1. Current IMAG Meeting
  2. 2023 Agenda
  3. Oral Presentation 2 (IMAG-AND Futures)

Oral Presentation 2 (IMAG-AND Futures)

1:40-2:00 pm              “Integrating machine learning in multiscale modeling for blood flow and platelet-mediated thrombosis initiation”

Peng Zhang, Stony Brook University

Peng Zhang photoBIO: Peng Zhang is a research scientist at Stony Brook University. He received his Ph.D. in Applied Mathematics from Stony Brook University, after completing his M.S. in Parallel Computing and B.S. in Mathematics from Nankai University with honors. His research focuses on the development of efficient and accurate multiscale modeling (MSM) and machine learning (ML) approaches for modeling the blood flow and platelet mediated thrombosis. He published 30+ papers in applied mathematics, high performance computing, biomedical engineering, plus three book chapters. He has five awarded patents in US and China. He received two XSEDE Research Awards in 2014 and 2015. He presented 20+ lectures at international conferences and seminars such as BMES and SB3C. His key contributions in multiscale modeling include: (1) Development of multiscale particle-based modeling framework for blood flow and platelet activation, aggregation and adhesion, by interfacing coarse grained molecular dynamics (CGMD) and dissipative particle dynamics (DPD). (2) Development of a semi-unsupervised learning system for platelet segmentation at the submicron resolution and a ML model for synthesizing the sparse segmentation data to predictive model for enhancing the model assessment. (3) Development of multiple time stepping algorithms to handle 3–4 orders of magnitude disparity in the temporal scales between DPD and CGMD. The numerical experiments demonstrated 3000x reduction in computing time over standard methods for solving multiscale models. (4) Development of ML-based methods for adapting time step sizes to the underlying biomedical events in massive multiscale simulations. The computing times can be further cut by 20~75% automatically.

Peng Zhang.pdf172.58 KB

Comment

Comment

Can your model be adapted to different materials/chemistries as the interacting surfaces with the platelets - to predict propensity of the different surfaces to promote thrombosis or not

Submitted by David (not verified) on Tue, 03/17/2020 - 13:54

Comment

Yes. One of our ongoing efforts is the study of the platelet adhesion on different material surfaces by simulating the rolling and attaching dynamics of platelets.

Submitted by Peng Zhang (not verified) on Tue, 03/17/2020 - 14:08

Comment

David- adding to Peng's reply to you. the model can be modified to simulate interactions with different materials by modulating/adjusting the many parameters that characterize the platelets interactions with a surface (number of active receptors on the platelet membrane, Fg interaction characteristics, vWF interactions- with the surface on one side and with at attaching platelets on the other side, etc., etc)

Submitted by Danny Bluestein (not verified) on Tue, 03/17/2020 - 14:49

Comment

Beautiful model! A unique strength of your approach is feedforward up through the lengthscales. Could you say a few words about the current state of feedback in your model? For example, can one currently model how flow perturbations from platelet deformation affect the force-dependent binding kinetics of individual vWF molecules? Thanks! Guy Genin

Submitted by Guy Genin, Was… (not verified) on Tue, 03/17/2020 - 13:55

Comment

The model can feedback meaning that after a platelet changes its morphology, we coarse grained the CGMD-based platelet geometry via interactions with much larger-scale DPD flows. This is the way we change the platelet morphology (We have a special interface function that takes care of the CGMD-DPD interactions).

Different scenarios "single platelet attaching to the wall comparing with multiple platelets attaching to the wall", we plan to incorporate different binding kinetics of individual vWF molecules according to the local interaction of the aggregate with the wall.

Submitted by Peng Zhang (not verified) on Tue, 03/17/2020 - 14:24

Comment

Once aggregation processes have occurred, does the dynamic time step stay "stuck" as the fast time scale? 

Does the compute cost / complexity for the adaptive time step truly pay off over optimizing for a fixed time step size at that point? 

