Pi-Ensemble

A sugar transporter has an inward-facing structure and an outward-facing structure, both in the PDB. Hand its native sequence to a structure predictor and you get one of them back — usually the same one every time — and nothing along the path between. Recovering that path conventionally means microseconds of molecular dynamics, or retraining a generative model on MD trajectories that few systems have.

Pi-Ensemble, from the Shukla lab at the University of Illinois Urbana-Champaign, takes a different route: it interpolates in sequence space rather than in Cartesian coordinates. Each anchor backbone is passed through ProteinMPNN, which returns a per-residue probability distribution over amino acid identities compatible with that pose. The two distributions are mixed as P_λ(i,a) = λP₁(i,a) + (1−λ)P₂(i,a), the maximum-likelihood sequence is decoded, and that sequence — often far from the native one — is folded by a pretrained structure predictor. The predicted structure replaces one anchor and the cycle repeats, filling in the region between the endpoints. The native sequence is finally backmapped onto the generated backbones with cg2all side-chain repacking and OpenMM energy minimization.

The framework trains nothing. It contributes no weights of its own and runs frozen third-party checkpoints — ESM-3, Boltz, ESMFold2 or BioEmu-1 for structure generation — which the user installs separately. Its premise is that inverse-folding models assign different residue preferences to different backbone poses of the same protein, and that this sensitivity, usually treated as noise, encodes what steers a structure predictor off its default answer. It follows ESMDynamic as the same group's second route to protein dynamics without simulation.

#Key Features

• Interpolation in inverse-folding probability space: The mixing weight λ does not move atoms; it shifts which residue identities the prompt asserts and lets the structure predictor resolve the geometry. No noising, restraints or MSA subsampling involved.
• Two λ schedules: The serial schedule sweeps a fixed grid in both directions, replacing one anchor each round. The batch schedule instead solves for the critical λ values at which the decoded sequence flips at some position and samples the midpoints, so compute goes only to prompts that differ.
• Swappable frozen backends: Structure predictors and inverse-folding models are registered behind abstract interfaces and chosen in a YAML config, so a new checkpoint is a wrapper class, not a fork of the workflow.
• Seed generation for parallel MD: Clustering the ensemble yields starting structures for many short trajectories. On adenylate kinase, 298 seeds of 10 ns each explored more of the lid/NMP angle space, and converged faster, than two 1.5 μs runs from the anchors alone.

#Technical Details

The framework has no parameters of its own; effective capacity is whatever backend is registered. Reported runs use 20 interpolation rounds over λ ∈ [0,1], ProteinMPNN at T = 0.1, and ESM-3 at T = 0.7 with 8 sampling steps. Against 132 μs of reference MD on the Arabidopsis AtSWEET13 transporter, the RMSD from each generated structure to its nearest MD frame peaks at 1.5 Å, and the ensemble spans the gating-distance free-energy landscape, including occluded intermediates, with no prior knowledge of those coordinates. On an 18-protein benchmark spanning 83 to 527 residues and anchor-pair RMSDs of 3.32–13.97 Å, the ESM-3 serial variant passed the strict RMSD filter for above 90% of structures on eight of the 18 systems, with mean RMSD-axis coverage across 20 bins generally between 50% and 90%. Against 51 DEER distance distributions measured on T4 lysozyme, predicted and measured means agree to a mean absolute error of 2.32 Å, with a median distribution overlap integral of 0.64.

#Applications

The method suits any system with two known endpoint structures — active and inactive GPCR states, inward- and outward-facing transporters — where the question is what lies between them. Intermediates serve as hypothesis-generation material, as initial images for umbrella sampling or string-method calculations, and as seeds for distributed simulation: on glutamine-binding protein, 3.76 μs across 188 seeds reproduced the coverage of a 53.6 μs reference dataset, with denser sampling of the transition region.

#Impact

The authors position Pi-Ensemble as a conformational proposal mechanism, not an ensemble sampler: the structures are not guaranteed to trace a physical transition pathway, and their density is not a thermodynamic distribution, so free energies, populations and kinetics need downstream MD or experimental reweighting. It inherits its backends' failure modes: ProteinMPNN can emit low-complexity or poly-alanine prompts that induce artificial helices and disorder, and BioEmu structures grown from interpolated prompts were often discarded for steric clashes. It is least reliable when a transition involves extensive unfolding, ligand-dependent stabilization or a change in oligomeric state. Within those bounds it shows that pretrained inverse folding and structure prediction, coupled in a loop, produce behavior neither was trained for. The preprint awaits peer review.