Molecular noise modulates transitions in the cell-fate differentiation landscape Authors: Yujing Liu, Adriana Zanca, Michael P H Stumpf, Lucy Ham Published: 2026 ABSTRACT AbstractWaddington’s epigenetic landscape has become one of biology’s cornerstone metaphors, widely used both conceptually and computationally. Cell types are often associated with the qualitatively stable points or valleys of this landscape. In previous work, we showed that the molecular noise dominating sub-cellular dynamics can distort and profoundly reshape this landscape. In non-equilibrium systems, an equally profound question arises: to what extent does noise alter the transition paths between valleys in such dynamic landscapes? We tackle this question using a set of illustrative exemplars and show that noise gives rise to paths that differ substantially from the canonical least-action paths calculated under deterministic dynamics. We dissect the dynamics of these exemplars and determine the reactive density and transition currents, which show us, respectively, where and how transitions occur for different realizations of stochastic dynamics. Our analysis unambiguously demonstrates that reaction paths for stochastic dynamics diverge non-trivially from their deterministic least-action paths or simple barrier crossing models.Subject terms: Biophysics, Computational biology and bioinformatics, Mathematics and computing, Physics, Systems biology Introduction During early life and development, embryonic stem cells can differentiate into all the different cell types within the body1. Tissue-specific stem cells can generate all, or at least subsets of, the cells in a given tissue and are important during growth, tissue homeostasis, and tissue repair and regeneration2. The process by which an embryonic or tissue-specific stem cell ultimately gives rise to specific cell types is known as stem cell differentiation. This process is tightly regulated and involves a cascade of events that lead to the activation or repression of specific genes. These genes determine the characteristics of the specialized cell, such as its function, structure, physiological and spatial niche3–5. The genes and proteins that a cell expresses reflect, but also shape, all these biophysical, biochemical, and cellular physiological characteristics. As a convenient shorthand, we often identify the corresponding cellular states (such as expression patterns) with the phenotype of a cell6. In the context of cell differentiation, for example, the cell’s phenotype is often assumed to be characterized by its gene expression profile, which is influenced by both intrinsic factors (such as the cell’s genetic makeup) and extrinsic factors (such as signals from neighboring cells or the surrounding environment)7,8. One of the central questions in developmental biology is to understand how the correct cell types appear at the right place and time in the correct numbers and proportions. To attain a given cell type, a cell (or lineage) must therefore traverse a particular differentiation path (out of a vast number of such potential paths) in a self-organized manner through activation of the correct signalling and gene regulatory networks (GRNs)9. Decoding this complex biological transition process can help us identify key regulatory mechanisms of cell differentiation. This has fundamental importance and offers transformative potential for applications in medicine, tissue regeneration, and biotechnology. For example, it reveals how healthy cells transition into diseased states such as cancerous states6,10–12, or how to control transitions between different cellular states for tissue repair or organ regeneration13–15. The dynamics of cell differentiation16 have been studied predominantly through the lens of (deterministic) dynamical systems theory. This perspective originates from Waddington, who introduced the “epigenetic landscape" as a metaphor for cell differentiation. In this metaphor, the landscape consists of continuous hills and valleys, where each point in the landscape represents a cellular state. Cells undergoing differentiation can be conceptualized as marbles rolling down a hill from pluripotent states (cells that can give rise to all cell types) towards a stable, differentiated state in the valleys. Most subsequent modelling has focused on “quantifying” the landscape to determine the “energy" (where “energy" is not physical energy, but used as an analogy representing differentiation potential) of different cell states and the barriers between them17–19. Gene regulation, signalling, and effects from the microenvironment guide cells through the differentiation process and will render the landscape/manifold as dynamic rather than static20. The transition path for stem cell differentiation can be understood as the series of intermediate states or stages that a stem cell passes through as it changes from a less differentiated state to a more specialized state21. This path is often depicted by a trajectory in a suitable high-dimensional space representing the changing gene expression profiles of the cells throughout the course of differentiation16. In a dynamical system, a transition pathway can be understood as the route along which a cell transits from one stable attracting state to another on the potential landscape, or another suitable manifold. Cells are inherently noisy systems, with fluctuations arising from the stochastic production and degradation of molecules, as well as interactions with neighboring cells. This noise can profoundly reshape the structure of the epigenetic landscape, often in non-intuitive ways22. Traditionally, the landscape has been viewed as being shaped by deterministic forces, which are characterized by a quasi-potential. However, noise interacts with deterministic forces, changing the structure of the landscape. Such changes include creating or destroying attracting states and promoting or inhibiting transitions by changing the heights of barriers between attracting states. Exactly how noise changes transition pathways is still not well understood20,23. To study transitions between cellular states, the minimum action path (MAP) has been used under the assumption that a quasi-potential exists18. Motivated by the large deviation principle, the MAP measures the minimum energy that is required for a system to transition between attractor states24,25. It has been used to identify the most probable pathways between different cell states, effectively providing a deterministic approximation of rare stochastic transitions18,26–29. The advantage of this approach lies in its computational efficiency and adaptability to high-dimensional spaces. However, it has important limitations when analyzing the role of noise in transitions. Firstly, the MAP is only valid in the small noise limit. This assumption often does not hold in biological systems, where noise can be substantial and system-altering. Second, it offers only a coarse approximation by capturing the exponential scaling of transition rates, without resolving finer details of the dynamics (see Supplementary Note 1 Section 1). Most importantly, the MAP describes a single, optimal path and thus fails to