Research map / Probabilistic and scientific modeling / Molecules and proteins
Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression
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 ConfRover, an autoregressive model that encodes each MD frame with a frozen OpenFold encoder, models temporal dependencies with a Llama-style causal transformer, and decodes each frame with an SE(3) diffusion head, unifying trajectory simulation, time-independent sampling, and conformation interpolation.
Problem
Existing generative models for protein MD either ignore temporal dependencies between conformations or cannot directly sample time-independent conformations, and no single model handles trajectory simulation, ensemble sampling, and transition interpolation.
Approach
Factorize an MD trajectory as an autoregressive product over frames, encode frames into latents with a shared encoder, model inter-frame dependencies with a causal transformer in latent space, and generate each continuous conformation with an SE(3) diffusion decoder trained by denoising score matching; masking all context frames gives time-independent sampling and reordering frames gives interpola
Key results
- On the multi-start benchmark, ConfRover reaches Pearson 0.75 on C-alpha coordinate trajectory changes versus 0.56 for MDGen.
- In 100 ns simulation state recovery, ConfRover achieves F1 0.58 and recall 0.42 versus 0.42 and 0.29 for MDGen.
- Time-independent sampling is comparable to AlphaFlow and ConfDiff, with ConfRover best on several metrics among the three.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use autoregressive ConfRover instead of MDGen | Protein trajectory simulation · Multi-start | Cα trajectory correlation: 0.56 ± 0.03 -> 0.75 ± 0.01 | Helped |
| Use autoregressive ConfRover instead of MDGen | Protein trajectory simulation · Multi-start | PCA 2D trajectory correlation: 0.18 ± 0.01 -> 0.73 ± 0.01 | Helped |
| Use autoregressive ConfRover instead of MDGen | Protein trajectory simulation · 100 ns simulation | JSD: 0.56 ± 0.01 -> 0.51 ± 0.01 | Helped |
| Use autoregressive ConfRover instead of MDGen | Protein trajectory simulation · 100 ns simulation | Recall: 0.29 ± 0.01 -> 0.42 ± 0.00 | Helped |
| Use autoregressive ConfRover instead of MDGen | Protein trajectory simulation · 100 ns simulation | F1: 0.42 ± 0.01 -> 0.58 ± 0.00 | Helped |
| Use ConfRover rather than specialized ensemble-generation baselines | Time-independent conformation sampling · Time-independent generation | Per target RMSF r: 0.85 ± 0.01 -> 0.85 ± 0.00 | No clear effect |
| Remove the single-frame objective (ablation) | ConfRover · Time-independent generation | MD PCA W2: 1.47 ± 0.03 -> 1.43 ± 0.01 | No clear effect |
| Remove the single-frame objective (ablation) | ConfRover · Time-independent generation | Pairwise RMSD r: 0.51 ± 0.01 -> 0.48 ± 0.00 | Hurt |
| Remove the single-frame objective (ablation) | ConfRover · Time-independent generation | Per target RMSF r: 0.85 ± 0.00 -> 0.84 ± 0.01 | Hurt |
| Train with an interpolation objective and reorder conditioning frames (ablation) | ConfRover · Conformation interpolation | Directed transition toward the end state | Helped |
| Continue training with the interpolation objective (ablation) | ConfRover · Conformation quality | MolProbity score: 1.72 ± 0.38 -> 1.61 ± 0.51 | Helped |
Limitations
- Triangular updates in structural modules are computationally costly, limiting scaling to larger proteins and longer trajectories.
- Evaluation is limited to ATLAS single-chain proteins and few baselines exist for trajectory simulation and interpolation.
- Still falls short of MD 100ns oracle in state recovery and fidelity of conformational space.
Experimental setup
Datasets: ATLAS
Benchmarks: ATLAS
Metrics: Pearson correlation JSD recall F1 pairwise RMSD r RMSF r RMWD PCA W2 weak/transient/exposed contact J Ramachandran outliers
Baselines: MDGen AlphaFlow ConfDiff
Builds on: AlphaFlow ConfDiff MDGen OpenFold Llama
Closest papers by meaning
- TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles (NeurIPS 2025 · similarity 0.92)
- Generative Modeling of Molecular Dynamics Trajectories (NeurIPS 2024 · similarity 0.89)
- Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings (NeurIPS 2025 · similarity 0.88)
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.88)
- Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics (NeurIPS 2026 · similarity 0.88)
- Predicting directional flexibility in proteins (NeurIPS 2026 · 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.