CoChat
Sign Up
Cited by 124Protein Structure and DynamicsAdvanced Physical and Chemical Molecular InteractionsComputational Drug Discovery Methods

The DOI matches its Crossref record, with a small difference in title or year; 103 of the 104 references it cites were found in a registry or an index.

Martinize2 and Vermouth provide a unified framework for molecular topology generation

Open access
Kroon, Peter C, Grünewald, Fabian, Barnoud, Jonathan, van Tilburg, Marco, Brasnett, Chris, Souza, Paulo CT, Wassenaar, Tsjerk A, Marrink, Siewert J·2025·eLife
DOI 10.7554/elife.90627PMC PMC12634043

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Ongoing advances in force field and computer hardware development enable the use of molecular dynamics (MD) to simulate increasingly complex systems with the ultimate goal of reaching cellular complexity. At the same time, rational design by high-throughput (HT) simulations is another forefront of MD. In these areas, the Martini coarse-grained force field, especially the latest version (i.e. v3), is being actively explored because it offers an enhanced spatial-temporal resolution. However, the automation tools for preparing simulations with the Martini force field, accompanying the previous version, were not designed for HT simulations or studies of complex cellular systems. Therefore, they become a major limiting factor. To address these shortcomings, we present the open-source Vermouth python library. Vermouth is designed to become the unified framework for developing programs, which prepare, run, and analyze Martini simulations of complex systems. To demonstrate the power of the Vermouth library, the Martinize2 program is showcased as a generalization of the martinize script, originally aimed to set up simulations of proteins. In contrast to the previous version, Martinize2 automatically handles protonation states in proteins and post-translation modifications, offers more options to fine-tune structural biases such as the elastic network (EN), and can convert non-protein molecules such as ligands. Finally, Martinize2 is used in two high-complexity benchmarks. The entire I-TASSER protein template database as well as a subset of 200,000 structures from the AlphaFold Protein Structure Database are converted to CG resolution and we illustrate how the checks on input structure quality can safeguard HT applications.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.7554/elife.90627

