis, several automated solutions have been developed over the last decade. Algorithms like FlowSOM, Phenograph, support vector machines, and others have attempted to improve data analysis automation. Some success has been achieved but only when stringent standardization procedures have been implemented, when the panel employed is not altered, and the algorithmic settings have been fine-tuned to the specific data set. Despite following standardized protocols, high throughput flow cytometry often encounters issues with panels, whether in the form of compensation problems or the common requirement to alter a given reagent’s concentration or composition. Introduction of changes, even minute alterations, force the user to re-fine-tune and reinterpret results generated from algorithmic outputs. This negates the proposed time savings and introduces another analysis bottleneck requiring additional expertise. Hence the practical application of these tools to a wide variety of FCM data is abrogated.
In order to enable high throughput flow cytometry applications, overcome algorithmic shortcomings, eliminate the time sink requirement for repetitive fine-tuning for a variety of FCM data, we propose the implementation of augmented intelligence (AI); a unique suite of interactive algorithms specifically designed with flexibility and parallel processing power to address FCM data variability and the sheer volume of data. Though the AI engine is complex- it consists of several supervised and unsupervised integrated modules- there are three components that have been designed specifically to enhance accessibility to results derived from automated analysis solutions. A key component known as Autofocus, is a dynamic, integrated tool that enables the user to visualize automated gating outputs at biologically meaningful depths., The AI suite comes with a built-in auto-labeling mechanism to make output populations immediately human-understandable, reducing the time-consuming requirements of reviewing a multitude of 2D plots, entangled line graphs, or heatmaps. Finally, a machine learning module “learns” the prescribed heuristic of selected populations and applies them to other files despite significant disparities in population frequencies and reasonable discrepancies in fluorescence intensity. Significantly, the module nullifies the requirement to analyze concatenated files, permitting an AI-driven approach to high throughput screening, longitudinal studies and large data sets.
