Molecular Dynamics (MD) simulations are crucial for understanding the behavior of complex molecular systems over time. As MD data repositories grow, efficient similarity searching becomes increasingly important for identifying structural similarities, functional changes, recurring patterns, and molecular mechanisms across different simulations. Large-scale similarity search systems heavily rely on the use of embedding-based retrieval. While there are many embedding methods for static proteins, no robust embedding for MD exists yet.

The goal of this thesis is to propose, implement and evaluate a similarity search method using embedding methods for static proteins within molecular dynamics simulations containing proteins as the primary biomolecule.

In particular, the student will:

• Report on the current state-of-the-art embedding methods for static proteins and survey the state of available vector representations for MD simulations.
• Propose and implement a k-NN search method that can determine which simulations contain a protein of interest.
• Propose and implement a k-NN search method that can determine what are the most similar conformations in the lifetime of the simulation to the protein of interest.
• Propose and implement a method that transforms a set of static embeddings for each selected simulation frame into one “embedding” to represent the entire simulation. This method will then be used within a k-NN search system to find the most similar simulation.
• Optionally, the student will use MD tool GROMACS (gmx cluster, https://manual.gromacs.org/current/onlinehelp/gmx-cluster.html) to perform a clustering analysis to inform selection of suitable frames within the simulation.

All the proposed methods will be evaluated using domain-specific metrics for structural similarity (TM-Score) and simulation similarity (Path Similarity Analysis - PSA) and on a representative set (at least tens to a hundred, optionally more) of real-world MD simulations from MDRepo (https://mdrepo.org/). The methods will be available as Docker images and open-sourced.