Cites

103 of 103 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1. Fundamental stages in topology generation from atomistic structures. First, the provided input is parsed (step 1). Second, for every parsed residue, its atoms are identified and, if needed, atom names are corrected and missing atoms are added (step 2). Third, mappings are taken from the library and a resolution transformation to the required output resolution is performed (step 3). Fourth, intra-residue interactions are added from blocks taken from the library, and inter-residue interactions are added from links taken from the library (step 4). Fifth, optionally, post-processing is performed to add, for example an EN (step 5). Finally, the produced topology is written to output files (step 6).
Figure 1. Fundamental stages in topology generation from atomistic structures. First, the provided input is parsed (step 1). Second, for every parsed residue, its atoms are identified and, if needed, atom names are corrected and missing atoms are added (step 2). Third, mappings are taken from the library and a resolution transformation to the required output resolution is performed (step 3). Fourth, intra-residue interactions are added from blocks taken from the library, and inter-residue interactions are added from links taken from the library (step 4). Fifth, optionally, post-processing is performed to add, for example an EN (step 5). Finally, the produced topology is written to output files (step 6).CC BY · PubMed Central
Figure 2. Organization of the Vermouth library. The vermouth library defines 5 types of data structures (blue) to store molecular information and force field information. For convenience, it also defines two collection classes (orange) composed of several data-structure instances. Data structures are initiated or get input from parsers, which read 6 types of data files (see Appendix 1—table 1 for more details on file types). The central data structure(s) are Molecule and System. These are changed, updated, or transformed by so-called Processor classes, which take force field data as input. Parsers, data structures, and Processors only depend on three Python libraries as shown. At the moment, vermouth also exposes four types of writers (not shown here) to go along with the parsers (see Appendix 1—table 2).
Figure 2. Organization of the Vermouth library. The vermouth library defines 5 types of data structures (blue) to store molecular information and force field information. For convenience, it also defines two collection classes (orange) composed of several data-structure instances. Data structures are initiated or get input from parsers, which read 6 types of data files (see Appendix 1—table 1 for more details on file types). The central data structure(s) are Molecule and System. These are changed, updated, or transformed by so-called Processor classes, which take force field data as input. Parsers, data structures, and Processors only depend on three Python libraries as shown. At the moment, vermouth also exposes four types of writers (not shown here) to go along with the parsers (see Appendix 1—table 2).CC BY · PubMed Central
Figure 3. Illustration of atom recognition, mapping, and linking in topology generation. (a) To identify all atoms in the input molecule (black and orange) every residue in the molecule is overlaid with its canonical reference (blue and green). Atoms are recognized when they overlap with atoms in the reference (green). Atoms not present in the molecule are also identified (blue) and will be added later. Finally, atoms in the molecule not described by the canonical references are also labeled (orange) so that they may be identified later. (b) Identifying the terminal atoms that are not part of the canonical residues. The modification templates are depicted in blue and the atoms they match in orange. The cysteine does not participate since it does not carry any unexpected atoms and is depicted in gray for clarity. (c) Mappings (blue, orange, and green) describe a molecular fragment at two different resolutions and a correspondence between their particles. The correspondence is depicted approximately here. The mappings are applied to the molecule (black). (d) Example of applying a Link. The link depicted (dark blue) adds an angle potential over CG backbone beads.
Figure 3. Illustration of atom recognition, mapping, and linking in topology generation. (a) To identify all atoms in the input molecule (black and orange) every residue in the molecule is overlaid with its canonical reference (blue and green). Atoms are recognized when they overlap with atoms in the reference (green). Atoms not present in the molecule are also identified (blue) and will be added later. Finally, atoms in the molecule not described by the canonical references are also labeled (orange) so that they may be identified later. (b) Identifying the terminal atoms that are not part of the canonical residues. The modification templates are depicted in blue and the atoms they match in orange. The cysteine does not participate since it does not carry any unexpected atoms and is depicted in gray for clarity. (c) Mappings (blue, orange, and green) describe a molecular fragment at two different resolutions and a correspondence between their particles. The correspondence is depicted approximately here. The mappings are applied to the molecule (black). (d) Example of applying a Link. The link depicted (dark blue) adds an angle potential over CG backbone beads.CC BY · PubMed Central
Figure 4. Workflows for identifying protonation states or PTMs exemplified on protonated histidine. In route (a), the residue name of the protonated histidine extracted from the atomistic coordinates matches the residue name in the library and matches the fragment. Hence, the protonation state is correctly picked up. In route (b), the residue name matches that of neutral Histidine in the library. A mismatch of the fragments is recognized, and the extra hydrogen is labeled. Subsequently, by matching the extra hydrogen to a modification of the Histidine block, the protonated Histidine is recognized as neutral Histidine plus protonation modification, and the correct parameters for protonated Histidine are generated.
Figure 4. Workflows for identifying protonation states or PTMs exemplified on protonated histidine. In route (a), the residue name of the protonated histidine extracted from the atomistic coordinates matches the residue name in the library and matches the fragment. Hence, the protonation state is correctly picked up. In route (b), the residue name matches that of neutral Histidine in the library. A mismatch of the fragments is recognized, and the extra hydrogen is labeled. Subsequently, by matching the extra hydrogen to a modification of the Histidine block, the protonated Histidine is recognized as neutral Histidine plus protonation modification, and the correct parameters for protonated Histidine are generated.CC BY · PubMed Central
Figure 5. Example of automated identification of PTMs. CG Martini model of phosphorylated Tyrosine found in the EGFR kinase activation loop. The mapped structure of the phosphorylated residue is shown as beads overlying the atomistic structure.
Figure 5. Example of automated identification of PTMs. CG Martini model of phosphorylated Tyrosine found in the EGFR kinase activation loop. The mapped structure of the phosphorylated residue is shown as beads overlying the atomistic structure.CC BY · PubMed Central
