Self-assembly is a many-chain process, and an all-atom trajectory that resolves every backbone dihedral and every water orientation will usually exhaust its wall-clock budget before those chains have found each other. Coarse-graining buys the missing reach by deleting degrees of freedom. The MARTINI protein extension was introduced on exactly that argument: coarse-grained molecular modelling allows computer simulations to be run on length and time scales two to three orders of magnitude larger than atomistic simulations, providing a bridge between the atomistic and the mesoscopic scale [1].
The arithmetic behind that has three separable parts. MARTINI uses a 4-to-1 mapping, four non-hydrogen atoms grouped per bead, relaxed to 3-to-1 or even 2-to-1 for aromatics to replicate their shape better; water is explicit, with each water bead corresponding to four water molecules; and beads carry only net charges, the rest of the non-bonded interaction being handled by Lennard-Jones terms [2]. Removing those degrees of freedom lets the integrator take 20–40 fs steps where all-atom molecular dynamics is constrained to 1–2 fs. On top of that, the smoother coarse-grained potential energy surface makes molecules diffuse unrealistically faster, and an additional speed-up factor of 4 is often assumed for MARTINI, which together give up to 160 times fewer steps to simulate the same nominal time than all-atom molecular dynamics [2].
The molecular dynamics guide in this library covers the atomistic case: one chain, an all-atom potential, and the question of what conformational populations that potential predicts. Coarse-grained assembly simulation is a different instrument pointed at a different observable — the morphology many chains adopt, and where each chain sits in a partition — using a potential that no longer represents a hydrogen bond as a hydrogen bond.
What the beads were actually fitted to
MARTINI was parameterised in a systematic way, based on the reproduction of partitioning free energies between polar and apolar phases of a large number of chemical compounds; reproducing those free energies required increasing the number of possible interaction levels of the coarse-grained sites relative to the earlier lipid model, and a further extension allowed the simulation of planar (ring) compounds, including sterols [3]. A bead type is therefore a transferable statement about how a chemical group distributes between a polar and an apolar environment. It is not a statement about directionality, valence or geometry.
That transferability is testable, and it has failed in places that matter for peptides. Once the model could be tested against fully atomistic force fields, phenylalanine and proline were found to be too hydrophobic, and dimers formed by polar residues in apolar solvents did not bind strongly enough; those residues were reparameterised either by reassigning particle types or by introducing embedded charges, and some bonded terms were improved to give a more realistic length of α-helices and better numerical stability for polyalanine and glycine repeats, producing Martini version 2.2 [4]. Phenylalanine heading that list is not incidental: diphenylalanine recurs below as the established test case for short-peptide assembly, so that test case's aggregation is being scored on a bead type whose hydrophobicity had to be corrected.
Martini 3 is the current general-purpose release — a refined model with an improved interaction balance, new bead types and expanded ability to include specific interactions representing, for example, hydrogen bonding and electronic polarizability, allowing more accurate predictions of molecular packing and interactions across a range running from oil/water partitioning data to protein–protein and protein–lipid interactions [5].
Secondary structure goes in; it does not come out
The MARTINI backbone carries no explicit hydrogen bonding. What it carries is a bead whose polarity is set by an assigned secondary structure, and despite that lack the model can predict self-assembling sequences and structural features by introducing those secondary-structure-specific beads [6]. The assignment is an input. For short peptides an extended β-sheet encoding is typically used as that input, on the basis of experimental observation, and the assumption becomes increasingly unreliable beyond six to ten residues, where folded conformations begin to emerge [6].
The consequence is large enough to change conclusions. In hexapeptides and decapeptides simulated with MARTINI 2.2p, changes in the secondary-structure encoding significantly affected the predicted self-assembly behaviour, with aggregation propensity scores for the same peptide varying by up to one unit — shifting from fully dissolved (AP ≈ 1) to well-aggregated (AP > 2) in specific cases [6]. The route is amphiphilicity: shifting backbone polarity alters the overall amphiphilicity of the molecule, and because the magnitude and direction of that effect depend on side-chain polarity and peptide length, the resulting bias is highly sequence-specific and difficult to anticipate or correct systematically [6]. Those authors argue for reevaluating the conventional extended β-sheet encoding and for more native-like backbone representations [6].
Internal flexibility can be restored rather than assumed, at least in part. A set of pseudodihedral potentials describing the internal flexibility of MARTINI peptides was parameterised using atomistic simulations of tetrapeptides, octapeptides and longer peptides in solution as a reference; adding them significantly improved agreement with the contact maps computed from the atomistic simulations, and the modified model showed reasonable transferability of parameters for the amyloid- and elastin-like peptides examined [7]. The same improved model was applied to the β-sheet-forming octapeptides SNNFGAIL and (GV)4 in water and at a water–octane interface; at the interface, peptide adsorption occurred rapidly and the peptides aggregated spontaneously in favour of stretched conformers resembling β-strands [7].