capture the full range of possible transition dynamics. As illustrated in Fig. 1, multiple transition paths may contribute meaningfully to differentiation. The MAP alone cannot reveal the distribution or relative importance of alternate transition pathways, limiting its ability to describe the diversity and richness of real stochastic transitions. Example paths between cell states in a double-well potential system. a Transition trajectories begin at the initial state A indicated by the blue cell, transition towards the target state B, indicated by the red cell. Yellow paths represent example transition trajectories, while the red path represents the MAP of transitions. The yellow shaded area represents the distribution of transition pathways or reactive density in TPT. b Left panel: Simulated trajectories with light blue areas representing transition pathways. Blue and red shaded regions represent the cellular states A and B, respectively. Right panel: The stationary distribution of a cell-fate process (black curve) and an estimated transition pathway distribution (light blue shaded curve). Here, we employ transition path theory (TPT), which characterizes the statistical properties of transitions between two attracting states. TPT allows us to consider the full ensemble of transition paths connecting two attracting states. TPT has been successfully applied in many biological systems, including extracting transition pathways in protein folding, identifying parallel transition routes in epithelial-to-mesenchymal transition, and quantifying the likelihood of all possible transition trajectories between cellular states using single-cell data30–32. TPT was originally developed as a purely theoretical perspective. The introduction of RNA velocity enables the reconstruction of cellular landscapes from scRNA-seq data by deriving vector fields of cellular dynamics33. Recent studies have used TPT to characterize transition dynamics in single-cell datasets, revealing key insights into the reprogramming of gene expression. In particular, TPT has shown that interactions between cell communities intensify during cell-fate transitions, coordinating global gene expression changes and guiding cells toward distinct developmental trajectories34. Despite the broad applicability of TPT, few studies have examined how noise influences transitions within this framework. In this work, we systematically map the impact of noise on transition pathways using TPT. Our findings show that noise does more than modulate barrier heights between attractor states, it can fundamentally reshape the transition landscape, revealing alternative intermediate states and pathways that differ markedly from those observed in the absence of noise. Results We use potential landscapes alongside transition path theory to determine the influence of different noise structures on the paths between cell fates. We explore four exemplars to illustrate the merit of using such an approach to investigating cell fate. Our exemplars are formulated as stochastic differential equations (SDEs) of the form where Xt represents the state of the system at time t, μ(Xt) represents the deterministic forces, and g(Xt) is the diffusion term accounting for the intrinsic stochasticity. The dWt term refers to an increment of a Wiener process. Following Waddington, cell-fate dynamics are often modelled as gradient systems with a potential function U(x): However, we note that when noise is present, the long-term dynamics are more accurately captured by the stationary distribution, where areas of higher probability density can be understood as the attracting states in the landscape. The stationary distribution is often expressed as a Gibbs-like distribution governed by a potential landscape16,35, where ρ(x) is the stationary distribution and U~(x) can be considered as a quasi-potential function under the influence of noise. In one-dimension, an analytical general form of ρ(x) can be obtained in terms of U~(x) (see section “Methods”). Most cell-fate models in the form of Eq. (1) assume additive noise (i.e., g(Xt) = σ), even for biological systems where we know that this is unfounded23. To obtain an accurate description of cell fate, it is necessary to incorporate multiplicative noise, which captures state-dependent properties, into our models. A recent study has shown that different types of noise can influence the landscape in distinct ways. While additive noise can only quantitatively change the structure of the landscape (i.e., valleys become shallower and hills become flatter), multiplicative noise can qualitatively reshape and distort the landscape22. However, previous analyses provide only a static view of cellular states. The critical dynamics of noise driven transitions, specifically how long it takes for a cell to transition between states, which pathways are most likely, and where cells tend to dwell along these trajectories, remain unresolved. We quantify these transitions between different states via transition path theory (TPT), which was introduced to study rare transitions. Transition properties are evaluated in terms of forward and backward committor functions. For two states, A and B, the forward committor function q+(x) is defined as the probability that a cell currently at point x reaches state B before entering state A. Similarly, the backward committor function q−(x) is defined as the probability that a cell arriving at point x originated from state A rather than state B. We use Ω to denote the state space and ∂A, ∂B to denote the boundaries of states A and B. The committor functions can be written as where τA=inf{t∈R+∣Xt∈∂A} and τA-=sup{t∈R-∣Xt∈∂A} are the first hitting time and last exiting time of state A for a diffusion process (Xt)t≥0 and τB,τB- are defined analogously for state B. The functions q+(x) and q−(x) provide a useful statistical description of transition properties. For example, the probability distribution of a cell found in the transition from A to B at any random time point, known as reactive density36 can be expressed as: where ZR = ∫q+(x)q−(x)ρ(x)dx is the normalization constant. The reactive density tells us where in state space cells spend time during the transition as shown in Fig. 1. By analyzing the reactive density and other properties thereof, such as the transition rates, mean first passage time, and transition probability currents (see section “Methods”), we can comment on the properties of the transition paths between cellular states. Gradient systems in equilibrium We start by considering the system that is in equilibrium (see Supplementary Note 1, Section “Gradient systems in equilibrium”) with canonical one-dimensional double-well potential This system has two distinct local minima at x = ± 1, separated by a saddle point at x = 0. Therefore, the transition from one state to another can be understood as moving the state space variable x from one minimum to the other. For simplicity, we define the neighborhood with radius 0.2 around the minimum x = −1 as state A and x = 1 as state B and consider transitions from A to B. In one dimension, we can show that the forward committor function is where bA, chosen to take a value of −0.8 in