Elign Achieves Faster 3D Molecular Modelling Via Foundational Machine-Learning Force Fields

Researchers are tackling the challenge of generating realistic 3D molecular conformations, ensuring these models adhere to fundamental physical principles and produce stable structures. Yunyang Li, Lin Huang (from IQuestLab), and Luojia Xia, along with colleagues, present Elign, a novel post-training framework designed to overcome computational limitations in existing equivariant diffusion models. This work is significant because it amortises the costs of both expensive quantum mechanical calculations and repeated energy evaluations by leveraging pretrained machine-learning force fields and a new reinforcement learning approach, Force, Energy Disentangled Group Relative Policy Optimization (FED-GRPO). Consequently, Elign generates conformations with improved energies and stability, all while maintaining the speed of standard sampling methods.

Elign corrects molecular conformation biases with MLFFs

Scientists have developed Elign, a new post-training framework designed to enhance the generation of three-dimensional molecular conformations by addressing limitations in existing Diffusion models. These models, while capable of respecting Euclidean symmetries, often replicate biases present in their training data rather than accurately capturing the equilibrium distribution dictated by a high-fidelity Hamiltonian. The research team tackled this challenge by amortizing two significant computational costs: expensive quantum-chemical evaluations, such as Density Functional Theory (DFT), and the need for repeated energy queries during each sampling step. Elign achieves this by initially substituting DFT evaluations with a faster, pretrained foundational machine-learning force field (MLFF) to provide physical guidance.

Crucially, the team eliminated repeated run-time queries by shifting physical steering to the training phase, a novel approach to optimization. FED-GRPO incorporates both a potential-based energy reward and a force-based stability reward, optimizing and group-normalizing these independently to ensure accurate and stable molecular conformations. Experiments demonstrate that Elign generates conformations exhibiting lower gold-standard DFT energies and forces, alongside improved overall stability, representing a significant advancement in molecular modelling.

The core innovation of Elign lies in its ability to maintain the speed of unguided sampling during inference, as no energy evaluations are required at the generation stage. This is achieved through the amortization of computational costs, first by leveraging the efficiency of MLFFs in place of DFT calculations and then by embedding physical constraints directly into the diffusion model via post-training. By framing reward-guided diffusion as a stochastic optimal control problem, the researchers circumvent the need for differentiability assumptions that often plague molecular diffusion models with mixed discrete and continuous variables. This allows for more robust gradient estimation and improved performance.

Furthermore, the study builds upon recent advances in “foundation” MLFFs, trained on large and chemically diverse datasets, enabling accurate potential energy surface approximations across a wide range of molecular classes. These foundational MLFFs, combined with the FED-GRPO optimization strategy, allow Elign to generate physically plausible conformations with enhanced efficiency and accuracy. The work opens avenues for applications in computational chemistry, materials science, and drug discovery, where the generation of realistic and stable molecular structures is paramount, and promises to accelerate the design of novel compounds and materials.

Scientists Method

Scientists introduced Elign, a post-training framework designed to address computational bottlenecks in generative models for 3D molecular conformations. The research team focused on amortizing the costs associated with both expensive density functional theory (DFT) evaluations and the need for repeated energy queries during sampling. To achieve this, they replaced DFT calculations with a faster, pretrained foundational machine-learning force field (MLFF) to provide physical signals, effectively transferring the computational burden to the training phase. This initial step represents the first level of amortization within the Elign framework.

The study pioneered a novel approach by formulating reverse diffusion as a reinforcement learning problem, enabling the fine-tuning of the denoising policy. FED-GRPO leverages the inherent connection between MLFFs and diffusion models, building upon the effectiveness of the denoising objective used in both MLFF pretraining and diffusion model training. Experiments employed the QM9 and GEOM-Drugs datasets to validate the performance of Elign.

