Related Experiment Video
Updated: Oct 9, 2026
Formation of Ordered Biomolecular Structures by the Self-assembly of Short Peptides
Published on: November 21, 2013
DeepSAP: An Integrated Generative and Predictive Deep-Learning Framework for Controllable Design of Self-Assembling
Yuanming Song1,2, Yuhan Wu1,2, Yuqiu Jiang1,2
1MOE Key Laboratory of Laser Life Science & Institute of Laser Life Science, College of Biophotonics, School of Optoelectronic Science and Engineering, South China Normal University, Guangzhou 510631, China.
Motivation:
Self-assembling peptides (SAPs), particularly short sequences whose amino acid composition dictates their spontaneous organisation, are emerging as versatile building blocks in biomedicine and bio-optoelectronic materials. As the thermodynamic and kinetic behaviours of peptide self-assembly are affected by intrinsic factors, such as amino acid sequence, length, and hydrophobicity, the controllable design of SAPs with tailored properties is essential to achieve specific structures and functions for various application scenarios. However, existing computational approaches, which primarily focus on prediction or heuristic screening, lack conditional generative capability of SAPs.
Results:
We present DeepSAP, an end-to-end framework integrating conditional peptide generation with predictive screening for property-guided design of SAPs. By pretraining on large-scale peptide datasets and fine-tuning on limited SAP-specific data, the generative SAPGEN model transfers general sequence knowledge to the SAPs domain, enabling the conditional generation of SAPs in low-data regimes. Subsequently, we constructed a lightweight predictive model, SAPBRF, to further screen the generated peptides. Comparative analyses demonstrated that SAPGEN enabled reliable and precise control of the physicochemical properties of the generated sequence, outperformed baseline generative models in terms of efficiency and quality across multiple design tasks. The SAPBRF predictor achieved efficient and accurate screening with high predictive performance. Finally, the complete DeepSAP framework demonstrated strong generalisation under unseen physicochemical regimes, where peptides generated and screened by the pipeline consistently exhibited stable self-assembly behavior, as further verified by molecular dynamics simulations. Our findings indicate that DeepSAP enables the conditional generation of SAPs with desired properties, holding the potential for tailored design of functional biomaterials.
Availability And Implementation:
Source code for the DeepSAP model training and evaluation is available at: https://github.com/Paul-hihihopes/Deep_SAP_Design and at DOI: 10.5281/zenodo.21522690.
Supplementary Information:
Supplementary data are available at Bioinformatics online.
