IDPFold

An intrinsically disordered protein does not have a structure; it has a distribution. The standard way to obtain one is microseconds of all-atom molecular dynamics, run with an IDP-specific force field and reweighted against NMR or SAXS data to correct that force field's biases — hundreds of GPU hours per sequence, and only for the few dozen IDPs characterized well enough to reweight against. That is a hard constraint on proteins making up over 40% of eukaryotic proteomes.

IDPFold learns the distribution once, at training time, and then samples it for new sequences. It is a sequence-conditioned score-based diffusion model over protein backbone frames, developed in Hai-Feng Chen's group at Shanghai Jiao Tong University with collaborators at Fuzhou University, BioMap and Shandong University. Given a sequence, it draws hundreds of backbone conformations from a diffusion process conditioned on ESM-2 embeddings — no multiple sequence alignment, no experimental restraints, no per-target simulation. The preprint appeared on bioRxiv in May 2024 and was revised through September 2024; the peer-reviewed version was published in Advanced Science in October 2025.

The design bet is a two-stage curriculum. IDPs are structurally under-represented in the PDB, so the model is first pretrained on folded, well-resolved structures to learn backbone geometry, then fine-tuned on disordered ensembles to learn how far and how loosely real IDP chains extend. That second stage is what turns a folded-protein generator into an IDP one: without it, the paper shows generated ensembles collapsing toward over-compact states.

#Key Features

• MSA-free ensemble generation: Conformations are conditioned on ESM-2 embeddings alone, sidestepping alignments that are least informative precisely where sequences are poorly conserved.
• Backbone-level resolution: Unlike coarse-grained generators that emit only Cα traces, it produces full backbone frames, so dihedral angles, secondary chemical shifts and scalar couplings follow directly.
• Two-stage training curriculum: Pretraining on crystal and NMR structures supplies local geometry; fine-tuning on coarse-grained IDP trajectories supplies the global disorder statistics experimental structures lack.
• Structured-to-disordered sampling: One model samples both compact, partially folded states and extended disordered states, making it usable on chains that mix ordered domains with disordered regions.
• Practical sampling cost: A 300-conformation α-synuclein ensemble takes roughly 20 minutes on an A100, against hundreds of hours for the MD it is benchmarked against.

#Technical Details

The denoising network is a variant of the AlphaFold2 structure module: an initialization block plus four denoising blocks, each pairing Invariant Point Attention over nearby residues with a Transformer for long-range context (4 IPA layers, 2 Transformer layers, 8 heads, 256-channel single and 128-channel pair representations). Diffusion follows the SE(3)-equivariant frame parametrization used by FrameDiff, treating each residue's rotation on the SO(3) manifold and its translation under a VP-SDE, with the codebase adapted from Str2Str. The model has 17.8M parameters — small, because ESM2-650M supplies the sequence representation. Pretraining used 25,495 experimental structures (15,051 X-ray structures at ≤2.5 Å and 30% redundancy, plus 10,454 NMR conformations from 539 systems) for 1.12M iterations; fine-tuning used 77,600 all-atom conformations back-mapped from 3,880 IDRome coarse-grained trajectories longer than 256 residues, for 0.44M more.

On a 27-system held-out benchmark, IDPFold reached a radius-of-gyration error of −0.06, against −0.12 for idpGAN, −0.24 for AlphaFlow and −0.12 for AF-cluster, with the best backbone-level RMSDs on Cα (0.65 ppm) and Cβ secondary chemical shifts, J(HN,Hα) couplings and N-HN RDCs, at 0.95 conformation validity. Against four all-atom force fields on seven proteins with converged 30 μs trajectories, its Rg error was lower than all four and its chemical-shift accuracy comparable to a99SB-disp. Its weaknesses are documented: under-sampling of highly extended conformations for long chains, overestimated helical propensity in the α-synuclein N-terminus, and slightly lower local validity after fine-tuning.

#Applications

The model suits work where the quantity of interest is an ensemble average rather than a structure: estimating compactness for phase-separation studies, predicting NMR observables before committing spectrometer time, screening disordered regions across a proteome for expansion or residual structure, and generating starting ensembles for simulation. Needing only a sequence, it reaches the large majority of IDPs with no experimental restraint data, and its backbone-level output makes all-atom reconstruction a short energy-minimization step.

#Impact

IDPFold belongs to the shift from simulating IDP ensembles to generating them, alongside AlphaFlow, idpSAM and later sequence-conditioned generators such as IDPForge — a distinct model from a different group, despite the near-identical name. Its contribution is the demonstration that a small backbone diffusion model, pretrained on folded structures and fine-tuned on coarse-grained disorder data, can match all-atom force fields on experimental observables at a fraction of the cost, and its 27-system benchmark gives the subfield a shared yardstick. Caveats remain: the repository describes itself as a test version, training code is unpublished, checkpoints are distributed through a Google Drive folder, no license is attached, and evaluation covers single chains only.