MF-ProtDisMap

Inter-residue distance maps remain a useful intermediate representation in protein structure modelling: they are cheap to predict, they feed directly into folding and refinement pipelines, and — unlike the end-to-end coordinates produced by AlphaFold 2 — they expose a per-pair error estimate that downstream samplers can weight. Two families of features drive them. Coevolutionary signal, extracted from multiple sequence alignments, captures physical contacts implied by correlated mutation. Protein language model embeddings capture sequence semantics and long-range dependency, and work where alignments are shallow. Most methods commit to one or the other.

MF-ProtDisMap (Multi-Feature Protein Distance Map) is a real-valued distance predictor that uses both. Developed at Shandong Agricultural University and posted as a preprint in June 2025 before publication in the International Journal of Biological Macromolecules, it draws coevolutionary features from a frozen MSA Transformer and sequence features from a frozen ESM2, compresses both with group pooling, and fuses them in a trained module the authors call Diff-former — a diffusion process paired with triangular attention.

The design point is that the two frozen backbones are complementary rather than redundant, and the paper's ablations are the evidence: either feature set alone is clearly worse than the fusion, and the fusion is worse than the fusion plus Diff-former.

#Key Features

• Dual frozen backbones: Row and column attention maps from the MSA Transformer supply coevolutionary structure; ESM2 attention supplies sequence-level and long-range signal. Neither backbone is retrained, so the trainable footprint stays small.
• Group pooling: Raw attention tensors are large — L×L×144 per MSA attention type and L×L×1440 from ESM2 — and are reduced to L×L×12, L×L×24, and L×L×36 feature maps by grouped pooling rather than a lossy linear projection.
• Diff-former module: A forward/reverse diffusion process refines the MSA feature map and is combined with the original features by a residual connection, then passed through a triangular-attention Former network that enforces the geometric consistency constraints distance maps must satisfy.
• Real-valued rather than binned: Distances are regressed directly instead of being discretised into bins, and can be thresholded afterwards to recover binary contacts.
• Consistent gains across benchmarks: Improvements over single-feature variants hold on four independent test sets, including two CASP editions and CAMEO.

#Technical Details

MSAs are generated with HHblits 3.3.0 against UniRef30_2020_03 and subsampled to 64 rows before entering the MSA Transformer, whose 12 encoder layers with 12 heads each yield the attention maps used as coevolutionary features. Training uses the benchmark set curated by Yang et al. — 15,051 sequences reduced to 14,942 after excluding those longer than 1,022 residues to fit ESM2's context — split into 11,000 training and 3,942 validation proteins. The model was trained in PyTorch on a single Nvidia H800 GPU with the Adam optimiser, an initial learning rate of 1e-3, weight decay of 0.01, and batch size 1.

On the 90-protein 4.05_release test set MF-ProtDisMap achieves a mean absolute error of 2.20 Å, an RMSE of 3.40 Å, and an R² of 0.90, versus 2.56 Å MAE and 5.38 Å RMSE for ESMFold. Thresholded to contacts, it reaches ROC 84.56%, PR 81.01%, F1 72.86%, and recall 63.28%. Ablations on the same set give MAE 2.47 Å for MSA features alone, 2.43 Å for ESM2 features alone, 2.22 Å for the naive fusion, and 2.20 Å with Diff-former added. Generalisation was checked on 40 CASP13 targets, 138 CASP14 targets, and 129 CAMEO targets from June–September 2020.

#Applications

Predicted distance maps serve as restraints for structure modelling, and the authors pair the model with a reconstruction pipeline, ProtDisFold, that samples candidate conformations in parallel and ranks them by TM-score; reconstructions are demonstrated on targets including the wheat detoxification enzyme Fhb7-GST and the plant NLR immune receptor Sr35, with backbone dihedral angles taken from the native structures to assist folding. The predictor is most useful for teams that need explicit, interpretable geometric restraints — for refinement, for hybrid experimental modelling, or for scoring — rather than a single end-to-end structure.

#Impact

MF-ProtDisMap is a compact demonstration that alignment-derived and language-model-derived features remain complementary in 2025, and that diffusion can be used as a representation-refinement operator on a 2D feature map rather than as a generative sampler. Its reach is limited by the practicalities of release: no code or weights have been published, the data availability statement points only to the public benchmark datasets, and inference still requires an HHblits search against a large sequence database, so the method is not a drop-in alternative to single-sequence predictors.