DeepMDisPre

A protein that changes shape to do its job has more than one right answer, but a distance predictor is usually built as though it has one. Feed an alignment into a contact or distance network and it returns, for every residue pair, a probability distribution over distance bins — and an objective drawn from single crystal structures pushes that distribution toward one sharp peak. For a residue pair spanning a hinge between two domains, or a transporter caught between inward- and outward-facing states, that single peak is an average of two real distances and describes neither.

DeepMDisPre is built to keep the second peak. It is trained on a purpose-built corpus mixing proteins with a single known structure and proteins with several experimentally observed conformations, so it learns when to spread probability mass across separated distance bins instead of collapsing it. Tested on 114 proteins with multiple conformations, the predicted distributions came out more multi-peaked for flexible residue pairs than for rigid ones, and for two case-study proteins the authors rebuilt both conformations from the predicted distances rather than one averaged structure.

The method comes from Fujin Zhang and Guijun Zhang's group at Zhejiang University of Technology, published in IEEE/ACM Transactions on Computational Biology and Bioinformatics in 2024. It is also the distance engine inside the group's own PAthreader, which threads these profiles against a clustered structure database to find remote-homolog templates, and the two ship the same checkpoint. The lab later approached multiple conformations from the alignment side instead with AlphaFlex.

#Key Features

• Multi-peak distance output: every residue pair gets a full 37-bin distribution rather than one predicted distance, and the shape of that distribution is itself the flexibility readout.
• Mixed-conformation training corpus: single-structure and multi-conformation proteins are trained on together, which is what teaches the network that a split distribution can be the correct answer.
• Triangle updates with axial attention: pair features are refined by ingoing and outgoing triangular multiplicative updates and row/column attention before entering a deep dilated convolutional trunk.
• Frozen language-model features: a pretrained ESM MSA Transformer supplies per-residue embeddings and symmetrized row-attention maps as fixed inputs, never fine-tuned.
• Single-command inference: one shell script runs alignment search, feature extraction and prediction end to end from a FASTA file, writing the distance distribution as a compressed array.

#Technical Details

Inputs are concatenated to 447 channels per residue pair: 47 covariance and alignment-statistic features, 128 coevolution channels from an outer product over projected MSA embeddings, 128 channels of tiled one-dimensional features, and 144 symmetrized row-attention maps from the frozen esm_msa1b_t12_100M_UR50S MSA Transformer, which reads a fixed-size subsample of the query alignment. A projection reduces this to 64 channels; four attention blocks then apply outgoing and ingoing triangular multiplicative updates followed by an eight-head axial encoder layer, and 61 residual two-dimensional convolution blocks with dilations cycling through 1, 2, 4 and 8 follow. The output is symmetrized and passed through a 1×1 convolution to 37 bins covering 2–20 Å in 0.5 Å steps plus a final bin for pairs beyond 20 Å. Alignments come from HHblits against UniRef30_2020_03.

Evaluation spans 114 proteins with multiple conformations and 279 proteins with a single structure. On the single-structure set, average contact accuracy and the average TM-score of 3D models built from the predicted distances both improved on the comparison method reported in the paper. The released package covers inference only, from a checkpoint distributed as a tarball alongside the PAthreader downloads.

#Applications

The distributions are useful wherever one structure is not the answer: enumerating candidate conformations of transporters, hinged multi-domain enzymes and other proteins whose function depends on motion, and supplying distance restraints to a folding or threading pipeline that would otherwise commit to a single fold. Because the flexibility signal falls out of the peak structure, one run both produces modelling restraints and flags which residue pairs are worth probing experimentally. Practically it is a workstation tool rather than a web service: it needs a local UniRef30 copy, HHblits, and two vendored PSIPRED-derived binaries, and the sequence database it was configured against is pinned to a 2020 snapshot.

#Impact

The contribution that travels is the framing — reading multi-modality in a predicted distance distribution as signal about conformational heterogeneity rather than as prediction error. Its demonstrated reach so far is largely within its own lab, where PAthreader and the systems built on top of it consume its profiles. Distribution is thin: the weights are a plain tarball served from the PAthreader download page with no checksum or version tag, the repository carries no license file, and neither training code nor the training corpus was released, so the dataset cannot be audited and the network cannot be retrained from what is public.