>

Integrative Modeling: From Sparse Distances to Conformational Models

Combining DEER restraints with Rosetta and deep learning

Conformational changes are an essential component of the functional cycles of many proteins, and their characterization often requires an integrative approach (Sala et al., 2022). Distance distributions from double electron-electron resonance (DEER) spectroscopy provide unique restraints on the width of conformational ensembles, which makes them well suited to integrative structural modeling (Schiemann et al., 2021). Our work focuses on turning these distributions into models of alternative conformations and on defining how reliable those models are.

Diagram of three lines of evidence converging on one ensemble model: DEER distance distributions constrain the populations of coexisting states, cryo-EM structures constrain the structure of each state, and simulation and AlphaFold ensembles propose candidate states.
Each method constrains part of the ensemble and none constrains all of it. Solid outline: measured. Dashed outline: predicted. Schematic, not to scale.

DEER restraints in Rosetta

We described an approach for integrating DEER restraints into Rosetta to model alternative conformations from an initial experimental structure. A multilateration algorithm harnesses sets of interconnected spin-label pairs to identify the optimal rotamer ensemble at each residue that fits the DEER decay in the time domain. The algorithm yields distance distributions comparable to those of established analysis packages, with the advantage that fitting of the DEER decay and optimization of the rotamer ensemble are coupled. Applied to the protonation-dependent transition of the multidrug transporter PfMATE, the approach reached a deviation of less than 2 Å Cα root-mean-square deviation from the experimental inward-facing structure. By decreasing spin-label rotamer entropy, the method yields models that are both more accurate and more closely clustered (del Alamo et al., 2021).

ConfChangeMover

ConfChangeMover (CCM) is a method built into Rosetta that models conformational changes from sparse experimental data. It can rotate and translate secondary structural elements and modify their backbone dihedral angles in regions of interest. We benchmarked CCM on soluble proteins with simulated Cα-Cα distance restraints and on membrane proteins with sparse experimental DEER restraints, and in both benchmarks CCM outperformed state-of-the-art Rosetta methods. Because the Rosetta framework accepts a wide variety of experimental data, the approach extends beyond DEER (Sala et al., 2022).

DEERFold: guiding AlphaFold2 with distance distributions

DEERFold is a modified version of AlphaFold2 that incorporates experimental distance distributions into the network architecture. Using the OpenFold platform, we fine-tuned AlphaFold2 on structurally dissimilar proteins to model distance distributions between spin labels explicitly. We benchmarked DEERFold in switching the predicted conformations of a set of membrane transporters with experimental DEER distributions, and sparse sets of simulated distributions showed that the approach generalizes to a large set of water-soluble and membrane proteins. The intrinsic performance of AlphaFold2 substantially reduces the number of distributions, and the accuracy of their widths, needed to drive conformational selection, which increases experimental throughput (Wu et al., 2025).

A workflow for a new transporter

  1. Generate an ensemble of candidate conformations by MSA subsampling or SPEACH_AF.
  2. Choose spin-label sites that discriminate among the candidates.
  3. Measure DEER distance distributions in lipid nanodiscs under defined ligand and lipid conditions.
  4. Fit the time-domain data with a model-based analysis that returns components, fractional amplitudes and their uncertainty (Hustedt et al., 2021).
  5. Drive Rosetta CCM or DEERFold with the distributions, and compare the resulting models with cryo-EM structures and simulations.

Restraints can also disprove a model

DEER distances in nanodiscs combined with cysteine cross-linking showed that the multidrug transporter MdfA assumes a relatively flexible outward-closed/inward-closed conformation in a membrane. Neither the substrate nor protonation induced large-scale conformational changes. Instead, we identified a substrate-responsive lateral gate that is open toward the inner leaflet and closes on drug binding, which suggests a modified cycle that does not invoke the canonical elements of alternating access (Yardeni et al., 2020).

Training opportunities

Rosetta modeling with DEER restraints; fine-tuning and benchmarking of DEERFold on a new target.