Submitted by mathcancer on Tue, 03/17/2020 - 13:57

Comment

Great talk!! Research from Roger Kamm's group has shown that tethers are important in tumor metastasis (adhesion of circulating tumor cell to the vessel endothelium), and clusters of circulating tumor cells have different metastatic potential -- is it possible to adapt your modeling tools to simulate the (different) context of circulating tumor cells?  

Submitted by Shayn Peirce-Cottler on Tue, 03/17/2020 - 13:57

Comment

The time step does not "stuck". For example, if the aggregate dynamics suddenly change now, the Machine Learning tells the algorithm to keep adjusting to smaller time steps if suddenly the dynamics becomes faster.

We compared our Adaptive Time Stepping method with a fixed time step size: we achieved 20~75% reduction in computing time (previously we have developed a multiple time stepping (MTS) method that can achieve a much higher speed up [1].)

[1] Zhang, P., Zhang, N., Deng, Y., Bluestein, D., "A Multiple Time Stepping Algorithm for Efficient Multiscale Modeling of Platelets Flowing in Blood Plasma", Journal of Computational Physics, vol. 284, pp. 668-686, 1 March 2015. DOI: 10.1016/j.jcp.2015.01.004

Submitted by Peng Zhang (not verified) on Tue, 03/17/2020 - 14:33

Comment

Definitely, our method can be adapted to other processes like tethers in tumor endothelium. But We have not done this yet.

Submitted by Peng Zhang (not verified) on Tue, 03/17/2020 - 14:34

Comment

Can you please discuss the challenges associated with implementing active learning for synthesizing sparse in vitro data.

Submitted by nkkchem on Tue, 03/17/2020 - 13:58

Comment

The challenges associated with implementing active learning include: (1) the very limited amount of labeled data (human labeling is time consuming); (2) the labels from different human experts are also different (inconsistent labels); and (3) the high background noise from in vitro data (high noise). These challenges drove us to design the three layered learning system. Thank you.

Comment

Very interesting talk! The raw data in the model you used from are all collected in vitro if this model also fitted to in vivo data of intravital imaging?Thanks a lot.

Submitted by Qiaoya Lin (not verified) on Tue, 03/17/2020 - 14:01

Comment

We intentionally use the well-controlled in vitro data to validate our numerical model and help the model to learn in different scenarios - it is becoming more and more complex. We are not there yet - we believe in the future the model combined with machine learning will be able to cope with the in-vivo data. But this is a long way to go.

Comment

Thanks! Looking forward to your future work. I am doing intravital imaging of nanomedicine, and I have a lot of in vivo data that need to be stimulated.

Submitted by Qiaoya Lin (not verified) on Tue, 03/17/2020 - 15:05

Comment

Thank you very much for your questions and comments.

If you have further questions, please feel free to contact me at peng.zhang@stonybrook.edu

Peng Zhang, Ph.D.

Stony Brook University, NY

Submitted by penzhang on Tue, 03/17/2020 - 14:53

Comment

Nice article. Great information.
Insideaiml is one of the best platforms where you can learn Python, Data Science,
Machine Learning, Artificial Intelligence & showcase your knowledge to the outside world.
Visit us at https://insideaiml.com/home

Submitted by shilkaren (not verified) on Thu, 09/10/2020 - 08:03

Comment

<a href="https://albakh.ru">can i buy allegra without a prescription</a>

Submitted by Richardyve (not verified) on Fri, 10/23/2020 - 06:47

Comment

Meds information. What side effects can this medication cause? <a href="https://lisinopril17.top">can you buy generic lisinopril price</a> Everything information about drug. Read now.

Submitted by JosephVah (not verified) on Sat, 11/07/2020 - 09:19

Comment

Medication information for patients. Long-Term Effects. <a href="https://medicals.top">order lexapro pills</a> Everything about medication. Read information here.

Submitted by Georgegoons (not verified) on Mon, 12/14/2020 - 01:43