Temperature artifacts in protein structures bias ligand-binding predictions

Abstract

X-ray crystallography is the gold standard to resolve conformational ensembles that are significant for protein function, ligand discovery, and computational methods development. However, relevant conformational states may be missed at common cryogenic (cryo) data-collection temperatures but can be populated at room temperature. To assess the impact of temperature on making structural and computational discoveries, we systematically investigated protein conformational changes in response to temperature and ligand binding in a structural and computational workhorse, the T4 lysozyme L99A cavity. Despite decades of work on this protein, shifting to RT reveals new global and local structural changes. These include uncovering an apo helix conformation that is hidden at cryo but relevant for ligand binding, and altered side chain and ligand conformations. To evaluate the impact of temperature-induced protein and ligand changes on the utility of structural information in computation, we evaluated how temperature can mislead computational methods that employ cryo structures for validation. We find that when comparing simulated structures just to experimental cryo structures, hidden successes and failures often go unnoticed. When using structural information in ligand binding predictions, both coarse docking and rigorous binding free energy calculations are influenced by temperature effects. The trend that cryo artifacts limit the utility of structures for computation holds across five distinct protein classes. Our results suggest caution when consulting cryogenic structural data alone, as temperature artifacts can conceal errors and prevent successful computational predictions, which can mislead the development and application of computational methods in discovering bioactive molecules.

Fig. 1. Global and local structural responses to temperature. (A) Globally, structures at cryogenic temperatures (cryo; blue plot) are more variable and more compact than their room temperature (RT; red plot) equivalents, as shown by average unit cell (UC) volumes across 9 matched structures collected at both temperatures. (B) The isomorphous F_o − F_o map of the apo structure collected at cryo versus RT shows differences in the electron density (green mesh, positive difference electron density; red mesh, negative difference electron density) that indicate idiosyncratic temperature effects, especially around the ligand-binding site in the bottom lobe, indicated by the black dotted mesh in panel C (labeled LIG). (C) Occurrence of temperature-dependent rotamer differences across all 9 structures are projected onto the respective residues in the T4L apo structure; colored by temperature sensitivity of each residue across all 9 structure pairs: yellow for few structures, orange for several structures, and red for most structures showing temperature differences of the residue; white patches are Gly and Ala that do not have Chi angles; and grey patches show no rotamer change with temperature. (D) Locally, RT data of the L99A apo cavity reveal an alternative F-helix conformation (conf. B) in the F_o − F_c difference electron density maps (green and red mesh for positive and negative density, respectively; only cyan conformation A was included in refinement) that is not visible at cryo; 2mF_o − DF_c map shown as blue mesh; stick thickness represents relative occupancy. (E) All 8 ligand complexes show a shift in preferred orientation in response to temperature rather than due to ligand binding for at least 1 residue rotamer in the F-helix near the ligand-binding site. Ringer plots for selected residues, with rotamer differences at RT (red) versus cryo (blue) indicated by arrows.

Fig. 2. Temperature sensitivity of binding congeneric ligands. (A) Ringer plots compare rotamers for 2 proximal F-helix residues, Glu108 and Thr109, across 3 congeneric structures (apo, bound to toluene, and bound to iodobenzene) in response to temperature (cryo in blue, RT in red). Arrows indicate temperature-sensitive rotamers, and tildes indicate no major rotamer change. (B) Toluene's alternative ligand conformation at RT is indicated by the presence of green F_o − F_c difference density when only the major conformer is included in refinement and confirmed by an unbiased Polder OMIT map that excludes all ligands (here superimposed onto the map for clarity). (C) o-Xylene experiences a 0.41 Å RMSD shift upon changing temperature, while the overall protein structure differs by only 0.2 Å.

Fig. 3. Cryo artifacts misinform computational method validation. Computational Cringer plots derive histograms of rotamer populations of each residue, plotted as a frequency across residual dihedral angles iterated over all frames of an MD simulation. Cringer plots enable comparison to experimental Ringer plots to identify true positives (MD rotamers agree with both RT and cryo), true negatives (MD disagrees with both RT and cryo, which may agree or not), false positives (MD agrees with cryo, both differ from RT) and false negatives (MD agrees with RT, both differ from cryo). Shown here are selected examples of all 4 categories; more examples of false negatives and false positives are provided in the ESI (Fig. S10 and S11†).

Fig. 4. Temperature-induced structural differences affect docking performance. (A) AUC (shown as fractions) and (B) adjusted logAUC (%) enrichment plots from docking 98 known T4L–L99A binders against 3152 property-matched DUD-E decoys using (A) OEdock or (B) Autodock Vina and input structures with a closed and intermediate F-helix conformation; equilibrated structures were generated after 50 ns of MD simulations. (C and D) Docking poses of 2-ethoxyphenol (colored sticks) docked against (C) experimental and (D) equilibrated cryo and RT apo structures (as in A–B) are compared to PDB structure 2RB1 with RMSDs indicated.

Fig. 5. Temperature artifacts mislead computational validation across protein classes. Ringer–Cringer comparisons reveal hidden differences between experimental data collected at RT (red) or cryo and computational predictions. Several binding site residues are highlighted as false negative (FN), true negatives (TN) and false positives (FP) across four protein systems cytochrome C peroxidase (A), thrombin (B), protein tyrosine phosphatase 1B (C), galectin (D). Respective 2F_o − F_c electron density maps are shown at 1 sigma. See Fig. S45 for more examples.