Review Special Issues

Cellular mechanotransduction

  • Received: 21 August 2015 Accepted: 11 January 2016 Published: 20 January 2016
  • Cell adhesion and cell–cell contacts are pre-requisites for proper metabolism, protein synthesis, cell survival, and cancer metastasis. Major transmembrane receptors are the integrins, which are responsible for cell matrix adhesions, and the cadherins, which are important for cell-cell adhesions.  Adherent cells are anchored via focal adhesions (FAs) to the extracellular matrix, while cell-cell contacts are connected via focal adherens junctions (FAJs). Force transmission over considerable distances and stress focusing at these adhesion sites make them prime candidates for mechanosensors. Exactly which protein(s) within FAs and FAJs or which membrane component of ion channels sense, transmit, and respond to mechano-chemical signaling is currently strongly debated and numerous candidates have been proposed.

    Citation: José Luis Alonso, Wolfgang H. Goldmann. Cellular mechanotransduction[J]. AIMS Biophysics, 2016, 3(1): 50-62. doi: 10.3934/biophy.2016.1.50

    Related Papers:

    [1] Hyeontae Jo, Hwijae Son, Hyung Ju Hwang, Eun Heui Kim . Deep neural network approach to forward-inverse problems. Networks and Heterogeneous Media, 2020, 15(2): 247-259. doi: 10.3934/nhm.2020011
    [2] Qing Li, Steinar Evje . Learning the nonlinear flux function of a hidden scalar conservation law from data. Networks and Heterogeneous Media, 2023, 18(1): 48-79. doi: 10.3934/nhm.2023003
    [3] Yinlin Ye, Yajing Li, Hongtao Fan, Xinyi Liu, Hongbing Zhang . SLeNN-ELM: A shifted Legendre neural network method for fractional delay differential equations based on extreme learning machine. Networks and Heterogeneous Media, 2023, 18(1): 494-512. doi: 10.3934/nhm.2023020
    [4] Travis G. Draper, Fernando Guevara Vasquez, Justin Cheuk-Lum Tse, Toren E. Wallengren, Kenneth Zheng . Matrix valued inverse problems on graphs with application to mass-spring-damper systems. Networks and Heterogeneous Media, 2020, 15(1): 1-28. doi: 10.3934/nhm.2020001
    [5] Yinlin Ye, Hongtao Fan, Yajing Li, Ao Huang, Weiheng He . An artificial neural network approach for a class of time-fractional diffusion and diffusion-wave equations. Networks and Heterogeneous Media, 2023, 18(3): 1083-1104. doi: 10.3934/nhm.2023047
    [6] Wenkai Liu, Yang Liu, Hong Li, Yining Yang . Multi-output physics-informed neural network for one- and two-dimensional nonlinear time distributed-order models. Networks and Heterogeneous Media, 2023, 18(4): 1899-1918. doi: 10.3934/nhm.2023080
    [7] Lili Jia, Changyou Wang, Zongxin Lei . Synchronization of nonautonomous neural networks with Caputo derivative and time delay. Networks and Heterogeneous Media, 2023, 18(1): 341-358. doi: 10.3934/nhm.2023013
    [8] Feiyang Peng, Yanbin Tang . Inverse problem of determining diffusion matrix between different structures for time fractional diffusion equation. Networks and Heterogeneous Media, 2024, 19(1): 291-304. doi: 10.3934/nhm.2024013
    [9] L.L. Sun, M.L. Chang . Galerkin spectral method for a multi-term time-fractional diffusion equation and an application to inverse source problem. Networks and Heterogeneous Media, 2023, 18(1): 212-243. doi: 10.3934/nhm.2023008
    [10] Antonio Bilotta, Emilio Turco . Constitutive modeling of heterogeneous materials by interpretable neural networks: A review. Networks and Heterogeneous Media, 2025, 20(1): 232-253. doi: 10.3934/nhm.2025012
  • Cell adhesion and cell–cell contacts are pre-requisites for proper metabolism, protein synthesis, cell survival, and cancer metastasis. Major transmembrane receptors are the integrins, which are responsible for cell matrix adhesions, and the cadherins, which are important for cell-cell adhesions.  Adherent cells are anchored via focal adhesions (FAs) to the extracellular matrix, while cell-cell contacts are connected via focal adherens junctions (FAJs). Force transmission over considerable distances and stress focusing at these adhesion sites make them prime candidates for mechanosensors. Exactly which protein(s) within FAs and FAJs or which membrane component of ion channels sense, transmit, and respond to mechano-chemical signaling is currently strongly debated and numerous candidates have been proposed.


    This paper marks the first step toward a comprehensive study on deep learning architectures to solve forward-inverse problems for differential equations. Recent advances in deep learning show its capability to handle various types of model problems in many disciplines. In particular, deep learning techniques have been applied to understand data augmented differential equations. While most of such studies have been centered around heuristics and modeling prospectives, to the best of our knowledge, there is little to no theoretical analysis to confirm whether the deep learning architectures give rise to the correct solutions to the governing differential equations. An overreaching goal of this paper is to provide a comprehensive analysis of Deep Neural Networks (DNNs) to solve data-driven differential equations. This paper reports a novel architecture leveraging recent progress in deep learning techniques that solves forward-inverse problems for differential equations. The paper further includes a convergence analysis and its experimental results for our deep learning architecture.

    Forward-inverse problems (or inverse problems in short) for differential equations in this paper are related to data augmented differential equations. Namely, we consider equations of states for physical systems as governing differential equations and model parameters such as advection rates, reaction diffusion coefficients, for example, that need to be fitted by the given data. Hence numerical methods solving forward-inverse problems typically become constraint problems that require an ensemble of two steps, (1) solve the state equations, which is called the forward problems and (2) find the correct model parameters that fit the given data set. Inverse problem is an actively studied field and many numerical algorithms for the inverse problems are robust enough to handle sufficiently large data sets, see for example [1,5,15,17,19,25,26] references therein. However, such algorithms can be computationally expensive and they may be too sophisticated for non-experts in inverse problems to implement them. This calls for simplistic methods that unify two steps in solving forward-inverse problems.

