Kinetics-Optimized Enhanced Sampling Using Mean First Passage Times
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Kinetics-Optimized Enhanced Sampling Using Mean First Passage Times".
Title and authors: Kai: So we're looking at the paper "Kinetics-Optimized Enhanced Sampling Using Mean First Passage Times". The title itself tells you they are focusing on using molecular kinetics to optimize the sampling bias.
Mira: Exactly. It’s about taking that standard enhanced sampling idea, where you just add some energy bias, and instead of optimizing it based on thermodynamics, they're optimizing it based on how fast the system actually moves—the kinetics.
Lev: From my side, I think the implication is that we can skip a lot of unnecessary exploration in the phase space because we aren't blindly adding bias everywhere; we're steering toward the target state using kinetic information instead of just trying to flatten a potential surface.
Kai: It’s about making sure that every move the simulation makes is pointed in the right direction kinetically, which should save a ton of computational time when you’re dealing with very slow processes.
Mira: Right. So instead of running simulations for billions of steps just to find an energy minimum, they are using a kinetic observable called the mean first passage time, or MFPT, as their main goal function.
The paper's summary: Kai: The paper introduces this KOMBI algorithm, which stands for Kinetics-Optimized Markovian Bias Integration. It adaptively optimizes the bias based on molecular kinetics to minimize the mean first passage time to a target state.
Mira: They achieve this by using the relationship between biased and unbiased kinetics that comes from discretized Markov state models, established by something called the dynamic histogram analysis method, or DHAM. This lets them build those unbiased rates and free energy surfaces without just relying on standard weighted histogram analysis methods which only give you free energy profiles.
Lev: So DHAM gives them the framework to construct an unbiased Markov matrix from multiple biased trajectories, and they use that matrix to iteratively update the bias parameters based on minimizing that mean first passage time. That’s where most of the complex math is happening.
Kai: In each step, they look for a set of Gaussian kernel parameters—mu alpha, amplitude alpha, and sigma alpha—and they optimize those three parameters to find the minimal MFPT value from the last position to the local target state.
Mira: The main idea is that you’re not just adding random bias; you're designing a bias profile specifically tailored to minimize that mean first passage time, which is a key physical observable for this whole study.
Lev: So they are essentially trying to find the best set of Gaussian functions in the CV space—the biasing potential—that steers the system most efficiently toward its goal based on kinetic constraints.
Kai: They show this works across different one dimensional and two dimensional analytical potential-based models, as well as all atom simulations for things like sodium chloride dissociation and phosphate unbinding in Ras GTPase. That’s a broad scope for validation.
The paper's improvements: Mira: The core improvement is that this approach moves beyond just using energy to find a flat surface; they are optimizing the bias directly against the MFPT objective function. That’s the first time someone has done that adaptation based on kinetics alone for this purpose.
Lev: They use an analytical relationship between biased and unbiased rate or Markov matrices where each element of the biased rate matrix is proportional to an exponential involving both biases, which links those two kinetic descriptions together.
Kai: And they use the dynamic histogram analysis method, DHAM to build those unbiased Markov matrices from multiple biased trajectories, which is how they get that kinetic information needed for the optimization.
Mira: This lets you get unbiased rates and free energy surfaces and then you can feed that back into a new iteration to find a better bias profile that minimizes MFPT. It’s a feedback loop driven by kinetics.
Lev: So the improvement is moving beyond just using thermodynamics to explicitly incorporating molecular kinetics into the bias selection process, which is a significant shift in how we design these sampling methods.
Kai: It’s about creating an optimized pathway that doesn't waste time exploring directions in the CV space that aren't useful for reaching the target state.
Conclusion: Mira: To wrap this up, the paper "Kinetics-Optimized Enhanced Sampling Using Mean First Passage Times" proposes an MSM-based, kinetics-optimized enhanced sampling algorithm that optimizes the applied bias to minimize MFPT to a target state.
Kai: The key is that they adaptively update the bias based on an MSM analysis of the discretized phase space using the DHAM method. This iterative bias generation method significantly reduces the time required to reach target states in their test cases.
