DyME: A New Engine Automates Thousands of Protein Simulations for Molecular Design

Proteins are picky conversationalists. They recognize partners through a few interfacial amino acids whose identities, orientations, and surrounding water molecules decide whether two molecules bind tightly, weakly, or not at all. Researchers have long wanted to tweak these molecular handshakes deliberately—strengthening an antibody’s grip on a target, steering a peptide ligand toward one kinase and away from its close relatives, or designing mimetics that fool a biological interaction. The physics needed to do this properly is known: molecular dynamics (MD) simulations track every atom and reveal how mutations reshape binding at atomic detail. The problem has been scale. Each mutation set demands its own prepared, solvated, simulated, and analyzed system, and comparing results across dozens or hundreds of runs has meant weeks of manual labor. A team at TU Dresden’s Structural Bioinformatics group, led by Pedro M. Guillem-Gloria, Gloria Ruiz-Gómez, and M. Teresa Pisabarro, now reports in PLOS Computational Biology an integrated platform called DyME—Dynamic Mutagenesis Engine—that automates the entire pipeline, from generating thousands of mutant structures to delivering interactive comparative analysis through a web browser.

The core insight behind DyME is that high-throughput mutagenesis has outpaced the software built around it. Existing automation tools each cover only fragments of the workflow. Earlier pipelines such as one described by Chiappori and colleagues in 2009, and the more recent MutateX, probe saturation mutagenesis using energy-minimized structures and the FoldX force field—but they skip dynamics entirely. MDFit automates MD simulations through Desmond but requires a manually pre-built ligand library and performs no automated mutagenesis. StreaMD can launch simulations from existing input structures and assist post-processing with GROMACS, yet it too leaves large gaps. Static structure-based methods, the authors note, struggle to capture the conformational complexity that governs biomolecular recognition, which is why MD remains the gold standard, especially now that GPUs make large macromolecular simulations fast enough to imagine running them by the thousand. DyME was built to fill the void: a single, cohesive platform covering preparation, mutagenesis, simulation, feature extraction, storage, and comparative exploration of up to several thousand molecular systems.

Technically, DyME is a distributed system whose architecture resembles a well-organized factory. A user begins by uploading the 3D coordinates of a protein–protein or protein–DNA complex, prepared with pdb4amber to enforce compatibility with the Amber file conventions that standardize the platform. A guided web wizard, reminiscent of OpenMM-Setup, then collects everything the simulation engine needs: solvation model, ionic neutralization strategy, periodic boundary conditions, temperature, statistical ensemble, step size, equilibration and production lengths. The user designates one molecule in the complex as the “mutable object” and selects “anchor points”—interfacial residues whose alteration will be probed. DyME analyzes inter-molecular contacts and proposes a tentative list of anchor points, which the user can freely edit. Because exploration at multiple positions can explode combinatorially, the platform offers “anchor point clustering,” grouping nearby or functionally related positions and enumerating all singlet, doublet, and triplet permutations within each cluster. Before launching, DyME estimates the total number of mutant permutations and the required GPU-hours, giving researchers a realistic budget at a glance.

Once a project is defined, the platform’s producer-consumer machinery takes over. MD nodes deployed on GPU servers periodically poll a central MongoDB database for mutant records marked “pending.” When an idle GPU card and a waiting mutant coincide, the node builds the mutant’s 3D structure with Modeller, treating each mutation as a small homology-modeling problem that resolves steric clashes, refines rotamers, adjusts the backbone, and minimizes energy. An embedded template generator then drives Amber’s tleap tool to solvate, neutralize, and produce the topology and coordinate files OpenMM requires. Notably, DyME splits topologies into separate files for the mutable and non-mutable objects with atom counts kept consistent across every mutant—a seemingly mundane detail that is essential for valid per-residue energy decomposition later. Simulations run asynchronously, one per GPU, with a Langevin middle integrator by default; status codes written back to the database prevent race conditions and let other nodes pick up completed work immediately.

The third phase belongs to the “Scavenger nodes.” For every finished trajectory, they extract a battery of features automatically: root-mean-square deviations, binding free energies computed with MMPBSA.py, pairwise and per-residue energy decompositions, inter-molecular contact frequencies, and—unusually for tools in this space—solvent information. A dedicated water-site mapping module characterizes the water molecules that linger at the binding interface, cataloguing which protein atoms they contact and how long they reside there. Interfacial waters are far more than passive fillers; they mediate recognition and can determine specificity, yet they are notoriously difficult to track across large simulation sets. All scavenged data flows into the central database, where MongoDB’s native aggregation pipelines perform correlational queries across millions of records at speeds that make comparing hundreds of simulations genuinely interactive rather than an overnight batch job.

