ISOLDE: a physically realistic environment for model building into low-resolution electron-density maps

Abstract

This paper introduces ISOLDE, a new software package designed to provide an intuitive environment for high-fidelity interactive remodelling/refinement of macromolecular models into electron-density maps. ISOLDE combines interactive molecular-dynamics flexible fitting with modern molecular-graphics visualization and established structural biology libraries to provide an immersive interface wherein the model constantly acts to maintain physically realistic conformations as the user interacts with it by directly tugging atoms with a mouse or haptic interface or applying/removing restraints. In addition, common validation tasks are accelerated and visualized in real time. Using the recently described 3.8 Å resolution cryo-EM structure of the eukaryotic minichromosome maintenance (MCM) helicase complex as a case study, it is demonstrated how ISOLDE can be used alongside other modern refinement tools to avoid common pitfalls of low-resolution modelling and improve the quality of the final model. A detailed analysis of changes between the initial and final model provides a somewhat sobering insight into the dangers of relying on a small number of validation metrics to judge the quality of a low-resolution model.

Keywords: ISOLDE; model building; molecular dynamics; real-space refinement; visualization.

Figure 1

Common annotations and unit operations used during ISOLDE simulations. (a, b) Cis peptide bonds (marked with asterisks) are filled with a red trapezoid, while twisted peptide bonds (not shown) are filled in yellow. C^α atoms are coloured according to their current Ramachandran status, with outliers (arrowheads) appearing in maroon, marginal conformations shaded from maroon to yellow and preferred conformations shaded from yellow to green with increasing probability. Scripted cis–trans and peptide-plane flips act on the peptide bond N-terminal to the selected residue. (c, d) Flipping a peptide plane involves imposing temporary restraints on the φ and ψ dihedrals. Dihedral restraints are annotated by a ring-and-posts motif around each axial bond (marked with daggers), where the angle between the posts gives the current deviation from the target and the colour denotes the level of satisfaction of the restraint. (e, f) Secondary-structure restraints combine φ and ψ restraints, with distance restraints between O_n and N_n+4 and between C^α _n and C^α _n+2 displayed as purple dotted pseudobonds (marked with double daggers). (g, h, i) Previews of rotamer options (marked with section symbols) are shown in a thinner stick representation and cycle in order of probability for the given secondary structure. The chosen rotamer coordinates may be committed directly, but it is generally preferable to instead apply the target as dihedral restraints, allowing the atoms to approach the target conformation smoothly without risking clashes. Any heavy atom may be restrained to a given location with a user-defined spring constant (j) and/or tugged directly with the mouse or a three-dimensional input device (k). All panels are screen captures taken from the ISOLDE environment during rebuilding of the MCM2-7 model.

Figure 2

Ramachandran plots for general residues at key stages of rebuilding/refinement. (a) PDB entry 3ja8 was originally refined from a largely hand-built model with the aid of Ramachandran and rotamer restraints, achieving a MolProbity score of 2.51. Such restraints can be problematic when the nearest allowed region of the plot is not the true conformation for a given residue. (b) After energy minimization and 3000 MD steps in ISOLDE the number of outliers had increased substantially, suggesting that portions of the original model were indeed overfitted into energetically unfavourable conformations. (c) Extensive remodelling and restrained refinement yielded a final model with no Ramachandran outliers and an overall MolProbity score of 1.44.

Figure 3

Examples of regions where bulk remodelling was necessary. (a) Presumably owing to an erroneous sequence alignment to the homology template, Lys468 of chain 4 was missing, with the following 27 residues shifted in register by one residue to fill the gap. This was corrected using ISOLDE’s register-shift function prior to adding the missing residue in Coot. (b) Final conformation after refinement against the autosharpened map. (c) The zinc-finger domain in chain 6 is found in weak and fragmented density, and was originally modelled as a polyalanine trace with the cysteine residues out of position. (d) Autosharpening of the map significantly improved interpretability in this region, allowing hand-modelling of this domain into the canonical fold with the coordinated zinc ion present in the centre of the highest density peak.

Figure 4

Per-residue changes in (a) backbone torsion angles and (b) heavy-atom positions highlight the fact that modest improvements in validation statistics may involve extensive changes to the model. Insets denote the numbers of residues with large changes in each criterion (defined as >45° change in dihedral angle or >2 Å heavy-atom r.m.s.d.). ‘Combined’ gives the number of residues with a large change in at least one criterion. Annotations in (b) refer to a single-residue register shift in chain 4, residues 469–497 (asterisk), a poorly interpretable loop originally modelled into density complicated by the unmodelled C-terminal end of chain 6 (dagger) and the zinc-finger domain of chain 6, which was a polyalanine trace in the original model (double dagger).

Figure 5

Extensive rebuilding yields minimal changes in correlation to the map. Starting from the original model (blue), real-space refinement with the current version of phenix.real_space_refine reduced the overall masked FSC from 0.828 to 0.816, with some improvement in geometry. Rebuilding in ISOLDE followed by a tightly restrained phenix.real_space_refine run as described (black) increased the FSC to 0.821, while running phenix.real_space_refine with default settings on the result (green) further increased the FSC to 0.830 with a slight degradation in geometry. These changes are far smaller than those arising from choice of map sharpening or inclusion/exclusion of H atoms, for example. The vertical dotted line denotes the published resolution of the map.

Figure 6

The relationship between model quality and fit to data breaks down in deposited 3.5–4 Å resolution crystal (a) and cryo-EM (b) models. The crystal structure cohort was limited to models deposited from 2007 to the present, and R _free values for data at ≤3.75 Å and >3.75 Å resolution are offset by ±0.0015 for clarity. All EM models in the resolution range with masked CC ≥ 0.5 were included in the analysis.