Lev: It’s about facilitating a more direct trajectory from a local starting point to the target state within that phase space, demonstrating that KOMBI efficiently directs the system towards the global target without extensive exploration of non-essential directions in the CV space, yet providing a likely pathway.
Kai: This demonstrates that KOMBI is an efficient and accurate way to probe conformational changes that normally occur over long timescales using ordinary classical MD simulations.
Mira: It’s a novel, efficient and accurate way to probe those conformational changes that normally occur over long timescales using ordinary classical MD simulations.
Lev: So while the algorithm itself is quite smart, the limitation they state is that if you don't have the free energy surface of the system of the system beforehand, you have to collect that kinetic information from simulation data.
Kai: Right. That’s where we need those initial unbiased simulations to start building that kinetic foundation before we can apply these new tools.
Mira: It’s a necessary trade-off for getting this kinetic optimization working in the way they describe it, so it shows the current state of enhanced sampling where you still need some groundwork laid before you can apply these new tools.
Lev: So it’s a novel tool, but we still have to figure out how to make that iterative bias generation method stable enough for real hardware.
Kai: That’s the next step for us, figuring out the stability on our cryo-hardware when we try to run this thing.
Tiejun Wei, Balint Dudas, Edina Rosta
University College London
physics.bio-ph
Submitted: 2024-06-13
Updated: 2026-11-17
Comments: Revised manuscript accepted for publication in The Journal of Chemical Physics on 4 November 2026. 21 pages: 12-page main manuscript and 9-page supplementary material
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The gist: Molecular dynamics simulations have become essential in many areas of atomistic modelling from drug discovery to materials science Enhanced sampling methods provided avenues to access such crucial…
Key concepts
- Mean First Passage Time (MFPT)
- MFPT measures the average time it takes for a system to transition from a specific starting state to a designated target state. Minimizing this time is the primary goal of the algorithm, as it directly addresses the timescale problem in molecular dynamics simulations.
- Markov State Models (MSMs)
- MSMs are mathematical models that treat a complex system's dynamics as a Markov chain by discretizing its conformational space into different states. They allow researchers to quantify transitions between these states using a rate matrix, which is crucial for kinetic analysis.
- Dynamic Histogram Analysis Method (DHAM)
- DHAM is the theoretical framework used here to bridge the gap between biased and unbiased kinetics. It allows the creation of an unbiased Markov State Model from multiple biased simulation trajectories, providing accurate kinetic information needed for bias optimization.
Terminology
Summary
Molecular dynamics simulations have become essential in many areas of atomistic modelling from drug discovery to materials science Enhanced sampling methods provided avenues to access such crucial rare events, for example key slow conformational changes of biomolecules Here, we introduce a novel enhanced sampling algorithm that adaptively optimizes the bias based on the kinetics of the system for the first time We identify the optimal bias that minimizes a key physical observable, the mean first passage time (MFPT) from a starting state to a target state Our algorithm makes use of the relation between biased and unbiased kinetics obtained from discretized Markov state models (MSMs), as established in the dynamic histogram analysis method (DHAM) We demonstrate the applicability of the method for different 1D and 2D analytical potential-based model examples, NaCl dissociation in explicit water, and phosphate unbinding in Ras GTPase Our algorithm has excellent performance compared with state-of-the-art enhanced sampling methods in terms of the timescales required to reach the final state in the benchmarking systems
Pruning redundant reasoning tokens
The paper proposes a novel enhanced sampling algorithm called KOMBI (Kinetics-Optimized Markovian Bias Integration) which adaptively optimizes the bias based on molecular kinetics to minimize the mean first passage time (MFPT) to a target state This approach is signatured by a targeted approach to facilitate mapping rare events The optimization problem is formulated where the bias is defined as the summation of a set of Gaussian functions in the CV space, with parameters optimized to yield the minimal MFPT value The search for this optimal bias is defined by minimizing MFPT(xs → xf) given an unbiased rate matrix K
