Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models

arXiv:2607.06421 · physics.comp-ph · Submitted 2026-07-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models".

Mira: Here's the gist of the paper. Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models introduces Gradient-Based Free Energy Surface Optimization (GBFESO), an inverse design framework that uses a trained conditional diffusion model as a differentiable surrogate for the ensemble distribution <ref:2607.06421#pg2>.

Kai: This method optimizes the conditioning variables defining the system so that the generated ensemble reproduces a prescribed target free-energy surface <ref:2607.06421#pg4>.

Mira: Pruning redundant reasoning tokens The core idea of GBFESO is to treat a trained conditional diffusion model as a differentiable surrogate for the ensemble distribution <ref:2607.06421#pg2>.

Kai: After training, the diffusion model is frozen, and the conditioning variables defining the system are optimized so that the generated ensemble reproduces a prescribed target free-energy surface <ref:2607.06421#pg4>.

Paper summary: Kai: Thinking about the title Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models, it really captures how they’ve connected free energy surfaces, which govern state populations and transition barriers, with diffusion models for design. It’s about using a generative model to navigate that landscape.

Mira: I agree. What this work points toward is the shift from optimizing a single structure to controlling the entire statistical ensemble of conformations two <ref:2607.06421#pg2>. They are showing how you can use deep learning surrogates to guide that control process through backpropagation on distribution losses like KL divergence six.

Lev: From my perspective in error correction, the fact that they are optimizing parameters governing the full ensemble means you’re designing for robustness across a whole set of possibilities, not just one specific configuration two <ref:2607.06421#pg2>. It’s about designing systems that behave correctly even when perturbed.

Kai: So, it’s less about hitting one target structure perfectly and more about steering the underlying probability distribution toward the desired free energy surface using the learned mapping from the diffusion model three <ref:2607.06421#pg3>. That ability to extrapolate beyond training data is significant, showing it can generalize its design capability.

Mira: And while they are making progress on this framework, they’re also highlighting that for very large systems, we might need different distributional objectives than KL divergence to maintain scalability and robustness four. So the future direction looks like exploring those alternatives to keep this method practical for bigger things.

Lev: Exactly. The authors themselves noted that improving scalability and robustness in higher-dimensional settings will require alternative distributional objectives, such as Wasserstein-based metrics, to really push this beyond toy systems four. That’s a clear path forward for the next generation of molecular design tools.

Conclusion: Kai: So, this paper is about using gradient methods to design free energy landscapes for systems instead of just looking at one specific structure, right?

Mira: Yeah, basically they took a diffusion model and trained it to be like a map that tells you how to change the system parameters to get the free energy surface you want.

Kai: It’s about this gradient-based inverse design thing where you’re backpropagating through the diffusion sampling trajectory using a distribution loss, which is like, minimizing how much the generated ensemble differs from your target ensemble.

Mira: The authors are showing that even if you're not training on a specific protein structure, but on an entire collection of state ensembles across many different systems, you can still optimize those underlying design variables.

Kai: They did this on a four-bead system, and it worked pretty well for reproducing target landscapes in most test cases, even when the parameters you were tuning were outside what the model had seen before.

Mira: That's interesting because they’re using these complex neural networks to do something fundamentally thermodynamic—designing the energy landscape itself.

Kai: It suggests this conditional diffusion model approach could be a practical way to guide molecular or materials design by controlling the statistical properties of the output ensemble, not just picking one good structure.

Mira: But they also admit there are challenges, especially with very large networks where you might need different ways to measure that distributional agreement than just using the standard Kullback–Leibler divergence.

Kai: So it’s a promising first step for differentiable free energy surface design, but we still have to figure out how efficient this runs when we scale up to truly massive systems.

Mira: Exactly. That leads us into how they handled the specific setup of their four-bead system and what that tells us about the complexity of these models in practice.

Eli Zick, Dan Mendels

Wolfson Department of Chemical Engineering, Technion – Israel Institute of Technology

physics.comp-ph

Submitted: 2026-07-07

Updated: 2026-11-15

Code: https://github.com/MendelsResearchGroup/Gradient-Based-Free-Energy-Surface-Optimization

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 99/100

The gist: Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models introduces Gradient-Based Free Energy Surface Optimization (GBFESO)…

Key concepts

Gradient-Based Free Energy Surface Optimization (GBFESO)
This is an inverse design framework that treats a trained diffusion model as a differentiable tool. Instead of designing single structures, it optimizes the input conditions that govern the entire ensemble distribution to match a desired target free-energy surface.
Conditional Diffusion Model
A trained diffusion model acts as a differentiable mapping. It takes system-defining properties (the condition 'c') and generates an ensemble of samples that collectively define the corresponding free-energy surface, allowing the optimization process to be guided by this generative capability.
Ensemble-Level Objective
The method optimizes based on how well the generated ensemble matches a target distribution, not just individual conformations. This agreement is measured using metrics like Kullback–Leibler (KL) divergence, which quantifies the difference between two probability distributions.

Terminology

Summary

Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models introduces Gradient-Based Free Energy Surface Optimization (GBFESO), an inverse design framework that uses a trained conditional diffusion model as a differentiable surrogate for the ensemble distribution <ref:2607. p. 0

Where the paper says it

Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models Eli Zick1 and Dan Mendels1, ∗ 1The Wolfson Department of Chemical Engineering, Technion – Israel Institute of Technology, Haifa 32000, Israel (Dated: July 8, 2026) Free-energy surfaces govern the populations of metastable states and the barriers that control transitions between them, making their direct optimization a central challenge in molecular and materials design. In this work, we introduce Gradient-Based Free Energy Surface Optimization (GBFESO), an inverse design framework that uses a trained conditional diffusion model as a differentiable surrogate for the ensemble distribution.