Figure 6. Fine-tuning options for the elastic network. (a) ENs and backbone bonds within the human insulin dimer when generated with the molecule or all-option. The dimer consists of two chains colored in red and orange, which are connected by two disulfide bridges shown in purple. EN bonds are generated between the two chains and within the chains. (b) EN and backbone bonds within the insulin dimer when generated with the chain option. In this case, no elastic bonds are generated between the two chains. They are only connected by the disulfide bridge and non-bonded interactions. (c) EN within the FtsZ protein, when generated for both the intrinsically disordered tail domains (orange) and structural domain (red) (d) EN within the FtsZ protein when the EN is only generated within the structural domain by defining the EN unit as going from residues 12–320.
Figure 6. Fine-tuning options for the elastic network. (a) ENs and backbone bonds within the human insulin dimer when generated with the molecule or all-option. The dimer consists of two chains colored in red and orange, which are connected by two disulfide bridges shown in purple. EN bonds are generated between the two chains and within the chains. (b) EN and backbone bonds within the insulin dimer when generated with the chain option. In this case, no elastic bonds are generated between the two chains. They are only connected by the disulfide bridge and non-bonded interactions. (c) EN within the FtsZ protein, when generated for both the intrinsically disordered tail domains (orange) and structural domain (red) (d) EN within the FtsZ protein when the EN is only generated within the structural domain by defining the EN unit as going from residues 12–320.CC BY · PubMed Central
Figure 7. Ligands, cofactors, and polymers transformed to CG Martini level. (a) Flavin Reductase with two FMNs and one NDP cofactor bound in the reference AA state and mapped to Martini CG as indicated by the spheres. The inset shows a zoom onto FMN; (b) Lysozyme with benzene ligand bound in the reference AA structure and mapped to Martini CG resolution; (c) Crown ether with Martini beads shown on top of the AA structure; (d) Branched polyethylene at AA resolution (left) and Martini resolution (right) with the linear chain part shown in gray and the branches in yellow.
Figure 7. Ligands, cofactors, and polymers transformed to CG Martini level. (a) Flavin Reductase with two FMNs and one NDP cofactor bound in the reference AA state and mapped to Martini CG as indicated by the spheres. The inset shows a zoom onto FMN; (b) Lysozyme with benzene ligand bound in the reference AA structure and mapped to Martini CG resolution; (c) Crown ether with Martini beads shown on top of the AA structure; (d) Branched polyethylene at AA resolution (left) and Martini resolution (right) with the linear chain part shown in gray and the branches in yellow.CC BY · PubMed Central
Figure 8. Summary of the successes and failures of the high-throughput pipeline. We ran the pipeline on the 87084 structures from the template library used by the I-TASSER (Yang et al., 2015) protein prediction software, of which 73% could be converted with Martinize2. The other 26.4% failed mostly due to missing coordinates and unrecognized residues. For 100% of the converted structures, a GROMACS run input file (i.e. tpr-file) could be generated, and on all but 13 of the converted structures, an energy minimization could be performed.
Figure 8. Summary of the successes and failures of the high-throughput pipeline. We ran the pipeline on the 87084 structures from the template library used by the I-TASSER (Yang et al., 2015) protein prediction software, of which 73% could be converted with Martinize2. The other 26.4% failed mostly due to missing coordinates and unrecognized residues. For 100% of the converted structures, a GROMACS run input file (i.e. tpr-file) could be generated, and on all but 13 of the converted structures, an energy minimization could be performed.CC BY · PubMed Central
Figure 9. Two examples of problematic atomistic protein structures flagged by Martinize2. (a) The cysteine residue with too small O-O and O-C distances leads to superfluous bonds being recognized. (b) The incorrect interatomic distances in the histidine ring led to missing bonds (transparent), an erroneous O-N bond connecting the histidine to a neighboring asparagine. Additionally, a nitrogen atom is switched for an oxygen atom in asparagine.
Figure 9. Two examples of problematic atomistic protein structures flagged by Martinize2. (a) The cysteine residue with too small O-O and O-C distances leads to superfluous bonds being recognized. (b) The incorrect interatomic distances in the histidine ring led to missing bonds (transparent), an erroneous O-N bond connecting the histidine to a neighboring asparagine. Additionally, a nitrogen atom is switched for an oxygen atom in asparagine.CC BY · PubMed Central
Appendix 3—figure 1. Comparison of processing speeds between Martinize and Martinize2. 11 Protein structures with increasing size from a dataset of 200,000 structures were processed with Martinize and Martinize2 to record the computation time. The protein structures were spaced evenly by residue count.
Appendix 3—figure 1. Comparison of processing speeds between Martinize and Martinize2. 11 Protein structures with increasing size from a dataset of 200,000 structures were processed with Martinize and Martinize2 to record the computation time. The protein structures were spaced evenly by residue count.CC BY · PubMed Central
Appendix 4—figure 1. Example of finding all LCISs between graphs X and Y. Grayed out nodes are not used (they are excluded from the comparison by the shrinking step), but are depicted for clarity. Since nodes A1 and A2 in Y are symmetry equivalent, not all subgraphs are taken into account. Those that are excluded due to symmetry reasons are depicted in the box Symmetry pruned. Iteration 1: We try to find a subgraph isomorphism between X and Y. None is found. Iteration 2: Y is shrunk to produce the graphs depicted. We try to find subgraph isomorphisms between these and X. None are found. Iteration 3: all graphs from iteration 2 are shrunk further. Since a subgraph isomorphism can be found between at least one of these ({A1, A2, B}) and X, the algorithm terminates afterwards. To highlight how often the algorithm discovers that {A1, B} is subgraph isomorphic to X, it is shown in bold.
Appendix 4—figure 1. Example of finding all LCISs between graphs X and Y. Grayed out nodes are not used (they are excluded from the comparison by the shrinking step), but are depicted for clarity. Since nodes A1 and A2 in Y are symmetry equivalent, not all subgraphs are taken into account. Those that are excluded due to symmetry reasons are depicted in the box Symmetry pruned. Iteration 1: We try to find a subgraph isomorphism between X and Y. None is found. Iteration 2: Y is shrunk to produce the graphs depicted. We try to find subgraph isomorphisms between these and X. None are found. Iteration 3: all graphs from iteration 2 are shrunk further. Since a subgraph isomorphism can be found between at least one of these ({A1, A2, B}) and X, the algorithm terminates afterwards. To highlight how often the algorithm discovers that {A1, B} is subgraph isomorphic to X, it is shown in bold.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 104 references

