Refold: Refining Protein Inverse Folding with Efficient Structural Matching and Fusion

Yiran Zhu; Changxi Chi; Hongxin Xiang; Wenjie Du; Xiaoqi Wang; Jun Xia

cs.LGMar 15, 2026

Refold: Refining Protein Inverse Folding with Efficient Structural Matching and Fusion

Authors: Yiran Zhu, Changxi Chi, Hongxin Xiang, Wenjie Du, Xiaoqi Wang, Jun Xia

Organizations: AIMS Lab, HKUST (GZ), Guangzhou, China · North China Electric Power University, Baoding, China · Westlake University, Hangzhou, China · Hunan University, Changsha, China · University of Science and Technology of China, Hefei, China · Northwest Polytechnical University, Xi’an, China · HKUST, Hong Kong SAR, China

Abstract

Protein inverse folding aims to design an amino acid sequence that will fold into a given backbone structure, serving as a central task in protein design. Two main paradigms have been widely explored. Template-based methods exploit database-derived structural priors and can achieve high local precision when close structural neighbors are available, but their dependence on database coverage and match quality often degrades performance on out-of-distribution (OOD) targets. Deep learning approaches, in contrast, learn general structure-to-sequence regularities and usually generalize better to new backbones. However, they struggle to capture fine-grained local structure, which can cause uncertain residue predictions and missed local motifs in ambiguous regions. We introduce Refold, a novel framework that synergistically integrates the strengths of database-derived structural priors and deep learning prediction to enhance inverse folding. Refold obtains structural priors from matched neighbors and fuses them with model predictions to refine residue probabilities. In practice, low-quality neighbors can introduce noise, potentially degrading model performance. We address this issue with a Dynamic Utility Gate that controls prior injection and falls back to the base prediction when the priors are untrustworthy. Comprehensive evaluations on standard benchmarks demonstrate that Refold achieves state-of-the-art native sequence recovery of 0.63 on both CATH 4.2 and CATH 4.3. Also, analysis indicates that Refold delivers larger gains on high-uncertainty regions, reflecting the complementarity between structural priors and deep learning predictions.

Figures & tables

Figure 1. Template-based methods use database-derived structural priors to achieve high local precision but often show poor OOD generalization. Deep learning approaches generalize better to new backbones yet struggle with fine-grained local structure. Refold integrates these complementary strengths to improve both OOD generalization and local precision.

Figure 2. Overview of Refold. (a) We perform structural matching to search matched neighbors $k$ for the target backbone $\mathcal{S}$ . (b) These neighbors are organized into a Stacked Neighbor Alignment matrix $\mathbf{A}$ , where row 0 serves as a base anchor derived from the base logits $\operatorname*{argmax}(z_{\text{base}})$ . A Similarity-Weighted Fusion Module distills $\mathbf{A}$ into reference logits $z_{\text{ref}}$ and linearly fuses them with $z_{\text{base}}$ to form $p_{\text{fused}}$ . (c) To mitigate the adverse impact of low-quality neighbors, a Dynamic Utility Gate guided by the Inference-Time Statistics modulates the fusion, falling back to the base prediction when structural priors are unreliable.

Figure 3. Schematic of the Similarity-Weighted Fusion Module. From the base logits $z_{\text{base}}$ , we construct a Stacked Neighbor Alignment matrix $\mathbf{A}$ : row 0 serves as the base anchor $\operatorname{argmax}\!\left(z_{\text{base}}\right)$ , while rows $1,\dots,K$ contain aligned matched neighbor tokens. The module embeds $\mathbf{A}$ , applies row-wise smoothing to mitigate local noise, and aggregates structural priors via attention weighted by the reliability bias $\beta$ . This process produces reference logits $z_{\text{ref}}$ , which are then residually fused with the base logits to yield $z_{\text{fused}}=z_{\text{base}}+\lambda\cdot z_{\text{ref}}$ .

