Research map / Image generation / Autoregressive and latent generation

Protein Conformation Generation via Force-Guided SE(3) Diffusion Models

ICML 2024 · Image generation › Autoregressive and latent generation · approach: Diffusion for protein and antibody design · 12 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ConfDiff, a SE(3) protein backbone diffusion model with sequence-based classifier-free guidance and an intermediate force guidance network derived from MD (OpenMM) force fields, to generate Boltzmann-like protein conformations without MD training data.

Problem

Diffusion models for protein conformation generation do not incorporate physical prior knowledge (MD force fields), so sampled conformations deviate from the equilibrium Boltzmann distribution and can be high-energy.

Approach

Combine classifier-free guidance (sequence-conditional vs unconditional SE(3) score models) with an intermediate force guidance network that approximates the force of the marginal energy E_t, derived from the MD energy function (OpenMM), added to the reverse SDE score; trained without MD trajectories.

Key results

• On fast-folding proteins, ConfDiff-Force achieves PwD JS 0.29/0.27 and RMSE contact 3.25/3.38 versus EigenFold 0.53/0.56 and 6.18/6.22, and Str2Str-SDE 0.34/0.32.
• On BPTI, ConfDiff-Force attains the best RMSDavg 1.41 Å and RMSDcls3 2.12 Å, versus EigenFold 1.47 and 2.45 Å.
• ConfDiff-Force adds only negligible inference time over ConfDiff-Base (e.g., 104 s vs 71 s for 58-residue protein on a V100), while energy guidance nearly doubles memory.

Measured effects reported in the paper

Surprising findings

• Energy guidance (ConfDiff-Energy) did not improve over the base model on BPTI (RMSDavg 1.57 vs 1.55), while force guidance did.

Limitations

• Full-atom MD energy evaluation via OpenMM is computationally expensive, so a surrogate force network is needed.
• Models conditioned on sequence tend to predict folded structures and lack conformational diversity, limited by PDB diversity.
• Force guidance applied only to C-alpha translational components.

Experimental setup

Datasets: PDB single-chain structures fast-folding proteins MD trajectories (12 proteins) BPTI MD trajectories

Benchmarks: fast-folding proteins (12 proteins) BPTI metastable states

Metrics: JS distance (PwD, Rg, TIC, TIC-2D) Val-CA RMSE contact RMSF RMSD to cluster centers

Baselines: EigenFold Str2Str-SDE Str2Str-ODE ConfDiff-Energy

Builds on: Str2Str FrameDiff contrastive energy prediction classifier-free guidance ESMFold

Closest papers by meaning

• SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.89)
• Structure Language Models for Protein Conformation Generation (ICLR 2025 · similarity 0.89)
• Unlocking hidden biomolecular conformational landscapes in diffusion models at inference time (NeurIPS 2025 · similarity 0.88)
• Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.88)
• DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization (AAAI 2024 · similarity 0.88)
• Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.88)