this illustrative example, is the upper bound of the basin of attraction of state A and C∈R+ is the normalization constant that enforces q+(x) ∈ [0, 1], for all x in the state space excluding states A and B and their corresponding neighborhoods. The equilibrium property of the system gives the backward committor function as Therefore, the reactive density ρAB(x) can be obtained from Eq. (5) via numerical integration. We first consider the case of additive noise, g(x) = σ, with noise strength σ > 0. The reactive density, ρAB(x), given in Fig. 2a, describes the transition dynamics for increasing values of σ. The density ρAB(x) peaks at the saddle point, x = 0, indicating that this region around the saddle point serves as the transition state for the system. When the noise strength σ is low, the density is heavily localized at the saddle point with tails extending to its surroundings. That is to say, the system will be trapped at the transition state in the absence of noise, where reactive density acts like a Dirac delta function centered at the saddle point. The transition can occur only when there is some noise associated with the system. Noise breaks the stability of the transition state at the saddle point, allowing the cell to transition between states A and B. A transient change in the density peak around the saddle point is observed with increasing noise strength, as shown in Fig. 2a (left panel). This suggests that the stability of the transition state decreases with increasing noise strength initially, but subsequently begins to increase and stabilize. The stability of the transition state can be quantified using the value of the reactive density at the saddle point, shown in Fig. 2c. Our results reveal that the transition is maximally promoted when the noise strength is σ ≈ 0.43, and thereafter the stability of the transition state will gradually increase towards an asymptotic limit as the noise strength grows, corresponding to an increase in the reactive density. Reactive density, transition rate and stability analysis for the one-dimensional double-well potential with additive and multiplicative noise. a The reactive density, ρAB, (x) of the diffusion process under different types of noise. We define states A and B as neighborhoods of radius 0.2 around the corresponding stable points for the additive noise, and similarly for the intrinsic multiplicative noise, where the minima are shifted due to noise. Left panel: additive noise with g(x) = σ. Right panel: multiplicative noise with g(x)=σ4(x4-4x2+8). As noise strength σ increases, the maximum of ρAB(x) concentrates around the saddle point x = 0 for additive noise. While for multiplicative noise, the maximum is bifurcating towards two maxima around the saddle point. b The transition rate, νAB, and mean first passage time, E[τAB], in log10 scale for different form of noise. With the increase of noise strength σ, the transition rate is increasing monotonically, and the mean first passage time is decreasing monotonically. In the small noise limit, the system with multiplicative noise is less likely to have transitions than a system with additive noise. Conversely, for the large noise limit, the multiplicative noise system is more likely to have transitions. c The reactive density for additive noise at the saddle point (transition state) under different noise strength σ. The red dashed line represents the asymptote for large noise strength. The minimum value of the reactive density, corresponding to lowest stability, is attained at σ ≈ 0.43. In the case of multiplicative noise, we consider g(x)=σ4(x4-4x2+8), again with noise strength σ > 0. The introduction of multiplicative noise into the system can lead to a qualitative change in the structure of the landscape22. Specifically, here the positions of the attracting states are altered: the minima in the original landscape ρ(x) are pushed away from x = ±1 towards x=±2 as noise strength increases. We consider the neighborhoods with radius 0.2 around shifted minima as new attracting states and label them as states A and B, as before. Figure 2a (right panel) gives the reactive density of the multiplicative noise system across varying noise strengths. Similarly to the case of additive noise, with low noise strength, the transition state remains concentrated around the saddle point. However, rather than being tightly localized around the saddle point, the transition state region expands with increasing noise strength, covering a broader neighborhoods due to the interplay between deterministic forces and multiplicative noise. Moreover, unlike the case of additive noise, where the stability of the transition state exhibits transient changes with increasing noise strength, the stability of transition states continuously decreases as noise intensifies. The destabilization of the transition state leads to a bifurcation where the original transition state is split into two states symmetrically arranged around the original saddle point at x = 0. Thus, under high enough multiplicative noise, there are two transition states located at points equidistant from the original saddle point, which is no longer a transition state. Two quantities provide insight into the transition dynamics: the transition rate and the mean first passage time (mfpt). The transition rate, νAB, quantifies the mean number of transitions of the process from state A to state B as t → ∞ (see Eq. (32) of section “Methods”). The mfpt, E[τAB], is the average time the process takes to reach state B from state A (see Eq. (29) of section “Methods”). Analyzing the transition rate and mfpt for the considered additive noise and multiplicative noise terms (Fig. 2b) reveals that increasing noise strength, σ, generally promotes transitions. This finding agrees with a previous study22, which showed that larger noise strength can lower the transition barrier, therefore promoting transitions. Additionally, we find that when the noise strength is low, transitions are less likely to take place under multiplicative noise g(x)=σ4(x4-4x2+8) than additive noise. However, when noise strength is large enough, transitions occur more frequently with noise g(x). This behaviour further suggests that, while there are more transition states due to bifurcation under noise g(x), these transition states are less stable than the original one at the saddle point and therefore allow for easier transitions between states. We now formalize this intuition about how the structure of the reactive probability density changes under multiplicative noise (see Supplementary Note 1, Section 3 for details). Specifically, we have As a result, the mode of the reactive probability density function approaches that of this limiting distribution, which is governed by the multiplicative noise g(x). Gradient system in non-equilibrium We now extend our study to higher-dimensional systems where multiple transition pathways may exist, subject to the direction of probability flux. Here, we approximate the numerical solution of the forward committor function using the finite element method (see Supplementary Note 1, Section 5). We consider the two-dimensional double-well potential given by: which consists