Assessment of transferable forcefields for protein simulations attests improved description of disordered states and secondary structure propensities, and hints at multi-protein systems as the next challenge for optimization

Luciano A Abriata; Matteo Dal Peraro

doi:10.1016/j.csbj.2021.04.050

Assessment of transferable forcefields for protein simulations attests improved description of disordered states and secondary structure propensities, and hints at multi-protein systems as the next challenge for optimization

Comput Struct Biotechnol J. 2021 Apr 25:19:2626-2636. doi: 10.1016/j.csbj.2021.04.050. eCollection 2021.

Authors

Luciano A Abriata^{1

2

3}, Matteo Dal Peraro^{1

3}

Affiliations

¹ Institute of Bioengineering, School of Life Sciences, École Polytechnique Fédérale de Lausanne, CH-1015 Lausanne, Switzerland.
² Protein Production and Structure Core Facility, School of Life Sciences, École Polytechnique Fédérale de Lausanne, CH-1015 Lausanne, Switzerland.
³ Swiss Institute of Bioinformatics, CH-1015 Lausanne, Switzerland.

Abstract

Continuous assessment of transferable forcefields for molecular simulations is essential to identify their weaknesses and direct improvement efforts. The latest efforts focused on better describing disordered proteins while retaining proper description of folded domains, important because forcefields of the previous generations produce overly compact disordered states. Such improvements should additionally alleviate the related problem of over-stabilized protein-protein interactions, which has been largely overlooked. Here we evaluated three state-of-the-art forcefields, current flagships of their respective developers, optimized for ordered and disordered proteins: CHARMM36m with its recommended corrected TIP3P* water, ff19SB with the recommended OPC water, and the 2019 a99SBdisp forcefield by D. E. Shaw Research with its modified TIP4P water; plus ff14SB with TIP3P as an example of the former generation of forcefields. Our evaluation entailed simulations of (i) multiple copies of a protein that is highly soluble yet undergoes weak dimerization, (ii) a disordered peptide with low, well-characterized alpha helical propensity, and (iii) a peptide known to form insoluble β-aggregates. Our results recapitulate ff14SB-TIP3P over-stabilizing aggregates and secondary structures and place a99SBdisp-TIP4PD at the other end i.e. predicting overly weak intermolecular interactions despite reasonably predicting secondary structure propensities. In-between, CHARMM36m-TIP3P* still over-stabilizes aggregates but predicts residue-wise alpha helical propensities in solution slightly better than ff19SB-OPC, while ff19SB-OPC poses the best prediction of weak dimerization of the soluble protein still predicting aggregation of the β-peptides. This independent assessment shows that the claimed forcefield improvements are real, but also that a right balance between noncovalent attraction and repulsion has not yet been reached. We thus propose developers to consider systems like those tested here in their forcefield tuning protocols. Last, the good performance of CHARMM36m-TIP3P* further shows that tuning 3-point water models might still be an alternative to the more costly 4-point models like OPC and TIP4PD.

Keywords: Aggregation; Amber; Charmm; Force fields; Molecular dynamics; Molecular simulations; Peptide; Protein; Protein dynamics; Simulation; Solubility; Water simulation.