How the training signal is built
The algorithm iteratively updates the Markov matrix using the DHAM algorithm given all trajectories and their corresponding biases This is achieved by unbiasing all trajectories with biasing potentials used throughout all iterations to obtain an estimated Markov matrix Mestα For each iteration, a new set of Gaussian kernel parameters (μiα+1, Aiα+1, σiα+1) is found by minimizing the biased MFPT from the last position to the local target state
Where the gains come from
The gains are demonstrated across several benchmarking systems including a 1D model potential, a 2D model potential, and all-atom simulations of solvated NaCl and phosphate unbinding in Ras.GAP For the 1D model potential example, the system reached the target state in 39 ± 12. p. 34 ps using KOMBI, which is orders of magnitude faster than WT-MetaD or classical MD simulations In the 2D potential energy system, KOMBI reached the target state in 79. p. 385 ± 36.42 ps compared to 10. p. 06 ± 4.1 ns for WT-MetaD
Conclusions
The MSM-based, kinetics-optimized enhanced sampling algorithm significantly reduces the time required to reach target states in test cases The iterative bias generation method facilitates a more direct trajectory from the local starting point to the target state within the phase space This demonstrates that KOMBI efficiently directs the system towards the global target without extensive exploration of non-essential directions in the CV space, yet providing a likely pathway The algorithm provides a novel, efficient and accurate way to probe conformational changes that normally occur over long timescales using ordinary classical MD simulations This opens up new ways to explore and facilitate rare events in the selected CV space in a highly-efficient manner The KOMBI algorithm provides a novel, efficient and accurate way to probe conformational changes that normally occur over long timescales using ordinary classical MD simulations
REFERENCE
KINETICS-OPTIMIZED ENHANCED SAMPLING USING MFPTS
Tiejun Wei1, Balint Dudas1, and Edina Rosta1 University College London, Department of Physics and Astronomy London WC1E 6BS Correspondence: e.rosta@ucl.ac.uk Molecular dynamics simulations have become essential in many areas of atomistic modelling from drug discovery to materials science They provide critical atomic-level insights into key dynamical events experiments cannot easily capture However, their impact often falls short as the timescales of the important processes are inaccessible using standard molecular dynamics Enhanced sampling methods provided avenues to access such crucial rare events, for example key slow conformational changes of biomolecules However, the bias in enhanced sampling simulations is rarely optimized, and even if they are, the optimization criteria is based on the thermodynamics or Hamiltonian of the system, but do not directly consider molecular kinetics Here, we introduce a novel enhanced sampling algorithm that adaptively optimizes the bias based on the kinetics of the system for the first time We identify the optimal bias that minimizes a key physical observable, the mean first passage time (MFPT) from a starting state to a target state Our algorithm makes use of the relation between biased and unbiased kinetics obtained from discretized Markov state models (MSMs), as established in the dynamic histogram analysis method (DHAM) We demonstrate the applicability of the method for different 1D and 2D analytical potential-based model examples, NaCl dissociation in explicit water, and phosphate unbinding in Ras GTPase Our algorithm has excellent performance compared with state-of-the-art enhanced sampling methods in terms of the timescales required to reach the final state in the benchmarking systems Our findings provide a novel, kinetics-driven enhanced sampling strategy, signatured by a targeted approach to facilitate mapping rare events, with the potential for breakthrough applications in drug discovery and materials science
I. INTRODUCTION