of a double-well potential in the x-direction and a parabolic function in the y-direction. Due to the higher-order terms in the x-direction, the potential function is primarily influenced by variations along x-direction. The landscape again has two minima at (±1, 0) that are separated by a saddle point at the origin, (0, 0). The attracting states, A and B, are defined as the basins around the two minima of the probabilistic landscape at (±1, 0), respectively. Our analysis is conducted over the domain Ω = [−2.5, 2.5] × [−2.5, 2.5], ensuring sufficiently high potential values at the boundaries. This results in the stationary probability density function ρ(x, y) effectively vanishing around the boundaries of the domain. We start with the simpler case where only additive noise is present in the system for both the x- and y-directions where D1 is the diffusion matrix, and σ is noise strength. In this case, the structure of the landscape remains unchanged with states A and B at (±1, 0) separated by the saddle at (0, 0) (Fig. 3a, left panel). The stationary probability density function ρ(x, y) and the reactive density ρAB(x, y) for the two-dimensional double-well potential system with different noise terms. a The stationary probability density function with additive noise described by diffusion matrix D1 on the left and multiplicative noise described by D2 on the right. b The reactive density ρAB(x, y) for with additive noise on the left and multiplicative noise on the right. We define states A and B as neighborhoods of radius 0.2 around the corresponding stable points for the additive noise and similarly for the intrinsic multiplicative noise, where the minima are shifted due to noise. The white regions correspond to the attracting states A and B, respectively. The red arrows represent the probability current JAB of the transition processes. The larger the arrows, the higher the probability that cells will transition through these regions. Introduction of multiplicative noise into the system leads to the creation of new transition states symmetric around the target attracting state, as well as enabling new transition pathways towards the target attracting state. The noise strength σ is set to be 0.6 here for both additive noise and multiplicative noise. Note that when the additive noise is coupled with the gradient drift, the system is in equilibrium37 and consequently the backward committor function can be simplified to Figure 3b (left panel) gives the contour plot of reactive density ρAB(x, y) alongside the transition probability current (quantifying the flow of probability density), JAB, represented by red arrows, where the value of the flux is indicated by the arrow size. The larger the arrows, the higher the likelihood that cells will transition through these regions. For the mathematical definition of transition probability current JAB, see Eq. (27) of section “Methods”. One can see that the transition pathways in this system are mostly through the central region crossing the saddle point, which directly connects states A and B. This is represented by both the higher reactive density ρAB(x) around the central region and the higher probability current JAB crossing this region. Therefore, the area around the saddle point can be viewed as the transition state of this system. Additive noise does not change the transition dynamics or the topology of the reactive density function, similar to the previous section. Its influence on the transition dynamics is mainly in reducing the stability of the transition state with increasing noise strength, making it easier for a cell to transition from state A to B. This is represented by both a decrease in reactive density and an increase of flux through the transition state (see Supplementary Note 1, Section 7). Now we turn to the case of multiplicative noise where we take as the diffusion matrix in the system. While the intrinsic features of the landscape are maintained—the system again has two attracting states separated by a saddle point in between—multiplicative noise alters the position of the attracting states where minima are pushed away from (±1, 0) towards (±12,0) with increasing noise strength. Although the overall intrinsic appearance of the landscape is preserved, multiplicative noise now also introduces new transition pathways as well as new transition states. The change in transition pathways is illustrated in Fig. 3b (right panel). When additive noise is present, the system exhibits a dominant transition pathway that connects attracting states A and B through the saddle point. However, when multiplicative noise is present, the transitions into state B will not proceed directly through the original transition state. Instead, the original pathways are bifurcated into two distinct and equally weighted branches that vary in the y-direction, towards the target state. The two bifurcated pathways have equally weighted probability due to the symmetric multiplicative noise and potential. This result highlights the role of multiplicative noise in changing the transition dynamics of the system: it can create other transition routes and enable entrance into the desired attracting states through potentially different molecular spaces. The creation of new transition paths can be understood as resulting from the interplay between noise and the state-dependent dynamics of the system. Multiplicative noise introduces state-dependent fluctuations, which lead to a change in (“effective”) curvature of the local landscape as well as the local stability. As a result, the interaction between the drift force and stochastic perturbations suggests that it may be worthwhile exploring multiple potential pathways, rather than concentrating on the highest probability one. These findings have implications for understanding biological processes where noise is an intrinsic feature that depends on the cellular state. Previous studies17 assumed transitions take place following routes mostly controlled by deterministic forces. Here, we highlight the influence of noise on shaping the transition dynamics; this includes external or environmental sources of noise, as new transition pathways can also be formed when cells are placed in a noisy environment. Therefore, different noise environments can, in principle, lead to different transition states and transition pathways during cell differentiation. Reaction networks in equilibrium (Schlölgl model) So far, our study has focused on theoretical potential models where multiplicative noise g(x) is independent of the drift term, f(x). We now extend our analysis to biological systems where noise can appear as intrinsic or extrinsic. Regardless of the specific noise type, the system in one dimension remains in equilibrium. To demonstrate this, we study the bistable chemical reaction system introduced by Schlölgl where A and B are chemical species with constant concentrations a and b, respectively, and x is the concentration of the species X, which is allowed to change over time. This reaction system can be expressed as a gradient flow of a potential energy: where the parameters c3 = k2, c2 = (a + 3)k1, c1 = (ak1 + 2k2 + k4), c0 = bk3 are chosen such