Research map / Probabilistic and scientific modeling / Molecules and proteins
MarS-FM: Generative Modeling of Molecular Dynamics via Markov State Models
Kacper Kapuśniak, Cristian Gabellini, Michael M. Bronstein, Prudencio Tossou, Francesco Di Giovanni
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on MSM state-to-state transitions instead of fixed-lag frame pairs (MarS-FM vs MDGen) | MD emulator · MD-CATH 450K, 500 samples | Gyration radius KL: 1.13 -> 0.55 (vs MDGen-100) | Improved |
| Use MarS-FM instead of MDGen on MD-CATH folding free energy | MD emulator · MD-CATH 450K, 100 samples | Folding free energy MAE: 1.52 -> 1.25 (vs MDGen-20) | Improved |
| Sample MDGen in parallel to reduce autoregressive calls (ablation) | MDGen · MD-CATH 450K, 500 samples | Gyration radius KL: 1.13 -> 2.15 (MDGen-100) | Hurt |
| Shorten MDGen window from K=100 to K=20 (ablation) | MDGen · MD-CATH 450K, 500 samples | MSM JSD: 0.29 -> 0.43 | Hurt |
| Combine MarS-FM with MDGen-20 instead of MarS-FM alone (ablation) | MarS-FM · MD-CATH 450K, 500 samples | Gyration radius KL: 0.55 -> 0.98; secondary structure KL 0.93 -> 0.73 | Mixed |
| Reduce sample budget from 1000 to 100 conformations (ablation) | MarS-FM · MD-CATH 450K | Folding free energy MAE: 1.20 (1000) vs 1.25 (100); oracle 0.80 vs 2.40 | Mixed |
| Use MarS-FM instead of MDGen on tetrapeptides | MD emulator · Tetrapeptides | Macrostate MAE: 1.13 -> 0.63 | Improved |
| Use MarS-FM on tetrapeptides for torsion and TICA JSD | MD emulator · Tetrapeptides | TICA-0 JSD: 0.23 -> 0.21 | No clear effect |
| Use MarS-FM instead of MDGen for sampling speed | MD emulator · 159-residue protein, 500 conformations | Time (s): 100.87 -> 30.34 (vs MDGen-20) | Improved |
Closest papers by meaning
- Generative Modeling of Molecular Dynamics Trajectories (NeurIPS 2024 · similarity 0.92)
- Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics (NeurIPS 2026 · similarity 0.90)
- Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression (NeurIPS 2025 · similarity 0.90)
- Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional (ICML 2025 · similarity 0.90)
- TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles (NeurIPS 2025 · similarity 0.89)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 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.
