bioRxivBioinformaticsPreprint
PartiDesign: A structure-sequence co-design paradigm for practical all-to-all biomolecules design
Abstract
Engineering interactions among biomolecules offers a route to creating new biological functions. Although unified models have advanced biomolecular structure prediction and design, efficient and controllable generation across molecular modalities remains challenging. A key challenge is to couple three-dimensional structure generation with sequence or chemical-identity design across chemically diverse systems. Here, we introduce PartiDesign, a unified structure-sequence co-design model that extends the partially latent framework of La-Proteina to proteins, nucleic acids and small molecules. PartiDesign jointly generates molecular structures and their corresponding sequences or chemical identities, while providing fine-grained control over secondary structure, chemical bonding and amino acid preferences. Across diverse benchmarks for the de novo design of proteins, nucleic acids and small molecules against specified targets, PartiDesign achieved the highest estimated successful-design throughput among the evaluated methods on most benchmarks under the adopted compute-budget protocols. Its configurable constraints further support cyclic peptide and antibody design, allowing generation to be tailored to specific structural and chemical requirements. PartiDesign was developed using 128 Huawei Ascend 910B accelerators, each with 64 GB of memory. The source code is available at https://github.com/MegaGatlingPea/PartiDesign-official, and a freely accessible web server at https://cadd.zju.edu.cn/partidesign/ enables users to generate binding partners without programming expertise. Together, these capabilities provide a versatile and accessible framework for all-to-all biomolecular design, with applications in molecular engineering and drug discovery.
- assay: protein sequence
- assay: DNA methylation
- assay: proteomics
- research focus: drug response / drug discovery
- research focus: protein design
- approach: deep learning
- approach: transformer / foundation model
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- BoltzGen: Toward Universal Binder Design
Hannes Stark, Felix Faltings, … Tommi Jaakkola·bioRxiv·Nov 20, 2025
- BioPipelines: Accessible Computational Protein and Ligand Design for Chemical Biologists
Gianluca Quargnali, Pablo Rivera‐Fuentes·Computational and Structural Biotechnology JournalPeer-reviewed·Mar 1, 2026
- GeoFlow-V2: A Unified Atomic Diffusion Model for Protein Structure Prediction and De Novo Design
BioGeometry Team·bioRxiv·May 6, 2025
- BoltzDesign1: Inverting All-Atom Structure Prediction Model for Generalized Biomolecular Binder Design
Yehlin Cho, Martin Pačesa, … Sergey Ovchinnikov·bioRxiv·Apr 6, 2025
- Toward De Novo Protein Design from Natural Language
Fengyuan Dai, Shiyang You, … Fajie Yuan·NaturePeer-reviewed·Oct 21, 2026
- A high-level programming language for generative protein design
Brian Hie, Salvatore Candido, … Alexander W. Rives·bioRxiv·Dec 21, 2022