Topological potentials guiding protein self-assembly Authors: Ivan L A Spirandelli, Arnur Nigmetov, Dmitriy Morozov, Myfanwy E Evans, Jianhan Chen, Arne Elofsson, Jianhan Chen, Nir Ben-Tal, Jianhan Chen, Nir Ben-Tal, Jianhan Chen Published: 2026 ABSTRACT AbstractThe simulated assembly of molecular building blocks into functional complexes is central to computational biology and materials science. Protein-assembly simulations, driven by short-range nonpolar interactions, can in principle reach their biologically correct structures, but rugged energy landscapes often trap simulations in non-functional local minima. We introduce a long-range topological potential, quantified by weighted total persistence, and combine it with the morphometric approach to solvation free energy. Across four protein systems, this combination increases assembly success rates by up to sixteen-fold and enables assembly in cases that otherwise fail. Unlike previous topology-based approaches, our method uses topological measures as an active energetic bias rather than a descriptive tool. Depending only on atom geometry, the method extends in principle to other self-assembling systems, offering a general strategy for overcoming kinetic barriers in molecular simulations. Introduction In silico assembly of molecular building blocks into functional complexes is a central challenge in biology and materials science. A primary obstacle is the problem of kinetic trapping [1,2]: Even when the functional native structure represents the minimum of the energy function that governs the simulation, the vastness of the configurational landscape, containing many nonfunctional local minima, can prevent simulations from discovering it on a feasible timescale. A successful simulation methodology must therefore not only be based on an accurate energy function that identifies the native state, but also navigate its landscape efficiently to avoid dead-end assembly pathways. Here, we introduce a long-range potential based on the principles of computational topology to act as a guiding bias. Topological data analysis, particularly persistent homology [3,4], is utilized to solve optimization tasks in machine learning contexts [5,6], and has been widely used to analyze static and dynamic molecular data [7–10] and nanomaterials [11,12] post hoc. We use it here as an active component within the simulation of protein assembly. The physically motivated core of our simulations is the morphometric approach to solvation free energy [13,14]. It is an implicit solvent model [15,16] and a sophisticated implementation, describing non-polar contributions to the solvation free energy as a linear combination of four geometric measures: the excluded volume V, solvent-accessible surface area A, integrated mean curvature C, and integrated Gaussian curvature X of the solvent-accessible surface [17–19]. This method has proven highly accurate for calculating solvation properties of complex biomolecules [20–23] and assembles exotic geometries for generic solutes [24,25]. For simulation of macromolecules, we augment it with a soft-sphere overlap penalty L, yielding a total short-range geometric potential: The prefactors for the geometric terms are derived from fundamental measure theory [26], where p is pressure, σ is surface tension, and κ and κ― are bending rigidities [14]. The weight ϕ is empirically determined and, in combination with L, acts both as a simple model of repulsion and as a relaxation of the shape of the modeled molecules. The excluded volume V describes the volume inaccessible to solvent molecule centers. The surface area A captures the area available to solvent molecule centers as close to the solute as possible. The integrated mean curvature C measures the total bending of the solvent-accessible surface. The integrated Gaussian curvature X is a topological measure equal to 4π times the Euler characteristic of the union of balls, hence measuring connectivity, loops and voids of the configuration. We calculate the prefactors utilizing the White Bear functional mark II [26] from density functional theory, giving us values for a hard-sphere solvent depending on solvent radius rs and packing fraction η, which we set to 1.4Å and 0.3665 respectively to mimic the properties of water. As such, the effective attraction that drives the assembly of the solutes is entropic in origin. Previous work has shown that for tobacco mosaic virus (TMV) subunits, the assembly observed experimentally corresponds to the minimum of Fsol* obtained via Random Walk Metropolis simulations [27]. Random Walk Metropolis explores the configuration space by perturbing the current configuration to propose a new one. A proposal that lowers the energy is always accepted, whereas a proposal that raises the energy is accepted only with a probability that decreases as the energy increase grows, and is otherwise rejected. This makes it a powerful algorithm for exploring the configuration space and locating thermodynamically stable states. It does not, however, model dynamics, and no kinetic accessibility between successive states can be inferred from the sequence of configurations it produces. The morphometric approach is inherently short-ranged, as its geometric measures are defined by a local solvent probe (e.g., rs=1.4Å for water). Consequently, its energy landscape is largely flat until the subunits are in close contact, at which point it becomes rugged. A simulation driven only by Fsol* performs a largely random search until protein subunits interact at a close range. Here, the simulations find local minimizers that are sometimes difficult to escape from, preventing further exploration and correct assembly. This highlights a kinetic, rather than thermodynamic, hurdle that the model has to overcome in a simulation context. In physical systems long-range interactions such as electrostatics [16] and long-range hydration forces [28] pull the proteins together and correctly align them. The sum of these effects can be interpreted as a long-range shape matching effect. The biasing potential we introduce in the following does not have a direct physical interpretation but acts as a long-range shape matching potential and in this sense is a similar mechanism to long-range physical interactions. The calculation of morphometric measures already relies on weighted Alpha complexes [29–32], which naturally provide a multi-scale representation of shape known as a filtration (see Fig 1). This Alpha-filtration is constructed on the union of atom centers of the simulated molecules. Persistent homology is the tool for quantifying topological features within such a filtration. Alpha filtration and persistence. Weighted Alpha complexes for a system of four toy shapes in two dimensions at increasing values of α (A-E). Vertices are shown in black, edges in white and triangles in red; the dual intersection of the balls with the weighted Voronoi diagram is shown in blue. At α=0 the shapes carry their solvent-accessible radii (A); as α grows, the space between them is progressively filled (B-E). Panel F shows the resulting persistence barcode: each bar is a topological feature and runs from its birth b to its death d along the α-axis. Blue bars are zero-dimensional features (connected components) and green bars one-dimensional features (loops). The length of a bar, d−b, is its persistence, and the total persistence Pi is the sum of the lengths of all i-dimensional bars. We only show the features of finite persistence. We define a topological potential, 𝒯, as a weighted sum of total persistence measures: where Pi=∑(b,d)∈Dgmi(d−b) is the total persistence of all i-dimensional features (birth-death pairs) in the system’s persistence diagram Dgmi and the λi are scalar weights. Consider the persistence barcode in Fig 1F. As the scale α increases, the union of balls grows: initially separate components merge, and loops between the shapes form and later fill in. Each such feature is recorded as a bar from its birth b, the scale at which it appears, to its death d, the scale at which it disappears, and its persistence d−b measures the range of scales over which it survives. Concretely, the long green bar corresponds to the central loop that forms in panel B and persists until it is filled in panel E, while the short green bar corresponds to the short-lived loop that appears to the right of it in panel C and closes again by panel D. The total persistence Pi in Eq 2 is exactly the summed length of all finite persistence i-dimensional bars, so 𝒯 is a weighted readout of this barcode. The zero-dimensional persistence P0 is small when the distinct molecules are close together. P1 and P2 track the total persistence of loops and voids enclosed between the molecules. Note that no voids exist in the two-dimensional example in Fig 1. Total persistence of loops and voids can be large when few features persist for a long time or many features persist for a short time. The weights λi tune how strongly these terms are maximized or minimized during assembly. In the results section we determine promising candidates for the λi based on a grid search in which we explore for which parameters the topological potential alone obtains minima that are energetically close to true assemblies of TMV