Utilizing Machine Learning to Analyze the Effects of PI3K Inhibition on Shear-Mediated Platelet Kinematics and Adhesion

The introduction of medical devices, such as stents and grafts, into the body’s circulation is commonly used to restore or enhance physiological function disrupted by cardiovascular diseases. However, their surface chemistry promotes the adsorption of proteins, such as von Willebrand factor (vWF), a glycoprotein found in blood plasma, leading to platelet adhesion and aggregation, making these foreign materials prone to thrombosis. In the initial stages of clot formation, shear stress from blood flow causes platelets to flip on their edge before firmly adhering to vWF-coated surfaces and encouraging coagulation. One promising strategy to prevent flow-induced thrombosis is modulating PI3K, as previous literature suggests that it functions as a hub for mechanotransduction and that its inhibition impedes platelet coagulation. Therefore, it is crucial to analyze platelet behavior in the context of PI3K. Previously, platelet adhesion has been analyzed via flow cytometry, surface density, etc. Peak platelet flipping velocity during initial adhesion may help describe and predict trends in subsequent morphology change, stable adhesion, multi-platelet aggregation, and clot formation. Recent machine learning methods such as unsupervised image classification and segmentation have provided a standardized and reliable method for analysis[1]. This study aims to design a test bed for assessing the effects of drugs on platelet activity and integrate a novel machine-learning model to analyze platelet flipping under different flow conditions.<br/>Custom microchannels constructed of polypropylene, a hydrophobic polymer, simulated foreign insertions. These rectangular flow channels measured 1 mm in width and 100 µm in height, and were pre-coated with vWF. Blood samples were collected from healthy adult humans (IRB2024-00008), centrifuged, and filtered to obtain purified gel-filtered platelets (GFP). GFP were then separated into control and treatment groups, with the latter receiving TGX-221, a PI3K inhibitor. The flow channels were connected to a syringe pump, allowing GFP to flow under a constant shear of 15 dynes/cm2. Platelets were imaged with an inverted DIC microscope, recording 6-second videos of flipping incidents at 1000 frames per second. Video frames were segmented and binarized to format the data for automated analysis. A semi-unsupervised learning system, consisting of a series of convolutional neural networks that process image data to identify and characterize platelet morphology, was used. Using a policy and reward network to train the model allowed for increased accuracy in platelet image segmentation. These images were then used to measure other geometric parameters such as platelet surface area, major axis length, minor axis length, and thickness. Using these parameters, a graph illustrating the relationship between the time and the platelet’s rotational angle was obtained for each flip. After removing outliers and utilizing a wavelet denoising technique, a polyfit regression was applied to each graph to determine peak flipping velocity.<br/>Results revealed that PI3K inhibition in platelets caused faster flipping, with the average peak velocity increasing by 49.28% compared to untreated samples (n = 11), indicating reduced platelet adhesion and clotting. This suggests that PI3K inhibition potentially mitigates life-threatening thrombosis promoted by blood-contacting foreign materials. This study also highlights the efficacy of the experimental design as a platform for testing drug delivery on vascular surfaces, allowing for the application of artificial intelligence for standardized characterization of thrombotic events. These results are promising for future platelet and coagulation research, necessitating further sampling and investigation.<br/>We would like to acknowledge the Louis Morin Charitable Trust for their support in our research.<br/>[1] Sheriff, J., et al. <i>Ann Biomed Eng </i>49, 3452–3464 (2021).<b> </b>