Research works printed but not linked: 3

  1. Chung, 2010, Graph theory in the information age, Notices of the American Mathematical Society, 57, 726
  2. Kroon PC, Barnoud J, Brasnett C, Grunewald F, Peter M, Tadema M, Liu X, Araújo LB, marrink-lab, Tsjerk, Romumrn. 2025. Vermouth-martinize. Software Heritage. swh:1:rev:a3037bc2544daa0faa8ba105b9670a2659885d43. https://archive.softwareheritage.org/swh:1:dir:270aae9780a9b8f07c0c1411175f4e982bc0b539;origin=https://github.com/marrink-lab/vermouth-martinize;visit=swh:1:snp:155374469e37d66bb87fc5832a0e0f314e27ca85;anchor=swh:1:rev:a3037bc2544daa0faa8ba105b9670a2659885d43.
  3. marrink-lab. 2025. Martinize-examples. Software Heritage. swh:1:rev:78a81b00ff8bc9798386c442251a14f568e8f8eb. https://archive.softwareheritage.org/swh:1:dir:0869dc8a9197f72f91847dbb5739f97fb20576bd;origin=https://github.com/marrink-lab/martinize-examples;visit=swh:1:snp:85479e72dfea864eca7b37f1eb7390153820e6eb;anchor=swh:1:rev:78a81b00ff8bc9798386c442251a14f568e8f8eb.

Cited by

Sorted by citations

See all 124

Related

Metadata

Authors
Kroon, P. C., Grünewald, F., Barnoud, J., van Tilburg, M., Brasnett, C., Souza, P. C., Wassenaar, T. A., & Marrink, S. J.
Affiliations
  • University of Groningen (NL)
  • École Normale Supérieure de Lyon (FR)
Corresponding author
  • Grünewald, Fabian
  • Marrink, Siewert J
ORCID
Venue
eLife
Publisher
eLife Sciences Publications, Ltd
Published
2025-11-20
Volume
12
ISSN
2050-084X (ISSN-L)
DOI
10.7554/elife.90627
PMC
PMC12634043
OpenAlex
W7106126080
Cited by
124
References
104
Citations by year
2026: 73 · 2025: 22
Type
posted content
Language
en
Open access
Licence
Locations
  • eLife (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • University of Groningen research database (University of Groningen / Centre for Information Technology) (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion)
  • HAL (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University) (cc-by)
Funders
  • European Research Council (COMP-MICR-CROW-MEM)
  • Novo Nordisk (BOUNDLESS)
  • Klaus Tschira Foundation (Independent Postdoc)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Physical Sciences › Chemistry › Physical and Theoretical Chemistry › Advanced Physical and Chemical Molecular Interactions
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
Keywords
Martini coarse-grained force field, vermouth, high-throughput molecular dynamics, protonation states, post-translational modification, elastic networks, AlphaFold structures
MeSH terms
Proteins, Software, Molecular Dynamics Simulation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review