06421#pg2>. This method optimizes the conditioning variables defining the system so that the generated ensemble reproduces a prescribed target free-energy surface <ref:2607. p. 0
Where the paper says it

In this work, we introduce Gradient-Based Free Energy Surface Optimization (GBFESO), an inverse design framework that uses a trained conditional diffusion model as a differentiable surrogate for the ensemble distribution. After training, the diffusion model is frozen, and the conditioning variables defining the system are optimized so that the generated ensemble reproduces a prescribed target free-energy surface.

06421#pg4>.

Pruning redundant reasoning tokens

The core idea of GBFESO is to treat a trained conditional diffusion model as a differentiable surrogate for the ensemble distribution <ref:2607.06421#pg2>. After training, the diffusion model is frozen, and the conditioning variables defining the system are optimized so that the generated ensemble reproduces a prescribed target free-energy surface <ref:2607. p. 0

Where the paper says it

In this work, we introduce Gradient-Based Free Energy Surface Optimization (GBFESO), an inverse design framework that uses a trained conditional diffusion model as a differentiable surrogate for the ensemble distribution. After training, the diffusion model is frozen, and the conditioning variables defining the system are optimized so that the generated ensemble reproduces a prescribed target free-energy surface.

06421#pg4>. This optimization is carried out by backpropagating a distribution-level loss, based on kernel density estimates of the Kullback–Leibler divergence, through a deterministic diffusion sampling trajectory <ref:2607. p. 0
Where the paper says it

After training, the diffusion model is frozen, and the conditioning variables defining the system are optimized so that the generated ensemble reproduces a prescribed target free-energy surface. The optimization is carried out by backpropagating a distribution-level loss, based on kernel density estimates of the Kullback–Leibler divergence, through a deterministic diffusion sampling trajectory.

06421#pg6>. This shifts the inverse-design problem from targeting individual configurations such as a protein’s native state to tuning the parameters that govern the full conformational ensemble <ref:2607. p. 0
Where the paper says it

This shifts the inverse-design problem from targeting individual configurations such as a protein’s native state to tuning the parameters that govern the full conformational ensemble. Such an approach is particularly relevant because the behavior of many physical systems, including proteins, arXiv:2607.

06421#pg2>.

How the training signal is built

The conditional diffusion model serves as a differentiable mapping from a design condition c, representing a set of system-defining properties, to an ensemble of samples that collectively define the corresponding FES <ref:2607.06421#pg3>. The model is trained on a collection of state (conformational) ensembles spanning a diverse set of systems within the domain of interest <ref:2607. p. 1

Where the paper says it

After training on a collection of state (conformational) ensembles spanning a diverse set of systems within the domain of interest, the parameters of the generative neural network are frozen. At this stage, the conditional diffusion model is no longer used only to sample ensembles for a prescribed condition; instead, it is treated as a differentiable surrogate for the condition-dependent ensemble distribution.

06421#pg2>. The training objective involves predicting the noise injected into xt−1 to obtain xt, typically using a mean-squared error objective <ref:2607. p. 1
Where the paper says it

The training at each diffusion step is important also because the amount of corruption varies across the diffusion trajectory, requiring the model to adapt its denoising behavior at different noise levels. In practice, training is commonly performed using a simple mean-squared error objective on the predicted noise [7, 29]: MSE = 1 n Xn i=1 (yi − yˆi) 2 (5)

06421#pg6>.

Where the gains come from

The optimization is driven by an ensemble-level objective rather than by agreement between individual generated and target conformations <ref:2607.06421#pg4>. The required quantitative metric for this comparison is the Kullback–Leibler (KL) divergence, which can be approximated with Gaussian kernel density estimation (KDE) to form a differentiable loss <ref:2607. p. 2

Where the paper says it

A natural and widely used choice is the Kullback–Leibler (KL) divergence[30]: DKL(P ∥ Q) = Z p(x) log p(x) q(x) dx. (6) Without an explicit analytic form for the generated distribution, the KL divergence can be approximated with Gaussian kernel density estimation (KDE)[31], and can be written as a differentiable loss: DKL(ˆptarget ∥ pˆgen) ≈ 1 N X N j=1 log pˆtarget(xj ) pˆgen(xj ) , xj ∼ pˆtarget.

06421#pg6>. The final optimization objective is formulated as minimizing c∗ = arg min c∈C h DKL(ˆptarget(x)∥ pˆgen(x)) + λDKL(ˆpDDPM(x c)∥ pˆtarget(x))i <ref:2607. p. 2
Where the paper says it

(7) Here log p(x) ≈ log ˆp(x) = log" 1 N X N i=1 1 (2π) d/2h d × exp − ∥x − xi∥ 2 2h 2 , (8) and h is a bandwidth of the KDE. Given an input condition c, the final optimization objective can be written as c ∗ = arg min c∈C h DKL(ˆptarget(x) ∥ pˆDDPM(x | c)) + λDKL(ˆpDDPM(x | c) ∥ pˆtarget(x))i , (9)

06421#pg6>.

The method has been validated on one-dimensional Gaussian ensembles, demonstrating that both continuous and relaxed discrete conditioning variables can be optimized to recover target distributions, including those outside the training domain <ref:2607.06421#pg4>. For a four-particle Lennard–Jones toy peptide, GB-FESO successfully optimizes the interaction parameters to reproduce target free-energy landscapes in the majority of test cases <ref:2607. p. 0

Where the paper says it

We then apply the method to a four-particle Lennard–Jones toy peptide exhibiting multiple metastable conformational states. In this more physically motivated setting, GB-FESO successfully optimizes the interaction parameters to reproduce target free-energy landscapes in the majority of test cases, with optimization performed either in the full internal-coordinate space or in a reduced collective-variable representation.

06421#pg4>. The method can also handle discrete conditioning variables by employing continuous relaxation techniques such as binary logit relaxation or Gumbel–Softmax relaxation <ref:2607. p. 3
Where the paper says it

Discrete conditioning variables, however, are not differentiable and therefore require a continuous relaxation during optimization. One possible approach is binary logit relaxation[36], although several alternative methods for handling discrete variables have been proposed, including the Straight-Through Estimator (STE)[37] and the Gumbel–Softmax relaxation[38, 39].