that the deterministic system exhibits two stable states, allowing for transitions between them. When additive noise is incorporated into the system, the dynamics can be described by, For intrinsic multiplicative noise, we consider the chemical Langevin equation (CLE): where the noise term g(x) now depends on the deterministic drift f(x). Previous studies22 have shown how different noise types can change the structure of the landscape for the Schlölgl model. In the case of purely additive noise, the overall structure of the landscape is preserved, with increasing noise strength effectively rescaling the landscape and lowering the transition barrier between the two attracting states. In contrast, multiplicative noise causes a gradual change in landscape structure: while the bistable structure of the landscape is maintained at low noise strength, the landscape becomes increasingly skewed at higher noise levels, eventually forming a single state. This occurs because noise is significantly higher at the far right of the landscape compared to its strength around the origin, which strongly pushing cells into a single region near the origin. Here, we are concerned with transitions between cell states, and restrict our study to the cases where there are two states for cells to transition between. Attracting states A and B are therefore defined by the neighborhoods with radius 1.0 around the corresponding minima of the potential landscape (Fig. 4a). Regarding the transition dynamics, as shown in Fig. 4b, at low noise strength, the reactive density exhibits a similar qualitative shape in both cases of additive and multiplicative noise, with a peak located around x ≈ 2.9 and a monotonic decay for larger values of x, particularly beyond x ≈ 3.5. This indicates that, in the weak noise regime, the transition property is mainly determined by the landscape structure and is less sensitive to the specific noise structure. The stationary and reactive densities of the diffusion process of the Schlölgl model under different types of noise. We define states A and B as neighborhoods of radius 1.0 around the corresponding stable points for the additive noise, and similarly for the intrinsic multiplicative noise, where the minima are shifted due to noise. The left panels of a and b show the results for additive noise with g(x) = σ. The right panels of a and b show the results for multiplicative noise with g(x)=σ(c0+c1x2+c2x+c3x3). The shaded yellow regions in a correspond to transition regions between cell states, with corresponding reactive densities, showing transition path distributions, given in (b). Note that in the case of multiplicative noise, the locations of the cell states, and therefore the boundaries of the cell states, change, hence the graded yellow shading in the right panel of (a) (and different x-intercepts in (b)). As σ increases, the maximum of ρAB(x) concentrates around x ≈ 3.5 for additive noise while for multiplicative noise, the maximum is sheared and tends to be around the left with x ≈ 2.9. The parameters for the system are chosen as c0 = 0.4, c1 = 0.55, c2 = 0.20, c3 = 0.02. This results in a potential landscape with two attracting states for both additive noise and multiplicative noise. As the noise strength increases, the behaviour of the two systems diverges. For additive noise, the peak of the reactive density shifts towards x ≈ 3.5 and becomes approximately symmetric around this point, adopting a parabolic form centered at the midpoint. This is consistent with the theoretical equilibrium model for the one-dimensional double-well potential system. In contrast, under multiplicative noise, the peak of the reactive density remains located around x ≈ 2.9, while the overall shape becomes increasingly asymmetric, skewing the stability of the transition state towards the left side of the transition region. This result is similar to previous studies: the stronger noise at the right extremes effectively pushes cells away towards the region around the origin, and thereby reactive trajectories are biased towards the left side of the landscape, stabilizing the transition state closer to the origin and suppressing transitions towards the right extreme. Since the system is one-dimensional and in equilibrium, the limiting behavior of the reactive density function with high noise strength can be described by Eq. (9). In this case, the resulting shape of the transition state comes from the interplay between the parabolic potential and the state-dependent structure of the multiplicative noise, highlighting the implicit influence of intrinsic multiplicative noise in shaping transition dynamics. Reaction networks in non-equilibrium (toggle switch model) To further explore the effect of intrinsic noise on transition pathways in non-equilibrium chemical systems, we extend our analysis to a two-dimensional system. We examine the gene toggle switch model, which is used to describe binary cell-fate decisions. The system contains two genes, X and Y, that are self-activating while mutually inhibiting each other (Fig. 5a). The stationary probability density and reactive density for the toggle switch model. a A schematic of the toggle switch model, consisting of two mutually repressing genes X and Y that also self activate. b The stationary probability density function ρ(x, y) for the toggle switch model with additive noise (left) and multiplicative noise described by the CLE (right). c The reactive density ρAB(x, y) for the toggle switch model with additive noise (left) and multiplicative noise (right). We define states A and B as neighborhoods of radius 0.35 around the corresponding stable points for the additive noise, and similarly for the intrinsic multiplicative noise, where the minima are shifted due to noise. The white regions where arrows point outward and inward correspond to the attracting states A and B, respectively. The red arrows with magnitude represent the probability current JAB of the transition processes. The larger the arrows, the higher the probability that cells will transition through these regions. The introduction of multiplicative noise into the system leads to change in dominant pathway, from a winding movement into a heavily dominant one through the lower area of the transition state. The parameters for the system are a = 0.65, b = 1.2, k = 0.8, s = 0.6, n = 4, g0 = 0.2. The chosen parameters result in a landscape with three attracting states for both additive noise and multiplicative noise. The noise strength σ is set to be 0.45 here for both additive noise and multiplicative noise. The dynamics of this system are described by a general SDE of the following form: where a Hill function is used to model the self-activation and mutual inhibition. Here g0 is the basal synthesis rate for genes X and Y, a is their self-activation rate, b is their mutual inhibition rate, k is their self degradation rate, n is the number of regulators binding to the genes, and S represents the threshold of the Hill function38; this is assumed to be the same for both genes. The system is again considered with additive noise and multiplicative noise. For additive noise, the diffusion matrix is given by D= [σ200σ2].For multiplicative