Biomolecular simulations, including all atom molecular dynamics (MD) have become essential to provide unprecedented details into the dynamics of complex systems In particular, slow conformational changes in biomolecular systems play a crucial role in many biological functions, but are mostly unavailable experimentally, particularly the atomic level insights into the rate limiting transition states and corresponding mechanistic pathways MD simulations therefore find key applications in e.g., drug discovery, as powerful tools to provide important structural and energetic information for complex systems of biomedical interest [1,2] On the other hand, their impact often falls short due to the well-known sampling problem [3,4] While aiming to uncover pivotal rare events, these simulations frequently stall in local minima, consuming extensive computational resources, limiting the timescales that can be reached to study biomolecular processes Enhanced sampling algorithms are developed to overcome this dilemma, and access orders of magnitude longer timescales that would be beyond reach for current computational resources [5] A wide variety of simulation strategies were proposed to map pathways from an initial state based on adding a bias to the Hamiltonian of the system Beyond the classical umbrella sampling (US) [6], adaptive methods such as the finite temperature string method [7], its variant with coupled Hamiltonian replica exchange method [8], targeted MD [9], milestoning-based sampling [10], Adaptive Biasing Force [11,12] as well as metadynamics (MetaD) and its improved variants, welltempered metadynamics (WT-MetaD) [13–15] were successful to explore the conformational space by deploying and optimizing bias [16] In most of biasing strategies, thermodynamics-based bias is applied to the system For example, MetaD aims to explore the selected collective variable (CV) space and iteratively identify the bias that, when added to the original Hamiltonian, provides a flat surface [13] WT-MetaD includes additional parameters to enable a smooth biasing process [15] Additionally, machine learning-aided enhanced sampling methods are also gaining attention as promising avenue for CV selection and sampling [17–21] However, these algorithms are not directly optimal to reduce the timescale for the conformational change of interest, and the bias is not optimized for the underlying molecular kinetics To describe the underlying molecular kinetics of the system, Markov state models (MSM) were highly successful [3,22,23] These also find applications in both deep learning and reinforcement learning strategies, such as using the multi-armed bandit framework [24] or autoencoders [25] Alternatively, transition path sampling (TPS) methods [26–28] and finite temperature string method [7] also provide valuable insights into the kinetics of rare events [10,28,29] However, the lack of bias in these methods limits the timescales that can be reached [30] To design optimal kinetics-based bias, an analytical relationship is needed between the dynamics of the biased and unbiased systems However, the weighted histogram analysis method (WHAM) [31], which is typically used to estimate the unbiased free energy profiles from biased runs, does not provide kinetic information To enable the estimation of the unbiased kinetics, the recently developed dynamic histogram analysis method (DHAM) [22] provides the necessary theoretical framework to build the unbiased MSM from multiple biased trajectories
II. MARKOV STATE MODELS AND MEAN FIRST PASSAGE TIMES
The MSM model serves as a quantitative tool to analyze the dynamics of the system as a Markov chain It is common to use discrete states to represent different conformational clusters of the system The transition between the states can be explicitly expressed by a rate matrix K, or a transition probability matrix, M, the Markov matrix The rate matrix K contains the transition probabilities per unit time In the context of continuous time, it describes the progress of an initial population at time 0, p(t) ∈ [0,1]n to the population after a lag time t: p(t) = e(Kt)p(0) (1) where p(t) represents the population of n discrete states at time t The corresponding transition probability matrix or Markov matrix for lag time t is defined as M = e(Kt) A key kinetic quantity of the system is the MFPT matrix, which provides a measure of how quickly the system reaches the target state j starting from state i The MFPT matrix can be calculated from K (or analogously from M) as: tji = 1 / p jeq [(p eq1n T - K)jj-1 - (p eq1n T - K)ji-1] (2) where tji represents the MFPT from state i to state j, p jeq stands for the equilibrium probability of state j ∈ [1,..., n] with shape of [n, 1], and 1n T is an all-ones row vector Eq. p. 1 (2) quantifies the average time it takes for the system to first reach state j, starting from state i The aim is to identify the best u∗ bias that would provide an optimal biased rate matrix, Kbias α, which minimizes tfs MFPT given a starting state, s, to the final state f We envision that the optimal bias is determined iteratively based on the current sampling history, to allow progressive exploration of the phase space over several iterations, α We introduced a relation between unbiased and biased rate or Markov matrices [22], such that the ji-th element of a biased rate matrix Kbias α and unbiased rate matrix K have the following relationship: kjiα kji ∝ exp (-u jα - u iα / (2kBT)) (3) where u iα is the