vamp
play

(VAMP) Iden%fica%on of molecular order parameters and states from - PowerPoint PPT Presentation

Varia%onal Approach to Markov Processes (VAMP) Iden%fica%on of molecular order parameters and states from nonreversible MD simula%ons Fabian Paul Computer Tutorial in Markov Modeling 18-FEB-2020 Recap: the spectral theory of MSMs A Markov


  1. Varia%onal Approach to Markov Processes (VAMP) Iden%fica%on of molecular order parameters and states from nonreversible MD simula%ons Fabian Paul Computer Tutorial in Markov Modeling 18-FEB-2020

  2. Recap: the spectral theory of MSMs β€’ A Markov state model consists of: 1. a set of states 𝑑 ! !"#,…& 2. (condi8onal) transi8on probabili8es between these states π‘ˆ !' = β„™(𝑑 𝑒 + 𝜐 = π‘˜ ∣ 𝑑(𝑒) = 𝑗) …

  3. Markov state models: estimation Markov model es-ma-on starts with: β€’ grouping of geometrically [1] or kine-cally [2] related conforma-ons into clusters or microstates 1 2 microstates 3 [1] Prinz et al ., J. Chem. Phys. 134 , 174105 (2011) 3 [2] PΓ©rez-HernΓ‘ndez, Paul , et al. , J. Chem. Phys. 139 , 015102 (2013)

  4. Markov state models: estimation β€’ We then assign every conforma-on in a MD trajectory to a microstate. -me 𝑒 2 t 3 t 4 t 5 t 6 t 7 t t trajectory microstate 𝑑 1 2 3 2 3 1 3 We count transitions between microstates and tabulate them in a β€’ count matrix 𝐃 e. g. 𝐷 !! = 1 , 𝐷 !" = 1 , 𝐷 "# = 2 , … We estimate the transition probabilities π‘ˆ $% from 𝐃 . β€’ NaΓ―ve estimator: ) π‘ˆ $% = 𝐷 $% / βˆ‘ & 𝐷 $& β€’ Maximum-likelihood estimator [1] β€’ [1] Prinz et al ., J. Chem. Phys. 134 , 174105 (2011) 4 [2] PΓ©rez-HernΓ‘ndez, Paul , et al. , J. Chem. Phys. 139 , 015102 (2013)

  5. The spectrum of a Energy U ( i ) reversible T matrix β€’ The large eigenvalues of the transition matrix and their corresponding eigenvectors encode the information about the slow molecular processes. β€’ Flat regions of the eigenvectors allow to identify the metastable states. Prinz et al ., J. Chem. Phys. 134 , 174105 (2011)

  6. Both MSMs and TICA make use of the same spectral method The spectral method (working with eigenvalue and eigenvector) is not limited to Markov state models. β€’ Es;ma;on of MSMs π‘ˆ(𝜐) = 𝐷 !" (𝜐) 𝐷 ! β€’ In matrix nota;on 𝐔 𝜐 = 𝐃 0 #$ 𝐃(𝜐) β€’ Eigenvalue problem: 𝐔 𝜐 𝐰 = πœ‡π° ⇔ 𝐃 0 #$ 𝐃 𝜐 𝐰 = πœ‡π° ⇔ 𝐃 𝜐 𝐰 = πœ‡πƒ(0)𝐰 β€’ The last equa;on is known as the TICA problem. All equa;ons generalize to the case where 𝐃(0) and 𝐃 𝜐 are not count matrices, but correla;on matrices. β€’ The indices 𝑗, π‘˜ don’t longer refer to states but to features . Schwantes, Pande, J. Chem. Theory Comput. 9 2000 (2013) PΓ©rez-HernΓ‘ndez et al ,J. Chem. Phys. , 139 015102 (2013)