    The contributions of this paper are three-fold. First, the DNN architecture presented in this paper highlights its simplistic approach to handling forward-inverse problems simultaneously. Second, a rigorous analysis of the convergence of the DNN solutions to the actual solutions for the governing problems is provided. Third, numerical experiments validate the robustness of our simple architecture.

    The paper comprises the followings. A short overview of related works on data-driven differential equations and the problem formulation are presented in the rest of Section 1. The methodology including the DNN architecture and loss function is described in Section 2. Theoretical results are given in Section 3. Section 4 is devoted to the experiments done for the problem. In Section 5, we conclude the paper.

    There are many works to utilize an Artificial Neural Network (ANN) to solve Differential Equations (DEs) in place of using well established numerical methods such as finite difference, finite element, and finite volume methods. Those finite schemes are heavily depending on mesh-grid points, and they may become a hindrance when the state equations reside in a domain with complex geometry. As such, a mesh-free approximation using clever constructions of basis functions has been introduced, see for example [10,24], and references therein. The concept of using ANN to solve DEs can be related to mesh-free approximations as one can consider Multi-Layer Perceptrons (MLPs) as an approximation to solutions to DEs.

    The ANN structure was first introduced in [20]. Several studies follow to identify a class of functions that can be recognized by ANNs. The following is a brief overview and a few highlights of such studies. Cybenko [8] established sufficient conditions for which a continuous function can be approximated by finite linear combinations of single hidden layer neural networks with the same univariate function. About the same time, Hornik et al [11] showed that measurable functions can be approximated by multi-layer feedforward networks with a monotone sigmoid function as the authors called them "universal approximators". Cotter [6] extends the result of [11] to new architectures, and later Li [18] showed that a MLP with one hidden layer can approximate a function and its higher partial derivatives on a compact set.

    The concept of using ANN to solve DEs are not new, and it has gained much attention recently. Some of the highlights include the following. Lagaris et al [13] studied to solve DEs and PDEs using an ANN with architecture including 1 single layer and 10 units and they next extended in [14] their results to a domain with complex geometry. Jianyu et al [16] used ANN with a radial basis function as an activation function for Poisson equations. More recently, Berg et al in [4] used a DNN to solve steady (time-independent) problems for a domain with complex geometry in one and two space dimensions, and later in [3] they studied DNN architectures to solve augmented Poisson equations (inverse problems) including three space dimensions.

    The recent work by Raissi et al [22] can be perhaps closely related to our work in the sense that their DNN architectures (they called "continuous time models") resemble ours. We note however the aim of this paper is to establish a comprehensive study that includes a rigorous convergence analysis of DNNs to solve forward-inverse problems for DEs and PDEs. Since our convergence result is for linear equations at this point, we present our experiments for the well known linear equations as well. While our experiments cover simpler equations than those studied in [22], as their focuses were on architectures for "data-efficient and physics informed learning machines", we hope that our result can enhance the experiments shown in [22]. We believe our first comprehensive results can shed lights onto further studies toward more complex and complicated systems using machine learning architectures.

    We consider the equations of states as the following time dependent initial boundary value problems:

    $ Lpu=0, t(0,T], xΩ,
    $
    (1)
    $ Iu=f, xΩ,
    $
    (2)
    $ Bu=g, t(0,T], xΩ,
    $
    (3)

    where $ L_{p} $ is a differential operator, $ p $ is a set of model parameters, $ \Omega\subset\mathbb{R}^d $ is a bounded domain (for the position of the state $ u $), $ \partial\Omega $ is a boundary of $ \Omega $, $ f(x) $ is an initial distribution, $ I $ is an initial operator, $ B $ is a boundary operator, and $ g(t,x) $ is a boundary data. Next the governing equation is equipped with a set of observation data, which may be provided from actual experiments, as following:

    $ D={(ti,xj,uij)|i=1,2,...,n, j=1,2,...,m},
    $
    (4)

    where $ u_{ij} $ denotes the output value (observation from the experiment) at position $ x_{j}\in\Omega $ and time $ 0<t_i\leq T $ with the final time $ T $. $ n $ and $ m $ refer the numbers of time and spatial observations respectively. Since our experiments cover ODEs and PDEs with one and two spatial dimensions, $ x_j $ dependencies and the index $ j $ for $ x_j $ will be adjusted for each example. Since it is hard to find a realistic data, we sample the observation data from analytic (or numerical) solution of the equations.

    We apply DNNs to solve governing equations (1)-(3) and the observation data (4). Our loss function includes the governing equation, observations, initial and boundary conditions. The loss function is minimized by using the standard gradient descent method. The results show that our DNN architecture can handle much lesser numbers of observations compared to known numerical solvers such as finite difference, element and volume methods. The results presented in this paper demonstrate that our DNN architecture is perhaps the most simplistic way to solve forward-inverse problems, yet robust enough to handle many different cases. Furthermore, we establish the convergence result of forward-inverse problems for parabolic or hyperbolic linear PDEs with Dirichlet boundary conditions. Specifically we show that the sequence of DNN solutions converges to an analytic solution in $ L^{\infty}(0,T;H_0^{1}(\Omega)) $ sense.

    This section provides our DNN architecture and the mathematical formulation of the training procedure.

    The DNN architecture can be written as a repeated composition of some affine transformations and some nonlinear activation functions. We denote by $ u_N $ the DNN solution and assume that the architecture consists of $ L + 1 $ layers. The first layer takes $ (t, x) $ as an input, and the last layer gives the output value $ u_N(t,x) $. The $ L-1 $ layers between the first and the last layers are called hidden layers. We have used common nonlinear activation functions such as sigmoid, rectified linear units, and hyperbolic tangents through the DNN. Each neuron in the DNN contains a bias except for the input neurons. In two consecutive layers, the connections between neurons are written as weight matrices. Relations between $ (l-1)^{th} $ and $ l^{th} $ layers are defined by :

    $ zlj=Nl1kwljkσl1(zl1k)+blj,
    $
    (5)

    where

    $ z^{l-1}_{k} $: the value of $ k^{th} $ neuron in $ (l-1)^{th} $ layer

    $ N_{l-1} $: the number of neurons in $ (l-1)^{th} $ layer

    $ b^{l}_{j} $: the bias of $ j^{th} $ neuron in $ l^{th} $ layer

    $ w^{l}_{jk} $: the weights between $ k^{th} $ neuron in $ (l-1)^{th} $ layer and $ j^{th} $ neuron in $ l^{th} $ layer

    $ \sigma_{l-1} $ the activation function in $ (l-1)^{th} $ layer

    For convenience, we denote $ z^{0} = (z^{0}_{1},...,z^{0}_{N_{0}}) $ as $ (t,x) $ and $ z^{L} = (z^{L}_{1},...,z^{L}_{N_{L}}) $ as $ u_N(t,x) $, respectively. The values $ N_{l-1} $, $ L $, and the form of $ \sigma_{l-1} $ should be chosen before training. In the training procedure, we have to calculate the optimal weights and biases $ w^{l}_{jk} $, $ b^{l}_{j} $ which minimize a suitable loss function.