Researchers harnessed a foundational MLFF, trained on large, chemically diverse quantum-mechanical datasets, to approximate potential energy surfaces across a wide range of molecular classes. This MLFF provides accurate and efficient energy and force evaluations, crucial for guiding the diffusion process without incurring significant computational overhead. The team engineered a system where physical constraints are compiled into the diffusion model itself during a post-training stage, shifting computation away from inference and eliminating run-time costs. This second level of amortization ensures that inference remains as fast as unguided sampling, as no energy evaluations are required during generation.

The approach achieves lower gold-standard DFT energies and forces in generated conformations, while simultaneously improving their stability. The system delivers a significant advancement over existing methods that rely on runtime alignment or post-processing, which often introduce additional computational demands. By disentangling energy and force rewards and optimizing them with FED-GRPO, the study reveals a pathway to generating high-fidelity molecular conformations with unprecedented efficiency and accuracy. This technique enables the creation of physically grounded simulation pipelines for Molecular dynamics and equilibrium sampling.

Elign lowers energy and boosts stability

Scientists achieved a significant breakthrough in generative 3D molecular conformation modeling by introducing Elign, a post-training framework that substantially reduces computational costs. The work addresses the challenge of balancing Euclidean symmetries with the need for mechanically stable molecular structures, often hampered by biases inherited from semi-empirical training data. Experiments demonstrate that Elign generates conformations with demonstrably lower gold-standard DFT energies and forces, while simultaneously improving overall stability. Crucially, inference speed remains comparable to unguided sampling, as no energy evaluations are required during the generation process.

The team measured a substantial reduction in computational burden by replacing expensive Density Functional Theory (DFT) evaluations with a faster, pretrained machine-learning force field (MLFF) to provide physical signals. This substitution allows for physical guidance without the typical computational bottlenecks associated with high-fidelity Hamiltonian calculations. Furthermore, Elign eliminates repeated run-time queries by shifting physical steering to the training phase, a process achieved through formulating reverse diffusion as a reinforcement learning problem. These rewards were optimized and group-normalized independently, leading to improved performance. Measurements confirm that the energy-based potential shaping, utilizing the MLFF energy as a proxy for thermodynamic stability, facilitates credit assignment over long reverse diffusion horizons. The intermediate shaping reward is defined as rshape t := γ Ψ(St−1) −Ψ(St), where Ψ(St) represents the potential energy evaluated on reconstructed geometries.

Tests prove that the alignment objective encourages a terminal distribution that prioritizes physically valid structures. Specifically, the team derived a theoretical result demonstrating that the optimal terminal distribution, ρ⋆, is proportional to the prior distribution ρθpre, weighted by an exponential of the negative effective energy, as defined in Equation 5. The normalizer, Zφ, ensures proper probabilistic weighting. This result highlights the framework’s ability to generate conformations that align with both thermodynamic stability and mechanical equilibrium, as evidenced by the lower DFT energies and forces recorded in the experiments.

Elign stabilises molecules via reinforcement learning

Scientists have developed Elign, a new framework designed to improve the generation of 3D molecular conformations by diffusion models. The research addresses a key challenge in this field: ensuring generated structures adhere to physical principles and represent stable, realistic molecules. Elign achieves this by combining a pretrained machine-learning force field with a reinforcement learning approach to fine-tune the diffusion model’s denoising policy. This post-training framework amortizes computational costs by replacing expensive quantum chemical evaluations with faster machine-learned predictions and by shifting physical guidance from the sampling phase to the training phase.

Experiments demonstrate that Elign generates conformations with improved thermodynamic stability and mechanical equilibrium, as measured by gold-standard DFT energies and forces, while maintaining the speed of unguided sampling. The authors report achieving competitive results on the GEOM-drug dataset, demonstrating improvements in both atom stability and RDKit validity compared to existing methods. The authors acknowledge that further work is needed to improve alignment when using single-objective feedback, such as energy or force alone, and to explore performance with smaller preference models. Future research will focus on these areas to enhance the robustness and efficiency of the Elign framework. This work represents a step towards generating more reliable and physically plausible molecular conformations, which is crucial for applications in drug discovery and materials science.