Accelerating Quantitative Systems Pharmacology with Machine Learning - SMB 2021 скачать в хорошем качестве

Accelerating Quantitative Systems Pharmacology with Machine Learning - SMB 2021

Scientific machine learning (SciML) is the burgeoning field combining scientific knowledge with machine learning for data-efficient predictive modeling. We will introduce the Julia SciML ecosystem by describing some of its recent advances, showing how the GPU-accelerated differential equation solvers gave 175x acceleration on Pfizer's internal C-based QSP models and the 15,000x acceleration seen by the NASA Launch Services upon switching from Simulink to ModelingToolkit.jl. After describing the advances in differential equation solvers and automated model discovery, we will describe the JuliaSim simulation ecosystem and its ability to use continuous-time echo state networks (CTESNs) for automatically generating surrogates of highly stiff QSP models. This technique is shown to be validated on a wide variety of models by using CellML and SBML imports to automate the surrogate training process on ~1000 models. Using the Robertson chemical reaction network as an example case, we will see how multi-layer perceptrons (MLPs), recurrent neural networks (RNNs), Long short term memory networks (LSTMs), and physics-informed neural networks (PINNs) all fail to adequately train while only the CTESN succeeds in building a stable surrogate. Examples of accelerating simulations by over 560x over the Dymola Modelica compiler will showcase the scalability of the technique. The will showcase how JuliaSim composes with tools like Pumas to bridge QSP into clinical pharmacology. We will end by describing new adjoint techniques which are required to build neural ODE surrogates on stiff ODE models. Together this showcases the practical changes users of the JuliaSim ecosystem are seeing through scientific simulation From: Mathematical and computational methods to augment the reliability of biological models for better decision-making, SMB 2021 http://schedule.smb2021.org/MS06/MFBM... Related work: https://arxiv.org/abs/2105.05946 https://arxiv.org/abs/2010.04004 https://arxiv.org/abs/2103.05244 https://juliacomputing.com/products/j...