    The training procedure of the DNN is equivalent to the optimization problem of the loss function with respect to the DNN parameters. We denote the DNN solution by $ u_{N}(t,x) = u_{N}(t,x;w,b) $, where $ (w,b) $ are the set of weights and biases defined in (5). Denote the number of grid points of time, spatial variables, initial and boundary domains by $ N_t, N_x, I, B_t, B_x $ respectively. Now we define the loss function using (1),

    $ LossGE(w,b,p)=1T|Ω|[0,T]Ω|LpuN(t,x;w,b)|2dxdt1NtNxNt,Nxi,j=1|LpuN(ti,xj;w,b)|2,
    $
    (6)

    where the last approximated sum is obtained by sampling a grid point $ \{(t_i,x_j)|t_i\in[0,T],x_j\in\Omega,\text{ for }i = 1,...,N_t, j = 1,...,N_x\} $. Note that the reason that we define the above loss function is to find the optimal weights $ (w,b) $ which minimize $ Loss_{GE} $. However, (6) is not sufficient because it excludes information about initial and boundary conditions. Therefore we define two loss functions from (2)-(3)

    $ LossIC(w,b)=1|Ω|Ω|IuN(0,x;w,b)f(x)|2dx1IIi=1|uN(0,xi;w,b)f(xi)|2,
    $
    (7)
    $ LossBC(w,b)=1T|Ω|[0,T]Ω|BuN(t,x;w,b)g(t,x)|2dSdt1BtBxBt,Bxi,j=1|uN(ti,xj;w,b)g(ti,xj)|2.
    $
    (8)

    Combining all loss functions (6)-(8) is still not enough because the solution of DEs (1)-(3) could differ depending on the choice of the equation parameter $ p $. Due to this reason, we should make one additional loss function to calibrate $ p $ using the observed data (4).

    $ LossObs(w,b)=1|D|(ti,xj)D|uijuN(ti,xj;w,b)|2.
    $
    (9)

    Finally, we define the forward loss and total loss as a summation of all three, and four loss functions respectively. For the forward loss, the model parameter $ p $ is considered to be fixed.

    $ LossForward(w,b)=LossGE(w,b)+LossIC(w,b)+LossBC(w,b),
    $
    (10)
    $ LossTotal(w,b,p)=LossGE(w,b,p)+LossIC(w,b)+LossBC(w,b)+LossObs(w,b).
    $
    (11)
    Algorithm 1: Training
    1: procedure train(number of epochs)
    2:   Initialize the nerural network.
    3:   For number of epochs do
    4:     sample $ z^1, z^2,..., z^m $ from uniform distribution over $ \Omega $
    5:     sample $ z_I^1, z_I^2,..., z_I^m $ from uniform distribution over $ \{0\} \times\Omega $
    6:     sample $ z_B^1, z_B^2,..., z_B^m $ from uniform distribution over $ \partial\Omega $
    7:     sample k observation points $ z_O^1, z_O^2,..., z_O^k $
    8:     Find the true value $ u_j = u_p(z_O^j) $ for $ j=1,2,...,k $
    9:     Update the neural network by descending its stochastic gradient :
    $w,b[1mmi=1[Lp(uN)(zi)2+(uN(ziI)f(ziI))2+(uN(ziB)g(ziB))2]+1kkj=1(uN(zjO)uj)2]
    $

    10:   end for
    11: end procedure

     | Show Table
    DownLoad: CSV

    This section provides a theoretical proof that there exists a sequence of weights such that the corresponding sequence of DNN solutions converges to an analytic solution on any compact subset of the domain. We focus on the DEs (1)-(3) where the existence and the uniqueness of solutions are guaranteed. We establish the result in two steps. We first show that a sequence of DNN solutions converges to an analytic solution for the corresponding model parameters, called the forward problem. We next show that both the estimated parameter and the DNN solutions converge to the model parameter and the analytic solution simultaneously, called the inverse problem.

    For the forward problem, we fix the model parameter $ p $ and denote the analytic solution to (1)-(3) by $ u_p $. We also denote the DNN solution obtained by using Algorithm 1 by $ u_N $. In $ u_N $, activation functions $ \sigma $ are any non-polynomial functions in $ C^{k}(\mathbb{R}^n) $.

    Next we quote the following Definition 3.1 and Theorem 3.2 from [18]

    Definition 3.1. Let $ I_n = [0, 1]^n $ be the unit interval on $ \mathbb{R}^n $, $ k $ be a non-negative integer. Then we say a function $ f $ is contained in $ \widehat{C}^k(I_n) $ if $ f\in C^k(U) $ for some open set $ U $ containing $ I_n $

    Theorem 3.2. (Li, Theorem 2.1 in [18]) Let $ f\in\widehat{C}^{k}(I_n) $. Then, given $ \varepsilon>0 $, we can find parameters of a neural network $ u_N $, defined in (5) with $ L = 1 $, so that

    $ \|D^{\underline{k}}f-D^{\underline{k}}u_N\|_{L^{\infty}(I_n)} < \varepsilon $

    holds for any multi-index $ \underline{k} = (k_1, k_2, ..., k_n) $, $ |k_1|+|k_2|+...+|k_n|\leq k $, and $ k_i $'s are non-negative integers.

    Remark 1. Since the above result can be generalized to multi-layer architectures (for example, [11]) and to a general compact set $ K $ instead of $ I_n $, we may assume that the architecture contains only one hidden layer and the domain $ \Omega $ is $ I_n $ ($ \Omega = I_n $).