noise, the diffusion functions are defined through the CLE (see Supplementary Note 1, Section 2.3), with: In both cases, the system has three metastable states: two attracting states A and B, and an unstable transition state at the center (Fig. 5b). We study the transition dynamics from state A to state B through the central transition state. In the absence of noise, the MAP connects the two metastable states through the central unstable state; this follows the steepest descent path dictated by the potential gradient. When additive noise is introduced into the system, the overall structure of the landscape will be maintained. However, the curl force in the non-equilibrium system distorts the MAP from the steepest descent path and leads to a shifted pathway. Instead of following the steepest descent path, the probability current exhibits winding movements as it transitions through the transition state (Fig. 5c left panel). This observation under non-equilibrium conditions agrees with Wang et al.’s results17, demonstrating that the MAP deviates profoundly from the steepest descent path under the influence of the curl force in non-equilibrium systems. With the introduction of multiplicative noise formulated by the CLE to the system, the structure of the landscape is subtly altered from the original one while the intrinsic tristable structure is preserved: the three metastable states remain intact, without any bifurcations, destruction, or creation of the attracting states. However, the dynamics of the transitions are heavily reshaped by the intrinsic multiplicative noise (Fig. 5c). With additive noise, transitions between attracting states follow a curved, undulating path, with a relatively higher probability current coming out of state A and entering state B, while the current through the central transition state is reduced. This indicates that the transition state is relatively stable and transitions are less likely to occur when cells reach this transition state. In contrast, the probability current through the center of the transition state increases under intrinsic multiplicative noise. The higher currents through the transition state, compared with those coming from state A, suggest easier transitions through the transition state and highlight the decrease of stability at the transition state under intrinsic multiplicative noise. Additionally, the shape of the transition pathway shifts appreciably under multiplicative noise. Instead of traversing the straight path as under a purely deterministic regime, or the same undulating path as under additive noise, the transition pathway, manifested by the probability current, passes predominantly through the lower region of the state space. The closer the cell is to the lower region of the accessible state-space, the more likely it will move directly into attracting state B. This directional bias again arises from the coupling between the intrinsic multiplicative noise and the deterministic forces governing the system. These findings highlight the role of intrinsic multiplicative noise in reshaping the dynamics of transitions. While the fundamental properties of the system remain intact, intrinsic multiplicative noise can lead to completely different transition dynamics by modifying the overall structure of the landscape. Discussion Our study provides an intuitive perspective and understanding of how noise, particularly multiplicative noise, influences transition dynamics in equilibrium and non-equilibrium systems. We have here focused on exemplar systems that highlight typical behaviour and that can help us hone our intuition. Larger and more complex systems will exhibit similar behaviour, but in a manner where we cannot expect to disentangle the different contributing factors. While it is known that the interplay between noise and deterministic forces can change the structure of the landscape and give rise to qualitatively different features in the landscape, the influence of noise on transition dynamics, to the best of our knowledge, has not yet been studied in detail. By employing TPT in the context of Waddington’s landscape, we show that multiplicative noise can alter transition pathways by modifying the effective structure of the landscape, even when the number of attracting states (our proxy for different cell types) may be maintained. When studying cell-state and fate transitions, previous approaches have generally focused on the MAP, where transitions occur along minimal action paths. This framework can only provide information about transitions along a specific path and neglects the actual distribution of transition pathways. As we have seen, the MAP quickly becomes unrepresentative of the true dynamics: during cell differentiation, small random fluctuations result in deviation from the MAP that becomes more severe and pronounced as the noise strength increases. Intrinsic noise arising from the probabilistic nature of chemical reactions and gene expression is a hallmark of biological systems at the cellular, sub-cellular, and molecular levels. A simplistic approach to incorporating noise is to use additive noise, an approach often favoured for its simplicity. We demonstrate that additive noise can only lead to changes in the stability of the transition state with increasing noise strength, while maintaining the underlying structure of the transition state as well as the transition pathways. However, additive noise models neglect the coupling between intrinsic noise and deterministic drift forces. More realistic forms of noise must reflect the overall dynamics and system state. This is particularly relevant in cell differentiation, where noise is known to depend on the system state: noise tends to be lower around the attracting states and higher near the unstable states, such as saddle points. By varying the stochastic component of a fixed deterministic system using multiplicative noise (i.e. noise that depends on the system state), we demonstrate that transition states can bifurcate and be reshaped. Similarly, transition pathways can also be altered via multiplicative noise, even in cases where the transition states themselves are not. Intrinsic noise not only amplifies or accelerates transitions but also, in a very concrete sense, enhances the system’s capacity to explore alternate states while preserving its fundamental structure. By examining the toggle switch model derived from the chemical Langevin equation (Eq. 17), we showed that in this case, noise can alter the dominant transition pathway, even while preserving the underlying landscape structure and transition states. Only in the simplest cases is it possible to obtain analytic results in the presence of noise. When analytic results are not obtainable, particularly in high-dimensional state spaces, we must turn to numerical methods to approximate properties of interest, such as the stationary probability density function, reactive density function, and mean first passage time. Here, our results in two dimensions are obtained by solving for the forward committor function using finite elements. Our method is less effective in the small noise regime (i.e. σ≪1) as the density function