timewarp / 2AA-complete /README.md
andrewfoong's picture
Upload folder using huggingface_hub
82d923d verified

2AA-1-complete "Two Amino Acid" data set

This folder contains a data set of all-atom molecular dynamics trajectories for all 400 dipeptides, i.e. small proteins composed of two amino acids. This includes also the peptides missing in the other 2AA datasets. Each peptide is simulated using classical molecular dynamics and the water is simulated using an implicit water model.

For each protein two files are available:

  • protein-state0.pdb: contains the topology and initial 3D XYZ coordinates.
  • protein-arrays.npz: contains trajectory information.

NPZ Information

The NPZ file contains detailed information for a subset of simulation steps. There are T such frames and the NPZ file contains the following arrays:

  • 'time': (T,) array, simulation time in picoseconds.
  • 'energies': (T,2) array, each row containing [potential, kinetic] energies in kJ/mol.
  • 'positions': (T,num_atoms,3) array, positions in nm.
  • 'velocities': (T,num_atoms,3) array, velocities in nm/ps.
  • 'forces': (T,num_atoms,3) array, forces in kJ/(mol nm).

Dataset construction

The dataset was constructed in the following way:

  1. Construct initial protein using AmberTools's tleap program.
  2. Process each PDB file using pdbfixer, adding all missing hydrogen atoms. This is the state0.pdb file.
  3. For each all-atom PDB file, perform a molecular dynamics simulation: a.) Use OpenMM with the AMBER99 force field and implicit water model. b.) Perform an energy minimization (relaxation) from the initial extended configuration. c.) Use a second-order Langevin integrator at temperature to T=310K, friction=0.3/ps, timestep=1.0fs for 50,000 steps to equilibriate ("burn-in phase"). d.) Use a second-order Langevin integrator at temperature to T=310K, friction=0.3/ps, timestep=1.0fs for 5e6 steps to sample a trajectory ("sample phase").
  4. Save trajectory information to an arrays.npz file.
  5. Shuffle and partition the trajectories into a train set (200 trajectories), val (80 trajectories), and test (100 trajectories).

Credit and Authors

This dataset was created in October 2022 as part of Timewarp.