protein subunits. The initial boundaries of the grid are chosen based on observations about the total persistence measures as well as theoretical considerations. The grid is then extended to ensure that the most promising region of the scan lies away from the boundaries. The potential captures some shape information over arbitrary ranges, suggesting a synergistic strategy in which the two potentials play complementary roles. The long-range topological term, 𝒯, can create a guiding funnel within the energy landscape, pulling subunits into the correct general vicinity and orientation. The short-range geometric potential, Fsol*, then acts as a refiner, selecting the precise low-energy binding interface. In this work, we test this principle by simulating dimer and trimer assembly of various systems governed by a combined potential: We show that this approach, which combines local physical accuracy with global topological guidance, enhances the purely geometric model, increasing the simulation success rate for the TMV and SARS-CoV-2 ORF9b proteins by more than an order of magnitude and further making the successful assembly of a hepatitis B core protein and an extracellular fragment of the human CD40 ligand possible on the simulated timescales. This result validates the use of topology not just as a post-hoc analytical tool, but as an active biasing potential to solve kinetic challenges in molecular simulation. Results We begin by discussing how Fsol* and 𝒯 are correlated and analyze how persistence measures develop over the course of simulated assemblies driven only by Fsol*. We do this to motivate the use of topological measures in our simulations using the inherent relationships between geometric and topological measures. We then investigate the minimizers of 𝒯 for different combinations of (λ0,λ1,λ2). Finally, we compare different mixtures of Fsol* and 𝒯 and how the inclusion of the topological potential improves the simulation results. Geometry-topology correlations We compare the topological and geometric potential, calculating 𝒯 for a sequence of configurations of two TMV capsid protein subunits, obtained via a successful simulation run driven by the geometric potential alone. Here, successful means that the two subunits are assembled in a configuration similar to one observed experimentally, as the simulation reaches the energy minimum. Let Sca be an experimentally determined configuration of two correctly assembled protein subunits. Let S be a configuration obtained by simulation. We say that S and Sca are similar if dRMSD(S,Sca)<2.5Å. See Materials and Methods for an explanation of the distance measure dRMSD. A visualization of the correct assembly of two protein subunits is shown in Fig 3A. The simulation is driven by a Random Walk Metropolis algorithm; see Materials and Methods. Topological potential of correct assemblies versus 𝒯 minimizers. Agreement between correctly assembled reference states and the minimizers of the topological potential 𝒯 across the weight parameter space. Render of the two correctly assembled TMV subunits used as the ground truth in the two-subunit case (A). Renders of the three configurations of three TMV subunits that we consider correctly assembled (B-D). Heatmaps over the weight plane (λ1,λ2), with λ0=1.0 fixed (E,F). The color encodes dz=log(𝒯z(S)−𝒯z(Sz)), the logarithm of the difference in topological potential between a correct assembly S and the minimizer Sz found at weights z=(λ0,λ1,λ2). Darker colors indicate a smaller difference, i.e., minimizers whose topological potential lies closer to that of the correct assembly. Groups of topological potential minimizers corresponding to the marked regions of the heatmaps (G-N); the minimizers closest to the correct assembly are J and L. Calculating total persistence for a successful simulation run Fig 2A shows how Fsol*, Eq 1, changes over the course of approximately 1500 steps of a simulated assembly pathway. Here, one step corresponds to an accepted proposal of the Random Walk Metropolis algorithm, i.e., a change in the configuration of the simulated molecules. The signal is flat for large stretches of the simulation because the geometric potential is short-ranged and only exhibits changes in value when the solvent-accessible surfaces of the simulated proteins overlap. There are several small valleys in the energy landscape, which correspond to configurations where the subunits are close. They detach again several times before finding a sequence of moves leading to a steep decline in energy and a local minimum, corresponding to configurations that are close to experimentally observed assembly. Fig 2B shows the goodness-of-fit measure dRMSD that tracks how close to correct assembly the current configuration is. We can see that dRMSD is more stable in the regions where the subunits overlap and oscillates in the regions where Fsol* is flat. Around step i = 1000 there is a drop in dRMSD corresponding to the protein subunits being correctly aligned but not yet in contact. As Fsol* reaches its minimum, so does dRMSD, which fluctuates around a low value indicating proximity to the correct assembly. Comparing potentials. Comparing potentials of a sequence of states obtained via a simulation driven by the geometric potential Fsol*. The geometric potential itself (A). A goodness of fit measure with respect to the correct assembly of two TMV proteins (B). The weighted sum 𝒯=0.5P0−0.35P1−0.35P2 (C). The total persistence in dimension zero (D), dimension one (E) and dimension two (F). Fig 2D-2F shows the weighted total persistence measures for dimensions zero, one and two. The chosen weights are λ0=1.0,λ1=−0.7,λ2=−0.7. These weights correspond to the best performance increase achieved by mixing the geometric and topological potentials. The sum of all three, i.e., 𝒯(1.0,−0.7,−0.7), is shown in Fig 2C. The total persistence of the zeroth homology group is plotted in blue and is minimal as soon as the proteins are close together because then the zero-dimensional hole between the two rigid bodies is closed. It is interesting to note that P0 is flat where Fsol* is not and vice versa. In addition, it seems that P2 and Fsol* are correlated and that dRMSD and 𝒯(1.0,−0.7,−0.7) share similar behavior. To quantify these observations, we calculate the Pearson coefficient [33] r of the different functions. We list the Pearson coefficients of different measures of sequences of configurations of two TMV capsid proteins, in Table 1. In the table, 𝒯 refers to 𝒯(1.0,−0.7,−0.7). The Pearson coefficient measures the strength and direction of the linear relationship between two quantities, ranging from −1 (perfectly anti-correlated) through 0 (no linear relationship) to +1 (perfectly correlated). Here it is informative since it indicates that the topological potential is correlated to dRMSD, pointing toward an exploitable signal to drive simulations toward the assembled state. It furthermore indicates which signs for the weights for total persistence per dimension are likely appropriate. Table 1Linear correlations between different measures. Fsol* d RMSD 𝒯 P 0 P 1 P 2 Fsol* 1.00.710.860.38 −0.8 −0.91 d RMSD 0.711.00.760.44 −0.7 −0.75 𝒯 0.860.761.00.42 −0.98 −0.95 P 0 0.380.440.421.0 −0.26 −0.46 P 1 −0.8 −0.7 −0.98 −0.26 1.00.87 P 2 −0.91 −0.75 −0.95 −0.46 0.871.0 Linear correlations between different measures. Linear correlations that stand out in particular are the high correlations of 𝒯(1.0,−0.7,−0.7) to both dRMSD and Fsol*, which are higher than the correlation between dRMSD and Fsol* themselves. The correlation of dRMSD to 𝒯(1.0,−0.7,−0.7) is higher (in absolute terms) than to each individual persistence measure. Furthermore, there is a high negative correlation between Fsol* and P2 of −0.91. The measures highly correlated in Table 1 match the plots in Fig 2 that exhibit similar trends. A final key observation is that 𝒯(1.0,−0.7,−0.7) and the corresponding weighted persistence measures are at or close to their respective minima as the simulation reaches its assembled state. This in combination with the relatively high linear correlation between dRMSD and 𝒯(1.0,−0.7,−0.7) motivates further analysis of the topological potential. Optimizing persistence We investigate the self-assembly of two and three subunits of the TMV protein under minimization of topological potential 𝒯. Inferring promising weights for 𝒯 from the correlation of Fsol and dRMSD with the total persistence measures given in Table 1, we analyze the parameter space given by λ0=1.0, λ1∈[−2.0,1.0] and λ2∈[−2.0,1.0]. We run both Random Walk Metropolis and simulated annealing simulations in a grid with step size 0.1 in the parameter space. Let 𝒮 be the set of all minimal configurations obtained during one of the simulations. For each point in the parameter space, we then check which configuration in 𝒮 has the lowest topological potential. We then compare the topological potential of each minimizer to that of a correctly assembled state. We calculate Here, z=(λ0,λ1,λ2) refers to a point in the parameter space and Sz to the topological potential minimizer at z. 