Research map / Probabilistic and scientific modeling / Molecules and proteins
Unlocking hidden biomolecular conformational landscapes in diffusion models at inference time
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 ConforMix, an inference-time twisted-SMC guidance method with RMSD-based biasing and MBAR reweighting that lets diffusion structure models like Boltz-1 and BioEmu reveal hidden conformations and estimate free energies without retraining.
Problem
Diffusion-based biomolecular structure models typically produce a single dominant conformation, so sampling the conformational distribution (rare states, transitions, free energies) is inefficient with default sampling.
Approach
Apply twisted diffusion sampling (SMC with time-independent guidance potentials) as an inference-time method to bias structure generation away from the default prediction via RMSD potentials, filter unphysical samples, and reweight conditional samples to the unconditional distribution using MBAR, which is applied to diffusion models for the first time.
Key results
- On domain motion proteins, ConforMixRMSD-Boltz reaches worst-matched coverage 0.69 vs 0.33 for default Boltz sampling and 0.51 for the best baseline (CF-random-Boltz) at 50% RMSD threshold.
- On membrane transporters, worst-matched coverage is 0.33 for ConforMixRMSD-Boltz vs 0.13 default; on cryptic pockets 0.45 vs 0.15 default.
- ConforMix sampling takes roughly 3x the wall-clock time of default sampling, and filtering is needed to remove unphysical samples; ConforMix free energy estimates converge faster than unbiased BioEmu sampling.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ConforMixRMSD inference-time guidance instead of strongest listed baseline sampling | Boltz-1 · Domain motion, worst-matched reference | coverage at 50%: 0.51 (±0.17) -> 0.69 (±0.15) | Helped |
| Apply ConforMixRMSD inference-time guidance instead of strongest listed baseline sampling | Boltz-1 · Membrane transporters, worst-matched reference | coverage at 50%: 0.20 (±0.20) -> 0.33 (±0.23) | Helped |
| Apply ConforMixRMSD inference-time guidance instead of strongest listed baseline sampling | Boltz-1 · Cryptic pockets, worst-matched reference | coverage at 50%: 0.39 (±0.16) -> 0.45 (±0.18) | Helped |
| Apply ConforMixRMSD inference-time guidance instead of strongest listed baseline sampling | Boltz-1 · Fold switching, worst-matched reference | coverage at 50%: 0.27 (±0.23) -> 0.13 (±0.17) | Hurt |
| Apply ConforMixRMSD inference-time guidance instead of strongest listed baseline sampling | Boltz-1 · Domain motion, best-matched reference | coverage at 50%: 0.94 (±0.07) -> 0.97 (±0.04) | Helped |
| Apply ConforMixRMSD inference-time guidance instead of strongest listed baseline sampling | Boltz-1 · Membrane transporters, best-matched reference | coverage at 50%: 0.67 (±0.23) -> 0.79 (±0.19) | Helped |
| Apply ConforMixRMSD inference-time guidance instead of strongest listed baseline sampling | Boltz-1 · Cryptic pockets, best-matched reference | coverage at 50%: 0.97 (±0.05) -> 0.94 (±0.08) | Hurt |
| Apply ConforMixRMSD inference-time guidance instead of strongest listed baseline sampling | Boltz-1 · Fold switching, best-matched reference | coverage at 50%: 0.70 (±0.30) -> 0.67 (±0.23) | Hurt |
| Apply ConforMixRMSD sampling instead of default sampling | Boltz-1 · Domain motion proteins | fraction of bond-length outliers: 2.96 \times 10^{-3} -> 9.52 \times 10^{-2} | Hurt |
| Filter ConforMixRMSD samples (ablation) | ConforMixRMSD-Boltz · Domain motion proteins | fraction of bond-length outliers: 9.52 \times 10^{-2} -> 9.12 \times 10^{-2} | Helped |
Surprising findings
- MSA-based methods outperform ConforMixRMSD on fold switchers (AFCluster worst-matched 0.27 vs ConforMixRMSD 0.13), suggesting discrete input modulation better captures such rearrangements.
- Default Boltz sampling is dominated by a single conformation, yet the variance it does have often aligns with the direction of experimental domain motion but lacks magnitude.
Limitations
- Depends on the robustness of the underlying model's energy landscape; default Boltz rarely explores alternate states without guidance.
- Non-asymptotic SMC sampling errors and MBAR statistical uncertainty may bias results.
- Elevated rates of geometric outliers versus default sampling, requiring filtering or relaxation; RNA evaluation is preliminary.
Experimental setup
Datasets: domain motion proteins (Lewis et al.; OC23 by Kalakoti et al.) TP16 membrane transporters (Xie & Huang) cryptic pocket proteins (Lewis et al.) fold switchers (Porter et al.) SemiSWEET transporter hammerhead ribozyme MD simulations BioEmu local unfolding proteins
Benchmarks: domain motion set (n=38) membrane transporters (n=15) cryptic pockets (n=31) fold switching (n=15) SemiSWEET hammerhead ribozyme trp-cage-helix free energy
Metrics: coverage at 50% of reference-to-reference RMSD AUC of coverage RMSD TM-score fill ratio PCA variance explained Ramachandran/bond outlier fractions wall-clock time
Baselines: Default Boltz sampling AFCluster AFSample2 CF-random BioEmu
Builds on: Twisted Diffusion Sampling MBAR Boltz-1 BioEmu
Closest papers by meaning
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.90)
- Protein Conformation Generation via Force-Guided SE(3) Diffusion Models (ICML 2024 · similarity 0.88)
- Learning conformational ensembles of proteins based on backbone geometry (NeurIPS 2025 · similarity 0.88)
- Consistent Sampling and Simulation: Molecular Dynamics with Energy-Based Diffusion Models (NeurIPS 2025 · similarity 0.88)
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.88)
- DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization (AAAI 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.