What this licenses is narrow and useful. The trajectory can tell you whether a peptide held in a given conformational class aggregates, and how the resulting object is arranged. It cannot tell you that the peptide adopts that conformational class, because you supplied it. If the question is whether a sequence acquires β-strand character before or during assembly, this is the wrong instrument. The minimum defensible practice is to run the alternative encodings and report the spread, since the spread is the size of the assumption you are making.
Effective microseconds are not experimental microseconds
The factor of 4 above is a convention rather than a per-system measurement — it is assumed [2], and it is applied globally to a trajectory in which bead diffusion, chain reorientation, nucleation of a first contact and ripening of an aggregate have no particular reason to be accelerated by the same amount. What survives the convention is the ordering of events within one simulation, and comparison of like with like across simulations run under identical settings. What does not survive is an absolute rate, a lag time or a half-life read off the axis and set beside a kinetic experiment. State raw simulated time and say explicitly whether any scaling factor has been applied; an axis labelled in microseconds without that statement cannot be checked by a reader.
Over-aggregation, and the corrections aimed at it
Across Martini studies of membrane protein oligomerization, a significant fraction demonstrate unrealistic protein clusters whose formation is essentially an irreversible process, which suggests that protein–protein interactions are exaggerated in the model [8]. The demonstration ran in two directions: dimerization free energies of helical transmembrane peptides obtained with Martini were compared with those determined from FRET experiments, and the predicted structures of transmembrane domain dimers were compared with the corresponding structures resolved using NMR and found to be in poor agreement [8]. Slightly scaling down the Martini protein–protein interactions, in a manner that does not interfere with the other Martini interaction parameters, prevents the excessive, irreversible and non-selective aggregation; it does not improve the predicted dimer structures, which implicates a need for careful amino-acid-specific adjustment rather than a uniform reduction across all protein beads [8].
A separate analysis reached the association problem from statistical mechanics. That work compared MARTINI 2.0, Dry MARTINI and all-atom simulations mapped onto equivalent coarse-grained sites for certain lipid bilayers, and separately computed the potential of mean force for the association of two amphipathic helices in a lipid bilayer, which differed substantially from the all-atom projection, especially at shorter ranges; the MARTINI curve is much less repulsive and hence gives a higher probability of helix association in the MARTINI bilayer than occurs in the actual all-atom case [9]. The same work traced issues of accuracy and temperature transferability to an incorrect enthalpy–entropy decomposition of the potentials of mean force obtained from the temperature dependence of the pair distribution functions via the reversible work theorem [9]. A top-down model fitted to free energies can reproduce a free energy while distributing it wrongly between enthalpy and entropy, and a temperature scan is precisely the experiment that exposes it.
Some artefacts are structural properties of the building-block approach itself. The absence of specific cross Lennard-Jones parameters between different particle sizes may lead to artificially high free energy barriers in dimerization profiles; deviating too far from the standard bonded parameters degrades both solute partitioning behaviour and solvent properties; and too weak bonded force constants entail the risk of artificially inducing clustering [10]. That last point is worth holding onto whenever a peptide is being restrained loosely to hold a conformation: a soft restraint is not a neutral choice.
The bias does not always point the same way. In Martini 3 the protein-focused improvements seem to have compromised the model's ability to reproduce short peptide self-assembly: diphenylalanine does not even exhibit aggregation in version 3, although it did in version 2.1 [11]. Systematically exploring parameters for the aromatic side chains and charged backbone beads established a parameter set that effectively reproduces tube formation, and those adjustments also replicate the self-assembly of other di- and tripeptides and coassemblies; the diagnosis was overestimated hydrophilicity arising from charged termini, disruption of π-stacking due to insufficient planarity in aromatic groups, and a discrepancy in intermolecular distances between aromatic interactions and backbone–backbone interactions [11]. Considerable effort in the development of Martini has been expended to reduce the “stickiness” of amino acids so as to better simulate proteins in bilayers, and one account of the family benchmarks all 400 dipeptides of the 20 gene-encoded amino acids, in triplicate, across the three most recently released Martini versions and variations in their solvents, scored by aggregation propensity [12]. Version number and solvent model are part of your result, not part of the methods boilerplate.