𝒮exp is the set of assemblies observed experimentally. In the case of two subunits, 𝒮exp contains the single configuration shown in Fig 3A. In the case of three subunits, 𝒮exp contains the configurations shown in Fig 3B-3D. The resulting differences in topological potential dz for two and three subunits in the parameter space are plotted in Fig 3E and 3F. Dark colors correspond to the region in which the differences between ground truth and found minimizer are small. The bounded regions in the plots correspond to different groups of topological potential minimizers that we grouped by distance measurement dRMSD and then manually aggregated further. Phases for two subunits In the case of two subunits, we can distinguish four large regions. One in which the subunits overlap almost perfectly is a large region to the right of the phase diagram, corresponding to configurations as shown in subfigure H. The second group consists of configurations in which the protein subunits do not touch. They appear in the upper left and lower left corners of the phase diagram and in the center bottom. We note that on the top left the subunits appear to form a circle between them, while on the bottom they appear to form a ball. This is not surprising considering that the one-dimensional persistence features contribute relatively more to the top-left, while the two-dimensional features, to the bottom. The final region lies roughly in the center of the phase diagram and contains configurations that touch and overlap slightly. None of the minimizing configurations is correctly assembled, although individual simulations did converge to a state close to correct assembly, which implies that it is a local minimizer. Phases for three subunits The same images for three subunits are shown in Fig 3K-3N. Here, the region of non-touching subunits is much larger and roughly falls into three categories: configurations in which a circle forms between the subunits as in subfigure K; a ball, as in subfigure M; and finally a spiral configuration, as in subfigure N. On the right side of the phase diagram there is the region with overlapping subunits, which is a bit smaller compared to the phase diagram of two subunits. This is caused by the possibility of more persistence features in dimensions one and two, counteracting the pull of minimizing zero-dimensional persistence. We have again marked this region with H, but note that here, three subunits overlap instead of two. We again have a region in which the subunits touch but do not overlap completely, although it has almost disappeared. Interestingly, for this configuration two of the subunits are assembled in a way that corresponds to the correct assembly of two subunits, while the third one lies on top, with some overlap. Considering the difference dz between the correct assembly and the topological potential minimizers, we see that for two subunits the region J, where the subunits touch but do not overlap corresponds to a dark area, that is, a small difference between the ground truth and the topological potential minimizer. For three subunits, this region of small dz values lies exactly at the boundary between the regions of minimizing configurations that do and do not overlap. The single point in the parameter space at λ1=−0.4 and λ2=−0.9 where the minimizer of three subunits touches but does not overlap is also the point in the parameter space with the lowest distance between ground truth and the topological potential minimizer. Combining geometry and topology We will now compare Random Walk Metropolis simulations in which a combination of geometric and topological potential is minimized. We recall the combined potential, Eq 3: where μ∈[0,1] is the interpolation between geometric and topological potential. The measure in which we are interested is the success rate. Recall that a configuration S is correctly assembled if dRMSD(S,Sca)<2.5Å, for Sca denoting the experimentally observed assembly. Let 𝒮 be a set of minimal configurations of a set of simulations, and let be the set of successful simulations. We define the ratio ∣𝒮*∣∣𝒮∣ as the success rate of 𝒮. All simulations are run with the perturbation parameter σt=1.25, which is the standard deviation for the normal distribution from which the translation perturbations are drawn; see Materials and Methods. Fig 4A, shows the success rates of simulations driven by Fsol* alone, depending on temperature T and rotational perturbation parameter σR. Taking the value with the highest aggregated success rate in the temperature range, we set σR=0.2 for all simulations driven by different mixtures Ecomb. Comparison of simulation setups. Success rates for the pure morphometric approach, i.e., (λ0,λ1,λ2)=(0.0,0.0,0.0) depending on rotational perturbation parameter σR and simulation temperature T (A). A colorbar for the success rate for a mixture Ecomb=μFsol*+(1−μ)𝒯 (B). Heatmaps of the success rate depending on simulation temperature T and weighting μ of topological potential 𝒯, for parameters (λ0,λ1,λ2)=(1.0,−1.0,1.0) (C), (λ0,λ1,λ2)=(1.0,−0.7,−0.7) (D), (λ0,λ1,λ2)=(1.0,−0.4,−0.9) (E) and (λ0,λ1,λ2)=(1.0,0.0,0.0) (F). Weight-selection for the topological potential We select different combinations for triples of (λ0,λ1,λ2) defining 𝒯. We pick (1.0,−0.4,−0.9), which corresponds to the point in the parameter space where the minimizer of 𝒯 for three subunits has touching but non-overlapping protein subunits. We further select (1.0,−0.7,−0.7), which lies in the region of low distance between the correct assembly and the topological potential minimizer for two subunits. We chose (1.0, 0.0, 0.0) to see how much of the impact can be attributed to the topological potential simply pulling the proteins together. And finally we pick (1.0,−1.0,1.0) because it relates the topological potential to the integrated Gaussian curvature as: Here, α0 and αn are the lowest and highest values for which topological changes occur in the filtration of weighted Alpha complexes and ℬ(α) is the corresponding union of balls. See Materials and Methods for details. Scanning the parameter space for TMV dimerization Fig 4C-4F shows heat maps for the choices of (λ0,λ1,λ2) mentioned above, where we alter the simulation temperature T = 3,4,5,6 and the factor μ=0.1,0.2,...,0.9 that interpolates between Fsol* and 𝒯 to calculate the total energy as Ecomb=μFsol*+(1−μ)𝒯. For each point in the heatmap, we run 40 simulations, where each simulation is run for 50000 iterations. This corresponded to roughly 12 hour long simulations on a single core of a high performance cluster. We report the success rate uncertainty as a Wilson score 95% confidence interval [34], which expresses that given the sample size and success rate, we are 95% confident that its true value lies within this interval. We write the interval in square brackets following the success rate. For the pure morphometric approach, the best success rate we achieve is 5% [1.4, 16.5]. The highest success rate among mixtures of both the morphometric approach and the topological potential is 85% [70.9, 92.9] for 𝒯(1.0,−0.7,−0.7) (Fig 4D) at a simulation temperature of 5.0 and a weighting μ=0.5, as well as for 𝒯(1.0,−0.4,−0.9) (Fig 4E) at T = 5.0 and μ=0.3. This corresponds to a 16-fold increase in the success rate. The highest success rate for 𝒯(1.0,−1.0,1.0) (Fig 4C) is equal to 22.5% [12.3, 37.5]. It is found for T = 3.0 and μ=0.8. Although substantially lower than for the other variants, it is still a 4-fold increase over the pure morphometric approach. The combination of Fsol* with 𝒯(1.0,0.0,0.0) (Fig 4F) achieves its best success rate of 12.5% [5.5, 26.1] at T = 3.0 and μ=0.8. From this we infer that the improvements in success rate are not obtained only because minimizing total zero-dimensional persistence is pulling the subunits together. Analysis of a successful simulation of the combined potential Fig 5 illustrates a sample simulated assembly pathway of two subunits with a mixture at μ=0.5 and T = 2.0 for 𝒯=P0−0.7P1−0.7P2. The simulation reaches its minimum energy, a correctly assembled configuration, after 276 steps. The figures in subfigures A-C show Fsol*, 𝒯 and their mixture relative to the initial (dispersed) state. Subfigures D-F show the individual persistence measures. Subfigures J-L show the persistence barplots for three selected configurations: the initial state (i = 0), a state in which the subunits are correctly aligned but have a gap between them (i = 217) and finally the minimal energy state (i = 276), in which the subunits are correctly assembled. Renders of the corresponding configurations are shown in G-I. Analysis of final steps. Analysis of the final steps of a simulation driven by Ecomb=510Fsol*+510𝒯(1.0,−0.7,−0.7). All energy plots (A-F) show values relative to step 0. We see Fsol* as given by Eq 1 (A), 𝒯(1.0,−0.7,−0.7) (B), Ecomb (C), and the individual persistence measures (D-F). We show persistence bar plots of steps 0, 217 and 276 with zero-dimensional features in blue, one-dimensional features in green and two-dimensional features in red (J-L). The corresponding protein configurations are shown in (G-I). Note that Fsol* fluctuates around its initial value until step 217, while the topological potential steadily declines. In the beginning most of the decline is driven by the disappearing zero-dimensional hole, but after about 20 iterations the total persistence for dimension zero remains roughly constant, while the total