FLEXR GUI: a graphical user interface for multi-conformer modeling of proteins

Proteins are well known `shapeshifters' which change conformation to function. In crystallography, multiple conformational states are often present within the crystal and the resulting electron-density map. Yet, explicitly incorporating alternative states into models to disentangle multi-conformer ensembles is challenging. We previously reported the tool FLEXR, which, within a few minutes, automatically separates conformational signal from noise and builds the corresponding, often missing, structural features into a multi-conformer model. To make the method widely accessible for routine multi-conformer building as part of the computational toolkit for macromolecular crystallography, we present a graphical user interface (GUI) for FLEXR, designed as a plugin for Coot 1. The GUI implementation seamlessly connects FLEXR models with the existing suite of validation and modeling tools available in Coot. We envision that FLEXR will aid crystallographers by increasing access to a multi-conformer modeling method that will ultimately lead to a better representation of protein conformational heterogeneity in the Protein Data Bank. In turn, deeper insights into the protein conformational landscape may inform biology or provide new opportunities for ligand design. The code is open source and freely available on GitHub at https://github.com/TheFischerLab/FLEXR-GUI.

1. Introduction

Proteins are dynamic machines that change conformation to perform functions such as catalysis (Eisenmesser et al., 2005), allostery (Buhrman et al., 2010) and ligand binding (Krojer et al., 2020). These changes occur across timescales and can be global and slow, such as major domain or hinge motions, or local and fast, such as subtle changes in the rotameric state of individual amino acid residues (Henzler-Wildman & Kern, 2007). The protein conformational landscape can also be altered (St-Jacques et al., 2023) by changes in the environment, such as temperature (Stachowski et al., 2022; Wolff et al., 2023). Detailing the course of these conformational changes along the protein functional life cycle, for example through time-resolved crystallography (Pearson & Mehrabi, 2020; Kupitz et al., 2014), can reveal deeper biological insights (Kang et al., 2015; Buhrman et al., 2010; Johnson et al., 2019) and new features to leverage for drug design (Carlson, 2002; Kim et al., 2023).

X-ray crystallography is still the foremost method to determine three-dimensional protein crystal structures with high spatial resolution, often better than 2 Å. While structures solved using crystallography are often simplified to representations as single models, this ignores their dynamic nature and subtle yet important alternative states (Bourne et al., 2022). Rather, explicitly modeling multiple protein conformations that are present throughout the crystal and the resulting electron-density map promises richer insights into the protein conformational landscape. Yet, even at high resolutions (better than 2 Å), detecting these conformations is challenging since minor states are often masked by more predominate ones. Additionally, genuine signal may be obscured by noise (Lang et al., 2010) or mistaken as ordered water molecules, which can concurrently alternate with corresponding changes in the protein (Lang et al., 2014). Few tools exist that can disentangle whether these electron-density features represent signal or noise (Lang et al., 2010) or that are able to discern the specific combination of states that best explain the density (Riley et al., 2021; Burnley et al., 2012; Wankowicz et al., 2023).

We previously reported FLEXR, a tool that creates multi-state models by sampling electron density around side-chain dihedral angles and maps that signal to specific rotameric states (Stachowski & Fischer, 2023). Alternate states are then explicitly added to models using Coot building tools (Casañal et al., 2020). We showed that FLEXR finds previously unappreciated conformational states in ligand binding sites and produces multi-state models without compromising model quality (Stachowski & Fischer, 2023). FLEXR was first released as a command-line interface (CLI) tool. Here, we present a graphical user interface (GUI) with a point-and-click menu to run the entire FLEXR workflow. To show the ease and accessibility of the FLEXR GUI, we interrogated a typical use case, a `straight from the Protein Data Bank' (PDB) structure of the flexible therapeutic target KRAS. In less than a minute, FLEXR generated a multi-state model that not only reproduced all flexible sites in the deposited model but also revealed previously missed side-chain conformations. Having the GUI available as a Coot plugin allows the user to immediately and interactively interrogate FLEXR models with the extensive validation and modeling toolboxes available within Coot.

2. FLEXR GUI

2.1. Implementation, availability and installation

FLEXR is implemented in Python 3.9 and the GUI was designed using the toolkit GTK4 (GTK Team, 2023), the same one as used by Coot 1 (Emsley, 2023). Python was chosen because of its simplicity and the potential to integrate with other structural-biology programs, many of which are written in Python. Phenix (Liebschner et al., 2019) is required since it is used to run Ringer (Lang et al., 2010), which is available in the mmtbx library (https://cctbx.github.io/mmtbx/mmtbx.html). To ensure consistent column labels and follow the original Ringer protocol (Fraser et al., 2011), map coefficients are recalculated from deposited reflections with phenix.maps prior to running Ringer. FLEXR starts from a single-conformer model; if alternative side-chain conformations are present these are removed using phenix.pdbtools remove_alt_confs = True. Otherwise, all Phenix steps are run with default settings. At the click of a single button, the FLEXR GUI interfaces with these programs to perform the entire process from calculating map coefficients and computing electron densities to model building. Complete instructions for downloading, installing and using the GUI version of FLEXR are documented in the supporting information and at https://github.com/TheFischerLab/FLEXR-GUI. For the tutorial, Coot 1 (Emsley, 2023) should be installed using Homebrew (https://brew.sh/). FLEXR is not backwards compatible with earlier versions of Coot. The FLEXR GUI was tested on Macs with either Intel or silicon processor architectures running macOS Sonoma. Future development will focus on making FLEXR available on Linux and Windows operating systems.

2.2. Functional description

2.2.1. Evaluating electron density

The inputs to the FLEXR GUI are coordinates (PDB or mmCIF) and an MTZ file containing reflections and map coefficients. The model should be near deposition quality [Fig. 1(1)]. It is crucial that FLEXR has a solid starting point to explore all dihedral angles at maximum signal-to-noise ratio in the electron-density map. Ringer is used to sample electron density around each dihedral angle of each side chain in the input model (Lang et al., 2010) [Fig. 1(2)]. FLEXR uses the find_peaks function from the SciPy library (Virtanen et al., 2020) to find heightened levels of electron density above a user-defined threshold (default 0.3σ) [Fig. 1(3)]. The angle at which each electron-density peak occurs is recorded for each dihedral angle of a residue. If the `Ringer plotting' option is selected, these plots can be used to monitor peak-detection accuracy [Fig. 1(3)].

Figure 1 FLEXR workflow. (1) Users should start with a near deposition quality model that has solvent modeled in PDB or mmCIF format and an MTZ file containing reflections and map coefficients. (2) Ringer is used to compute electron density around side-chain dihedral angles. (3) Electron densities above a user-defined threshold (usually 0.3σ; dotted line) are evaluated for peaks (red stars). Peak intensities and rotation angles are recorded. (4) Peaks detected around the dihedral angles of each side chain are assembled across dihedral angles into possible rotamers. (5) These combinations are checked against the ideal rotamer library (Lovell et al., 2000) to reject outliers and to determine which rotamer the series of angles corresponds to. (6) Rotamers that pass this check are built into models using Coot. (7) The final FLEXR models are ready for manual inspection and refinement before deposition.

2.2.2. Determining rotamers

For each side chain, all possible combinations of angles for each peak detected in Section 2.2.1 are assembled across all dihedral angles [Fig. 1(4)]. Each combination is then cross-checked against the `ideal rotamer library' (Lovell et al., 2000) [Fig. 1(5)]. Combinations that have tolerable geometry (within 30°) are assigned a `rotamer name' from the matching entry; for instance, `p', `t' or `m' for the three ideal threonine rotamers. Where a measured rotamer matches multiple entries in the rotamer library, the rotamer that occurs most often, as surveyed in the ideal rotamer library, is selected for building.

2.2.3. Multi-conformer model building

Rotamers are automatically built into models using Coot building functions [Fig. 1(6)] and displayed directly in the Coot workspace. The output models should be further inspected and refined prior to deposition [Fig. 1(7)]. An update from the first release of FLEXR is that rotamers for a particular residue are added to the model in decreasing order of electron-density signal strength so that rotamers with stronger signal are added to the model first. The signal strength for a potential rotamer is the sum of the relative peak intensities across all dihedral angles. This is important as the new option `Max. number of alts/residue' allows users to define the maximum number of conformers built per residue for an entire structure (default = 3). By setting this constraint to a lower value, fewer conformers will potentially be added overall and the model will have fewer low-signal conformers. Likewise, setting this constraint to a higher value will allow more conformers to be added to the model, which will have lower signal and typically be refined to lower occupancy. After inspecting the FLEXR model, the GUI makes it easy for users to test different settings and produce an internally consistent model that maximally captures weak density features of interest.

2.3. Graphical user interface

FLEXR was initially released as a tool available only through the CLI (Stachowski & Fischer, 2023). To expand FLEXR's use, the GUI version presented here gives users a point-and-click method to access some of the same options and outputs available in the CLI version (Fig. 2; see also Section S1.2 of the supporting information, point 6). This overcomes potential bottlenecks of the CLI version, which requires several independent steps including providing pre-computed Ringer densities and providing a single-conformer model as input. The GUI option seamlessly performs these steps, given that the user has Phenix installed. The GUI is loaded as a Coot 1 extension where it is available in the top menu bar. The GUI was designed to match the look and feel of Coot so that it is intuitive to crystallographers familiar with Coot. Users can start FLEXR in the GUI by simply providing a coordinate file (PDB or mmCIF) and an MTZ file containing map coefficients and reflections. As part of the workflow, map coefficients will first be recalculated with Phenix to ensure consistent column labeling and then Ringer will compute electron densities. Alternatively, users have the option to supply pre-computed Ringer densities, for instance if they want to use a different map such as an OMIT map to minimize model bias (Terwilliger et al., 2008). The interface also gives users options to control some of the FLEXR decision making via the following parameters: the electron-density threshold for peak finding (default = 0.3σ), the maximum number of conformers to build for each residue (default = 3), whether to add atoms with alternative conformations for the entire residue or just the side chain starting at the Cα atom (default = ALL), and the option to produce plots of electron-density values with peak-finding information (default = False). FLEXR produces a multi-state model and a single-state model, both of which are shown in Coot alongside the original model. By using the many validation and modeling tools native to Coot, users can then inspect and further alter FLEXR models or repeat the process with different settings. A tutorial on using the FLEXR GUI is provided in the supporting information and on GitHub.

Figure 2 The FLEXR GUI. Once loaded, the FLEXR extension is available from the top menu bar in Coot 1 (red arrow).

3. Results

To illustrate the utility of the GUI, we applied FLEXR to a recently reported structure of KRAS-G12D bound to an inhibitor (PDB ID 8txg, resolution 1.5 Å; Cheng et al., 2023). In addition to its therapeutic significance (Punekar et al., 2022), KRAS is known to be structurally dynamic (Ostrem et al., 2013) and has four residues modeled with alternative conformations at a resolution of 1.5 Å. KRAS features a P-loop containing residues G12 and G13, which are typically mutated in cancers, and has two highly dynamic Switch I and II regions that control substrate and inhibitor binding, respectively (Ostrem et al., 2013). Using the GUI, we tested whether FLEXR could reproduce alternative conformations in the deposited model and find additional sites of flexibility. We downloaded coordinates and the 2F_o − F_c map coefficients and reflections directly from the PDBe (https://www.ebi.ac.uk/pdbe/) using the `Fetch PDB & Map using EDS…' function in Coot. We ran FLEXR from the GUI using default settings, which took ∼24 s to recalculate map coefficients, 3 s to calculate Ringer electron-density values and 13 s to build the FLEXR multi-conformer model. The entire process to produce a FLEXR model took ∼40 s. If desired, the user can also create plots of Ringer values, which in this case took an additional 50 s. Refinement often improves the fit of the terminal atoms of ideal rotamers in the electron density and yields occupancy information for the relative populations of alternative states. To this end, both deposited and FLEXR models were refined with three macrocycles of Phenix using default settings and without solvent updating.

The FLEXR model retained R values nearly identical to the re-refined deposited model (Table 1). The FLEXR model had no rotamer outliers but a heightened clashscore. Clashes like these might be from misplaced water molecules or rotamers (Table S1 and Fig. S2 of the supporting information) or water molecules that should be carefully modeled with alternative positions by the user (Richardson et al., 2018). FLEXR reproduced flexibility at all 4 previously modeled multi-conformer sites and found an additional 12 sites that are missing in the deposition, totaling 16 total residues with alternative conformations [Figs. 3 and 4(a)]. Of these, 10 residues were modeled with 2 alternative conformations and 6 residues were modeled with 3 conformations (Fig. 3). One example of FLEXR reproducing a deposited alternative conformation is Cys118 [Fig. 4(b)]. FLEXR discovered new areas of flexibility throughout the structure, most notably at areas near the 2 binding sites [Fig. 4(a)], including Lys42 [Fig. 4(c)], Ser136 [Fig. 4(d)], Asp105 [Fig. 4(e)] and Lys5 [Fig. 4(f)]. FLEXR has the option to add atoms for alternative conformations beginning at the Cα atom or for all atoms in a residue, including backbone atoms. The utility of this is captured in the cases of Ser136 [Fig. 4(d)] and Asp105 [Fig. 4(e)]. In both cases, FLEXR found a second conformation and corresponding backbone conformational heterogeneity. A general benefit of the ability to build all atoms of the residue is that the strain of neighboring residues may be relieved upon refinement, leading to backbone heterogeneity and improved geometry (Stachowski & Fischer, 2023). For Lys5, the negative peak in the F_o − F_c map and insufficient 2F_o − F_c density suggest that the conformation in the deposited structure is incorrect [Fig. 4(f)]. FLEXR found an additional conformation at this position and the misplaced rotamer from the deposited model was corrected during refinement. Alternative conformations with low occupancies like Lys42 conformer C [Fig. 4(c)] should be removed from the model before deposition and the remaining two rotamers adjusted to better satisfy the density. Overall, FLEXR increased the structural heterogeneity of 8txg from ∼2 to ∼10% of residues being modeled as flexible.

Table 1
Quality metrics for deposited, re-refined and FLEXR models

As an example, using the FLEXR GUI, we investigated a recently reported structure of human KRAS-G12D bound to an inhibitor (8txg, resolution 1.5 Å). After running FLEXR with default settings and re-refining with three macrocycles with default settings in Phenix, we obtained a final model with global quality metrics similar to those in the re-refined deposited model (Chen et al., 2010).

	Deposited	Deposited-refined	FLEXR
Rwork	0.183	0.180	0.178
Rfree	0.205	0.210	0.208
Rgap	0.022	0.030	0.030
Clashscore	1.42	1.78	6.07
Rotamer outliers (%)	0	0	0
Rama-Z (whole)	−0.37	−0.56	−0.53

Figure 3 Summary plots produced by FLEXR for 8txg. A Venn diagram showing how often specific residues have alternative conformations modeled between the original model and the FLEXR model (left). A plot showing the number of alternative conformations per residue in the FLEXR model (right).

Figure 4 FLEXR reveals flexibility across KRAS-G12D. (a) Applying FLEXR to the KRAS-G12D structure (8txg, resolution 1.5 Å) revealed flexible sites along the Switch I and P-loop regions and adjacent to the Switch II region. This image was rendered in PyMOL (Schrödinger, New York). (b) Cys118 is an example of an alternative conformation that was present in the deposited model and reproduced by FLEXR. Examples of new conformational states include (c) Lys42, (d) Ser136, (e) Asp105 and (f) Lys5. Arrows in (d) and (e) indicate backbone conformational heterogeneity discovered by FLEXR. Fo − Fc maps are shown as green and red mesh and are contoured at ±3σ, respectively. 2Fo − Fc maps are shown as dark blue mesh and are contoured at 1σ. Maps shown here were calculated from the deposited model as provided by PDBe. Maps calculated from the refined FLEXR model are shown in Fig. S1.

4. Conclusions

Here we describe a GUI for FLEXR that is available as a Coot 1 plugin. This approach has three advantages over the previously reported CLI version. First, the GUI connects users with a point-and-click approach to running FLEXR, which will accommodate those who are unfamiliar with command-line tools. Second, the GUI version integrates tools from Phenix, namely Ringer, to create a seamless workflow for producing multi-state models starting from a refined model. Third, implementing the GUI within the Coot framework means that models produced by FLEXR are immediately available for further inspection and analysis using the comprehensive suite of tools available in Coot. Using a recent KRAS structure, we illustrated how the FLEXR GUI can quickly and easily produce multi-state models. Future implementations will expand FLEXR's compatibility with different map types and with performing single-conformer modeling. Ultimately, we anticipate that the FLEXR GUI will be a useful tool to retrospectively and prospectively interrogate protein structures for flexibility, which is often missed otherwise yet is key to understanding biology and informing ligand discovery and design.

5. Related literature

The following references are only cited in the supporting information for this article: Smart et al. (2012) and Yamashita et al. (2023).