On soluble chains the needed correction has run the other way. Martini 3 underestimates the global dimensions of intrinsically disordered proteins and multidomain proteins compared with small-angle X-ray scattering data, and increasing the strength of interactions between protein and water by about 10% gave improved agreement with the SAXS data, a more accurate description of self-association of IDPs and folded proteins, and better agreement with paramagnetic relaxation enhancement data for most IDPs, while simulations with the revised force field still showed deviations from experiment for some systems [13]. That work is on disordered and multidomain proteins rather than short synthetic peptides, and the correction should not be assumed to carry across.
Box size and molecule count are results-determining parameters
Supramolecular peptide assembly simulations differ from protein simulations in two respects that are easy to under-report: there is no experimentally validated initial structure to start from, and the simulation size, the number of molecules and the concentration matter to the outcome [2]. The worked instance is again diphenylalanine, where the box side was increased from 12.5 to 24.5 nm and the number of dipeptides from 300 to 1600 in order to replicate the formation of tubes [2]. A morphology that fails to appear in a small box is not evidence that the morphology does not form; it may be evidence that the box could not accommodate it. Report box edge, molecule count and concentration as three separate numbers, because two can be held fixed while the third changes the answer.
When a coarse-grained result needs backmapping
Backmapping converts a coarse-grained configuration to atomistic coordinates by geometric projection and subsequent force-field-based relaxation, with the initial model relaxed through several short cycles of energy minimization and position-restrained molecular dynamics, and the atomistic backbone reconstructed from the coarse-grained model by a dedicated algorithm; mapping definitions exist for MARTINI to and from GROMOS, CHARMM and AMBER [14]. The standard sequential workflow begins with the assembled state formed through coarse-grained simulation, backmaps it into all-atom resolution, and re-equilibrates the structure at all-atom resolution in a short simulation [2].
Backmap when the question turns on something a bead cannot carry: side-chain registry within a sheet, a hydrogen-bond network, the geometry of aromatic stacking. Those are the same quantities implicated in the short-peptide diagnosis of Martini 3 described above, which is a reason to treat them as suspect by default rather than as incidental detail. Note also what the exercise settles. A successful backmap yields an atomistic model consistent with the coarse-grained geometry it was given; it becomes a test of that geometry only if the atomistic simulation runs long enough, and without restraints, to show whether the packing survives.
What this method answers well
Partitioning is the strong suit, because partitioning free energy is the fitted observable. The protein extension was validated on exactly that: potentials of mean force for each amino acid as a function of distance from the centre of a dioleoylphosphatidylcholine bilayer, amino acid association constants, the partitioning of a series of model pentapeptides, and the partitioning and orientation of WALP23 in DOPC bilayers and of a series of KALP peptides in dimyristoyl- and dipalmitoylphosphatidylcholine bilayers, all showing good agreement with results obtained from atomistic models [1]. In a systematic investigation of five series of polyalanine–leucine peptides in DPPC bilayers, the fraction of peptides partitioned at the interface increased with decreasing peptide length and decreasing leucine content, demonstrating that the model discriminates partitioning behaviour arising from subtle differences in amino acid composition [1]. Simulating the concentration-dependent formation of transmembrane pores by magainin produced disordered toroidal pores, in line with atomistic simulation studies [1].
Screening for aggregation is the other. Coarse-grained molecular dynamics was used to rapidly screen all 400 dipeptide combinations and predict their ability to aggregate as a potential precursor to self-assembly; the simulation protocol and scoring method allow a rapid determination of whether a given sequence is likely to aggregate under aqueous conditions, and systems showing strong tendencies in the initial screen were taken into longer simulations that agreed well with dipeptide self-assembly or aggregation already reported in the literature [15]. The approach then covered all 8,000 possible tripeptides, with filters applied to select candidates that simultaneously optimize the apparently contradicting requirements of aggregation propensity and hydrophilicity, which produced a set of design rules; a number of the peptides were subsequently synthesized and characterized, including the first reported tripeptides able to form a hydrogel at neutral pH [16].
Two limits on how far a score of that kind travels. Aggregation propensity quantifies aggregation rather than self-assembly, and takes higher values for insoluble sequences [2] — an insoluble peptide and an ordered nanostructure both score high, and the parameter does not separate them. And any such score is conditional on every choice enumerated above: encoding, force-field version, solvent model, box edge, molecule count. A ranking within one consistent set of conditions is defensible. An absolute threshold transported between studies is not.
Report, at minimum: force-field version and any local modifications, solvent model, the secondary-structure encoding used and the spread of the result across alternative encodings, box edge and molecule count and concentration, raw simulated time with any scaling factor stated explicitly, and whether the final morphology was backmapped and how long it was held at atomistic resolution without restraints. A coarse-grained assembly result is a conditional statement, and those are its conditions.