PPubMed19 Dec 2025Cited 1×

SurfFold: a unified model for protein inverse folding by integrating surface and structural information

Darong Li, Lian Shen, Meijia Song, Deyi Li, Juan Liu, Xiangrong Liu, Jonathan Wren

Abstract

Proteins play an indispensable role in biological systems, including providing structural support, catalyzing metabolic reactions, mediating molecular transport, transmitting cellular signals, and supporting immune defense. Therefore, precise protein design is crucial for advancing applications in drug discovery, biocatalysis, materials science, agricultural biotechnology, and environmental remediation (). Protein design generally comprises two core aspects: structure design and sequence design. In recent years, considerable progress has been achieved in predicting protein structures from amin

A figure from SurfFold: a unified model for protein inverse folding by integrating surface and structural information — fig. from the paper

§ The Valyu brief

The paper, in five minutes

Read from the full text via Valyu Contents

SurfFold combines protein structural and surface data, achieving state‑of‑the‑art inverse‑folding accuracy.

What’s new

i.Integrates backbone, side‑chain, and surface representations for inverse folding
ii.Introduces a representation‑alignment module using 8‑nearest‑neighbor aggregation and bidirectional cross‑attention
iii.Employs E(n)‑equivariant GNNs and geodesic convolutions to extract structural and surface features
iv.Uses a one‑shot decoder adapted from PiFold for efficient sequence generation

How they did it

SurfFold extracts backbone and side‑chain features with EGNNs and generates surface points via dMaSIF, processing them with geodesic convolutions. A representation‑alignment module aligns surface and structural embeddings using nearest‑neighbor mapping and cross‑attention, producing a fused representation that a one‑shot MLP decoder converts into the predicted amino‑acid sequence.

What they found

CATH4.2 test set: perplexity 2.95 ± 0.01, sequence recovery 62.05 % ± 0.54
CATH4.2 short‑chain recovery 56.18 % ± 1.27, single‑chain 54.38 % ± 0.34
TS50 benchmark: perplexity 2.90 ± 0.05, recovery 64.16 % ± 0.27
TS500 benchmark: perplexity 2.74 ± 0.02, recovery 69.32 % ± 0.40
Improves over second‑best model by ~1.3 % on full CATH4.2 test set

Caveats

✕Evaluated only on protein datasets; not tested on nucleic acids or protein‑nucleic acid complexes
✕Performance depends on quality of generated surface representations, which may limit applicability to low‑resolution structures

§ The Valyu brief

Reading the full paper and taking notes. This takes a few seconds…

§ Ask this paper

Ask a question about this paper

Valyu reads the full text and answers from what the paper actually says.

§ Related work in other archives

Similar papers in all four archives

The same idea searched in arXiv, PubMed, bioRxiv and chemRxiv at once, so a method can lead you to where it is being applied.

AarXiv4 found

PubMed

Searching the other archives…