Title:	InBio: AI-enabled design of functional protein-semiconductor periodic nanostructures
DNr:	Berzelius-2026-220
Project Type:	LiU Berzelius
Principal Investigator:	Amijai Saragovi <amijai.saragovi@fysik.lu.se>
Affiliation:	Lunds universitet
Duration:	2026-09-01 – 2027-03-01
Classification:	10203
Homepage:	https://www.saragovi.science/
Keywords:

Abstract

The requested computational resources will support InBio, a funded WASP–WISE NEST project entitled "AI-enabled design of functional protein–semiconductor periodic nanostructures for the circular economy." The project aims to develop AI-designed self-assembling protein architectures that template the growth of semiconductor nanostructures with programmable geometry and nanoscale precision. Achieving these objectives requires extensive GPU-accelerated computation for iterative protein design, structural prediction, and optimization of large multicomponent protein assemblies. The requested GPU resources will be used to: (1) Design programmable self-assembling protein architectures comprising multiple interacting protein components using state-of-the-art generative protein design methods. (2) functional protein assemblies by incorporating semiconductor-templating motifs, selective pores, and modular functional domains while maintaining structural stability, symmetry, and cooperative assembly. (3) Predict and validate large multicomponent protein assemblies using state-of-the-art AI-based structure prediction methods, followed by iterative optimization of sequence and assembly geometry. (4) Screen and optimize thusands of candidate assemblies through large-scale parallel computational workflows to identify experimentally tractable protein templates for semiconductor biofabrication. These workflows rely on modern AI frameworks, including RFdiffusion, ProteinMPNN, LigandMPNN, Chai-I, AlphaFold 3, ESM-based protein language models, and custom deep-learning pipelines. The computational complexity of these methods increases rapidly with assembly size and the number of interacting protein components, making GPUs with large on-board memory essential. Access to the Berzelius Hopper system (NVIDIA H200 GPUs) will provide the memory capacity required for the design and structure prediction of large multicomponent protein assemblies that cannot be efficiently handled on conventional GPU hardware. These resources are essential for delivering the computational objectives of the InBio project within the planned project period.