PAthreader

Two proteins can share a fold and almost no sequence. When that happens, a profile-to-profile search like HHsearch — the template channel inside AlphaFold 2 — has nothing to lock onto, and the target is modelled without a template. That matters because AlphaFold 2's accuracy tracks template quality: across 551 benchmark targets, the share of models reaching TM-score 0.9 fell from 88% where a good template existed to 49.8% where the best one scored below TM-score 0.7.

PAthreader answers this by matching structures rather than sequences. It threads a query against PAcluster80, a master structure database built by clustering the PDB together with the AlphaFold Protein Structure Database at 80% structural similarity, and compares two profile types that survive sequence divergence: a multi-peak inter-residue distance profile predicted for the query by the group's in-house DeepMDisPre network, and a histogram structure profile extracted from each cluster seed. A three-track alignment — sequence alignment, residue-pair alignment, and profile alignment — searches for the query-template correspondence that satisfies those distance constraints, and a trained convolutional network with axial self-attention reranks the survivors. The identified templates are then handed to AlphaFold 2 in place of HHsearch's.

The method comes from Kailong Zhao and Guijun Zhang's group at Zhejiang University of Technology, with Stan Z. Li at Westlake University, published in Communications Biology in March 2023. A second result rides on the same machinery: because remote homologs of a protein are its evolutionary siblings, the regions that recur most consistently across them are read as folding intermediates, giving a predicted folding order from templates alone.

#Key Features

• Structure-space threading: alignment is scored between predicted distances and observed structure profiles, so a template with 15% sequence identity is as findable as one with 60%.
• AlphaFold DB as a template library: predicted models supply fold families the PDB never sampled, lifting average template TM-score by 23.4% on the hardest targets in an ablation.
• A learned reranker: pDMScore predicts DMScore, a global fold-similarity metric, from physical and geometric features of the aligned structure, complementing the alignment score.
• Drop-in template channel: the templates replace HHsearch's inside an otherwise unchanged AlphaFold 2, requiring no retraining of the folding network.
• Folding-pathway readout: per-residue frequency scores across the identified templates segment the chain into fragments that fold early and fragments that fold late.

#Technical Details

PAcluster80 holds 56,805 clusters covering 106,275 PDB chains and 100,912 AlphaFold DB models — the pLDDT ≥ 90 subset of the 564,449 predicted structures then available for 48 organisms — clustered by TM-align at a TM-score threshold of 0.8. Just over half the AlphaFold DB models (55.7%) fell into the 34,701 PDB-seeded clusters, and the remainder formed 22,105 clusters with no PDB representative. Query distance profiles come from DeepMDisPre, run on HHblits alignments against UniRef30, and are multi-peak, so flexible regions contribute several plausible distances rather than one. The pDMScore network voxelizes the aligned template, mixes 3D and 1D convolutions into a 128-channel feature map, and processes it with three axial multi-head attention blocks (8 heads, alternating over rows and columns) followed by 15 residual convolution blocks predicting per-residue-pair Cα deviations; templates are ranked by a linear combination of pDMScore and the alignment score.

On 551 nonredundant SCOPe 2.07 proteins with templates above 30% identity excluded, PAthreader templates averaged TM-score 0.725, 12.2% above HHsearch (0.646) and 5.2% above LOMETS3 (0.689). In a continuous three-month CAMEO blind test it ranked first among public servers, with structure-modelling TM-scores of 0.882, 0.900 and 0.876 across the three reporting windows against 0.878, 0.896 and 0.869 for AlphaFold 2. Inference averaged 1.25 hours per target on ten CPU cores.

#Applications

PAthreader suits pipelines that already run AlphaFold 2 but stall on orphan or multi-domain targets, where a better template is worth more than a deeper alignment: on multi-domain proteins its templates averaged 12.3% above HHsearch, and correct domain orientation is exactly what end-to-end predictors lose. The authors applied it to 17 SARS-CoV-2 proteins and to folding-pathway predictions for 37 proteins, 7 of which match published hydrogen-exchange and NMR experiments. A free web server and a downloadable package with both pretrained checkpoints serve one-off and batch use.

#Impact

The work is an early demonstration that the AlphaFold database is usable as a template library rather than only an endpoint, and that structural clustering makes it searchable at scale. The group's own DeepAssembly uses PAthreader templates in place of HHsearch's and reports a 14.5% reduction in inter-domain distance error from the swap. Practical adoption is constrained: the pipeline needs local copies of multi-gigabyte databases plus HHblits, PyRosetta and TM-align, PAcluster80 is pinned to fixed PDB and AlphaFold DB snapshots, the code carries no license file, and the folding pathways predicted for 30 human proteins remain experimentally unverified.