DS-ProGen: A Dual-Structure Deep Language Model for Functional Protein Design

TL;DR

DS-ProGen tackles inverse protein folding by integrating backbone geometry and surface features into a unified autoregressive design framework. Using a dual-branch encoder and a GPT-2–style decoder, it achieves state-of-the-art sequence recovery on PRIDE (61.47%) and demonstrates strong capacity for ligand and ion interaction design, while maintaining structural fidelity as reflected in RMSD and TM-Score analyses. The work highlights the benefit of multimodal structural encoding for meeting both global topology and local chemical constraints, with implications for drug discovery and synthetic biology. By combining geometry driven global cues with local surface chemistry, DS-ProGen advances functional protein design and sets the stage for multi-task, structure-guided design workflows across large biological datasets.

Abstract

Inverse Protein Folding (IPF) is a critical subtask in the field of protein design, aiming to engineer amino acid sequences capable of folding correctly into a specified three-dimensional (3D) conformation. Although substantial progress has been achieved in recent years, existing methods generally rely on either backbone coordinates or molecular surface features alone, which restricts their ability to fully capture the complex chemical and geometric constraints necessary for precise sequence prediction. To address this limitation, we present DS-ProGen, a dual-structure deep language model for functional protein design, which integrates both backbone geometry and surface-level representations. By incorporating backbone coordinates as well as surface chemical and geometric descriptors into a next-amino-acid prediction paradigm, DS-ProGen is able to generate functionally relevant and structurally stable sequences while satisfying both global and local conformational constraints. On the PRIDE dataset, DS-ProGen attains the current state-of-the-art recovery rate of 61.47%, demonstrating the synergistic advantage of multi-modal structural encoding in protein design. Furthermore, DS-ProGen excels in predicting interactions with a variety of biological partners, including ligands, ions, and RNA, confirming its robust functional retention capabilities.

Figures (12)

• Figure 1: Design protein sequences based on protein structure (backbone) and surface features.
• Figure 2: DS-ProGen integrates backbone and surface structural information to enable functional protein design. The backbone geometric encoder extracts geometric features from the N, C, and C$_\alpha$ atoms of protein structures. The surface feature encoder processes atomic types, surface points, and curvature features. The backbone and surface embeddings are projected into a unified hidden space and combined to form a single representation with the target sequence length. Then, a dual-structure fusion design decoder auto-regressively generates/designs the protein sequence conditioned on the structural context.
• Figure 3: Design cases for DS-ProGen and baselines (Length is 403). A lower RMSD indicates greater structural similarity, while a lower Identity Sequence Similarity suggests stronger diversity in the designed sequences.
• Figure 4: Structural alignment between ground truth structures (blue) and predicted structures (light brown) folded from protein sequences generated by DS-ProGen, visualized across four representative cases. Each panel highlights the global fold alignment and a zoomed-in view of the functional binding pocket, showing key residues together with the corresponding ligand or ion (e.g., ATP, GTP, $\mathrm{Cu}^{2+}$). The predicted sequences lead to highly accurate structural models, with high TM-scores, low RMSDs, and strong recovery rates, demonstrating DS-ProGen’s ability to preserve both global topology and local biochemical specificity.
• Figure 5: Topology-based splitting strategy for training set construction.