becomes singular at the attracting states, leading to numerical limitations when computing densities directly. However, the MAP, which is appropriate in small noise regimes, may provide some insight about transitions in this case, even if this approach cannot provide the distribution of transition paths. Furthermore, we note that the notion of a cell type is diffuse. In general, cell states can be thought of as stationary points of deterministic systems and a region around the stationary points in which the corresponding stochastic systems accumulate. The choice of neighborhood around attracting states significantly influences transition dynamics (see Supplementary Note 1, Section 5). Intuitively, a large choice for the basin boundary around the attracting states may encompass the area of the actual transition states, leading to different transition behaviors. In our approach, neighborhoods are defined based on specific quasi-energy levels of the corresponding deterministic systems. For practical applications involving experimental data, clustering methods or other boundary selection techniques may be necessary or preferable to accurately specify the attracting states in the system. Our study uses the Itô integration convention, as is common in the field of biological applications of stochastic processes39. However, different conventions may lead to distinct dynamics and potentially different global landscape structures40. In Supplementary Note 1, Section 6, we present the reactive density for the double-well potential system under Stratonovich (α = 1/2) and anti-Itô (α = 1) conventions. We show that different choices of stochastic integration conventions modify the spatial localization and relative dominance of transition states, resulting in qualitatively different transition dynamics. From a biological perspective, this suggests that cells may transiently reside in different intermediate states for different durations and thereby exhibit distinct cell-fate transition patterns depending on how stochasticity manifests. An avenue for future studies could consider the benefits of using A-type conventions, which preserve key deterministic information, including positions of stationary points and Lyapunov equivalence41. In our study, the attracting states A and B are defined based on the noise-induced shifts of the potential minima. Under A-type conventions, transition path theory quantities may be derived without explicitly accounting for such deviations in the positions of the minima40,42. Our study has focused on exemplar systems for computational and illustrative purposes. However, cell differentiation typically occurs in a high-dimensional state (typically gene expression, currently considered the most conveniently experimentally available) space. Here, transition pathways may exhibit more intricate changes under the influence of noise. We note that recent advances have introduced novel techniques for computing the statistical properties of TPT in higher dimensions43,44, which provide a foundation to extend our study to higher-dimensional systems. An extension of our work to experimental data is possible in principle; however, we have to note that our current findings regarding noise reshaped transition pathways should be viewed as theoretical predictions. While our framework demonstrates these effects in silico, direct experimental validation remains an open frontier. The main challenge of this lies in inferring both the deterministic force and noise from experimental data. RNA velocity allows us to infer deterministic forces from scRNA-seq data, but it typically assumes additive noise33. To account for multiplicative noise and potentially test our predictions empirically, recent studies employ a score-matching method with neural networks, enabling the inference of deterministic force in the presence of multiplicative noise from experimental data45,46. New methods of inferring gene regulatory dynamics from single-cell data that make sophisticated use of the single-cell transcriptomic data manifold offer promising routes to construct models that can be used to validate and explore these transition pathways in future data-driven studies47,48, especially if experimental setups were to make use of flexible reporter assays that allow us to dissect relative contributions from extrinsic and intrinsic noise49. Probability and randomness tend to defy our intuition, and their effects become even more unpredictable in the context of nonlinear systems, even for the simple systems described here. At the cellular, sub-cellular, and molecular level, the effects of randomness are ubiquitously felt, including during development and cell-fate decision making processes. What we have argued here is that, as soon as biochemically relevant noise (CLE for example) is present, the path a system takes through the relevant gene expression space can differ profoundly from what we see for the conventional deterministic case. There can be great variability among transition paths, and this variability may lead to cells exploring alternative fates. As systems become more high-dimensional, more local minima in the quasi-landscape, or more locally stable fixed points, tend to arise. Understanding how developmental systems, and complex stochastic systems more generally, manage to balance stochasticity with phenotypic or overall robustness requires careful analysis within an appropriate mathematical framework. Relying on convenience—for example, by using deterministic approaches or assuming purely additive noise—can obscure the underlying dynamics and be misleading. Our analysis clarifies the mathematical role of noise in shaping state transitions, providing deeper insight into cell-fate decisions within the landscape framework. Methods Quantifying Waddington’s landscape The behaviour of dynamical systems with noise can naturally be described by an SDE (Eq. (1), is here restated for to keep the presentation self-contained): where Xt represents the state of the system (for GRNs, typically defined in terms of molecular concentrations) at time t, μ(Xt) represents the deterministic forces governed by the GRNs and g(Xt) is the diffusion term accounting for the intrinsic stochasticity. The term dWt refers to increment of a Wiener process which satisfies E[dWt] = 0 and E[dWtdWs] = δ(t − s). For gradient systems (i.e. μ(Xt) = −∇ U(Xt)), the stationary distribution can often be expressed as the Gibbs-like distribution16,35, where ρ(x) is the stationary distribution and U~(x) is the potential function of the landscape under the influence of noise. In one dimension, the general form of stationary distribution can be expressed as22: with where Z is the normalization constant. Therefore, one can approximate the quasi-potential landscape U~(x) from the stationary distribution ρ(x) by conducting simulations of specific models or collecting experimental data and performing density estimation via e.g. Eq. (3)50 (note that in one dimension every dynamical system is necessarily a gradient system). It is well known that the introduction of a state-dependent noise component g(x) leads to subtle but important differences in reconstructing the potential landscape