Autoencoder Collective Variables from MD

Train an autoencoder on an apo MD trajectory, compare its latent space with the holo trajectory, check reconstruction by RMSF, and export the model as PLUMED collective variables for enhanced sampling.

by Temple Compute

Run Autoencoder Collective Variables from MD on Horus

W-25 · AutoEncoders for MD Analysis

Domain: BioExcel Building Blocks

Overview

Trains an autoencoder (biobb_pytorch) on an apo MD trajectory from MDDB (default: PDB 6W9C, MCV1900210), compares its latent space with a holo trajectory of the same protein (default MCV1900212), checks reconstruction quality by RMSF, and exports the model as a PLUMED collective variable. Both trajectories are fitted and featurized as Calpha Cartesian coordinates. The workflow has 16 tasks.

Pipeline

fetch_train_trajectory -> fit_train_trajectory -> featurize_train_trajectory    Apo trajectory
build_model -> train_model                                                      Autoencoder, trained on apo
fetch_test_trajectory -> fit_test_trajectory -> featurize_test_trajectory       Holo trajectory
evaluate_model -> reconstruct_trajectory                                        Latent space, reconstructed holo
make_ndx_apo / make_ndx_holo                                                    Calpha index files
gmx_rmsf_apo / gmx_rmsf_holo / gmx_rmsf_holo_recon                              RMSF profiles
make_plumed                                                                     PLUMED CV and bias input

Inputs / Outputs

Inputs are two MDDB projects of the same protein: the apo (training) and holo (test) trajectories. The autoencoder hidden layers are fixed (encoder 933, 747, 561, 375, 189; decoder mirrored); the input and output layers follow the protein automatically (3 x Calpha atoms of the featurized apo trajectory). Outputs (results/):

{apo,holo}.pdb/.xtc, *_fit.xtc, *.pt, *_stats.pt: raw, fitted and featurized trajectories
model.pth, apo_trained_model.pth, model_training_metrics.npz: untrained and trained model, training curves
apo_eval_results.npz: evaluation and latent-space results
holo_recon.xtc/.pdb: reconstructed holo trajectory
{apo,holo}.ndx, *_rmsf.xvg, holo_recon_rmsf.xvg: index files and RMSF profiles
plumed_model.ptc, plumed.dat, features.dat: PLUMED collective-variable model and input files

Parameters

| id | Default | Meaning | |---|---|---| | train_mddb_project | MCV1900210 | MDDB project of the apo trajectory the autoencoder is trained on. | | test_mddb_project | MCV1900212 | MDDB project of the holo trajectory projected onto the latent space; same protein as the training project. | | n_frames | 10000 | Frames downloaded per trajectory, from frame 1 (100-10000). | | n_cvs | 3 | Latent space size, i.e. number of collective variables (1-10). | | max_epochs | 100 | Maximum training epochs (1-2000). | | batch_size | 128 | Training batch size (8-4096). |