bias for state i in iteration α, kB is the Boltzmann constant, and T is the temperature of the simulation We consider the system with its Hamiltonian H = U(x), where x ∈ R 3N represents the 3D coordinates of the N atoms of the system, and U represents the overall potential energy We define a Markov transition probability matrix typically in a reduced space, via a d-dimensional CV space: ξ(x) ∈ R d As bias, in this work we use a sum of Ngauss Gaussian functions in the CV space, with the i-th Gaussian kernel Giα centered at μiα, with amplitude Aiα and width σiα, such that: ubias α(ξ) = ∑ Giα(μiα,Ai α, σi α; ξ) i=1 (4) The Gaussian parameters (μiα, Ai α, σi α) can then be optimized such that they yield the minimal MFPT value tfs from the starting state s to the final state f The search for the optimal bias ubias α of the system is a simple minimization problem as defined below: ubias α = argmin Ubias(ξ) to minimize MFPT(xs → xf)K (5)
III. RESULTS
We first demonstrate the algorithm for a 1D model potential derived rate matrix K The original FES is constructed using an analytical potential U (Fig. 1, black line) that is discretized into n states The transition rates in the unbiased rate matrix are restricted to neighboring states only, where all the non-neighboring transition rate are set to 0: K i,i±1 = Aexp (-(U(i±1)-U(i)) 2kBT) (7) We calculated the optimal bias for different starting states (Fig. p. 2 1, symbols) to the target state (green star in Fig. p. 2 1a) The results show that the total bias together with the original free energy profile yields a downhill profile from the respective starting points The optimal bias takes up the inverse shape of the original profile with an added slope, ensuring also that the system is unlikely to turn backwards in 1D, further away from the target position, justifying the objective function used as the MFPT: MFPT(xs→xf) A 2D example is shown in Fig. p. 3 2, indicating that this approach is generalizable to higher dimensions
IV. CONCLUSIONS
In this paper, we propose an MSM-based, kineticsoptimized enhanced sampling algorithm that optimizes the applied bias to minimize the MFPT to a target state By adaptively updating the bias based on an MSM-based analysis of the discretized phase space using the DHAM method [22], our approach significantly reduces the time required to reach the target states in our test cases These improvements are demonstrated by orders of magnitude smaller MFPTs for the 1D and 2D analytical potentials and all-atom simulations of solvated NaCl Furthermore, our iterative bias generation method facilitates a more direct trajectory from the local starting point to the target state within the phase space, demonstrating that the KOMBI algorithm efficiently directs the system towards the global target without extensive exploration of non-essential directions in the CV space, yet providing a likely pathway We also provided a practical example using our novel algorithm on the Ras-Ras.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Kinetics-driven enhanced sampling strategy for rare events: The algorithm
KOMBI (Kinetics-Optimized Markovian Bias Integration)
can be used to identify an optimal bias that minimizes mean first passage time from a starting state to a target state, as defined by the objective function MFPT(xs → xf). -
Kinetic information integration into Markov State Models: The system can utilize the dynamic histogram analysis method (DHAM) and its extensions to build unbiased Markov matrices from multiple biased trajectories, yielding unbiased rates/Markov matrices and the free energy surface (FES).
-
Adaptive biasing for phase space exploration: The algorithm iteratively updates the bias by minimizing MFPT to a local target state, allowing for
progressive exploration of the phase space over several iterations
and ensuring thatthe system is driven towards the target state and is less likely to extensively sample other metastable states.
-
Benchmarking complex biochemical processes: The improved system can be applied to simulate events like NaCl dissociation in explicit water or phosphate unbinding in Ras GTPase, achieving orders of magnitude smaller mean first passage times compared to state-of-the-art methods like WT-MetaD.
Related papers
- Supercoiling Drives DNA-Protein Co-Condensation
- Small-angle solution scattering: from fundamental theory to practical approximations
- Interfacial Resistance and Preferential Internal Mixing of Phase-Separated Condensates
- The effect of limb kinematics on the speed of a legged robot on granular media
- Rapid prediction of organisation in engineered corneal, glial and fibroblast tissues using machine learning and biophysical models
- Viral Lattice Theory: A Biophysical Framework for Collective Virion Motion