persistence values for P1 and P2 decline. Interestingly, P1 is almost as low as in the beginning as in the end, caused by a single very persistent feature. As the proteins get pulled together its value increases and then steadily declines. The same can be observed for P2, although much less pronounced. This matches the observation from the phase diagram in Fig 3, where configurations similar to the initial configuration are minimizers in parts of the phase diagram adjacent to the point in the parameter space we chose. After the initial period, where the proteins are pulled together, the decrease in topological potential is mainly driven by P1 and supported by P2. We refer to the sequence of steps before the simulation attains its minimum, but after the solvent-accessible surfaces begin to overlap as the final phase. In the final phase, both the topological and geometric potentials rapidly decline, while the final gap between the protein subunits is closed. Considering just the topological potential, we observe that the final phase accounts for one third of the total decline, whereas the other two thirds occur during topologically driven alignment of the proteins. Looking at the persistence diagrams in subfigures J-L, we observe that more persistence features appear as the subunits get closer to the correct assembly and that the persistence features are born at lower α values and persist for a shorter time period. Analyzing the impact of the topological potential in different simulation phases To quantify the impact of the topological potential at different stages of the assembly process, we compare the performance of Fsol* to Ecomb=12Fsol*+12𝒯(1.0,−0.7,−0.7) in a set of simulations that are initialized as the first state in the final phase of successful simulation runs. We take 20 such states for both potentials, which are closest in dRMSD to the correct assembly. We then start ten simulations from each of the 40 states for both potentials. Each simulation proposes 10000 steps. The following Table 2 shows the success rates for different combinations of initial states (Fsol* touch and Ecomb touch) and the energies driving the assembly (Fsol* finish and Ecomb finish). We compare the values of the combined potential at two different temperature values to verify the impact of potential detachment and realignment of the subunits. Table 2Comparing success rates. Comparing success rates with confidence intervals of Fsol* and Ecomb on first states of final phases of successful simulations for both potentials.Fsol* touchEcomb touchFsol* finish, T = 3.01% [0.3,3.6]6% [3.5,10.2]Ecomb finish, T = 3.025% [19.5,31.4]63% [56.1,69.4]Ecomb finish, T = 5.036% [29.7,42.8]54% [47.1,60.1] Comparing success rates. Comparing success rates with confidence intervals of Fsol* and Ecomb on first states of final phases of successful simulations for both potentials. The key takeaway from the table is that the topological potential is useful in both stages of the assembly process. That is, during long-distance alignment and short-range docking. Fsol* simulations initialized on states taken from successful Fsol* simulations find the correct assembly again in 1% of the cases. Fsol* simulations initialized on states taken from successful Ecomb simulations correctly assemble in 6% of the cases. This implies that alignment by Ecomb is better than alignment by Fsol*. We simulate Ecomb for both T = 3.0 and T = 5.0 and see that in both cases simulations are more likely to succeed on the set of states that were aligned by Ecomb. Interestingly, the lower temperature performs better in the Ecomb aligned states. We assume that this is caused by the selected states being comparatively good and that a higher temperature yields a higher chance to move away from them than a lower temperature. What is a drawback for well-aligned states is a feature for the low-quality alignments that appear during pure Fsol* simulations. Here, the higher temperature performs better, because moving away from the initialized alignment and realignment driven by Ecomb is favorable. Overall, we conclude that the topological potential is a useful long-range augmentation of the morphometric approach because it draws the proteins together in a way that assists assembly and because it further supports assembly at close range. Improvements for other systems To show that this approach can be generalized, we apply it to simulate dimerization of the human hepatitis B virus core protein [35], dimerization of an accessory protein of the SARS-CoV-2 virus that hinders the immune response [36], and trimerization of an extracellular fragment of the human CD40 ligand, relevant to T cell function [37]. All of these structures are assemblies of homodimers and homotrimers, respectively, which means that they are composed of two or three identical protein chains. Fig 6 shows the renderings of the final assembled states found via simulations driven by Simulated protein assemblies. Different simulated protein assemblies. Hepatitis B virus core protein (A), SARS-CoV-2 ORF9b (B) and fragment of the Human CD40 Ligand (C). Subunit geometry is given by the A-chains as specified in protein data base entries 4BMG, 6Z4U and 1ALY respectively. The simulation results are at a distance to the experimental assemblies of dRMSD≈1Å. The rendered assemblies are therefore visually indistinguishable from the experimentally determined structures, which is why we do not show the latter separately. For systems of two subunits we run simulations for the pure morphometric approach at T∈{2.0,3.0,4.0}; for combinations with 𝒯(1.0,−1.0,1.0) at μ=0.8 and T∈{2.0,3.0,4.0}; and for combinations with 𝒯(1.0,−0.4,−0.9) at μ=0.3 and T∈{4.0,5.0,6.0}. For the system with three subunits, that is, the human CD40 ligand, we run simulations at T = 5.0. For each combination of parameters we run 40 simulations with 50000 iterations for two subunits and 90000 iterations for three subunits. Table 3 shows the highest success rate found for any temperature and the potentials described above. We include the previously discussed values of TMV dimerization for comparison. Table 3Success rates across diverse protein systems. The success rate (sr) of simulations driven by the combined geometric-topological potential (Ecomb) is compared against the purely geometric potential (Fsol*) across four diverse protein systems. Wilson score 95% confidence intervals are given in brackets.ProteinAtoms persr [Wilson CI]sr [Wilson CI]sr [Wilson CI]Systemsubunit Fsol* 810Fsol*+210𝒯(1.0,−1.0,1.0) 310Fsol*+710𝒯(1.0,−0.4,−0.9) TMV12065.0% [1.4, 16.5]22.5% [12.3, 37.5]85.0% [70.9, 92.9]Hep. B11340.0% [0.0, 8.8]17.5% [8.7, 32.0]80.0% [65.2, 89.5]ORF9b6392.5% [0.4, 12.9]7.5% [2.6, 19.9]42.5% [28.5, 57.8]CD4011130.0% [0.0, 8.8]0.0% [0.0, 8.8]5.0% [1.4, 16.5] Success rates across diverse protein systems. The success rate (sr) of simulations driven by the combined geometric-topological potential (Ecomb) is compared against the purely geometric potential (Fsol*) across four diverse protein systems. Wilson score 95% confidence intervals are given in brackets. The key takeaways from the table are that the combined potentials perform better than the morphometric approach in all cases. In particular, for dimerization of the hepatitis B core and trimer assembly of the human CD40 ligand fragment, we observe an increase in the success rate from 0% for the pure morphometric approach to 80.0% and 5.0% respectively, for the potential given by 310Fsol*+710𝒯(1.0,−0.4,−0.9). The success rate of dimerization of the SARS-CoV-2-ORF9b protein increases from 2.5% to 42.5% for the same potential. Judging by the values for the pure morphometric approach, the tobacco mosaic virus subunits appear easier to assemble, probably due to the high geometric complementarity compared to the hepatitis B core protein, while simultaneously not being as intricately interlocked as the SARS-CoV-2-ORF9b protein. An animation of the simulated assembly of a CD40 trimer is included in the supporting information (SI). For all systems except CD40, the confidence interval of the best combined potential lies entirely and substantially above that of the pure morphometric approach, indicating that the improvement is not an artifact of the finite number of runs. It is notable that the selection of parameters that yielded the best results for the tobacco mosaic virus also yields strong improvements for other systems. The fact that it performs better for the hepatitis B core than the SARS-CoV2-ORF9b protein is likely caused by the existence of other states that are energetically close to correct assembly. Indeed, simulations of the SARS-CoV-2-ORF9b protein with energy Ecomb=310Fsol*+710𝒯(1.0,−0.4,−0.9) are the only ones in which the lowest energy state found is not the native assembly, but a state close by, with dRMSD=4.5Å. However, the minimum of the pure morphometric approach corresponds to the correct assembly and therefore larger values for μ ensure that correctly assembled configurations are obtained as the minimizer. In all other cases in which we observed at least one simulation that obtained a minimum corresponding to the biologically correct assembly, this structure is also minimizing Ecomb. As such, the combined potential is thermodynamically