[Paper Review] Protein Conformation Generation via Force-Guided SE(3) Diffusion Models

Yan Wang, Lihao Wang+5 more|arXiv (Cornell University)|Mar 21, 2024

Protein Structure and Dynamics12 citations

TL;DR

ConfDiff uses SE(3) diffusion with sequence-conditioned guidance and a force-guided intermediate term to generate diverse, high-fidelity protein conformations that align with Boltzmann distribution, outperforming state-of-the-art baselines on fast-folding proteins and BPTI.

ABSTRACT

The conformational landscape of proteins is crucial to understanding their functionality in complex biological processes. Traditional physics-based computational methods, such as molecular dynamics (MD) simulations, suffer from rare event sampling and long equilibration time problems, hindering their applications in general protein systems. Recently, deep generative modeling techniques, especially diffusion models, have been employed to generate novel protein conformations. However, existing score-based diffusion methods cannot properly incorporate important physical prior knowledge to guide the generation process, causing large deviations in the sampled protein conformations from the equilibrium distribution. In this paper, to overcome these limitations, we propose a force-guided SE(3) diffusion model, ConfDiff, for protein conformation generation. By incorporating a force-guided network with a mixture of data-based score models, ConfDiff can generate protein conformations with rich diversity while preserving high fidelity. Experiments on a variety of protein conformation prediction tasks, including 12 fast-folding proteins and the Bovine Pancreatic Trypsin Inhibitor (BPTI), demonstrate that our method surpasses the state-of-the-art method.

Motivation & Objective

• Motivate the need for efficient sampling of protein conformational landscapes beyond traditional MD and single-structure predictions.
• Develop a diffusion-based generative framework that respects physical priors from MD energy to sample low-energy conformations.
• Leverage sequence-conditioned guidance and force-based intermediate guidance to balance conformation quality and diversity.
• Evaluate on fast-folding proteins and BPTI to demonstrate improved fidelity and Boltzmann-consistent sampling.

Proposed method

• Represent protein backbones as SE(3) frames per residue and perform diffusion on translations and rotations with separate SE(3) dynamics.
• Train a baseline unconditional score model and a sequence-conditional score model, combining them via classifier-free guidance during sampling.
• Introduce energy guidance using a mid-diffusion intermediate energy function E_t and train a network to approximate it via CEP-style loss.
• Develop an intermediate force guidance network to approximate the intermediate force E_t' and inject it into reverse-time sampling to bias toward lower energy states.
• Infer force guidance in the translational components only, with an interpolation form h_ψ(x_t,t) to stabilize training and satisfy boundary conditions.
• Note: Training uses DSM loss; energy/force networks are trained separately from the base score model.

Experimental results

Research questions

• RQ1 Can force- and energy-guided diffusion on SE(3) produce protein conformations that are both diverse and faithful to the Boltzmann distribution?
• RQ2 How does classifier-free sequence conditioning affect the balance between conformational diversity and fidelity?
• RQ3 Do intermediate energy and force guidance improve sampling quality compared with purely data-driven diffusion baselines?
• RQ4 How do the proposed ConfDiff variants compare to state-of-the-art diffusion models on standard protein conformation benchmarks?

Key findings

Models	JS distance (down)	Val-CA (up)	RMSE contact (down)	RMSF	PwD (Å)	Rg (Å)	TIC	TIC-2D
EigenFold	0.53/0.56	0.52/0.55	0.50/0.50	0.64/0.66	0.15/0.08	6.18/6.22	1.6/1.1
Str2Str-SDE	0.34/0.32	0.30/0.24	0.39/0.38	0.56/0.58	0.97/0.98	3.68/4.01	7.8/8.0
Str2Str-ODE	0.37/0.38	0.33/0.30	0.40/0.39	0.57/0.59	0.96/0.97	4.14/4.36	6.4/6.3
ConfDiff-Base	0.29/0.27	0.25/0.22	0.36/0.37	0.52/0.52	0.89/0.91	3.61/3.57	6.1/5.9
ConfDiff-Energy	0.34/0.34	0.31/0.29	0.39/0.40	0.54/0.56	0.97/0.97	3.65/3.80	7.1/6.1
ConfDiff-Force	0.29/0.27	0.26/0.24	0.38/0.38	0.54/0.54	0.97/0.98	3.25/3.38	6.2/5.7

• ConfDiff with force guidance achieves competitive or improved metrics compared with baselines, showing reduced energy and maintained diversity.
• On fast-folding proteins, ConfDiff-Force and ConfDiff-Base achieve lower or comparable JS distances and higher Val-CA accuracy relative to EigenFold and Str2Str variants.
• Energy guidance and force guidance both steer samples toward lower energy conformations, enhancing Boltzmann-consistent sampling.
• Classifier-free guidance provides a tunable trade-off between sample quality and diversity, enabling better control over generated ensembles.
• Across benchmarks (including BPTI), the force-guided approach demonstrates improvements in structural validity and distributional similarity to MD-derived ensembles.