06421#pg6>.

The success in the toy peptide study showed that optimization can be performed in either the full internal-coordinate space or in a reduced collective-variable representation, such as using Harmonic Linear Discriminant Analysis (HLDA) <ref:2607.06421#pg6>. This reduced space approach was found to substantially mitigate computational demands for larger systems <ref:2607. p. 7

Where the paper says it

Moreover, the KDE-based approximation of the KL divergence objective can give rise to a rough optimization landscape, hindering convergence. As shown in this work, however, performing free-energy optimization in a reduced space of collective variables obtained through dimensionality-reduction methods, such as HLDA, can substantially mitigate these issues.

06421#pg6>. The results suggest that conditional diffusion models provide a practical route for ensemble-level molecular and materials design <ref:2607. p. 7
Where the paper says it

In this more physically motivated setting, the method successfully optimizes the interaction strengths to reproduce target free-energy surfaces in the large majority of test cases, including targets whose conditioning parameters lie outside the training region of the diffusion model. These results suggest that conditional diffusion models provide a practical route for ensemble-level molecular and materials design, particularly in settings where the desired outcome is naturally expressed as a distribution rather than as a single optimized structure.

p. 7
Where the paper says it

In this more physically motivated setting, the method successfully optimizes the interaction strengths to reproduce target free-energy surfaces in the large majority of test cases, including targets whose conditioning parameters lie outside the training region of the diffusion model. These results suggest that conditional diffusion models provide a practical route for ensemble-level molecular and materials design, particularly in settings where the desired outcome is naturally expressed as a distribution rather than as a single optimized structure.

06421#pg4>.

The paper concludes that optimizing the conditioning variables of a frozen generative model can effectively steer generated ensembles toward target thermodynamic behavior and target free-energy landscapes <ref:2607.06421#pg4>. The framework represents a promising first step toward differentiable free-energy-surface design <ref:2607. p. 7

Where the paper says it

Furthermore, improving scalability and robustness in higher-dimensional settings may require alternative distributional objectives, such as Wasserstein-based metrics [46]. Overall, the proposed framework represents a promising first step toward differentiable free-energy-surface design.

06421#pg4>. However, challenges remain regarding computational efficiency for very large networks and the potential roughness of the optimization landscape <ref:2607. p. 7
Where the paper says it

For very large neural networks and/or high-dimensional free-energy landscapes, straightforward backpropagation through the diffusion sampling trajectory may become computationally inefficient or unstable. Moreover, the KDE-based approximation of the KL divergence objective can give rise to a rough optimization landscape, hindering convergence.

06421#pg6>.

The authors note that performing free-energy optimization in a reduced space of collective variables obtained through dimensionality-reduction methods, such as HLDA, can substantially mitigate these issues <ref:2607.06421#pg6>. Furthermore, improving scalability and robustness in higher-dimensional settings may require alternative distributional objectives, such as Wasserstein-based metrics <ref:2607. p. 7

Where the paper says it

As shown in this work, however, performing free-energy optimization in a reduced space of collective variables obtained through dimensionality-reduction methods, such as HLDA, can substantially mitigate these issues. Furthermore, improving scalability and robustness in higher-dimensional settings may require alternative distributional objectives, such as Wasserstein-based metrics [46].

06421#pg4>. The code and dataset scripts used to generate and analyze the data supporting this article are openly available in Zenodo at https://doi. p. 8
Where the paper says it

Thus, the optimized condition is evaluated according to the final KL loss between the generated and target internal-coordinate ensembles. DATA AVAILABILITY The code and the dataset scripts used to generate and analyze the data supporting this article are openly available in Zenodo at https://doi.

org/10.5281/zenodo. p. 8
Where the paper says it

DATA AVAILABILITY The code and the dataset scripts used to generate and analyze the data supporting this article are openly available in Zenodo at https://doi. 5281/zenodo.

21075939 <ref:2607. p. 0
Where the paper says it

This shifts the inverse-design problem from targeting individual configurations such as a protein’s native state to tuning the parameters that govern the full conformational ensemble. Such an approach is particularly relevant because the behavior of many physical systems, including proteins, arXiv:2607.

06421#pg6>. The development version of the code is available on GitHub at https://github. p. 8
Where the paper says it

5281/zenodo. The development version of the code is available on GitHub at https://github.

com/MendelsResearchGroup/Gradient-Based-Free-Energy-Surface-Optimization and is released under the MIT License <ref:2607. p. 8
Where the paper says it

The development version of the code is available on GitHub at https://github. com/MendelsResearchGroup/ Gradient-Based-Free-Energy-Surface-Optimization and is released under the MIT License.

06421#pg6>. The authors acknowledge support from the Israel Science Foundation (ISF) under grant number 1181/24 <ref:2607. p. 8
Where the paper says it

com/MendelsResearchGroup/ Gradient-Based-Free-Energy-Surface-Optimization and is released under the MIT License. ACKNOWLEDGMENTS The authors acknowledge support from the Israel Science Foundation (ISF) under grant number 1181/24.

06421#pg6>. The paper is 2607. p. 0
Where the paper says it

This shifts the inverse-design problem from targeting individual configurations such as a protein’s native state to tuning the parameters that govern the full conformational ensemble. Such an approach is particularly relevant because the behavior of many physical systems, including proteins, arXiv:2607.

06421 <ref:2607. p. 0
Where the paper says it

This shifts the inverse-design problem from targeting individual configurations such as a protein’s native state to tuning the parameters that govern the full conformational ensemble. Such an approach is particularly relevant because the behavior of many physical systems, including proteins, arXiv:2607.

06421#pg4>.