sound in the sense that, for sufficiently large μ, it correctly distinguishes the ground truth as the energy minimum. Discussion Our study demonstrates that a topological potential, defined as the weighted sum of total persistence measures of proteins in a solvent, substantially improves the success rate of simulated protein assemblies. Acting as a long-range complement to the short-range morphometric approach to solvation free energy, it enables assembly in systems that fail under the morphometric potential alone and increases success rates by more than an order of magnitude for other systems, reaching 85% for TMV dimers with optimized parameters. We analyze the relationship between the topological potential 𝒯 and the morphometric solvation term with a soft-sphere constraint Fsol* in depth for the assembly of TMV dimers. Parameters optimized for TMV translate well to other systems, particularly those with protein subunits of comparable atomic size, yielding high success rates without system-specific tuning. While the topological potential does not yet have a physical interpretation, one specific parameterization, using the alternating sum of total persistence measures, can be mathematically linked to the Gaussian curvature term in the morphometric approach. This may be understood as capturing topological changes in an expanding molecular surface, potentially connected to solvation shell organization or other long-range, shape-dependent effects. In practice, the computational cost of calculating the geometric and topological potentials scales linearly in the number of protein subunits, or equivalently the number of atoms. The bottleneck for assembling larger systems is therefore not the cost of a single potential evaluation, but the number of iterations required to reach an assembled configuration. As the configuration space grows exponentially with the number of subunits, this number of iterations increases rapidly with system size; we observed such a rapid increase already for the purely short-ranged morphometric approach [27]. The improved efficiency observed when integrating the topological potential as a simulation strategy facilitates the assembly of larger systems. Coarse-graining [38] by reducing the number of modeled atoms per subunit is a feasible approach to increase the accessible system size further, as long as sufficient shape-specificity is retained that allows for the alignment of geometric locks and keys. In general, improvements in the running time of the topological potential, through coarse-graining, optimization or parallelization, would make it feasible to simulate flexible molecules and larger assemblies of tens to hundreds of subunits. While all systems studied here are homoassemblies of at most three subunits, the method does not depend conceptually on either restriction, and preliminary work including ligands and proteins shows it extends well. The calculation of the topological potential 𝒯 only depends on the geometry of the atoms and the choice for the parameters λi. Its integration into a simulation-governing energy function offers a general computational strategy to overcome rugged energy landscapes in shape assembly simulations. Its performance is demonstrated here across diverse protein systems. The method recovers the thermodynamically accurate states observed in nature with high probability given enough iterations to scan the configuration space. Hence it can be utilized to find low energy states of arbitrary, in particular unknown, molecular assemblies. It retains physical realism in the form of thermodynamic accuracy, but does not model dynamics like molecular dynamics (MD) simulations and therefore does not allow for the inference of kinetic accessibility. However, similarly to MD methods the approach does not depend on prior knowledge of the target states similar to the modeled system. This is in contrast to deep-learning methods like AlphaFold3, whose accuracy on docking tasks has recently been shown to depend strongly on the similarity of the target to the training set [39]. While the mechanism by which the topological potential guides assembly is not yet fully understood, the potential itself is transparent: it is a weighted sum of the total persistence of a weighted alpha complex, a visualizable geometric construction governed by atom geometry and three parameters λi. In this sense it is not a black box in the way a deep neural network is, whose behavior is determined by billions of parameters and is correspondingly difficult to interpret. The topological potential by itself can be viewed as a long-range shape-alignment and matching potential, making it useful for simulations in computational biology and materials science. Materials and methods Parameters for the morphometric approach For a hard-sphere solvent the physical prefactors for the four terms in the morphometric solvation free energy can be computed using the White Bear functional Mark II [26]. Each term is given as a function depending on the packing fraction η and the solvent radius rs: Here β is the inverse thermodynamic temperature, which we set to 1.0J-1, since we are only interested in relative energy levels. Mimicking the properties of water we set rs=1.4Å and η=0.3665 (the approximate fluid density relevant to biological processes). This yields the values of p≈0.182JÅ3, σ≈−0.131JÅ2, κ≈0.112JÅ and κ―≈−0.03J. In Equation 1 we augment the morphometric approach by a linear overlap penalty L to energetically penalize overlapping protein subunits as well as to relax the rigidity of the shapes. The weighting ϕ for L is determined empirically [27]. Persistent homology of Alpha filtrations The geometric and topological measures are derived from the weighted Alpha complex of the atomic centers. The Alpha complex is a subcomplex of the weighted Delaunay triangulation that provides a mathematically precise and computationally efficient representation of the union-of-balls model of a molecule [29,40]. For a set of atomic centers P={p1,...,pn}⊂ℝd and a corresponding set of weights W={w1,...,wn}⊂ℝ, the construction begins with the weighted Voronoi diagram. The i-th weighted Voronoi cell is defined as where πp,w(x)=||x−p||2−w is the power distance. The collection of weighted Voronoi cells is the weighted Voronoi diagram denoted by 𝒱(P,W). Under the assumption that no d + 2 Voronoi cells intersect in a common point, i.e., the points in P are in general position, the dual to the weighted Voronoi diagram is the weighted Delaunay triangulation. Now let W={w1,...,wn}⊂ℝ be a set of weights assigned to the points in P, where each weight wi is the squared atomic radius (wi=ri2). Let B(pi,wi) be the ball centered at point pi with radius ri=wi. We denote The weighted Alpha complex is defined as i.e., the dual of the weighted Voronoi cells intersected with the balls B(pi,wi). Inclusion-Exclusion formulas to compute the geometric measures V, A, C and X, are based on this incidence information [40–44]. These formulas are implemented in the AlphaMol program [32]. We also use the weighted Alpha complex to compute the overlap penalty L, by iterating over the edges whose vertices correspond to different proteins and summing up the overlap of the corresponding balls. Let α∈ℝ. We define B(pi,wi,α) as the ball centered at pi with radius ri=wi+α, where B(pi,wi,α)=∅ if α<−wi. We get that for points x with equal power distance to balls i and j: which is equivalent to ||x−pi||2−wi=||x−pj||2−wj as the α cancels out. Therefore, the weighted Voronoi diagram is the same for all values of α. We write and get a nested sequence of complexes for critical values of alpha α0≤...≤αm. The critical values are the lowest values for which a simplex in the weighted Delaunay Triangulation appears in the corresponding weighted Alpha complex. Given a filtration F=𝒜0⊂...⊂𝒜m and homomorphisms, induced by the inclusion maps, fpi,j:Hp(𝒜i)→Hp(𝒜j), the images are the p-dimensional persistent homology groups. Their ranks are the p-dimensional persistent Betti numbers. They count the p-dimensional cycles that are not boundaries. Given a class γ∈Hp(𝒜i) we say γ is born at 𝒜i if γ∉Hpi−1,i. We say it dies at 𝒜j if fpi,j−1(γ)∉Hpi−1,j−1 and fpi,j(γ)∈Hpi−1,j. We call the interval (αi,αj) the birth-death interval of γ and αj−αi its persistence. We refer to the collection of all p-dimensional birth-death intervals as the p-dimensional persistence diagram and denote it by Dgmp [45]. We use the Oineus software package [46] to calculate persistent homology. Simulation methods and parameters We simulate configurations of protein subunits using variants of a Random Walk Metropolis [47]. Each configuration is represented by a state x∈(SO3×ℝ3)m, where m is the number of simulated molecules. The geometries of the systems we study are taken from the protein data base entries with identifiers 6R7M [48], 4BMG [35], 6Z4U [36] and 1ALY [37], for TMV, hepatitis B, SARS-CoV2 and the CD40 ligand respectively. For atomic radii we use the ProtOr radii [49], which account for hydrogen atoms by giving radii for atomic groups instead of individual atoms. Each simulation is run in a bounding box of size 120Å. These bounds are not checked for each individual atom but for each center of mass of the simulated molecules. A perturbation