    Theorem 3.3. For a non-negative integer $ k $, assume that the highest order of linear operator (1) is $ k $ and $ u_p\in\widehat{C}^k(I_n) $, and the activation function $ \sigma(x) $ and its ($ k $-th order) derivatives are continuous and non-polynomials. Then, there exists $ \{m_j, w_{j}, b_j\}_{j = 1}^\infty $ such that a sequence of the DNN solutions with $ m_j $ nodes, denoted by $ \{u_j(w_j, b_j)\}_{j = 1}^{\infty} $ satisfies

    $ LossForward(wj,bj)0asj
    $
    (12)

    Proof. Let $ \epsilon > 0 $ be given. By Theorem 3.2, there exists a neural network $ u_{j}(x) = \sum_{i = 1}^{m_j} w_i^1 \sigma({w_i^2}x + b_i) $ such that $ \|D^{\underline{k}}u_p - D^{\underline{k}}u_{j}\|_{\infty} < \epsilon $, where $ {\underline{k}} $ is a non-negative multi-index up to differentiability of $ u_p $. By integrating $ |L_pu_{j}|^2\text{, } |Iu_{j}-f|^2\text{, } |Bu_{j} - g|^2 $ over $ [0,T] \times \Omega\text{, } \Omega\text{, } [0,T] \times \partial\Omega $ respectively, we obtain the desired result.

    Remark 2. The assumption $ u_p\in\widehat{C}^k(I_n) $ in Theorem 3.3 is a strong condition. Since we can also approximate it by a sequence of compactly supported smooth functions in $ I_n $, we can extend the assumption to a general Sobolev space.

    The Theorem 3.3 states that we can always find parameters of a DNN architecture which can reduce $ Loss_{Forward} $ if the DE has a smooth analytic solution. However, since we can not directly use information of an analytic solution, we next show that the DNN architecture equipped with parameters which minimize $ Loss_{Forward} $ converges to an analytic solution in Theorem 3.4.

    Theorem 3.4. Let $ L_{p} = \partial_{t}+L $ in (1) be a second order parabolic operator and $ Bu = 0 $ in (3) be a Dirichlet boundary condition. Also we define the DNN solution $ u_{j} = u_{j}(t,x;w_j,b_j) $ with $ m_j $ nodes and the corresponding loss $ Loss_{Forward}(w,b) $. Then, $ Loss_{Forward}(w,b)\rightarrow 0 $ implies

    $ uj(t,x;wj,bj)upinL([0,T];H10(Ω)),
    $
    (13)

    where $ u_p $ is a solution to (1)-(3)

    Proof. First we assume that the activation function satisfies the Dirichlet boundary condition by replacing it with $ b(x)\sigma(t,x) $, where $ b(x) $ is an arbitrary smooth function that satisfies $ b = 0 $ on $ \partial \Omega $. By evaluating $ u_p-u_{j} $ in (1)-(3), we have the following

    $ t(upuj)+L(upuj)=εj(t,x), t(0,T], xΩ,I(upuj)=ηj(x), xΩ,B(upuj)=0, t(0,T], xΩ
    $

    Then the energy estimates for the second order parabolic equation (see Theorem 5, Chapter 7 in [9]) are applied to obtain that

    $ esssup0tTupuj(,t)H10(Ω)+upujL2([0,T];H10(Ω))+t(upuj)L2([0,T];L2(Ω))C(εmL2([0,T];L2(Ω))+ηmH10(Ω)),
    $
    (14)

    where the constant $ C $ in (14) depends only on $ \Omega, T $ and the coefficients of $ L $. Note that the right hand side in (14) is equivalent to $ Loss_{Forward}(w,b) $. This shows that the sequence of DNN solutions converges to the analytic solution when $ Loss_{Forward}(w,b)\rightarrow 0 $.

    Remark 3. The convergence result also holds when $ L_{p} = \partial_{tt}+L $ is a second order hyperbolic operator, see [9].

    Definition 3.5. Let $ P $ be the set of all possible model parameters. Define $ S : = \{u_p\text{ }|\text{ }p \in P\} $ be the set of solutions corresponding to each model parameter $ p \in P $.

    Definition 3.6. We say the observation set $ D_p $ is $ clear $ if for any $ u_{p}, u_{q} \in S $, $ u_{p}|_{D_{p}} = u_{q}|_{D_{p}} $ if and only if $ p = q $

    Theorem 3.7. Let $ Loss_{total}(w,b,p) $ be the total loss defined in (11) and let the observation set $ D_{p} $ with $ p \in P $ be given and clear. We assume that for given $ \{\epsilon_j\}_{j = 1}^{\infty} $ with $ \epsilon_j \rightarrow 0 $, there exists $ (m_j, w_j, b_j, p_j) $ such that $ Loss_{Total}(w_j,b_j,p_j) < \epsilon_j $, and the parameter set $ \{p_j\} $ is contained in $ P $, then

    $ uj(wj,bj)upa.e. andpjpasj
    $
    (15)

    Proof. We first divide the total loss into $ Loss_{Total} = Loss_{Forward, p_{j}} + Loss_{Obs, p_{j}} $. For $ j $ fixed, we set $ u_{N_{j,k}}(w_{j,k},b_{j,k}) $ as a DNN solution with $ k\geq j $ nodes, where $ (w_{j,k}, b_{j,k}) $ is a minimizer of $ Loss_{Forward} $ defined in Definition 12. Then, Theorem 3.4 implies $ u_{N_{j,k}}\rightarrow u_{p_j} $ as $ k\rightarrow\infty $. Also, $ Loss_{Obs, p_{j}}\rightarrow 0 $ implies $ p_j \rightarrow p $ by definition of $ D_p $