--- Page 9 ---

9

The resulting dump files were postprocessed into internal coordinates rather than Cartesian coordinates <ref:2607. p. 7

Where the paper says it

The simulations were performed in reduced Lennard– Jones units with Langevin dynamics at low temperature t = 0. The resulting dump files were postprocessed into internal coordinates rather than Cartesian coordinates.

06421#pg6>. For every saved frame, the target vector was x = (r12, r23, r34, θ123, θ234) <ref:2607. p. 7
Where the paper says it

The resulting dump files were postprocessed into internal coordinates rather than Cartesian coordinates. For every saved frame, the target vector was x = (r12, r23, r34, θ123, θ234),

06421#pg6>. The model learns a conditional ensemble distribution pθ(x ε1, ε2), where changes in the condition correspond to changes in the interaction parameters that define the four-bead system <ref:2607. p. 8
Where the paper says it

The corresponding (ε1, ε2) values were extracted directly from the LAMMPS pair coefficients and used as the conditioning variables. Toy-peptide DDPM and inverse optimization Within this representation, the model therefore learns a conditional ensemble distribution pθ(x | ε1, ε2), where changes in the condition correspond to changes in the interaction parameters that define the four-bead system.

06421#pg6>.

--- Page 10 ---

10

The optimization is performed in two stages. In the first stage, the condition is optimized with a relatively large learning rate while stochastic perturbations are added to help escape local minima <ref:2607. p. 8

Where the paper says it

Optimization is performed in two stages. In the first stage, the condition is optimized with a relatively large learning rate while stochastic perturbations are added to help escape local minima.

06421#pg6>. In the second stage, if the KL loss remains above a chosen threshold, the condition is further refined with a lower learning rate and without additional perturbations <ref:2607. p. 8
Where the paper says it

In the first stage, the condition is optimized with a relatively large learning rate while stochastic perturbations are added to help escape local minima. In the second stage, if the KL loss remains above a chosen threshold, the condition is further refined with a lower learning rate and without additional perturbations.

p. 8
Where the paper says it

In the first stage, the condition is optimized with a relatively large learning rate while stochastic perturbations are added to help escape local minima. In the second stage, if the KL loss remains above a chosen threshold, the condition is further refined with a lower learning rate and without additional perturbations.

06421#pg6>. This procedure is used to improve convergence toward a condition whose generated ensemble matches the target distribution <ref:2607. p. 8
Where the paper says it

In the second stage, if the KL loss remains above a chosen threshold, the condition is further refined with a lower learning rate and without additional perturbations. This procedure is used to improve convergence toward a condition whose generated ensemble matches the target distribution.

p. 8
Where the paper says it

In the second stage, if the KL loss remains above a chosen threshold, the condition is further refined with a lower learning rate and without additional perturbations. This procedure is used to improve convergence toward a condition whose generated ensemble matches the target distribution.

06421#pg6>. The purpose of the optimization is not necessarily to recover the exact target values of ε1 and ε2 <ref:2607. p. 8
Where the paper says it

5 and 3600-4096 samples per optimization step and KL loss calculation. The purpose of the optimization is not necessarily to recover the exact target values of ε1 and ε2.

06421#pg6>. Since, in the broader case, different interaction parameters might produce similar ensemble statistics, success is instead defined by the agreement between the generated and target distributions <ref:2607. p. 8
Where the paper says it

The purpose of the optimization is not necessarily to recover the exact target values of ε1 and ε2. Since, in the broader case, different interaction parameters might produce similar ensemble statistics, success is instead defined by the agreement between the generated and target distributions.

06421#pg6>.

--- Page 11 ---

11

The model is trained to predict the Gaussian noise added during the forward diffusion process, using the standard DDPM noise-prediction mean-squared error objective <ref:2607.06421#pg6>. The diffusion timestep is encoded using a 64-dimensional sinusoidal embedding, while the timestep and condition embeddings are mapped through 128-dimensional conditioning layers <ref:2607. p. 8

Where the paper says it

Each block contains layer normalization, FiLM scale-and-shift modulation, two linear layers, SiLU activation, and dropout with probability 0. The diffusion timestep is encoded using a 64-dimensional sinusoidal embedding, while the timestep and condition embeddings are mapped through 128-dimensional conditioning layers.

p. 8
Where the paper says it

Each block contains layer normalization, FiLM scale-and-shift modulation, two linear layers, SiLU activation, and dropout with probability 0. The diffusion timestep is encoded using a 64-dimensional sinusoidal embedding, while the timestep and condition embeddings are mapped through 128-dimensional conditioning layers.

06421#pg6>. At each diffusion step, the noisy internal-coordinate vector, the diffusion timestep, and the condition vector are provided to the network <ref:2607. p. 8
Where the paper says it

The diffusion timestep is encoded using a 64-dimensional sinusoidal embedding, while the timestep and condition embeddings are mapped through 128-dimensional conditioning layers. At each diffusion step, the noisy internal-coordinate vector, the diffusion timestep, and the condition vector are provided to the network.

p. 8
Where the paper says it

The diffusion timestep is encoded using a 64-dimensional sinusoidal embedding, while the timestep and condition embeddings are mapped through 128-dimensional conditioning layers. At each diffusion step, the noisy internal-coordinate vector, the diffusion timestep, and the condition vector are provided to the network.

06421#pg6>. The network is trained to predict the Gaussian noise added during the forward diffusion process, using a simple mean-squared error objective on the predicted noise <ref:2607. p. 8
Where the paper says it

At each diffusion step, the noisy internal-coordinate vector, the diffusion timestep, and the condition vector are provided to the network. The network is trained to predict the Gaussian noise added during the forward diffusion process, using the standard DDPM noise-prediction mean-squared-error objective.

06421#pg6>.

--- Page 13 ---

13

The authors acknowledge support from the Israel Science Foundation (ISF) under grant number 1181/24 <ref:2607.06421#pg6>. [3] Rives, A. p. 8

Where the paper says it

1126/science. aat2663 [3] Rives, A.

et al. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. p. 8
Where the paper says it

aat2663 [3] Rives, A. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences.

