Collagen as a stress test and a tool for improvement of glycine and proline conformations in biomolecular force fields

Main text

Molecular simulations are an important instrument in the tool chest of biophysics and have continuously increased in prominence. Central to the accuracy and reliability of their results is the force field, which determines the strength of all intra- and intermolecular interactions and thus how physically a biomolecular system will behave during simulation. All commonly used force fields represent biomolecules in a grossly simplified fashion; this is necessary for simulations to run fast enough to reach biologically relevant timescales without requiring prohibitive computational resources. Perhaps unintuitively, this makes them highly complex entities, containing thousands of numerical parameters and a finely tuned cancellation of errors that allows them to produce accurate results despite their simplistic functional form. Over a time span of several decades, increases in computer power have led to the discovery of weaknesses in the force fields, which only came to light when studying larger systems and/or longer timescales than what were formerly within computational reach. These discoveries have, in turn, been a continuous driving force for improvements in the force fields. Unfortunately, properly identifying weaknesses in force fields requires substantial know-how and dedication; one could argue that this is an important bottleneck in the quest for better biomolecular force fields.

This issue of Biophysical Journal contains an instance of the discovery of a possible weakness in biomolecular force fields that is a potential opportunity for improvement. In their paper (1), Pantelopulos and Best have carefully assessed how key properties of collagen are reproduced by several versions of three common biomolecular force fields: AMBER, CHARMM, and GROMACS. Specifically, they performed multiple simulations of a solvated triple helix consisting of repeated proline-hydroxyproline-glycine sequences (POG₁₀); a crystallized leaflet consisting of such helices is visualized in Fig. 1. For this common collagen model, they compared backbone dihedral angles (and ring puckerings, in the case of Pro) to their respective distributions in a survey of the PDB crystallographic databank. In addition, they evaluated the dynamics of their model system by calculating small-angle x-ray scattering form factors and ¹⁵N-NMR spin-relaxation rates and comparing the results to available experimental data (2,3,4). For the crystal structure data, the AMBER force fields performed the best, the CHARMM force fields showed small deviations, and GRO54A7 showed large deviations. Conversely, small-angle x-ray scattering form factors were reproduced by the CHARMM force fields similarly well (better for q > 0.22 nm⁻¹ in the present author’s opinion) compared to the AMBER force fields. The GROMACS force fields considered in this analysis again showed large deviations. Similar trends are also present for the ¹⁵N-NMR spin relaxation rates, where R₁ is particularly well reproduced by CHARMM36m and the AMBER force fields; R₂ shows relatively large deviations for all force fields, which the authors are tentatively interpreting as a failure to accurately reproduce slower motions such as molecular tumbling. Driven by these observations, the authors proposed an improvement to the CHARMM36m force field, which they dubbed CHARMM36mGP to reflect the fact that the CMAP cross-term is scaled by 0.5 for glycine and proline. This modification improved its reproduction of the aforementioned structural data from the PDB and the ¹⁵N-NMR R₁ relaxation rate. In addition, it was shown to improve upon CHARMM36(m) with respect to reproducing experimental ³J_HH NMR coupling parameters in the GPGG peptide (5), which were originally used as validation for the latter force field’s CMAP term (6). This opens the intriguing possibility that CHARMM36mGP (or a more fine-tuned future iteration thereof) may also be an improvement for proteins other than collagen.

Figure 1.

Visualization of POG₉ triple helix, obtained by replicating the unit cell of PDB: 3B0S (7), which served as a template for the staring conformations in the Pantelopulos and Best paper. The crystal turned out to be organized in leaflets. A single leaflet is visualized, consisting of parallel triple helices that are alternatingly located slightly above and below the overall plane of the leaflet. The backbones of the three peptides forming the “upper” helices are depicted as ribbons colored yellow and two shades of orange, respectively; similarly, the “lower” helices consist of three peptides whose backbones are depicted in three shades of green. The skeletal structures of the peptides are depicted with a “licorice” representation using the common CPK color convention (except that carbon atoms are gray instead of black for clarity); the positions of crystal water oxygen atoms are marked by pink spheres.

As a general validation of the results obtained for POG₁₀, x-ray backbone dihedral angles of four more representative collagen-like peptides were reproduced in solution with CHARMM36mGP, AMBER ff99sb^∗-ILDN, and AMBER ff19sb. While this yielded somewhat different results for each peptide sequence, overall, it was concluded that all three force fields are equally good choices for collagen-like structures. As these results are roughly in line with those obtained on POG₁₀, where a broader panel of force fields was considered, the authors list a few additional AMBER variants that are probably reasonable choices. Conversely, some other force fields produced substantial deviations; if those were used in the past for collagen simulations, the results might require careful reconsideration.

A scientist’s work is never done, and this is especially true in force field validation and development. No matter how elaborate any given study is, there always remain questions about whether the outcome might have been different if additional simulations would had run, additional sources of experimental target data considered, or other possible tweaks to the force fields evaluated… In the present author’s opinion, a significant weakness in the Pantelopulos and Best study is the fact that the crystal structures are largely the only direct source of geometrical (as opposed to dynamical) data. While the authors rightly argue that x-ray statistics of water-soluble proteins generally correspond well to solution data, it remains to be seen whether this also holds for a poorly soluble fibrillar biopolymer such as collagen. Indeed, Fig. 1 reveals relatively tight interactions with neighboring triple helices, which may induce conformational preferences that differ from a single triple helix in solution. Of course, it is far easier to identify a weakness in the selection of target data than to address it by finding other sources of target data—even more so if they need to be matched with feasible simulation protocols. Typically, questions like this are addressed in follow-up studies wherein the community converges to a consensus about which tweaks to the force field have the best trade-off between accuracy on specific test cases and general applicability. Therefore, while the present study by Pantelopulos and Best is surely not the last word on the treatment of Gly and Pro in biomolecular force fields (or CHARMM in particular), it does open a new avenue toward potential enhancements that may very well end up improving simulation results for not only collagen but for proteins in general.

Declaration of interests

The author declares no competing interests.

Editor: James Gumbart.