depending on the stochastic integration convention that is adopted41. The most commonly used conventions: the Itô (α = 0), Stratonovich (α = 0.5), and anti-Itô interpretations are parameterized by α ∈ [0, 1], which specifies the point at which the diffusion process Wt is evaluated within each discretization interval. Specifically, for a partition 0 = t0 < t1 < ⋯ < tn = T, the stochastic integral is defined by evaluating the noise at tk(α)=tk-1+α(tk-tk-1). We note that an α-type SDE can be converted into an equivalent Itô SDE through the addition of a correction drift term. Specifically, the Fokker-Planck equation takes the form: where h(x)=∑j=1d∑k=1mgjk(x)∂gik∂xj is the noise-induced drift arising from the choice of stochastic convention. Throughout the main text, we adopt the Itô convention (α = 0), which has been widely used in the biological context. For completeness, some results for the Stratonovich and anti-Itô conventions are shown in Supplementary Note 1, Section 6. Transition path theory Transition properties are evaluated in terms of a forward committor function q+(x), which is the probability that a cell starts in state A and reaches state B before returning to state A. We use Ω to denote the state space and ∂A, ∂B to denote the boundaries of states A and B. The forward committor function can be written as where τA=inf{t∈R+∣Xt∈∂A} is the first hitting time of a diffusion process (Xt)t≥0 that starts at a point x ∈ ΩAB = Ω⧹(A ∪ B) and similarly for τB. One can show that q+(x) is the solution of the Dirichlet problem51,52 with where A is the generator of the diffusion process defined as where μ is the deterministic force and D = ggT is the diffusion matrix of the process. We use the finite element method to numerically mesh the space to solve Eq. (25) and approximate the forward committor function. Similarly, the backward committor function q−(x), which can be solved by the time-reversed generator, is defined as the probability that the diffusion process arriving at a point x originates from within state A. If the process is time-reversible, then we have q+(x) = 1 − q−(x). By analyzing the statistical properties of the committor functions—including reactive density, transition currents, transition rate, and mean first passage time—we can draw conclusions about the transition paths between cell states. We say that the diffusion process (Xt)t≥0 is in transition from state A to state B if it was previously in state A and will transition to state B before it returns to state A. The probability density function of such transition processes can be defined as: where the indicator function IR(t) is 1 if it is in transition and 0 otherwise. By definition, ρAB(x) is exactly the probability that at point x, Xt will visit B first before A and that Xt comes from A, restricted by the probability density function of the whole system at point x. Therefore, recall that the probability distribution of the transition pathways is36 (Eq. (5)): where ZR = ∫q+(x)q−(x)ρ(x)dx is the normalization constant that defines the probability that the process is found in the transition from A to B at any random time point. Note that we refer to ρAB(x) as the reactive density to distinguish it from the stationary density function ρ(x) and to make the terminology consistent with existing literature36,51,52. The probability flow associated with the transition process can be studied via the probability current, JAB, determined by the transition path process as36,53, Under potential theory, the stationary distribution and probability current define the intrinsic structure of the potential landscape; we refer to this as the transition current. Cellular states with a higher value of density ρ can be understood as the attracting states in the landscape. A higher transition current at point x can be understood as a higher flow of probability through point x. Therefore, cellular states with higher transition currents are less stable and more likely to be shortlived. The transition current, therefore, tells us how transitions between cell states occur. Regions with higher reactive density can be understood as transition states in the landscape. Transition states with low transition probability current can be viewed as transition traps of the process, as it is more difficult to transition out of these states. Since we are interested in the effect of noise on the expected time of transitions, we define as the transition rate which quantifies the mean number of transitions until time T, where NAB denotes the total number of transitions from state A to state B in [0, T]. We also define as the mean first passage time (mfpt) from A to B where τB,i+ is the ith entrance time to B and τA,i- is the ith exit time from A. We then have where TAB is the total time that a trajectory is in transition from A to B. The normalization constant ZR, defined previously, can be interpreted as It follows from the definition of the current traversing any surface S separating A and B that the transition rate νAB is the total flux moving from state A to state B through the surface, i.e.,53,54 where n⃗ is the unit normal pointing outwards from the surface S. To study transitions between cellular states in cell differentiation under different types of noise, we use TPT and quantify the statistical properties of the transition pathways described above. We use Eq. (5) to quantify the distribution of transition pathways and Eq. (27) to quantify flows. Additionally, Eqs. (29–32) are used to assess transition times under different noise conditions. Supplementary information Supplementary Information Acknowledgements We thank the members of the Theoretical Systems Biology group for their interest and support. Y.L., A.Z., & M.P.H.S. acknowledge funding through an Australian Research Council Laureate Fellowship to M.P.H.S. (FL220100005). L.H. acknowledges funding through the ARC Centre of Excellence for the Mathematical Analysis of Cellular Systems. Author contributions Y.L., A.Z., M.P.H.S., and L.H. conceived the study. Y.L. conducted the simulations and drafted the manuscript. All authors reviewed the manuscript and gave final approval for publication. Data availability The data that support the findings of this article are openly available in Liu, Y.: yujingll/transition pathway under noise: Transition pathways under noise. Zenodo (2025). https://doi.org/10.5281/zenodo15671167. Competing interests M.P.H.S. is co-founder, director, shareholder, and Chief Scientific Officer of Cell Bauhaus PTY LTD. The remaining authors declare that they do not have any commercial or financial relationships that could be construed as a potential conflict of interest. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Michael P. H. Stumpf, Email: mstumpf@theosysbio.org Lucy Ham, Email: lucy.ham@unimelb.edu.au. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41540-026-00762-8. References Associated Data Supplementary Materials Supplementary Information Data Availability Statement The data that support the findings of this article are openly available in Liu, Y.: yujingll/transition pathway under noise: Transition pathways under noise. Zenodo (2025). https://doi.org/10.5281/zenodo15671167.