A Hybrid AI-Model for Resolving Molecular Structures Integrating Sequence and Cryo-Electron Microscopy Data

Short description of the innovation study

Novel de novo protein structure prediction models (Alphafold3, Boltz), which surpass Alphafold 2 in prediction accuracy, incorporate a generative AI component (diffusion). However, this diffusion component can result in protein structure hallucinations, which are difficult to detect computationally and can be disastrous for drug development applications.

This study focuses on the development and training of a hybrid de novo protein structure prediction model built on top of the Boltz/OpenFold3 framework, capable of using cryoEM data alongside protein sequence data as input. This study, led by PUXANO, is conducted in collaboration with CSIC (for image processing expertise) and (for high-performance computing expertise). Through the development of this model, the Harmony Team aims to combine the speed of de novo protein structure prediction with the confidence of experimental protein structure determination.

Organisations involved: 

End User: Puxano

HPC provider: IT4I

Domain expert: CSIC

IT4I is a member of the NCC Czech Republic.