for a molecule consists of a translation and a rotation. The translational perturbation vector is generated by drawing three independent samples from a normal distribution 𝒩(0,σt2) with σt=1.25Å, and this vector is added to the molecule’s current position. The rotational perturbation is generated by first creating a three dimensional random vector v, whose components are also drawn independently from a normal distribution, 𝒩(0,σR2), with σR=0.2. This vector v represents an element of the Lie algebra so(3). It is then mapped to a rotation in SO(3) via the exponential map. This resulting small, random rotation is then composed with the molecule’s existing orientation to yield the new candidate orientation. Assuming the current state of the simulation is state xi, we pick one of the simulated molecules at random and perturb it to generate a candidate state x′. This is also known as component-wise or block-wise Random Walk Metropolis [50]. The probability of accepting x′ as the new state of the simulation is determined as: where ΔE=E(x′)−E(xi) is the difference in energy of the states. If x′ is accepted, we set xi+1=x′. Otherwise we set xi+1=xi. T is a simulation temperature, which regulates the probability of accepting states with higher energy. The energy used here is Ecomb as in Eq 3 for different parameters μ and λi. The prefactors of Fsol* are given by the White Bear prefactors [26] computed for packing fraction η=0.3665 and solvent radius rs=1.4Å, mimicking the properties of water. For the simulated annealing (SA) [51] simulations we change T over the course of a simulation by letting it decay to zero according to a quadratic annealing schedule: where I is the total number of iterations at which the temperature reaches zero and k is the current step in the decay period. Once T = 0.0 is reached, we reheat the system several times to scan for nearby local minima. In cases in which we did not know which initial temperature to start a simulation with, we run 12 shorter RWM simulations and check whether the total acceptance rate is above or below some target acceptance rate. If it is below the target acceptance rate, we multiply the initial temperature for the next simulation by a factor of 1.5, if it is above the target acceptance rate we multiply the initial temperature for the next simulation by a factor of 0.5. After the 12 simulations are finished we pick the initial temperature closest to our target acceptance rate. For RWM simulations we choose a target acceptance rate of 0.2 and for SA simulations we choose a target acceptance rate of 0.8. The morphometric approach is short ranged and additive. As such it is not necessary to recompute its energetic contribution of the whole system in every step, but only the connected components of molecules that have changed. We determine the connected components in each step by constructing a graph that has each molecule as a node and an edge whenever the bounding spheres of two molecules overlap. Let idx be the index of the molecule we moved for the proposal of simulation step i + 1. Let 𝒞i be the set of connected components from step i and let 𝒞i+1 be the set of connected components from step i + 1. For each component c∈𝒞i+1 there are three cases to consider. Case 1: The connected component c is an element of 𝒞i and idx∉c. Then the energy of c remains unchanged and we can just copy the energy value from the last iteration. Case 2: The connected component c is an element of 𝒞i and idx∈c. Then the energy of c changes and we have to recompute it. Case 3: The connected component c is not an element of 𝒞i. Then the energy of c is changed and we need to recompute. By passing the energy value of a single molecule to the algorithm we can further optimize this procedure and not recalculate the energetic contributions of components that consist of only one molecule. To ensure a fair comparison between different topological potential setups, we defined a relative contribution measurement, c𝒯. This measurement quantifies the energetic impact of the topological term relative to the geometric term during the last stage of the assembly process. We take the set of 800 simulations, the data shown in Fig 4A are based on. For every simulation i we take two states Smini, the minimal energy state, and Sdispi, the state immediately before the solvent-accessible surface areas of the protein subunits start to interact, before Smini is reached. We then calculate an average of how much of the change in energy in the final phase in these simulations is contributed by 𝒯 in relation to Fsol* as: We find that c𝒯(1.0,−0.7,−0.7)≈1.12, c𝒯(1.0,−0.4,−0.9)≈0.96 and c𝒯(1.0,−1.0,1.0)≈0.61, i.e., the values are reasonably similar for the different setups, ensuring comparability. Note that calculating c𝒯(1.0,0.0,0.0) is not very informative because zero dimensional persistence has little impact in the final stage of a simulation. Relation between the topological potential and the integrated Gaussian curvature Consider a set of atom centers P={p1,...,pn} and van der Waals radii R={r1,...,rn}. The set of weights for the solvent radius rs are the weights of the weighted Alpha complex used to compute the morphometric measures. Now let ℬ(α)=⋃j=1mBj(α) where Bj(α) is the ball centered at pj with radius wj+α. If wj+α<0, we define Bj(α)=∅. Because the Alpha complex 𝒜α is homotopy equivalent to the union of balls ℬ(α), their Euler characteristics are equal [40]. The sequence of values, α0<...<αk for which topological changes occur, when moving from 𝒜(αi) to 𝒜(αi+1), are the filtration values of the Alpha filtration on which we calculate persistent homology. Here The Euler characteristic χ of a union of balls ℬ(α) is related to its integrated Gaussian curvature X (integrated over the boundary) by a factor of 4π. That is: Then by the Euler-Poincaré formula [45] we get: That is, the alternating sum of total persistence per dimension is equal to the integral of α0 to αn over the Euler characteristic χ(ℬ(α)). We subtract 1 in the integral because the connected component representing the whole molecule persists forever, and we do not count it in the alternating sum of the total persistence values. Therefore: This choice of parameters can thus be seen as a range extension of one of the terms of the morphometric approach in a mathematically rigorous sense. We note that for α≤0, we expect the integral to be roughly constant across different configurations, since the atom centers within a protein subunit are usually closer together than those between subunits. For λ0*=λ0−1, λ1*=λ1+1 and λ2*=λ2−1 we have that: meaning the topological potential 𝒯, can always be decomposed into: Hence, while the morphometric approach has the integrated Gaussian curvature X(ℬ(0)) as a component, the topological potential contains the integral over the Euler characteristic of ℬ(α) over the range of α values in which topological changes occur in the Alpha filtration. Ground truths from multiple A-chains The data provided in the protein data base files 4BMG, 6Z4U and 1ALY contain geometric information for the protein assemblies discussed in Fig 6 and Table 3. The dimers or trimers are assemblies of several proteins of the same type, which have subtle differences in the geometry of each chain due to the non-rigid nature of the proteins or measurement imprecision. We take the A-chain of all entries as the geometric template in the simulations we run. The ground truth to which we compare simulation results is obtained by superimposing the A-chain onto the positions and orientations of the other chains of the dimer or trimer. Simulation evaluation Let Q,U,V,W⊂ℝ3k be rigid transformations of some underlying template X0⊂ℝ3k representing a molecule with k atoms. Now assume that Q and U are a pair of molecules in some configuration 𝒳 and that V and W are a pair of molecules in some configuration 𝒴. Let (RQ→V,tQ→V) be the optimal rigid-body transformation that superimposes molecule Q onto molecule V. This transformation is found by solving the minimization problem: This problem can be solved efficiently using the Kabsch algorithm [52]. Using this transformation, we define a pairwise RMSD, dpair. This metric quantifies the structural deviation of the molecule, U, from its target, W, after the transformation (RQ→V,tQ→V) has been applied. Because we assume all molecules are perfect rigid-body instances of the same template, the initial alignment of Q onto V results in zero error. The metric is therefore defined solely by the deviation of the second molecule: Note that the sum is now over the k atoms of the second molecule only, and the normalization factor is k. To handle the ambiguity of local assignment within the pair, the final pairwise distance, dminpair, is the minimum of the two possible alignments, Q to V or Q to W: Finally, the total distance dRMSD between two configurations, 𝒳={X1,...,Xm} and 𝒴={Y1,...,Ym}, is the average of these pairwise distances, minimized over permutations π∈Sm: where the sum goes over all unique pairs of molecules in the configuration. Intuitively, dRMSD finds the optimal global matching of molecules. For each matched pair, it determines the best alignment based on one of the molecules and then measures the resulting RMSD of the entire pair. A dRMSD value of zero