Supporting Data for SelfTICA

This repository provides supporting data, trained models, and simulation inputs for the paper:

Contrastive Learning of Dynamical Representations for Collective Variable Discovery

SelfTICA is a self-supervised framework for learning slow dynamical representations from molecular simulation trajectories. The learned representations are subsequently analyzed with time-lagged independent component analysis (TICA) to extract collective variables (CVs) for enhanced sampling.

The repository contains tutorials, PLUMED interfaces, frozen TorchScript models, and simulation files for all benchmark systems reported in the paper.

---

Repository Structure

tutorials

Step-by-step tutorials for training SelfTICA collective variables, including both feed-forward neural network and dual-cutoff graph neural network frameworks.

plumed_pytorch_gnn

PLUMED interface for deploying PyTorch-based graph neural network CVs in molecular dynamics simulations.

---

tri-well

Files for the tri-well potential system using PLUMED ves_md_linearexpansion.

• models: frozen TorchScript models
• run_unbiased: input files and trajectories from unbiased simulations at different temperatures
• run_biased: input files for biased simulations at (k_B T = 0.6)

---

alanine

Files for alanine dipeptide in vacuum using GROMACS.

• data: topology files
• models: frozen TorchScript models
• run_biased_multi: input files for multithermal simulations
• run_biased_nn: biased simulations using feed-forward neural network CVs
• run_biased_gnn: biased simulations using graph neural network CVs

---

chignolin

Files for chignolin folding in explicit water using GROMACS.

• data: topology and force-field files
• models: frozen TorchScript models
• run_biased_explore: input files for OPES-Explore simulations using different CVs

---

calixarene

Files for OAMe–G2 host–guest binding in explicit water using GROMACS.

• data: topology and force-field files
• models: frozen TorchScript models
• run_unbiased: input files for unbiased simulations in both bound and unbound states
• run_biased_gnn: biased simulations using GNN-based SelfTICA CVs
• run_biased_ref: biased simulations using reference CVs, such as the coordination number (h) and (V_2)

---

fen2

Files for catalytic dissociation of (\mathrm{N_2}) on Fe(111) surfaces using LAMMPS.

• data: topology files
• models: frozen TorchScript models
• run_initial: initial biased simulations used to generate training data
• run_biased_explore: OPES-Explore simulations using different CVs

---

Code and Tutorials

The collective variables used in this work were trained using the mlcolvar library, a Python framework for developing machine-learning collective variables for enhanced sampling simulations.

A frozen code archive, mlcolvar-selftica.zip, is provided in this repository to reproduce the SelfTICA training workflows reported in the paper. This archive contains the version of mlcolvar used for training SelfTICA CVs in the benchmark systems.

The tutorials directory provides step-by-step examples for training SelfTICA collective variables, including both feed-forward neural network and dual-cutoff graph neural network frameworks.

The plumed_pytorch_gnn directory provides the interface required to deploy PyTorch-based GNN collective variables in PLUMED-enhanced molecular dynamics simulations.

---