Multiscale Modeling of Platelet Mediated Thrombosis Initiation and Machine Learning (ML) for experimental validation and computational efficacy
Presenting author: Danny Bluestein
Additional authors: Peng Zhang, Jawaad Sheriff, Prachi Gupta, Changnian Han, Peineng Wang, Ziji Zhang, Marvin J. Slepian, Yuefan Deng
Stony Brook University, Stony Brook, NY, US; Sarver Heart Center, University of Arizona, Tucson, AZ, US.
BACKGROUND: We developed a multiscale model (MSM) incorporating dissipative particle dynamics (DPD) and coarse-grained molecular dynamics (CGMD), to describe mechanotransduction events triggered by blood flow in cardiovascular pathologies which may induce initiation of thrombosis via flow-induced platelet activation1-6. This model, tightly coupled to extensive in vitro results of platelet flow1,2, mechanical properties3,4, and shape change5, has been expanded to describe shear-mediated platelet aggregation6 and adhesion.
METHODS: Two machine learning (ML) approaches were developed. [ML-1] Deep Learning for Modeling in-vitro data: DIC microscopy images of aggregating and adhering platelets were captured at up to 1000 fps, and processed to obtain geometric and physical parameters, and input into a neural network model to predict inter-platelet contact area (Fig. 1). [ML-2] Increasing computational efficacy by Adaptive Discretization in Massive Multiscale Modeling: adapting temporal scales to diverse spatial scales with a novel adaptive time-stepping (ATS) algorithm that adapts time stepsizes to the underlying biophysical phenomena. Mesoscale DPD blood flow is simulated with μs-timescale and microscale CGMD platelets are modeled with ns-to-ps timescales. Conceptually, ATS ML corresponds to coarse-graining in time.
RESULTS: Our model describes biophysical properties of platelets. Mean normalized contact area model predictions and in vitro results (0.094±0.021 and 0.092±0.021, respectively) demonstrates that our ML-1 model accurately predicts the contact area for aggregated platelets, and is used to validate the in silico results. The ATS algorithm was compared with traditional single time-stepping (STS) algorithm (Fig 1, ML-2). Computing times using ATS for different phases were cut by 20~75%.
CONCLUSIONS: Our computationally affordable, highly resolved, and validated multiscale modeling framework provides a predictive platform to describe platelet activation, aggregation, and adhesion down to the nanoscales. Our novel ML models can be used to validate simulation predictions and improve the modeling efficiency. Ongoing simulations and experiments evaluate aggregation events with multiple platelets and incorporate GPIbα-vWF interactions for adhesion.
Acknowledgements: This project was funded by the NIH (U01 HL131052, R21 HL096930-01, U01 EB012487, DB).
References:  Zhang, P., et al, Cell Mol Bioeng, 2014.  Gao, C., et al, J Comput Phys, 335:812-827, 2017.  Zhang, P., et al, J Biomech, 2017.  Zhang, N. et al, J Comput Phys, 2014.  Pothapragada, S., et al, Int J Numer Meth Biomed Engng, 2015.  Gupta, P., et al, Cell Mol Bioeng, 2019. (under revision)