  7. VAC and VAMP

  8. Varia%onal approach to conforma%onal dynamics VAC (Rayleigh-Ritz for classical dynamics) Any autocorrelation is bounded by the system-specific number ! πœ‡ , that is related to the 𝑒 by ! πœ‡ = 𝑓 !"/$ % . system-specific autocorrelation time Μ‚ &!" πœ” 𝑦 𝑒 βˆ‘ % πœ” 𝑦(𝑒 + 𝜐) = πœ”, Tπœ” ' ≀ ! acf πœ”; 𝜐 : = πœ‡ &!" πœ” 𝑦 𝑒 πœ”, πœ” ' βˆ‘ % πœ” 𝑦(𝑒) β€’ The maximum is achieved if πœ” is an eigenfunction of T . Proof : Expand πœ” in an (orthonormal) eigen-basis of T : ) > 0 πœ” 𝑦 = βˆ‘ ( 𝑑 ( 𝜚 ( 𝑦 , πœ”, πœ” ' = βˆ‘ ( 𝑑 ( ) πœ‡ ( βˆ’ < ) ! ) πœ‡ ( βˆ’ ! πœ”, Tπœ” ' βˆ’ ! πœ‡ πœ”, πœ” ' = < 𝑑 ( 𝑑 ( πœ‡ = < 𝑑 ( πœ‡ ≀ 0 ( ( ( β€’ If ! πœ‡ is max πœ‡ ( the largest of T’s eigenvalues, the inequality holds. ( β€’ Result can only be zero if 𝑑 ( = 0 for 𝑗 β‰  π‘˜ and πœ‡ * = max πœ‡ ( β‡’ πœ” 𝑦 ∝ 𝜚 +,- 𝑦 ( β€’ Remark: the variational approach generalizes to the optimization of multiple eigenfunctions. ! . πœ‡ is replaced by the sum of the eigenvalues 𝑆 . = βˆ‘ (/0 πœ‡ ( NoΓ©, NΓΌske, SIAM Multiscale Model. Simul. 11 , 635 (2013)

  9. Interpretation of variational principle good test func0on bad test func0on 1. Pick some test func0on 𝝍 !"#! 𝐲 and pick some test conforma0ons 𝐲 $,&'&!() distributed according to equilibrium distribu0on 𝜌 Ξ© 2. Propagate 𝐲 $,&'&!() with the the MD integrator. Call result 𝐲 $,*&'() . 3. Correlate 𝝍 !"#! 𝐲 &'&!() with 𝝍 !"#! 𝐲 *&'() . $ βˆ‘ !"# 𝝍 𝐲 !,&'&()* ./ 𝝍 β‹… 𝝍 𝐲 !,+&')* ./ 𝝍 score= $ βˆ‘ !"# 𝝍 𝐲 !,&'&()* ./ 𝝍 β‹… 𝝍 𝐲 !,&'&()* ./ 𝝍

  10. Gradient-based optimization of function parameters Parameters πͺ of 𝝍 '()' 𝐲; πͺ can be op-mized with gradient-based techniques. Make use of the gradient of the VAC or VAMP score, the gradient of the test func-on and off-the-shelf op-mizers such as ADAM or BFGS.

  11. Reversible dynamics β€’ In equilibrium , every trajectory is as probable as its time-reversed copy β„™ 𝑑 𝑒 + 𝜐 = π‘˜ and 𝑑 𝑒 = 𝑗 = β„™ 𝑑 𝑒 + 𝜐 = 𝑗 and 𝑑 𝑒 = π‘˜ β„™ 𝑑 𝑒 + 𝜐 = π‘˜ ∣ 𝑑 𝑒 = 𝑗 β„™ 12 (𝑑 𝑒 = 𝑗) = β„™ 𝑑 𝑒 + 𝜐 = 𝑗 ∣ 𝑑 𝑒 = π‘˜ β„™ 12 (𝑑 𝑒 = π‘˜) 𝜌 ! π‘ˆ !' = 𝜌 ' π‘ˆ '! β€’ In mathematician’s notation 𝐟 ! , π”πŸ ' 3 = 𝐟 ' , π”πŸ ! 3 where 𝐲, 𝐳 3 = βˆ‘ ! 𝑦 ! 𝑧 ! 𝜌 ! β€’ 𝐔 is a symmetric matrix w.r.t. to a non-standard scalar product. β€’ 𝐔 has real eigenvalues and eigenvectors (linear algebra I). Prinz et al ., J. Chem. Phys. 134 , 174105 (2011)

  12. The problem with nonreversible systems & πœ‡ $ where πœ‡ ! are the true eigenvalues. β€’ 𝑆 & = βˆ‘ $*! β€’ For nonreversible dynamics 𝐟 ! , π”πŸ ' 3 β‰  𝐟 ' , π”πŸ ! 3 β€’ There might not even be a well-defined 𝝆 . β€’ Eigenvalues and eigenvectors will be complex. β€’ Varia8onal principle doesn’t work. acf(πœ”) ≀ ( πœ‡ ∈ β„‚ makes no sense. One can’t order complex numbers on a line. β€’ Op8miza8on of models not possible β€’ Feature selec8on not possible β€’ Is there any way to fix this? Can we maybe find some other operator that is related to dynamics and that is symmetric?