guarantees that two configurations are identical up to a global rigid transformation. Codebase The code we developed to generate the data presented in this study is available on github [53]. Raw simulation trajectories and analysis scripts are publicly available on Zenodo [54]. Supporting information Animation of the simulated assembly of a trimer of the human CD40 ligand. (MP4) Acknowledgments We acknowledge Patrice Koehl for the sharing and explanation of his program AlphaMol. We thank Gero Friesecke for discussions on virus self-assembly. This research used the computational cluster resource provided by the Center for Information Technology and Media Management at the University of Potsdam, and the Lawrencium computational cluster resource provided by the IT Division at the Lawrence Berkeley National Laboratory. Data Availability The simulation code is publicly available on GitHub. All data needed to evaluate the conclusions are present in the paper and the Supporting Information. Raw simulation trajectories and analysis scripts are available on Zenodo (DOI: https://doi.org/10.5281/zenodo.18624903). Funding Statement This work was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under TRR 109 “Discretization in Geometry and Dynamics,” Project-ID 195170736, awarded to ME (https://www.dfg.de). IS was further supported by a fellowship of the German-American Fulbright Commission (Fulbright Germany; https://www.fulbright.de). AN and DM were supported by the U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research, Mathematical Multifaceted Integrated Capability Centers (MMICCs) program, under Contract No. DE-AC02-05CH11231 (https://science.osti.gov/ascr). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. References 21 May 2026 Topological potentials guiding protein self-assembly Dear Dr. Spirandelli, Thank you for submitting your manuscript to PLOS Computational Biology. After careful consideration, we feel that it has merit but does not fully meet PLOS Computational Biology's publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Jul 21 2026 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at ploscompbiol@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pcompbiol/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: * A letter that responds to each point raised by the editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. This file does not need to include responses to formatting updates and technical items listed in the 'Journal Requirements' section below. * A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. * An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. We look forward to receiving your revised manuscript. Kind regards, Jianhan Chen, Ph.D. Academic Editor PLOS Computational Biology Arne Elofsson Section Editor PLOS Computational Biology Additional Editor Comments: Apologies for the delay. We were waiting for a third reviewer who did not submit the report after the due date despite several reminders. Both reviewers clearly appreciate the innovation and importance of this work, while raising various minor and major concerns on a range of aspects. I will greatly appreciate a revised manuscript with all points thoroughly addressed. Thank you. Journal Requirements: 1) Please ensure that the CRediT author contributions listed for every co-author are completed accurately and in full. At this stage, the following Authors/Authors require contributions: Ivan Lothar Arbo Spirandelli. Please ensure that the full contributions of each author are acknowledged in the "Add/Edit/Remove Authors" section of our submission form. The list of CRediT author contributions may be found here: https://journals.plos.org/ploscompbiol/s/authorship#loc-author-contributions 2) We ask that a manuscript source file is provided at Revision. Please upload your manuscript file as a .doc, .docx, .rtf or .tex. If you are providing a .tex file, please upload it under the item type u2018LaTeX Source Fileu2019 and leave your .pdf version as the item type u2018Manuscriptu2019. 3) Please upload all main figures as separate Figure files in .tif or .eps format. For more information about how to convert and format your figure files please see our guidelines: https://journals.plos.org/ploscompbiol/s/figures 4) We have noticed that you have uploaded Supporting Information files, but you have not included a list of legends. Please add a full list of legends for your Supporting Information files after the references list. 5) Please amend your detailed Financial Disclosure statement. This is published with the article. It must therefore be completed in full sentences and contain the exact wording you wish to be published. - State the initials, alongside each funding source, of each author to receive each grant. For example: "This work was supported by the National Institutes of Health (####### to AM; ###### to CJ) and the National Science Foundation (###### to AM)." - State what role the funders took in the study. If the funders had no role in your study, please state: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.". If you did not receive any funding for this study, please simply state: u201cThe authors received no specific funding for this work.u201d Reviewers' comments: Reviewer's Responses to Questions Comments to the Authors: Please note here if the review is uploaded as an attachment. Reviewer #1: The authors study protein self-assembly using a computational approach that combines a morphometric geometric solvation model with a long-range topological potential derived from persistent homology. The central idea is to use weighted total persistence as an energetic bias to guide assembly pathways and reduce kinetic trapping in rugged energy landscapes. The authors apply the approach to several systems, including TMV dimers, hepatitis B core protein, SARS-CoV-2 ORF9b, and CD40 ligand, and report substantial improvements in assembly success rates compared to the geometric model alone. The manuscript presents an interesting and potentially useful approach for improving simplified assembly simulations. The reported increases in assembly success rates are substantial, and the paper is generally clearly written. However, several points need to be addressed before publication. 1. The physical interpretation of the topological potential remains unclear. While the authors briefly relate a specific parameter choice to Gaussian curvature, the manuscript does not provide clear physical intuition regarding whether the topological term represents an effective approximation of omitted long-range interactions (e.g., electrostatic or hydrophobic effects) or functions primarily as an algorithmic bias that reshapes the energy landscape to reduce kinetic trapping. Relatedly, the manuscript frames protein assembly simulations as being dominated by short-range nonpolar interactions, whereas many assembly processes are strongly influenced by electrostatics, hydrogen bonding, and other longer-range effects. Since electrostatic interactions are omitted entirely, it remains unclear how the proposed topological potential compares to more conventional implicit-solvent or coarse-grained models that include electrostatics, and whether the topological term provides fundamentally new information or instead compensates for omitted physical interactions. 2. The manuscript focuses on relatively small homo-oligomeric systems, primarily dimers and trimers, rather than full viral capsids or larger supramolecular assemblies. Thus, the reported increases in “assembly success rates” correspond to improved formation of small native-like oligomeric complexes rather than full virus assembly. It would be useful for the authors to discuss more explicitly whether the proposed topological potential is expected to extend to larger-scale coarse-grained assembly simulations involving tens to hundreds of subunits, such as HBV or TMV capsid assembly. In particular, it remains unclear whether persistent homology can be computed efficiently using coarse-grained capsomers or bead-based representations rather than atomistic coordinates, whether the topological potential is expected to guide global shell closure or suppress malformed intermediates during capsid assembly, and how the computational cost is expected to scale with increasing system size. 3. Relatedly, the present model treats protein subunits as rigid bodies. In contrast, conformational flexibility often plays an important role in viral assembly pathways and polymorphism, including in coarse-grained models capable of assembling complete viral shells. See, for example, “From Disorder to Icosahedral Symmetry: How Conformation-Switching Subunits Enable RNA Virus Assembly,” Science Advances 11, eady7241 (2025). The authors should discuss whether the proposed topological framework could eventually be combined with flexible coarse-grained models and whether it is intended as a complementary guiding potential for improving existing coarse-grained assembly simulations rather than as a standalone assembly model. 