    In this section, we provide experimental results based on several differential equations including 1D transport equation, 2D heat equation, 2D wave equation, and the Lotka-Volterra system. For each equation, we have calculated an analytic (if possible) or numerical solution with fixed model parameters in order to generate a small amount of true solution points which will be regarded as the observation points. We apply our DNN model to find an approximated solution and the optimal equation parameter at the same time. In this step, we have used a neural network with variable depth and width, and ReLU activations. We used the Adam optimizer [12,23] with $ (\beta_1, \beta_2) = (0.9, 0.999) $ in order to find the minimizer $ w, b, $ and $ p $ defined in (11). Also, for higher order derivatives in (1) we have applied the reduction of order technique to express it as a system of differential equations. This step dramatically reduces the computational cost for calculating (6). For example, the second-order PDE $ u_{xx} = f $ can be replaced by $ v_x = f $ together with the equation $ v = u_x $. That is, we derive two first-order PDEs $ v_x = f, u_x = v $ from one second-order PDE $ u_{xx} = f $, then the output layer of DNN should be changed into $ (u,v) $. This method is applied to 2D Heat and Wave equations. Also, different types of activation functions are used depending on the behavioral characteristics of the governing equations. Finally, we provide two differences between 1) actual and model output values and 2) actual and calculated model parameters. Observation points were calculated from analytic (transport), series (heat, wave), numerical (Lotka-Volterra) solution and sampled randomly among them.

    In the rest of this section, we present the experimental results. For each figure, top left, top right figures show our neural network solution and the analytic solution respectively. Bottom left shows the absolute error between the neural network solution and the analytic solution. Bottom right figure shows the convergence of estimated parameters to the real parameters. We have implemented our method by using Pytorch [21], which is one of the most famous machine learning library. We first present the detailed experimental settings. Table 1 and 2 show the summarized information of the number of grid points and DNN architectures respectively.

    Table 1.  Information of grid and observation points.
    Data Generation
    Grid Range Number of Grid Points Number of Observations
    1D Transport $ (t,x) \in [0,1]\times[0,1] $ $ 17 \times 100 $ 17
    2D Heat $ (t,x,y) \in [0,1]\times[0,1]\times[0,1] $ $ 100 \times 100 \times 100 $ 13
    2D Wave $ (t,x,y) \in [0,1]\times[0,1]\times[0,1] $ $ 100 \times 100 \times 100 $ 61
    Lotka-Volterra $ t \in [0,100] $ 20,000 40

     | Show Table
    DownLoad: CSV
    Table 2.  Neural network architecture.
    Neural Network Architecture
    Fully Connected Layers Activation Functions Learning Rate
    1D Transport 2(input)-128-256-128-1(output) ReLU $ 10^{-5} $
    2D Heat 3(input)-128-128-1(output) Sin, Sigmoid $ 10^{-5} $
    2D Wave 3(input)-128-256-128-1(output) Sin, Tanh $ 10^{-5} $
    Lotka-Volterra 1(input)-64-64-2(output) Sin $ 10^{-4} $

     | Show Table
    DownLoad: CSV

    1D transport equation consists of

    $ tu+axu=0,u(0,x)={sin40.25π(x0.1)if 0.1x0.50otherwise,
    $
    (16)
    Figure 1.  Network architecture.
    Figure 2.  Experimental result for 1D transport equation.

    where $ t \in [0, \infty), x\in \mathbb{R} $, and $ a = \pi/10 (\simeq 0.314...) $. We have generated the observations from the analytic solution, by method of characteristics, of (16).

    $ tu=a2(xxu+yyu),u(t,0,y)=u(t,1,y)=0,u(t,x,0)=u(t,x,1)=0,
    $
    (17)

    where $ t \in [0, \infty), (x,y) \in [0,1]\times[0,1] $, and $ a = 1 $. We have generated the observations from the partial sum of the series solution, by separation of variables, of (17). We present two figures with different initial conditions.

    $ ttu=a2(xxu+yyu),u(0,x)=xy(1x)(1y),tu(0,x,y)=0,u(t,0,y)=u(t,1,y)=0,u(t,x,0)=u(t,x,1)=0,
    $
    (18)
    Figure 3.  Experimental result for 2D heat equation with $ u(0,x,y) = x(1-x)y(1-y) $.
    Figure 4.  Experimental result for 2D heat equation with $ u(0,x,y) = 1 \text{, if } (x,y) \in \Omega, 0 \text{, otherwise} $.

    where $ t \in [0, \infty), (x,y) \in [0,1]\times[0,1] $, and $ a = 1 $. We have generated the observations from the partial sum of the series solution, by separation of variables, of (18)

    $ u(t)=αuβuv,v(t)=δuvγv,u(0)=1,v(0)=1,
    $
    (19)
    Figure 5.  Experimental result for 2D wave equation.
    Figure 6.  Experimental result for Lotka-Volterra equation.

    where $ t \in [0, \infty) $, and $ \alpha = 1, \beta = 0.4, \delta = 0.4, \gamma = 0.1 $. We have generated the observations from a numerical solution by the Runge-Kutta method of (19). We used the $ sin $ function as the activation function for Lotka-Volterra system. Considering the periodic nature of the solution, the periodic activation function is a natural choice.

    In this section, we address the Courant$ \mbox{-} $Friedrichs$ \mbox{-} $Lewy (CFL) condition [7] which is a necessary condition while solving certain partial differential equations numerically. However, we claim that violating the CFL condition has no bad effect of our algorithm. To clarify this, we compare the results of transport equation with three different Courant numbers which all violate the convergence condition. As we can see in figure 7, our method converges well regardless of CFL condition. Note that our experiments in section 4.2 give nice results while the settings violate the well known stability condition called the Von Neumann stability condition.

    Figure 7.  Experimental result for CFL condition.

    First we summarize our theoretical results. For linear differential equations, we have shown that the DNN solution can reduce the proposed loss function as much as we want. A key point in the proof is a continuation of the Theorem 2.1 in [18] which states the fact that a linear combination of the dilated activations can approximate the target function and the derivative of such a linear combination can approximate its derivative in $ L_\infty $ sense. Next we have proved that the DNN which minimizes the loss converges to an analytic solution for linear parabolic or hyperbolic equations. In this step, we have applied basic energy estimates for each equation. Theoretical results for the inverse problem is also included as a continuation of the forward problem. We provide numerical experiments which show that our method indeed works accurately. We emphasize that our method can easily be implemented without any background knowledge about numerical analysis (for example, stability conditions) but about some libraries for implementing neural networks. Although it performs well for fundamental DEs, it might be hard to apply it to more complex equations. We have recognized that the error between a NN solution and an analytic/numerical solution is slightly increasing depending on time.

