Skip to main content
Back to timeline
bioRxivSource publication:

FoldMapper breaks inverse folding's circular sequence search: shape-retrieved homologs lift best-of-100 recovery to 60.0% on TED

Synopsis

Protein inverse folding must design an amino-acid sequence that folds into a given backbone, yet using evolutionary information has required searching a multiple sequence alignment with the very sequence being designed, a circular dependency. FoldMapper instead retrieves structural homologs by shape with Foldseek and accumulates their amino-acid identities, weighted by bit score, into a per-residue structural position-specific frequency profile; a multiscale graph encoder then encodes that profile on a sparse local graph before a denser structure-conditioned message-passing network decodes the sequence. On TED structures it reaches 60.0% best-of-100 sequence recovery, above the 43.2% to 56.3% range of the other tabulated methods, with 20.1% sequence diversity.

Source-provided article image: FoldMapper: Protein Inverse Folding Model with Structure-Derived Evolutionary Profiles and a Multiscale Graph Encoder
Fig. 1 · bioRxiv · Page 4

Interpretation

Inverse folding models can now take evolutionary information from the structure itself, without first having a sequence to build a multiple sequence alignment from. Earlier deep-learning methods used backbone geometry alone, while evolutionary statistics were read from a multiple sequence alignment searched with a query sequence; in inverse folding that query is the sequence being designed, so the dependency is circular. On TED structures FoldMapper reaches 60.0% best-of-100 sequence recovery, against 43.2% to 56.3% for the other tabulated methods, and its sequence diversity is 20.1%, above ESM-IF at 12.6% and LM-Design at 16.1%.

The route is to retrieve structural homologs by shape first, then pool their amino-acid identities into a per-residue frequency profile that seeds decoding. Foldseek searches in a mode that compares only 3Di states and excludes amino-acid identity, so retrieval itself carries no sequence bias; hits are accumulated into an L×21 count matrix weighted by bit score and normalized row by row into a structural position-specific frequency (SPSF) profile. The multiscale graph encoder builds two k-nearest-neighbour graphs over the same backbone: a sparse k=30 graph encodes the profile into node features in Stage 1, and a denser k=48 graph refines them on geometry before autoregressive decoding; removing the profile drops PDB recovery from 64.9% to 52.2%.

Perspective

Evaluation covers 10,000 TED domains and 2,126 PDB chains in TEDMonoBench, 576 heterodimer complexes in PPIBindBench, and 758 RFdiffusion de novo backbones (30 monomers, 571 binders, 157 symmetric oligomers); the results describe performance on these benchmarks and settings.

With no homolog hit the profile is the zero vector, recovery falls to 25%, and the model falls back to structure-only prediction; peptide binders change by at most 2 percentage points across ablations, which the authors attribute to short fragments rarely retrieving informative homologs; on symmetric oligomers even FoldMapper's mean and best scTM differ by 0.374, so rare foldable samples dominate the result.