nileshsarkar-ai/protein-folding-experiments

Protein Folding Experiments

Molecular-dynamics trajectories, steered-MD pulling data, and ESMFold structure predictions from the Nilesh Sarkar — Protein Folding Lab.

Companion code: github.com/nileshsarkar-ai/ProteinFolding

This dataset holds the large binary artifacts (trajectory .dcd files, serialized OpenMM System/State XMLs) that are too big for the GitHub repo. The GitHub repo keeps all code, figures, logs, CSVs, and small structures, plus .hfptr pointer files that reference the files here.

---

Contents

fc_bg_tier1/                         Force-Conditioned Boltzmann Generator — Tier 1
│                                    Titin I27 (PDB 1TIT) mechanical unfolding via SMD
├── prepared/
│   ├── system.xml                   OpenMM System (forces, masses) — solvated I27, 85,750 atoms
│   ├── state.xml                    Equilibrated positions + velocities + box
│   └── topology.pdb                 Topology of the solvated system
├── smd_run01/
│   ├── traj.dcd                     SMD trajectory, pulling velocity 0.01 nm/ps  (294 MB)
│   └── final_state.xml              Final OpenMM state
├── smd_vfast/
│   ├── traj.dcd                     SMD trajectory, pulling velocity 0.1  nm/ps  (30 MB)
│   └── final_state.xml
└── smd_vslow/
    ├── traj.dcd                     SMD trajectory, pulling velocity 0.001 nm/ps (2.9 GB)
    └── final_state.xml

(Chignolin equilibrium-MD trajectory is added once the 1 µs run completes.)

---

Experiment 1 — FC-BG Tier 1: titin I27 mechanical unfolding

Constant-velocity steered molecular dynamics (SMD) on the 27th immunoglobulin domain of titin (I27 / I91), the canonical AFM force-spectroscopy protein.

• System: I27 (PDB 1TIT, Improta et al., Structure 1996), solvated in TIP3P with 0.15 M NaCl, AMBER ff14SB, 85,750 atoms, elongated 7×7×18 nm box.
• Method: harmonic spring between N-terminal and C-terminal Cα, reference position moved at constant velocity. Spring stiffness 5000 kJ/mol/nm².
• Velocities: 0.1, 0.01, 0.001 nm/ps (smd_vfast, smd_run01, smd_vslow).
• Engine: OpenMM 8.5, OpenCL platform, NVIDIA A100.

Key result: peak unfolding force 2316 pN at r = 9.04 nm (Δ = 24 Å above native) for smd_run01. This is 12× the AFM value (200 pN; Carrion-Vazquez et al., PNAS 1999) — consistent with Bell-Evans loading-rate inflation, since SMD pulls ~7 orders of magnitude faster than an AFM cantilever.

Caveat: the trajectories are non-equilibrium pulling runs. Recovering the equilibrium free-energy surface F(L) requires Jarzynski averaging over an ensemble of trajectories at fixed velocity — a single trajectory per velocity is dissipation-dominated. See REVIEWER_NOTES.md in the code repo.

Experiment 2 — ESMFold benchmark + PISCES50 calibration

(Predictions & scores are small — they live in the GitHub repo, not here.) ESMFold sequence→structure prediction on 8 small proteins and a 50-protein PISCES-style calibration sweep. Headline: ESMFold per-residue confidence (pLDDT) correlates with TM-score accuracy at Pearson r = 0.71 (n = 40).

Experiment 3 — Chignolin equilibrium folding

1 µs of explicit-solvent MD on CLN025 chignolin at 340 K, tracking native contact fraction Q, RMSD, and radius of gyration. Trajectory added on completion.

---

Usage

from huggingface_hub import hf_hub_download
import mdtraj as md

dcd = hf_hub_download(
    repo_id="nileshsarkar-ai/protein-folding-experiments",
    filename="fc_bg_tier1/smd_run01/traj.dcd",
    repo_type="dataset",
)
top = hf_hub_download(
    repo_id="nileshsarkar-ai/protein-folding-experiments",
    filename="fc_bg_tier1/prepared/topology.pdb",
    repo_type="dataset",
)
traj = md.load(dcd, top=top)
print(traj)   # 300 frames of I27 being pulled apart

Or the whole dataset:

huggingface-cli download nileshsarkar-ai/protein-folding-experiments \
    --repo-type dataset --local-dir ./protein-folding-experiments

---

Citation / provenance

Generated by the Nilesh Sarkar Protein Folding Lab, May 2026, on a single NVIDIA A100. Force fields: AMBER ff14SB + TIP3P. Structure prediction: ESMFold v1 (Lin et al., Science 2023). MD engine: OpenMM 8.5.

License

Proprietary — all rights reserved. See the LICENSE file in the code repository.