    For future directions, we may consider two problems. First, we can use more complicated neural network architectures such as CNN, RNN. Since we have dealt with time dependent PDEs, the combination of CNN and RNN would be a great choice for modelling. Second, the theoretical results for non-linear PDEs should be explored. The convergence results of our work are only applicable to linear PDEs. However, as in the experiment for the Lotka-Volterra system, our method is successful in approximating solutions even for non-linear systems. We hope proper convergence results for non-linear systems to be explored.

    This work was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF-2017R1E1A1A03070105, NRF-2019R1A5A1028324)

    [1] Wang JHC, Thampatty BP (2006) An introductory review in cell mechanobiology. Biomechan Model Mechanobiol 5: 1–6. doi: 10.1007/s10237-005-0012-z
    [2] Osmanagic-Myers S, Dechat T, Foisner R (2015) Lamins at the crossroads of mechanosignaling. Genes Dev 29: 225–237.
    [3] Bausch AR, Schwarz US (2013) Cellular mechanosensing: Sharing the force. Nat Mat 12: 948–949. doi: 10.1038/nmat3791
    [4] Luo T, Mohan K, Iglesias PA, et al. (2013) Molecular mechanisms of cellular mechanosensing. Nat Mat 12: 1064–1071. doi: 10.1038/nmat3772
    [5] Wang N, Tytell JD, Ingber DE (2009) Mechanotransduction at a distance: mechanically coupling the extracellular matrix with the nucleus. Nat Rev Mol Cell Biology 10: 75–82.
    [6] Goldmann WH (2012a) Mechanotransduction in cells. Cell Biol Int 36: 649–652.
    [7] Shao X, Li Q, Mogilner A, et al. (2015) Mechanical stimulation induces formin-dependent assembly of a perinuclear actin rim. Proc Nat Acad Sci USA122: E2595–2601.
    [8] Jalali S, del Pozo MA, Chen KD, et al. (2001) Integrin-mediated mechanotransduction requires its dynamic interaction with specific extracellular matrix (ECM) ligands. Proc Natl Acad Sci USA 98: 1042–1046. doi: 10.1073/pnas.98.3.1042
    [9] Steinwachs J, Metzner C, Skodzek K, et al. (2015) Three-dimensional force microscopy of cells in biopolymer networks. Nat Methods [in press].
    [10] Geiger B, Spatz JP, Bershadsky AD (2009) Environmental sensing through focal adhesions. Nat Rev Molecular Cell Biology 10: 21–33. doi: 10.1038/nrm2593
    [11] Jaalouk DE, Lammerding J (2009) Mechanotransduction gone awry. Nat Rev Molecular Cell Biology 10: 63–73. doi: 10.1038/nrm2597
    [12] Grashoff C, Hofman BD, Brenner MD, et al. (2010) Measuring mechanical tension across vinculin reveals regulation of focal adhesion dynamics. Nature 466: 263–267. doi: 10.1038/nature09198
    [13] Honarmandi P, Lee H, Lang MJ, et al. (2010) A microfluidic system with optical laser tweezers to study mechanotransduction and focal adhesion recruitment. Lab Chip 11: 684–694.
    [14] Fabry B, Klemm AH, Kienle S, et al. (2011) Focal adhesion kinase stabilizes the cytoskeleton. Biophys J 101: 2131–2138. doi: 10.1016/j.bpj.2011.09.043
    [15] Goldmann WH (2014) Mechanosensation: a basic cellular process. Progress in Molecular Biology and Translational Science 126: 75–102. doi: 10.1016/B978-0-12-394624-9.00004-X
    [16] Dent JE, Devescovi V, Li H, et al. (2015) Mechanotransduction map: simulation model, molecular pathway, gene set. Bioinformatics 31: 1053–1059. doi: 10.1093/bioinformatics/btu776
    [17] Goldmann WH (2016) Role of vinculin in cellular mechanotransduction. Cell Biol Int [in press].
    [18] Janoštiak R, Pataki AC, Brabek J, et al. (2014) Mechanosensors in integrin signaling: the emerging role of p130Cas. Eur J Cell Biol 93: 445–454. doi: 10.1016/j.ejcb.2014.07.002
    [19] Ezzell RM, Goldmann WH, Wang N, et al. (1997) Vinculin promotes cell spreading by mechanically coupling integrins to the cytoskeleton. Exp Cell Res 231: 14–26. doi: 10.1006/excr.1996.3451
    [20] Mierke CT, Kollmannsberger P, Zitterbart DP, et al. (2010) Vinculin facilitates cell invasion into three-dimensional collagen matrices. J Biol Chem 285: 13121–13130. doi: 10.1074/jbc.M109.087171
    [21] Wozniak MA, Chen CS (2009) Mechanotransduction in development: a growing role for contractility. Nat Rev Molecular Cell Biology 10: 34–42. doi: 10.1038/nrm2592
    [22] Bays JL, Peng X, Tolbert CE, et al. (2014) Vinculin phosphorylation differentially regulates mechanotransduction at cell-cell and cell-matrix adhesions. J Cell Biol 205: 251–263. doi: 10.1083/jcb.201309092
    [23] Martinac B (2004) Mechanosensitive ion channels: molecules of mechanotransduction. J Cell Sci 117: 2449–2460. doi: 10.1242/jcs.01232
    [24] Ingber DE (2006) Cellular mechanotransduction: putting all the pieces together again. FASEB J 20: 811–827. doi: 10.1096/fj.05-5424rev
    [25] Haswell ES, Phillips R, Rees DC (2011) Mechanosensitive channels: what can they do and how do they do it? Structure 19: 1356–1369.
    [26] Delmas P, Hao J, Rodat-Despoix L (2011) Molecular mechanisms of mechanotransduction in mammalian sensory neurons. Nat Rev Neurosci 12: 139–153. doi: 10.1038/nrn2993
    [27] Nomura S, Yamamoto TT (2000) Molecular events caused by mechanical stress in bone. Matrix Biology 19: 91–96. doi: 10.1016/S0945-053X(00)00050-0
    [28] Hoffman BD, Grashoff C, Schwartz MA (2011) Dynamic molecular processes mediate cellular mechanotransduction. Nature 475: 316–23. doi: 10.1038/nature10316
    [29] Luo T, Mohan K, Iglesias PA, et al. (2013) Molecular mechanisms of cellular mechanosensing. Nat Materials 12: 1064–1071. doi: 10.1038/nmat3772
    [30] Leerberg JM, Gomez GA, Verma S, et al. (2014) Tension-sensitive actin assembly supports contractility at the epithelial zonula adherens. Curr Biology 24: 1689–1699. doi: 10.1016/j.cub.2014.06.028
    [31] Milllward-Sadler SJ, Salter DM (2004) Integrin-dependent signal cascades in chondrocyte mechanotransduction. Ann Biomed Eng 32: 435–446. doi: 10.1023/B:ABME.0000017538.72511.48
    [32] Shivashankar GV (2011) Mechanosignaling to the cell nucleus and gene regulation. Ann Rev Biophysics 40: 361–378. doi: 10.1146/annurev-biophys-042910-155319
    [33] McCain ML, Parker KK (2011) Mechanotransduction: the role of mechanical stress, myocyte shape, and cytoskeletal architecture on cardiac function. Eur J Physiol 462: 89–104.