Proceedings of the National Academy of Sciences, 118, e2016239118, 2021 <ref:2607. p. 8
Where the paper says it

Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proceedings of the National Academy of Sciences, 118, e2016239118, 2021.

06421#pg6>. [4] Salman, S. p. 8
Where the paper says it

https://doi. 2016239118 [4] Salman, S.

N., Shteingolts, S. p. 8
Where the paper says it

2016239118 [4] Salman, S. , Shteingolts, S.

A., Levie, R. p. 8
Where the paper says it

, Shteingolts, S. , Levie, R.

, and Mendels, D. p. 8
Where the paper says it

, Levie, R. , and Mendels, D.

p. 8
Where the paper says it

, Levie, R. , and Mendels, D.

Evaluating the use of a machine learning simulator for structure–property prediction: A case study on disordered elastic networks <ref:2607. p. 8
Where the paper says it

, and Mendels, D. Evaluating the use of a machine learning simulator for structure–property prediction: A case study on disordered

06421#pg6>. [5] Shteingolts, S. p. 8
Where the paper says it

2016239118 [4] Salman, S. , Shteingolts, S.

A., Salman, S. p. 8
Where the paper says it

https://doi. 2016239118 [4] Salman, S.

N., and Mendels, D. Enabling structure-only initialization and out-of-distribution generalization in GNN-based molecular dynamics simulators <ref:2607. p. 9
Where the paper says it

, and Mendels, D. Enabling structure-only initialization and out-of-distribution generalization in GNN-based molecular dynamics simulators.

06421#pg6>. [6] Sanchez-Gonzalez, A. p. 9
Where the paper says it

org/abs/2605. 09495 [6] Sanchez-Gonzalez, A.

, Godwin, J. p. 9
Where the paper says it

09495 [6] Sanchez-Gonzalez, A. , Godwin, J.

, Pfaff, T. p. 9
Where the paper says it

, Godwin, J. , Pfaff, T.

, Ying, R., Leskovec, J. p. 9
Where the paper says it

, Pfaff, T. , Leskovec, J.

, and Battaglia, P. p. 9
Where the paper says it

, Leskovec, J. , and Battaglia, P.

Learning to simulate complex physics with graph networks <ref:2607. p. 9
Where the paper says it

, and Battaglia, P. Learning to simulate complex physics with graph networks.

06421#pg6>. [7] Ho, J., Jain, A., and Abbeel, P. p. 9
Where the paper says it

html [7] Ho, J. , and Abbeel, P.

Denoising diffusion probabilistic models <ref:2607. p. 0
Where the paper says it

INTRODUCTION Recent progress in machine learning has expanded the methodological basis of molecular engineering, particularly through models that learn directly from large-scale sequence and structural data[1–6]. Within this landscape, Denoising Diffusion Probabilistic Models (DDPMs)[7] have emerged as a useful generative framework for protein ensembles.

06421#pg6>. [8] Jumper, J. p. 9
Where the paper says it

cc/paper/2020/ file/4c5bcfec8584af0d967f1ab10179ca4b-Paper. pdf [8] Jumper, J.

, Evans, R. p. 9
Where the paper says it

pdf [8] Jumper, J. , Evans, R.

, Pritzel, A. p. 9
Where the paper says it

, Evans, R. , Pritzel, A.

, Green, T. p. 9
Where the paper says it

, Pritzel, A. , Green, T.

, Figurnov, M. p. 9
Where the paper says it

, Green, T. , Figurnov, M.

, Ronneberger, O. p. 9
Where the paper says it

, Figurnov, M. , Ronneberger, O.

, Tunyasuvunakool, K. p. 9
Where the paper says it

, Ronneberger, O. , Tunyasuvunakool, K.

, Bates, R. p. 9
Where the paper says it

, Tunyasuvunakool, K. , Bates, R.

, Žídek, A. p. 9
Where the paper says it

, Bates, R. , Žídek, A.

, Potapenko, A. p. 9
Where the paper says it

, Žídek, A. , Potapenko, A.

, Bridgland, A. p. 9
Where the paper says it

, Potapenko, A. , Bridgland, A.

, Meyer, C. p. 9
Where the paper says it

, Bridgland, A. , Meyer, C.

Kohll S.A. Ballard J. Evans R Pritzel A Green T Figurnov M Ronneberger O Tunyasuvunakool K Bates R Žídek A Potapenko A Bridgland A Meyer C Kohl S.A. Ballard AJ Evans R Pritzel A Green T Figurnov M Ronneberger O Tunyasuvunakool K Bates R Žídek A Potapenko A Bridgland A Meyer C Kohl S.A. Ballard D Highly accurate protein structure prediction with AlphaFold Nature 596:583–589, 2021 <ref:2607. p. 9
Where the paper says it

Highly accurate protein structure prediction with AlphaFold. Nature, 596:583–589, 2021.

06421#pg4>. [9] Lewis, S. p. 9
Where the paper says it

https://doi. 1038/s41586-021-03819-2 [9] Lewis, S.

, Hempel, T. p. 9
Where the paper says it

1038/s41586-021-03819-2 [9] Lewis, S. , Hempel, T.

, Jimenez-Luna, J. p. 9
Where the paper says it

, Hempel, T. , Jimenez-Luna, J.

, Gastegger, M. p. 9
Where the paper says it

, Jimenez-Luna, J. , Gastegger, M.

, Xie, Y., Foong, A. p. 9
Where the paper says it

, Gastegger, M. , Foong, A.

Y. K., Satorras V. p. 9
Where the paper says it

, Foong, A. , Satorras, V.