Model	Perplexity $\downarrow$		Recovery $\uparrow$		CATH version
Model	Short	Full	Short	Full	4.2	4.3
GraphTrans ( Ingraham et al., 2019 )	8.39	6.63	0.28	0.36	$\checkmark$
StructGNN ( Ingraham et al., 2019 )	8.29	6.40	0.29	0.36	$\checkmark$
ESM-IF † ( Hsu et al., 2022 )	8.18	6.44	0.31	0.38		$\checkmark$
GCA ( Tan et al., 2022 )	7.09	6.05	0.33	0.38	$\checkmark$
GVP ( Jing et al., 2021 )	7.23	5.36	0.31	0.39	$\checkmark$

Table 1. Results comparison on the CATH dataset. “Short” refers to proteins with length

L<100

, and “Full” corresponds to the complete test set. Sequence recovery is reported as a fraction (0–1). Values in parentheses indicate the absolute improvement (Recovery

\uparrow

) or reduction (Perplexity

\downarrow

) compared to the base model on the specific dataset version. Best results are bolded.

Model	Test Set	Recovery $\uparrow$	Perplexity $\downarrow$
ProteinMPNN	TS50	0.54	3.93
ProteinMPNN	TS500	0.58	3.53
Refold	TS50	0.60 (+0.06)	3.78 (-0.15)
Refold	TS500	0.63 (+0.05)	3.51 (-0.02)

Table 2. OOD Generalization performance on TS50 and TS500 test sets. We employ Refold (Frozen) trained on CATH 4.3. Best results are bolded.

Uncertainty Regime	Base Rec.	Refold Rec.	$\Delta$
Low ( $H<1.5$ )	0.78	0.83	+0.05
Mid ( $1.5\leq H<2.3$ )	0.34	0.55	+0.21
High ( $H\geq 2.3$ )	0.18	0.45	+0.27
Overall (All)	0.44	0.61	+0.17

Table 3. Uncertainty-stratified recovery analysis based on base model predictive entropy

H_{j}

\Delta

denotes the absolute recovery gain. Note that gains are most pronounced in high-entropy regimes.

Figure 4. Site-wise recovery states (sampled proteins). We sample $N$ proteins and visualize the first 50 residues. Each cell denotes a residue-level transition from the Base model to Refold: Neg (wrong $\rightarrow$ wrong), Pos (correct $\rightarrow$ correct), Neg $\rightarrow$ Pos (wrong $\rightarrow$ correct, correction), and Pos $\rightarrow$ Neg (correct $\rightarrow$ wrong, error). Corrections (Neg $\rightarrow$ Pos) tend to appear in localized segments, while error events (Pos $\rightarrow$ Neg) are sparse, suggesting that Refold performs targeted, structure-aligned refinements rather than indiscriminate perturbations.

Source Contribution
Configuration	Recovery	Perplexity
Full Model (Ours)	0.61	3.77
w/o Refold (Baseline)	0.44	5.94
w/o base model (Priors Only)	0.52	6.40
Mechanisms
w/o TM-score Bias	0.58	4.21

Table 4. Ablation studies on ProteinMPNN and CATH 4.3.

Row	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	…
Base	M	S	S	E	G	Y	S	G	E	S	S	S	H	C	A	P	E	P	G	C	T	R	G	T	S	S	R	P	R	…
Refold	M	S	S	Y	G	W	S	G	L	S	C	D	H	C	A	P	E	P	G	C	T	R	G	T	C	S	R	P	R	…
Nbr	K	H	Q	Y	G	W	Q	G	L	Y	C	D	K	C	I	P	H	P	G	C	V	H	G	I	C	N	E	P	W	…

Table 5. Token-level comparison of Base, Refold, and matched neighbor (Nbr) sequences for the 2kb9.A case. Correct predictions are colored green, errors red, and Nbr positions highlighted in yellow indicate structure-guided corrections derived from the matched neighbor.

Figure 5. Sensitivity analysis of the number of matched neighbors

K

. The performance gain saturates around

K=10

, indicating that the model aggregates structural priors effectively without requiring a large number of neighbors.