  13. A possible solu;on: VAMP Varia%onal approach to Markov processes β€’ Introduce the β€œbackward” transition matrix 𝐔 + ∢= 𝐃 𝑂 ,! 𝐃 βˆ’πœ = 𝐃 𝑂 ,! 𝐃 - 𝜐 i.e. estimate MSM/TICA from time-reversed data, where 0 𝐷 $% βˆ’πœ : = 8 𝑔 $ 𝑦 𝑒 βˆ’ 𝜐 𝑔 % 𝑦(𝑒) .*/ 0 𝐷 $% 𝑂 : = 8 𝑔 $ 𝑦(𝑒) 𝑔 % (𝑦(𝑒)) .*/ β€’ Introduce the forward-backward transition matrix 𝐔 1+ ≔ 𝐔𝐔 + and 𝐔 +2 : = 𝐔 3 𝐔 β€’ Can show that 𝐔 1+ and 𝐔 +2 are symmetric without any reference to a stationary vector (symmetry is built into the matrices). β€’ Eigenvalues and eigenvectors of 𝐔 43 and 𝐔 34 are real. β€’ They fulfill a variational principle 𝐃 ,!/" 0 𝐃 𝜐 𝐃 N ,!/" ≀ 𝑆 Wu, NoΓ©, J. Nonlinear Sci. 30 , 23 (2020) Klus, S. et al, J. Nonlinear Sci., 28 , 1 (2018)

  14. Cross-validation β€’ The model parameters (in this example parameters of the line and steepness of the transition) were optimized for a particular realization of the dynamics. β€’ Didn’t we say that the eigenfunctions and eigenvalues were an intrinsic property of the molecular system? β€’ So the eigenfunctions should be the same if we repeat the analysis with a second simulation of the same system.

  15. Cross-valida;on β€’ The model parameters (in this example parameters of the line and steepness of the transi-on) were op-mized for a par-cular realiza-on of the dynamics. β€’ Didn’t we say that the eigenfunc-ons and eigenvalues were an intrinsic property of the molecular system? β€’ So the eigenfunc-ons should be the same if we repeat the analysis with a second simula-on of the same system.

  16. Cross-validation β€’ Ideally, we want to tell if the solu-on is robust at a single glance by measuring the robustness with one number. β€’ The VAMP score or VAC score (also called GRMQ 1 ) lends itself to this task. β€’ Keep all the trained model parameters fixed (here the line parameters and the steepness of the transition), plug in new data and recompute the test autocorrelation. β€’ The test autocorrelation will be lower in general, which means that the original model was fit to noise ( overfit ). [1] McGibbon, Pande, J Chem Phys., 142 124105 (2015)

  17. Cross-validation β€’ Repor-ng a test-score that was computed from independent realiza-ons is the gold standard. β€’ Independent realiza-ons can be expensive to sample. β€’ Do the approximate k -fold (hold-out) cross-valida-on. β€’ Split all data into training set and test sets. β€’ Op0mize the model parameters with the training set and test the parameters with test sets. β€’ Repeat for k different divisions of the data. k -fold cross-valida-on can be tricky with highly autocorrelated -me β€’ series data!

  18. Applica;ons

  19. Application: feature selection β€’ varia%onal principle: the higher the score the be3er β€’ Compare test scores for different selec%ons of molecular features. Which selec%on gives best score? distances? dihedrals? contacts? side chain flips? rigid body approximation? chemical intui0on?

  20. Application: feature selection Scherer et al J. Chem. Phys. 150 , 194108 (2019)

  21. Application: ion channel non- equilibrium MD Analysis of MD simulation data of the "controversial” direct- knock-on conduction mechanism in the KcsA potassium channel. Ions a constantly inserted at one side of the membrane and deleted at the other side. Paul et al , J. Chem. Phys. MMMK , 164120 (2019). Fig1 and data: KΓΆpfer et al., Science , 346 , 352 (2014).

  22. Applica:on: ion channel non- equilibrium MD By clustering in the VAMP space, we identified 15 different states that differ structurally near the selectivity filter and differ in their conductivity.

Recommend


More recommend