The front end—DyME’s Toolbox for Comparative Analysis—is where the platform becomes a design instrument. The Mutant Explorer presents an interactive ranked table of every simulated variant, color-graded from most favorable to least favorable binding free energy relative to the wild type, alongside statistical summaries including percentiles of the energy distribution. Double-clicking any mutant loads its data into context-aware widgets: a 3D explorer that superposes mutant structures on the wild type and renders interfacial water-sites as spheres sized by residence time, an RMSD plot for stability checks, per-residue energy bar charts at anchor points, and a pairwise contact heat map that reveals mutant-specific interactions at a glance. An Interactive Energy Explorer can even display energetic contributions from the perspective of either binding partner, exposing whether a mutation creates new interaction sites on the counterpart molecule. Filters allow users to demand combinations of up to three point mutations; if no simulated mutant matches the criteria, DyME can add the missing combination to the library and queue it for simulation on the fly.

Two widgets deserve special mention because they target the practical goals of engineering. The Water Site Explorer, in one click, lists interfacial water-sites, the residues and atoms they bridge, and the combined residence times of the waters involved—turning an analysis that typically demands custom scripting into an exploratory click-through. The Specificity Finder addresses the classic trade-off in molecular recognition: designing a binder that is stronger for one receptor but weaker for another. It compares two DyME projects that share the same mutable ligand but differ in receptor, and it surfaces up to fifty mutants common to both, ranked by the differential effect. In the published demonstration, a triplet mutation (M4Y, P5S, L8P) raised predicted affinity for one protein receptor from −27.5 to −31.8 kilocalories per mole while simultaneously weakening it for another from −26 to −16.8—an unambiguous specificity switch identified without manual data wrangling.

To validate the platform, the team turned to a classic benchmark in protein mimicry: the recognition of proline-rich peptides by SH3 domains, small signaling modules that bind poly-proline II motifs. The peptide 3 bp-1 (APTMPPPLPP) binds the SH3 domains of the Abl and Fyn tyrosine kinases with identical low affinity (a dissociation constant of 34 micromolar), and earlier structure-based design work had produced the peptide p41 (APSYSPPPPP), which showed a twenty-fold affinity increase for Abl-SH3 and a ten-fold decrease for Fyn-SH3. Replicating that mutagenesis rationale in DyME—defining every position of 3 bp-1 as an anchor point—the simulated binding free energies correlated with the experimentally reported Kd values, and the water-site mapping recapitulated water-mediated interactions that had been identified experimentally at the Abl and Fyn SH3 interfaces. The test data, including raw simulation inputs and outputs, has been deposited on Zenodo for community scrutiny.

DyME is open source, distributed under Docker and Apptainer containers, and scales horizontally by adding worker nodes to whatever computational infrastructure a lab already possesses. It even accepts non-standard molecular constituents such as synthetic amino acids through custom Amber libraries, widening its reach into chemical biology and peptidomimetic design. The authors frame it as a foundation for something larger: coupling large-scale MD with machine learning to rapidly identify and predict exploitable recognition features. If the bottleneck in computational protein engineering has been the fragmentation of tools across every step of a simulation campaign, DyME’s message is that the workflow can now be treated as a single, queryable dataset—one in which the question “which mutation gives me the binding profile I want?” becomes an interactive query rather than a research project of its own.

Subject of Research: A molecular dynamics software platform for high-throughput mutagenesis and protein recognition engineering

Article Title: DyME: An MD-based engine exploiting HTP mutagenesis for protein engineering and recognition mimicry

Article References: Guillem-Gloria, P. M., Ruiz-Gómez, G., & Pisabarro, M. T. (2026). DyME: An MD-based engine exploiting HTP mutagenesis for protein engineering and recognition mimicry. PLOS Computational Biology, 22(10), e1014221. https://doi.org/10.1371/journal.pcbi.1014221

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014221

Keywords: molecular dynamics, protein engineering, mutagenesis, protein-protein interactions, binding free energy, open-source software, high-throughput screening, SH3 domains, peptide design, interfacial water, molecular recognition, computational biology