Disordered proteins get all-atom ensembles in one diffusion pass, matching NMR and SAXS without reweighting

A diffusion model that is not trained per sequence now yields all-atom conformational ensembles of disordered proteins, and handles folded domains and disordered regions inside one model. Structure predictors such as AlphaFold and RoseTTAFold assign low confidence to disordered proteins and regions, while existing ensemble generators often need back-mapping from coarse-grained representations or reweighting of a structural pool. On 32 single-chain disordered test sequences with NMR and/or SAXS data, IDPForge reaches a normalized total X-EISD score of 0.89, versus 0.57 for IDPConformerGenerator, 0.58 for IDPFold, 0.58 for CALVADOS, 0.39 for STARLING and 0.20 for idpGAN.

Generation starts from random noise and denoises residue rigid frames step by step, reusing ESMFold's folding block and structure module while taking a noisy structure as an extra input. The model carries no multiple sequence alignment module and is not retrained for each new sequence, unlike approaches such as DynamICE that train per sequence or per set of experimental restraints. Training used 1882 disordered sequences and 10,074 CASP12 sequences, 198,399 conformers in total, of which 188,200 came from disordered sequences; test sequences share less than 30% identity with training sequences.

Experimental data can bias sampling at inference time through gradients, without training an extra classifier. For alpha-synuclein and Sic1, unguided IDPForge trails other methods on PRE error, a problem shared by CALVADOS, STARLING and a99SB-disp simulations. Guiding with PRE restraints only for residue-number differences greater than 10 lowers alpha-synuclein PRE mean absolute error from 4.116 Å to 2.565 Å and Sic1 from 3.503 Å to 1.978 Å, while other data types stay in good agreement.

Given a folded template, the model partially denoises only the disordered region, sampling terminal and linker disordered segments while keeping the folded domain. Few methods model disordered regions in the context of folded domains at all-atom resolution, and stitching-based approaches screen a pre-generated pool of single-chain conformations. Across ABL2, the SLC26A9 STAS domain, human prion protein, IL-6, PRAME20 and Cullin-1, IDPForge and CALVADOS both sample folded domains with average RMSD to templates of at most 2 Å, whereas AFflecto introduces very little structural variation in folded domains.

The demonstrated scope is 32 single-chain disordered test sequences with NMR and/or SAXS data, plus six disordered-region systems with folded domains; the latter cover only terminal and linker disordered segments, and folded domains come from experimental structures or AlphaFold2 predictions rather than de novo prediction. The model is released as an open-source resource with code and test ensembles.

The authors note the model was not trained on sequences longer than 200 amino acids and that computational cost scales quadratically for larger disordered proteins and regions; quantifying the spatial relationship between linker disordered segments and neighboring folded domains remains limited, and the test set includes only terminal and linker disordered segments. Unguided PRE errors for alpha-synuclein and Sic1 are relatively large, a problem shared by the other methods.