Model	Params	Latency	Overhead
ProteinMPNN	1.7M	13.41ms	–
Refold	3.5M	18.22ms	+35.9%
KWDesign	4.1B	656.24ms	+4793.7%

Table 6. Computational efficiency analysis on an NVIDIA A800 GPU. Latency is reported per protein with a batch size of 16. Overhead denotes the relative increase in latency compared to the base model ProteinMPNN

Explore similar work

Sep 1, 2026

SymFold: Synergizing Evolutionary and Structural Priors for Accurate Protein Inverse Folding

Protein inverse folding aims to recover amino acid sequences for a given 3D protein structure, underpinning broad applications such as enzyme engineering and drug discovery.Current methods often follow a serial pipeline, in which a structure encoder predicts a coarse sequence, which is then refined by protein language models (PLMs). However, because PLMs only perform post-hoc sequence edits, the refinement is bounded by the quality of upstream predictions.Thanks to recent multimodal protein language models (MPLMs), we could directly encode structure to generate sequences with pretrained structural knowledge, but we observe that they are not effective for inverse folding. Therefore, we introduce a symmetric dual-path architecture that both leverages PLMs for pretrained sequence evolution knowledge and MPLMs for pretrained structural knowledge to iteratively guide protein sequence generation.Through extensive experiments across standard protein inverse folding benchmarks, our method achieves state-of-the-art performance, surpassing prior approaches, and ablation studies validate the rationale of our symmetric design, revealing a promising direction for the community.

Handong Wang, Jiaxin Qi, Baisheng Lai +1

Computer Network Information Center, Chinese Academy of Sciences · University of Chinese Academy of Sciences

Protein Language Models Protein Design

Ask PDF

Sep 28, 2026

SPINET: Sheaf Protein Inverse Folding Network

Proteins change shape as they function, yet most inverse folding models predict amino acid sequences from a single, fixed backbone. A central challenge in protein engineering is to design proteins that undergo specific motions, which requires accounting for how their structures change over time. This motivates inverse protein folding conditioned on protein motion. We introduce SPINET, which predicts sequences from molecular dynamics trajectories. It uses cellular sheaves to represent residue interactions within each frame and recurrent units to integrate information across frames, then predicts all amino acids in a single pass. We evaluate SPINET on mdCATH and ATLAS, where it outperforms all evaluated static and ensemble baselines in sequence recovery. On mdCATH, it achieves 56.7% top-1 recovery, compared with 44.5% for the strongest static baseline and 40.7% for the strongest ensemble baseline. We also evaluate whether the predicted sequences are compatible with conformations sampled along the target trajectory. On mdCATH, they achieve a median TM-score of 0.760, and structural recovery favors target conformations over unrelated decoys for 99.5% of test domains.

Jens Lundsgaard, Colin Mikulski, Zhixuan Yan +1

Department of Computer Sciences, University of Wisconsin–Madison, Madison, WI, USA · Department of Mathematics, University of Wisconsin–Madison, Madison, WI, USA · Department of Biomedical Engineering, University of Wisconsin–Madison, Madison, WI, USA +4

Protein Design

Ask PDF

Mar 6, 2026

Property-driven Protein Inverse Folding With Multi-Objective Preference Alignment

Protein sequence design must balance designability, defined as the ability to recover a target backbone, with multiple, often competing, developability properties such as solubility, thermostability, and expression. Existing approaches address these properties through post hoc mutation, inference-time biasing, or retraining on property-specific subsets, yet they are target dependent and demand substantial domain expertise or careful hyperparameter tuning. In this paper, we introduce ProtAlign, a multi-objective preference alignment framework that fine-tunes pretrained inverse folding models to satisfy diverse developability objectives while preserving structural fidelity. ProtAlign employs a semi-online Direct Preference Optimization strategy with a flexible preference margin to mitigate conflicts among competing objectives and constructs preference pairs using in silico property predictors. Applied to the widely used ProteinMPNN backbone, the resulting model MoMPNN enhances developability without compromising designability across tasks including sequence design for CATH 4.3 crystal structures, de novo generated backbones, and real-world binder design scenarios, making it an appealing framework for practical protein sequence design.

Xiaoyang Hou, Junqi Liu, Chence Shi +3

BioGeometry · School of Computer Science, Peking University · Mila - Québec AI Institute +3

Protein Design Preference Alignment Preference Optimization

Ask PDF