4. The topological potential depends on weighting coefficients (λ0,λ1,λ2), which appear to be selected empirically through parameter scans, but the procedure used to choose these parameters remains somewhat unclear. The authors should clarify how the λi\lambda_iλi values are selected, whether they are optimized systematically or motivated by physical considerations, and how sensitive the results are to these choices. The apparent transferability across systems is interesting, but a more systematic robustness analysis would improve the conclusions. Similarly, the implementation of Eq. (1) remains somewhat vague. The manuscript should more clearly state the numerical values used for the morphometric parameters, how these parameters are selected, and whether any of them are adjusted during the simulations. This clarification is particularly important for readers unfamiliar with morphometric approaches. 5. The reported improvements in assembly success rates (e.g., from 5% to 85%) are striking. However, the statistical interpretation of these values could be improved by including confidence intervals or uncertainty estimates, discussion of run-to-run variability, and additional context regarding the dependence on simulation parameters such as temperature and perturbation size. Since “success rate” is the primary performance metric of the manuscript, a more detailed quantitative discussion would improve the interpretation of the results. Minor Comments 1. The description of persistent homology may be difficult for a broad audience. A more intuitive explanation in the main text would improve accessibility. 2. Figure 1 should more clearly explain the connection between the alpha filtration and the resulting topological potential. 3. The discussion of the morphometric solvation model should include a brief intuitive explanation of the integrated mean curvature C and integrated Gaussian curvature X of the solvent-accessible surface. 4. The manuscript states that simulations are driven by a Random Walk Metropolis algorithm, but the explanation in the main text is brief. Adding a short description of how moves are proposed and accepted/rejected would improve readability. 5. The definition of total persistence should include a simple illustrative example explaining the physical or geometric meaning of the birth (b) and death (d) values within the filtration. 6. The manuscript introduces Pearson correlation coefficients between different quantities. A short explanation of what the Pearson coefficient measures, and why it is informative in this context, would improve clarity for a broader audience. 7. The distinction between thermodynamic accuracy and kinetic accessibility is important and could be emphasized earlier in the manuscript. 8. Some notation associated with persistence measures and weighting coefficients could be simplified for readability. Reviewer #2: the review is uploaded as an attachment ********** Have the authors made all data and (if applicable) computational code underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data and code underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data and code should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data or code —e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: Yes ********** PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: Yes: Vysakh Ramachandran [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] Figure resubmission: While revising your submission, we strongly recommend that you use PLOS’s NAAS tool (https://ngplosjournals.pagemajik.ai/artanalysis) to test your figure files. NAAS can convert your figure files to the TIFF file type and meet basic requirements (such as print size, resolution), or provide you with a report on issues that do not meet our requirements and that NAAS cannot fix. After uploading your figures to PLOS’s NAAS tool - https://ngplosjournals.pagemajik.ai/artanalysis, NAAS will process the files provided and display the results in the "Uploaded Files" section of the page as the processing is complete. If the uploaded figures meet our requirements (or NAAS is able to fix the files to meet our requirements), the figure will be marked as "fixed" above. If NAAS is unable to fix the files, a red "failed" label will appear above. When NAAS has confirmed that the figure files meet our requirements, please download the file via the download option, and include these NAAS processed figure files when submitting your revised manuscript. Reproducibility: ?> Submitted filename: Review_Questions.pdf 22 Jul 2026 Submitted filename: response_to_reviewers.pdf 11 Aug 2026 Dear Dr. Spirandelli, We are pleased to inform you that your manuscript 'Topological potentials guiding protein self-assembly' has been provisionally accepted for publication in PLOS Computational Biology. Before your manuscript can be formally accepted you will need to complete some formatting changes, which you will receive in a follow up email. A member of our team will be in touch with a set of requests. Please note that your manuscript will not be scheduled for publication until you have made the required changes, so a swift response is appreciated. IMPORTANT: The editorial review process is now complete. PLOS will only permit corrections to spelling, formatting or significant scientific errors from this point onwards. Requests for major changes, or any which affect the scientific understanding of your work, will cause delays to the publication date of your manuscript. Should you, your institution's press office or the journal office choose to press release your paper, you will automatically be opted out of early publication. We ask that you notify us now if you or your institution is planning to press release the article. All press must be co-ordinated with PLOS. Thank you again for supporting Open Access publishing; we are looking forward to publishing your work in PLOS Computational Biology. Best regards, Jianhan Chen, Ph.D. Academic Editor PLOS Computational Biology Nir Ben-Tal Section Editor PLOS Computational Biology *********************************************************** Reviewer's Responses to Questions Comments to the Authors: Please note here if the review is uploaded as an attachment. Reviewer #1: The authors have addressed all my concerns and comments, and those of the other referee. Thus, I recommend the paper for publication. Reviewer #2: All of my concerns have been addressed in the revised manuscript. The authors have responded appropriately to the reviewers' comments, and I believe the manuscript has improved significantly. I have no further concerns and recommend that the manuscript be accepted for publication. ********** Have the authors made all data and (if applicable) computational code underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data and code underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data and code should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data or code —e.g. participant privacy or use of data from a third party—those must be specified. Reviewer #1: Yes Reviewer #2: Yes ********** PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy. Reviewer #1: No Reviewer #2: Yes: Vysakh Ramachandran PCOMPBIOL-D-26-00899R1 Topological potentials guiding protein self-assembly Dear Dr Spirandelli, I am pleased to inform you that your manuscript has been formally accepted for publication in PLOS Computational Biology. Your manuscript is now with our production department and you will be notified of the publication date in due course. The corresponding author will soon be receiving a typeset proof for review, to ensure errors have not been introduced during production. Please review the PDF proof of your manuscript carefully, as this is the last chance to correct any errors. Please note that major changes, or those which affect the scientific understanding of the work, will likely cause delays to the publication date of your manuscript. Soon after your final files are uploaded, unless you have opted out, the early version of your manuscript will be published online. The date of the early version will be your article's publication date. The final article will be published to the same URL, and all versions of the paper will be accessible to readers. For Research, Software, and Methods articles, you will receive an invoice from PLOS for your publication fee after your manuscript has reached the completed accept phase. If you receive an email requesting payment before acceptance or for any other service, this may be a phishing scheme. Learn how to identify phishing emails and protect your accounts at https://explore.plos.org/phishing. Thank you again for supporting PLOS Computational Biology and open-access publishing. We are looking forward to publishing your work! With kind regards, Sharmila Kamatchi PLOS Computational Biology | Carlyle House, Carlyle Road, Cambridge CB4 3DN | United Kingdom ploscompbiol@plos.org | Phone +44 (0) 1223-442824 | ploscompbiol.org | @PLOSCompBiol Supplementary Materials Animation of the simulated assembly of a trimer of the human CD40 ligand. (MP4) Submitted filename: Review_Questions.pdf Submitted filename: response_to_reviewers.pdf Data Availability Statement The simulation code is publicly available on GitHub. All data needed to evaluate the conclusions are present in the paper and the Supporting Information. Raw simulation trajectories and analysis scripts are available on Zenodo (DOI: https://doi.org/10.5281/zenodo.18624903).