G., Abdin O. Veeling B. S. Zaporozhets I. p. 9
Where the paper says it

, Veeling, B. , Zaporozhets, I.

et al Scalable emulation of protein equilibrium ensembles with generative deep learning Science 389(6761):eadv9817, 2025 <ref:2607. p. 0
Where the paper says it

This shifts the inverse-design problem from targeting individual configurations such as a protein’s native state to tuning the parameters that govern the full conformational ensemble. Such an approach is particularly relevant because the behavior of many physical systems, including proteins, arXiv:2607.

p. 9
Where the paper says it

Scalable emulation of protein equilibrium ensembles with generative deep learning. Science, 389(6761):eadv9817, 2025.

06421#pg4>. [10] Watson, J. p. 9
Where the paper says it

1126/science. adv9817 [10] Watson, J.

L., Juergens, D. p. 9
Where the paper says it

adv9817 [10] Watson, J. , Juergens, D.

, Bennett N. p. 9
Where the paper says it

, Juergens, D. , Bennett, N.

R. Trippe B. L., Yim J. Eisenach H. p. 9
Where the paper says it

, Trippe, B. , Eisenach, H.

E., Ahern W. Borst A J Ragotte R J Milles L F et al De novo design of protein structure and function with RFdiffusion Nature 620:1089–1100, 2023 <ref:2607. p. 9
Where the paper says it

De novo design of protein structure and function with RFdiffusion. Nature, 620:1089–1100, 2023.

06421#pg4>. [14] Liu, Y. p. 9
Where the paper says it

press/v202/yim23a. html [14] Liu, Y.

, Yu, Z., Lindsay R. p. 9
Where the paper says it

html [14] Liu, Y. , Lindsay, R.

J., Lin G. Chen M. Sahoo A. Hanson S M ExEnDiff: An experimentguided diffusion model for protein conformational ensemble generation PRX Life 3:023013, 2025 <ref:2607. p. 9
Where the paper says it

ExEnDiff: An experimentguided diffusion model for protein conformational ensemble generation. PRX Life, 3:023013, 2025.

06421#pg4>. [15] Kara C. p. 9
Where the paper says it

1103/PRXLife. 023013 [15] Kara, C.

, Hoppe T., Angelis E. p. 9
Where the paper says it

, Hoppe, T. , Angelis, E.

Schreiner J M. p. 9
Where the paper says it

, Angelis, E. , Schreiner, J.

Bauer S. Dittadi A. Olsson S Minimum-excesswork guidance: Score-based sampling with experimental data or sparse restraints Journal of Chemical Theory and Computation 2026 <ref:2607. p. 9
Where the paper says it

Minimum-excesswork guidance: Score-based sampling with experimental data or sparse restraints. Journal of Chemical Theory and Computation, 2026.

06421#pg6>. [16] Novak B. p. 9
Where the paper says it

https://doi. 6c00080 [16] Novak, B.

, Lotthammer J. p. 9
Where the paper says it

6c00080 [16] Novak, B. , Lotthammer, J.

M., Emenecker R. p. 9
Where the paper says it

, Lotthammer, J. , Emenecker, R.

J., and Holehouse A. p. 9
Where the paper says it

, Emenecker, R. , and Holehouse, A.

S Accurate predictions of disordered protein ensembles with STARLING Nature 652:240–250, 2026 <ref:2607. p. 9
Where the paper says it

Accurate predictions of disordered protein ensembles with STARLING. Nature, 652:240–250, 2026.

06421#pg4>. [17] Stenstrom O. p. 9
Where the paper says it

https://doi. 1038/s41586-026-10141-2 [17] Stenstrom, O.

, Diehl C., Modig K., and Akke M. p. 9
Where the paper says it

, Modig, K. , and Akke, M.

Ligandinduced protein transition state stabilization switches the binding pathway from conformational selection to induced fit Proceedings of the National Academy of Sciences 121(14):e2317747121, 2024 <ref:2607. p. 9
Where the paper says it

Ligandinduced protein transition state stabilization switches the binding pathway from conformational selection to induced fit. Proceedings of the National Academy of Sciences, 121(14):e2317747121, 2024.

06421#pg6>. [18] Spoendlin F. p. 9
Where the paper says it

1073/ pnas. 2317747121 [18] Spoendlin, F.

C., Fernandez-Quintero M. p. 9
Where the paper says it

2317747121 [18] Spoendlin, F. , Fernandez-Quintero, M.

L., Raghavan S. p. 9
Where the paper says it

, Fernandez-Quintero, M. , Raghavan, S.

S. R., Turner H. p. 9
Where the paper says it

, Raghavan, S. , Turner, H.

L., Gharpure A. p. 9
Where the paper says it

, Turner, H. , Gharpure, A.

, Loeffler J. p. 9
Where the paper says it

, Gharpure, A. , Loeffler, J.

R., Wong W. K., Bujotzek A. p. 9
Where the paper says it

, Loeffler, J. , Bujotzek, A.

, Georges G. p. 9
Where the paper says it

, Bujotzek, A. , Georges, G.

, Ward A B et al Predicting the conformational flexibility of antibody and T cell receptor complementarity-determining regions Nature Machine Intelligence 7(10):1755–1767, 2025 <ref:2607. p. 9
Where the paper says it

Predicting the conformational flexibility of antibody and T cell receptor complementarity-determining regions. Nature Machine Intelligence, 7(10):1755–1767, 2025.

06421#pg4>. [19] Nussinov R. p. 9
Where the paper says it

https:// doi. 1038/s42256-025-01131-6 [19] Nussinov, R.

, Regev C., and Jang H. p. 9
Where the paper says it

, Regev, C. , and Jang, H.

Leveraging conformational ensembles in allosteric drug discovery Trends in Pharmacological Sciences 47(3):276–289, 2026 <ref:2607. p. 9
Where the paper says it

Leveraging conformational ensembles in allosteric drug discovery. Trends in Pharmacological Sciences, 47(3):276–289, 2026.

06421#pg4>.