    [34] Frank D, Frey N (2011) Cardiac Z-disc Signaling Network. J Biol Chem 286: 9897–9904. doi: 10.1074/jbc.R110.174268
    [35] Paluch EK, Nelson CM, Biais N, et al. (2015) Mechanotransduction: use the force(s). BMC Biology 13: 47. doi: 10.1186/s12915-015-0150-4
    [36] Ross RS (2004) Molecular and mechanical synergy: cross-talk between integrins and growth factor receptors. Cardiovascular Res 63: 381–390. doi: 10.1016/j.cardiores.2004.04.027
    [37] Vogel V, Sheetz MP (2009) Cell fate regulation by coupling mechanical cycles to biochemical signaling pathways. Curr Biol Cell Biol 21: 38–46. doi: 10.1016/j.ceb.2009.01.002
    [38] Dupont S, Morsut L, Aragona M, et al. (2011) Role of YAP/TAZ in mechanotransduction. Nature 474: 179–185.
    [39] Goldmann WH, Auernheimer V, Thievessen I, et al. (2013) Vinculin, cell mechanics and tumour cell invasion. Cell Biol Int 37: 397–405. doi: 10.1002/cbin.10064
    [40] Kaminski A, Fedorchak GR, Lammerding J (2014) The cellular mastermind(?)– Mechanotransduction and the nucleus. Progress in Molecular Biology and Translational Science 126: 157–203. doi: 10.1016/B978-0-12-394624-9.00007-5
    [41] Wang N, Tytell JD, Ingber DE (2009) Mechanotransduction at a distance: mechanically coupling the extracellular matrix with the nucleus. Nat Rev Mol Cell Biology 10: 75–82. doi: 10.1038/nrm2594
    [42] Alenghat FJ, Ingber DE (2002) Mechanotransduction: All signals point to cytoskeleton, Matrix, and Integrins. Sci StKE 119: pe6.
    [43] Auernheimer V, Lautscham LA, Leidenberger M, et al. (2015) Vinculin phosphorylation at residues V100 and Y1065 is required for cellular force transmission. J Cell Sci 128: 3435–3443.
    [44] Goldmann WH (2002) Mechanical aspects of cell shape regulation and signaling. Cell Biol Int 26: 313–317. doi: 10.1006/cbir.2002.0857
    [45] Janoštiak R, Brábek J, Auernheimer V, et al. (2014) CAS directly interacts with vinculin to control mechanosensing and focal adhesion dynamics. Cell Mol Life Sci 71: 727–44. doi: 10.1007/s00018-013-1450-x
    [46] Samarel AM (2005) Costameres, focal adhesions, and cardiomyocyte mechanotransduction. Am J Physiol Heart Circ Physiol 289: H2291–H2301. doi: 10.1152/ajpheart.00749.2005
    [47] Butcher DT, Alliston T, Weaver VM (2009) A tense situation: forcing tumour progression. Nature Rev. Cancer 9: 108–122. doi: 10.1038/nrc2544
    [48] Goldmann WH (2012b) Mechanotransduction and focal adhesions. Cell Biol Int 36: 649–652.
    [49] Vogel V, Sheetz MP (2006) Local force and geometry sensing regulate cell functions. Nat Rev Molecular Cell Biology 7: 265–275. doi: 10.1038/nrm1890
    [50] Wang HB, Dembo M, Hanks SK, et al. (2001) Focal adhesion kinase is involved in mechanosensing during fibroblast migration. Proc Nat Acad Sci USA 98: 11295–11300. doi: 10.1073/pnas.201201198
    [51] Bendig G, Grimmler M, Huttner IG, et al. (2006) Integrin-linked kinase, a novel component of the cardiac mechanical stretch sensor, controls contractility in the zebrafish heart. Genes Dev 20: 2361–2372.
    [52] Shih YRV, Tseng KF, Lai HY, et al. (2011) Matrix stiffness regulation of integrin-mediated mechanotransduction during osteogenic differentiation of human mesenchymal stem cells. J Bone Miner Res 26: 730–738. doi: 10.1002/jbmr.278
    [53] Mehta PK, Griendling KK (2006) Angiotensin II cell signaling: physiological and pathological effects in the cardiovascular system. Am J Physiol Cell Physiol 292: C82–C97. doi: 10.1152/ajpcell.00287.2006
    [54] Schwartz MA, Assoian RK (2001) Integrins and cell proliferation: regulation of cyclin-dependent kinases via cytoplasmic signaling pathways. J Cell Sci 114: 2553–2560.
    [55] McBeath R, Pirone DM, Nelson CM, et al. (2004) Cell shape, cytoskeletal tension, and RhoA regulate stem cell lineage commitment. Cell 6: 483–495.
    [56] Bertrand AT, Ziaei S, Ehret C, et al. (2014) Cellular microenvironments reveal defective mechanosensing responses and elevated YAP signaling in LMNA-mutated muscle precursors. J Cell Sci 127: 2873–2884. doi: 10.1242/jcs.144907
    [57] Yuan JM, Chyan AL, Zhou HX, et al. (2008) The effect of macromolecular crowding on the mechanical stability of protein molecules. Protein Sci 17: 2156–2166. doi: 10.1110/ps.037325.108
    [58] Ladoux B, Nelson WJ, Yan J, et al. (2015) The mechanotransduction machinery at work at adherens junctions. Integr Biol 7: 1109–1119. doi: 10.1039/C5IB00070J
    [59] Agrawal S, Agrawal A, Doughty B, et al. (2003) Cutting edge: different toll-like receptors agonists instruct dendritic cells to induce distinct responses via differential modulation of extracellular signal-regulated kinase-activated protein kinase and cFos. J Immunol 171: 4984–4989. doi: 10.4049/jimmunol.171.10.4984
    [60] Chiu JJ, Chien S (2011) Effects of disturbed flow on vascular endothelium: pathophysiological basis and clinical perspectives. Physiol Rev 91: 10.1152.
    [61] Maroto R, Raso A, Wood TG, et al. (2005) TRPC1 forms the stretch-activated cation channel in vertebrate cells. Nat Cell Biol 7: 179–185. doi: 10.1038/ncb1218
    [62] Davies PF, Tripathi SC (1993) Mechanical stress mechanisms and the cell: an endothelial paradigm. Cir Res 72: 239–245. doi: 10.1161/01.RES.72.2.239
    [63] Burkholder TJ (2008) Mechanotransduction in skeletal muscle. Front Biosci 12: 174–191.
    [64] Benavides DT, Egli M (2014) Calcium's Role in Mechanotransduction during muscle development. Cell Physiol Biochem 33: 249–272. doi: 10.1159/000356667
    [65] Schwartz MA, Simone DW (2008) Cell adhesion receptors in mechano-transduction. Curr Opion Cell Biol 20: 551–556. doi: 10.1016/j.ceb.2008.05.005
    [66] Ingber DE (2003) Mechanobiology and diseases of mechanotransduction. Ann Med 35: 1–14.
    [67] Schreiner SM, Koo PK, Zhao Y, et al (2015) The tethering of chromatin to the nuclear envelope supports nuclear mechanics. Nat Comm 6: 7159.
    [68] Engler AJ, Kumar S (2014) Mechanosensation. Progress in Molecular Biology and Translational Science, Academic Press. 126: 1–384.
    [69] Isermann P, Lammerding J (2013) Nuclear mechanics and mechanotransduction in health and disease. Curr Biol 23: R1113–1121. doi: 10.1016/j.cub.2013.11.009
  • This article has been cited by:

    1. Hyung Ju Hwang, Jin Woo Jang, Hyeontae Jo, Jae Yong Lee, Trend to equilibrium for the kinetic Fokker-Planck equation via the neural network approach, 2020, 419, 00219991, 109665, 10.1016/j.jcp.2020.109665
    2. Jae Yong Lee, Jin Woo Jang, Hyung Ju Hwang, The model reduction of the Vlasov–Poisson–Fokker–Planck system to the Poisson–Nernst–Planck system via the Deep Neural Network Approach, 2021, 55, 0764-583X, 1803, 10.1051/m2an/2021038
    3. George Em Karniadakis, Ioannis G. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, Liu Yang, Physics-informed machine learning, 2021, 3, 2522-5820, 422, 10.1038/s42254-021-00314-5
    4. Sung Woong Cho, Hyung Ju Hwang, Hwijae Son, Traveling Wave Solutions of Partial Differential Equations Via Neural Networks, 2021, 89, 0885-7474, 10.1007/s10915-021-01621-w
    5. Hyung Ju Hwang, Hwijae Son, Lagrangian dual framework for conservative neural network solutions of kinetic equations, 2022, 15, 1937-5093, 551, 10.3934/krm.2021046
    6. Sunwoo Hwang, Seongwon Lee, Hyung Ju Hwang, Neural network approach to data-driven estimation of chemotactic sensitivity in the Keller-Segel model, 2021, 18, 1551-0018, 8524, 10.3934/mbe.2021421
    7. Zhicheng Zhu, Jia Hao, Jin Huang, Biao Huang, BC-PINN: an adaptive physics informed neural network based on biased multiobjective coevolutionary algorithm, 2023, 35, 0941-0643, 21093, 10.1007/s00521-023-08876-4
    8. Min Sue Park, Cheolhyeong Kim, Hwijae Son, Hyung Ju Hwang, The deep minimizing movement scheme, 2023, 494, 00219991, 112518, 10.1016/j.jcp.2023.112518
    9. Hwijae Son, Sung Woong Cho, Hyung Ju Hwang, Enhanced physics-informed neural networks with Augmented Lagrangian relaxation method (AL-PINNs), 2023, 548, 09252312, 126424, 10.1016/j.neucom.2023.126424
    10. Wai M. Kwok, George Streftaris, Sarat C. Dass, Laplace based Bayesian inference for ordinary differential equation models using regularized artificial neural networks, 2023, 33, 0960-3174, 10.1007/s11222-023-10289-1
    11. Hwijae Son, Minwoo Lee, Continuous probabilistic solution to the transient self-oscillation under stochastic forcing: a PINN approach, 2023, 37, 1738-494X, 3911, 10.1007/s12206-023-0707-z
    12. Hwijae Son, Hyunwoo Cho, Hyung Ju Hwang, Physics-Informed Neural Networks for Microprocessor Thermal Management Model, 2023, 11, 2169-3536, 122974, 10.1109/ACCESS.2023.3329562
    13. Amina Hassan Ali, Norazak Senu, Ali Ahmadian, Efficient solutions to time-fractional telegraph equations with Chebyshev neural networks, 2024, 99, 0031-8949, 115210, 10.1088/1402-4896/ad7c93
    14. Kaixuan Zheng, Huiyong Guo, Jin Di, Chen Liu, Quantitative health evaluation of frame structures using multi-physics enhanced neural networks with boundary constraints and damage consistency and its experimental study, 2025, 72, 23520124, 108281, 10.1016/j.istruc.2025.108281
  • Reader Comments
  • © 2016 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(10379) PDF downloads(1939) Cited by(42)

Figures and Tables

Figures(1)

Other Articles By Authors

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog