ExEnDiff

The RS peptide is a 24-residue intrinsically disordered region from a serine/arginine-rich splicing factor, and it has no fold. Point a unified protein ensemble sampler at it and the sampler returns helix-like structures anyway, because it learned conformational space from the Protein Data Bank — a catalogue of proteins that held still long enough to be resolved. The same bias yields over-compact ensembles and misplaced free-energy minima across disordered systems. ExEnDiff supplies the sampler with the one thing the PDB cannot: a measurement of what the ensemble actually looks like in solution.

The mechanism is a correction term inside the reverse diffusion process. A score-based sampler produces a conformation by integrating a reverse-time stochastic differential equation, with a learned score function supplying the drift at each step. ExEnDiff decomposes the conditional score by Bayes' rule into that learned term plus a measurement likelihood, then approximates the likelihood using manifold constraint sampling: a harmonic restraint penalizing disagreement between the observed value and the same observable evaluated on the posterior-mean structure at that timestep. Each trajectory is nudged toward configurations consistent with the measurement while the pretrained weights stay frozen, so no retraining and no per-protein optimization loop are involved.

ExEnDiff was built by researchers at the Flatiron Institute, Purdue University, and Northwestern University. It first appeared under the name EGDiff, was renamed in the second version of its preprint, and was published in PRX Life. It shares an objective with BioEmu-1 and AlphaFlow — Boltzmann-weighted ensembles rather than one structure — but reaches it by conditioning a frozen sampler on experimental data rather than training on simulation trajectories.

#Key Features

• Experimental observables as a physical prior: Guidance accepts end-to-end distance, radius of gyration, per-residue helix percentage, and RMSD to a folded reference — quantities read off FRET, small-angle X-ray and neutron scattering, circular dichroism, and NMR chemical shifts.
• Inference-time guidance over frozen weights: The correction term reshapes the sampling trajectory, not the score network, so one pretrained checkpoint serves every target.
• Sampler-agnostic formulation: The derivation applies to any diffusion-based ensemble sampler; the released implementation wires it to Str2Str, and any observable expressible as a differentiable function of atomic positions can be added as a new operator.
• Cryo-EM images as a guidance channel: A configuration-to-image operator lets noisy 2D density projections act as the observable, a conditioning modality the underlying sampler cannot otherwise accept.

#Technical Details

The guidance layer sits on Str2Str, a zero-shot conformation sampler whose score network is built from invariant point attention blocks; ExEnDiff itself introduces no new trained parameters, and the restraint strength is set per observable by a single constant. Evaluation covers ten proteins spanning 24 to 234 residues — two ordered (BBA, Villin) and eight disordered (ACTR, drkN SH3, NTail, Aβ40, PaaA2, Ash1, RS peptide, ELF3) — with ground truth from long unbiased molecular dynamics, including a 325 μs Villin trajectory from D. E. Shaw Research, REST2 simulations for ELF3, and a 1.76 ms RS-peptide trajectory generated on Folding@home and released with the paper. A composite sampler score sums Jensen-Shannon distances over end-to-end distance, radius of gyration, and solvent-accessible surface area with residue-level deviations in RMSF and helix content. Guidance improves that score on every protein tested, with radius of gyration plus secondary structure the strongest combination because it constrains a global and a local feature at once. Generating 100 conformations takes under 300 seconds on a single A100, and cryo-EM guidance holds up at a signal-to-noise ratio of 1, closely tracking the noise-free case.

#Applications

The natural use is turning a biophysical measurement into a structural ensemble: a lab with SAXS, FRET, or CD data on a disordered region gets atomistic conformations consistent with it in minutes rather than millisecond-scale simulation time. The Chignolin T8P experiment shows a second use — resolving how a destabilizing mutation reshapes an ensemble globally rather than only at the mutated site — which matters for interpreting variants in flexible regions. Because the generated ensembles cover conformational space broadly, they also seed short MD runs used to train data-driven collective variables, and supply an informative prior for cryo-EM analyses that would otherwise need converged simulation to reweight against.

#Impact

ExEnDiff argues that ensemble samplers can be corrected rather than retrained: experimental restraints, long used in MD, transfer cleanly to the reverse diffusion of a generative model at negligible cost. Code is MIT-licensed and the RS-peptide trajectory is deposited on Zenodo under CC BY. Two caveats bound the results: every guiding measurement is synthetic, derived from the same ground-truth MD used for evaluation, and the cryo-EM channel is a proof of concept on simulated images. The distributed checkpoint is the pretrained backbone rather than a new model of its own, hosted on Google Drive rather than a model registry.