--- Page 15 ---

15

The internal-coordinate representation of such a system is 5-dimensional and can be written as: x = (r12, r23, r34, θ123, θ234) <ref:2607.06421#pg6>. Ground-truth reference ensembles have been generated using molecular dynamics simulations in LAMMPS<ref:2607. p. 4

Where the paper says it

The internal-coordinate representation of such a system is 5-dimensional and can be written as: x = (r12, r23, r34, θ123, θ234), where r denotes distances and θ denotes bond angles. Ground-truth reference ensembles have been generated using molecular dynamics simulations in LAMMPS[41].

06421#pg6>. The corresponding (ε1, ε2) values were extracted directly from the LAMMPS pair coefficients and used as the conditioning variables <ref:2607. p. 8
Where the paper says it

These fivedimensional internal-coordinate samples were used as the training data for the conditional diffusion model. The corresponding (ε1, ε2) values were extracted directly from the LAMMPS pair coefficients and used as the conditioning variables.

06421#pg6>. The resulting 5-dimensional internal-coordinate samples were used as the training data for the conditional diffusion model <ref:2607. p. 8
Where the paper says it

9 where the three r values are adjacent bond distances and the two θ values are signed bond angles. These fivedimensional internal-coordinate samples were used as the training data for the conditional diffusion model.

06421#pg6>.

--- Page 18 ---

18

The first parameter, ε1, was assigned to bonded-neighbor pairs, (1, 2), (2, 3), (3, 4), while the second parameter, ε2, was assigned to nextnearest-neighbor pairs <ref:2607.06421#pg6>. The remaining long-range pair (1, 4) was kept fixed at ε3 = 0. p. 7

Where the paper says it

The first parameter, ε1, was assigned to bonded-neighbor pairs, (1, 2), (2, 3), (3, 4), while the second parameter, ε2, was assigned to nextnearest-neighbor pairs, (1, 3), (2, 4). The remaining long-range pair (1, 4) was kept fixed at ε3 = 0.

5 <ref:2607.06421#pg6>. Each LAMMPS input file therefore defines one trajectory and one condition vector, ε = (ε1, ε2) <ref:2607. p. 7
Where the paper says it

The remaining long-range pair (1, 4) was kept fixed at ε3 = 0. Each LAMMPS input file therefore defines one trajectory and one condition vector, ε = (ε1, ε2).

06421#pg6>. The resulting dump files were postprocessed into internal coordinates rather than Cartesian coordinates <ref:2607.06421#pg6>.

--- Page 19 ---

19

The model learns a conditional ensemble distribution pθ(x ε1, ε2), where changes in the condition correspond to changes in the interaction parameters that define the four-bead system <ref:2607.06421#pg6>. The diffusion timestep is encoded using a 64-dimensional sinusoidal embedding, while the timestep and condition embeddings are mapped through 128-dimensional conditioning layers <ref:2607.06421#pg6>. At each diffusion step, the noisy internal-coordinate vector, the diffusion timestep, and the condition vector are provided to the network <ref:2607.06421#pg6>.

--- Page 23 ---

23

The optimization is performed in two stages. In the first stage, a high learning rate with stochastic perturbations are added to escape local minima <ref:2607. p. 8

Where the paper says it

Optimization is performed in two stages. In the first stage, the condition is optimized with a relatively large learning rate while stochastic perturbations are added to help escape local minima.

06421#pg6>. In the second stage, if the KL loss remains above a chosen threshold, the condition is further refined with a lower learning rate and without additional perturbations <ref:2607.06421#pg6>. This procedure is used to improve convergence toward a condition whose generated ensemble matches the target distribution <ref:2607.06421#pg6>.

--- Page 3 ---

3

The objective is not merely structural accuracy or the optimization of specific properties, but the controlled optimization of the ensemble distribution itself <ref:2607.06421#pg2>. GB-FESO employs a conditional diffusion model as a differentiable generator of conformational ensembles <ref:2607. p. 1

Where the paper says it

For example, the relative free-energy differences between a system’s metastable states, the overall shape of these states, or the barriers separating them. GB-FESO employs a conditional diffusion model as a differentiable generator of conformational ensembles.

06421#pg4>. The training is performed on a collection of state (conformational) ensembles spanning a diverse set of systems within the domain of interest <ref:2607. p. 1
Where the paper says it

, its physical structure, chemical composition or amino acid sequence), to an ensemble of samples that collectively define the corresponding FES. After training on a collection of state (conformational) ensembles spanning a diverse set of systems within the domain of interest, the parameters of the generative neural network are frozen.

06421#pg2>.

--- Page 4 ---

4

The optimization is driven by an ensemble-level objective rather than by agreement between individual generated and target conformations <ref:2607.06421#pg4>. This requires a quantitative, and ideally differentiable, metric that measures agreement between the generated and target ensemble distributions <ref:2607. p. 2

Where the paper says it

Thus, our goal is to match the ensemble distribution by backpropagating from a distribution-level loss, rather than from a comparison of individual samples. This requires a quantitative, and ideally differentiable, metric that measures agreement between the generated and target ensemble distributions.

06421#pg6>. A natural choice is the Kullback–Leibler (KL) divergence <ref:2607. p. 2
Where the paper says it

This requires a quantitative, and ideally differentiable, metric that measures agreement between the generated and target ensemble distributions. A natural and widely used choice is the Kullback–Leibler (KL) divergence[30]: DKL(P ∥ Q) = Z p(x) log p(x) q(x) dx.

06421#pg6>. Without an explicit analytic form for the generated distribution, the KL divergence can be approximated with Gaussian kernel density estimation (KDE) <ref:2607. p. 2
Where the paper says it

A natural and widely used choice is the Kullback–Leibler (KL) divergence[30]: DKL(P ∥ Q) = Z p(x) log p(x) q(x) dx. (6) Without an explicit analytic form for the generated distribution, the KL divergence can be approximated with Gaussian kernel density estimation (KDE)[31], and can be written as a differentiable loss: DKL(ˆptarget ∥ pˆgen) ≈ 1 N X N j=1 log pˆtarget(xj ) pˆgen(xj ) , xj ∼ pˆtarget.

