Mac-Diff

Mac-Diff (Modal-aligned conditional Diffusion) is a score-based diffusion model that generates conformational ensembles for proteins it has not seen during training. It was developed by groups at East China Normal University, the East China University of Science and Technology, and Fudan University, posted to bioRxiv in February 2025 and published in Nature Machine Intelligence in 2026.

Structure prediction has largely solved the problem of predicting one structure per sequence, but many proteins function precisely by interconverting between states — allosteric switching, ligand-induced rearrangement, and disorder-to-order transitions are all invisible in a single predicted model. Molecular dynamics simulation captures this behavior but is expensive enough that adequate sampling of even a modest protein consumes substantial compute.

Mac-Diff's central argument is about what to condition on. Earlier conditional diffusion models for ensemble generation — including systems built on structure predictors — draw their sequence representation from a folding model such as AlphaFold 2, ESMFold, or OmegaFold. Those representations were optimized to produce a single most-probable structure, so they may be biased toward the dominant conformation and carry less information about the alternatives. Mac-Diff instead conditions on embeddings from a protein language model, which was trained by masked prediction across sequence databases and encodes evolutionary, structural, and functional signal without any single-structure objective steering it.

#Key Features

• LAMA-attention: The Locality-Aware Modal Alignment attention module restricts each residue-pair representation to attend only to residues likely to interact with that pair biologically, replacing the loose global cross-attention borrowed from text-to-image models with a physically motivated attention field.
• Language-model conditioning: Sequence conditioning comes from ESM-2 rather than a structure predictor, avoiding representations that were shaped by a single-conformation training objective.
• Two trained variants: Mac-Diff-MD, trained on molecular dynamics trajectories, targets conformational distributions; Mac-Diff-PDB, trained on experimental structures, targets alternative discrete states.
• Zero-shot on new sequences: Test proteins were rigorously excluded from training cohorts, and the model generates ensembles for unseen sequences without any per-protein fitting.
• Sampling speed: Ensemble generation runs over 50 times faster than the molecular dynamics simulation it substitutes for.

#Technical Details

Mac-Diff performs iterative denoising on protein backbone geometry (the target view) guided continuously by the sequence representation (the conditional view), with LAMA-attention bridging the two. Two training regimes were used. Mac-Diff-MD trained on 1,674 molecular dynamics trajectories — 371 from GPCRmd and 1,303 non-membrane proteins from the ATLAS dataset. Mac-Diff-PDB trained on Protein Data Bank structures, which offer far greater sequence and conformational diversity than available MD data. For the PDB variant, separate models were trained for sequences of length 1-128 and 128-256 to control inference cost, with ESM-2 650M and ESM-2 3B supplying residue embeddings.

On a 12-protein fast-folding benchmark, ensembles were scored by Jensen-Shannon divergence against reference MD along three axes: pairwise Cα distance distribution (JS-PwD), radius of gyration (JS-Rg), and the top-two time-lagged independent components (JS-TIC). Mac-Diff produced greater conformational diversity than competing methods at comparable distributional accuracy, with JS-TIC error 38%, 24%, and 64% lower than the best competitor on α3D, Protein B, and Homeodomain respectively. It recovered conformational substates of BPTI that had been observed only in a 1 ms simulation, and predicted both the open and closed states of adenylate kinase with TM-scores above 0.8 to each experimental structure. On the 40-protein Cfold40 benchmark, Mac-Diff-PDB showed the highest conformational diversity and the best group-level residue flexibility among the methods compared, which included MSA subsampling, AF-cluster, and AlphaFlow.

#Applications

Mac-Diff suits structural biologists and computational chemists who need the conformational landscape rather than a single model: identifying cryptic or transient pockets for structure-based drug design, characterizing allosteric mechanisms, and estimating per-residue flexibility. Its speed advantage over simulation makes it practical as a triage step, generating candidate states that focused MD runs can then refine. Protein engineers assessing whether a design will populate an intended state, and researchers studying intrinsically flexible regions where a single predicted structure is misleading, are the other natural users.

#Impact

Mac-Diff's peer-reviewed appearance in Nature Machine Intelligence, and its systematic case that language-model embeddings beat structure-predictor embeddings as a conditioning signal, make it a substantive contribution to a fast-moving area. The LAMA-attention design is also a broader point about transferring architectures across domains: cross-attention that works for text-to-image generation is a poor fit where the two modalities are related by concrete spatial constraints. Limitations remain honest in the paper's own reporting — on protein-level residue flexibility Mac-Diff-PDB scored below MSA subsampling and AlphaFlow-MD, and on JS-TIC it trailed a 100 µs simulation even while beating it on the other two metrics. Code is MIT-licensed but inference-only: sampling and evaluation scripts ship with no training loop or loss definition, the checkpoint directory is an empty placeholder, and weights come from a Google Drive link, so the model can be run but not retrained. Neither paper version is openly licensed — the preprint reserves all rights, the journal article is CC BY-NC-ND 4.0.