06421#pg6>.

--- Page 5 ---

5

Given an input condition c, the final optimization objective can be written as c∗ = arg min c∈C h DKL(ˆptarget(x)∥ pˆgen(x)) + λDKL(ˆpDDPM(x c)∥ pˆtarget(x))i <ref:2607.06421#pg6>. The optimization is carried out by backpropagating a distribution-level loss through a deterministic diffusion sampling trajectory <ref:2607. p. 0

Where the paper says it

After training, the diffusion model is frozen, and the conditioning variables defining the system are optimized so that the generated ensemble reproduces a prescribed target free-energy surface. The optimization is carried out by backpropagating a distribution-level loss, based on kernel density estimates of the Kullback–Leibler divergence, through a deterministic diffusion sampling trajectory.

06421#pg6>.

--- Page 8 ---

8

The method successfully optimizes the interaction strengths to reproduce target free-energy surfaces in the large majority of test cases, including targets whose conditioning parameters lie outside the training region of the diffusion model <ref:2607.06421#pg4>. The results suggest that conditional diffusion models provide a practical route for ensemble-level molecular and materials design <ref:2607.06421#pg4>.

--- Page 9 ---

9

The optimization trajectory shown in Fig. p. 4

Where the paper says it

The simulations show that, for this simple test case, GB-FESO is consistently able to identify the optimal conditioning variables corresponding to the predefined target ensembles in both the continuous and discrete settings. The optimization trajectory shown in Fig.

2 illustrates several intermediate ensembles generated throughout the optimization process <ref:2607. p. 4
Where the paper says it

The optimization trajectory shown in Fig. 2 illustrates several intermediate ensembles generated throughout the optimization process.

06421#pg6>. Notably, the procedure also demonstrates the ability to extrapolate beyond the training domain, with the DDPM converging to mean and variance values that are not present in the training data while still generating the corresponding target distributions <ref:2607. p. 4
Where the paper says it

2 illustrates several intermediate ensembles generated throughout the optimization process. Notably, the procedure also demonstrates the ability to extrapolate beyond the training domain, with the DDPM converging to mean and variance values that are not present in the training data while still generating the corresponding target distributions.

06421#pg6>.

--- Page 10 ---

10

The continuous version of the experiment, where mean and variance are represented directly as continuous conditioning variables, serves as a baseline for evaluating the optimization procedure <ref:2607.06421#pg6>. The discrete formulation is important because many practical molecular or material design problems involve conditioning variables that are not naturally continuous <ref:2607. p. 4

Where the paper says it

This split is used to test whether optimization of the condition vector could move the generated ensemble between distinct regions of the learned conditional distribution. This discrete formulation is important because many practical molecular or material design problems involve conditioning variables that are not naturally continuous.

06421#pg6>.

--- Page 13 ---

13

The simulations show that, for this simple test case, GB-FESO is consistently able to identify the optimal conditioning variables corresponding to the predefined target ensembles in both the continuous and discrete settings <ref:2607.06421#pg6>. The optimization trajectory shown in Fig. 2 illustra…

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Gradient-Based Free Energy Surface Optimization (GB-FESO) can be used for ensemble-level inverse design by optimizing system design variables such as interaction energies, chemical composition or amino acid sequence so that the generated ensemble approaches a prescribed target distribution. This shifts the problem from targeting individual conformations to tuning parameters governing the full conformational ensemble.

  2. The system can be trained on a conditional diffusion model to serve as a differentiable surrogate for the ensemble distribution. This allows optimization by backpropagating through a deterministic diffusion sampling trajectory to update design variables while keeping model parameters frozen, enabling the generation of ensembles matching target free-energy surfaces.

  3. The method can be applied to engineer free-energy landscapes in physically motivated systems, such as a four-particle Lennard–Jones toy peptide, by optimizing interaction parameters like LJ pair coefficients with epsilon1 representing interactions between nearest neighbors and epsilon2 representing interactions between next-nearest neighbors. This allows the system to reproduce target free-energy landscapes in the majority of test cases.

  4. GB-FESO can handle both continuous and discrete conditioning variables; continuous variables are optimized directly, while discrete variables are handled via binary logit relaxation, such as Gumbel–Softmax relaxation, providing a differentiable route toward selecting an appropriate category.

  5. The improved AI system can perform ensemble-level molecular and materials design by steering generated ensembles toward target thermodynamic behavior and target free-energy landscapes. This is particularly useful when the desired outcome is expressed as a distribution rather than a single optimized structure, such as in catalysis or ligand recognition.

Abstract

Free-energy surfaces govern the populations of metastable states and the barriers that control transitions between them, making their direct optimization a central challenge in molecular and materials design. In this work, we introduce Gradient-Based Free Energy Surface Optimization (GB-FESO), an inverse design framework that uses a trained conditional diffusion model as a differentiable surrogate for the ensemble distribution. After training, the diffusion model is frozen, and the conditioning variables defining the system are optimized so that the generated ensemble reproduces a prescribed target free-energy surface. The optimization is carried out by backpropagating a distribution-level loss, based on kernel density estimates of the Kullback-Leibler divergence, through a deterministic diffusion sampling trajectory. We first validate GB-FESO on one-dimensional Gaussian ensembles, demonstrating that both continuous and relaxed discrete conditioning variables can be optimized to recover target distributions, including those outside the training domain. We then apply the method to a four-particle Lennard-Jones toy peptide exhibiting multiple metastable conformational states. In this more physically motivated setting, GB-FESO successfully optimizes the interaction parameters to reproduce target free-energy landscapes in the majority of test cases, with optimization performed either in the full internal-coordinate space or in a reduced collective-variable representation. These results establish GB-FESO as a promising first step toward an ensemble-level inverse design framework for molecular systems with prescribed thermodynamic and kinetic behavior.

Sources

Related papers