Fold-Conditioned Protein Sequence Generation From 3D Voxel Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for inverse protein design face challenges in accurately modeling the complex sequence-fold relationship due to heterogeneous latent domains and inefficient fold representation, leading to low generalization and accuracy in designing proteins that fold into desired 3D structures.
Innovation Solution
A transformer-based generative framework that learns a joint sequence-fold embedding by representing 3D protein structures as voxel densities and using transformer models to encode and decode protein sequences, enabling a unified latent space for sequence and fold information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If separate sequence encoding and fold encoding are used, then the model can process sequence and structure information independently, but the heterogeneous latent domains increase the difficulty of learning the sequence-fold relationship
Solution Approach 1:
The patent merges separate sequence encoding and fold encoding into a unified joint embedding space. The transformer encoder processes both sequence tokens and fold representation tokens together, allowing the model to learn correlations between sequence and structure in a unified latent space, thereby resolving the heterogeneity issue while maintaining architectural modularity
2Device complexity
If hand-designed fold representation methods are used, then the model structure is simple, but the representation does not capture the complete original fold space resulting in low generalization capacity
Solution Approach 1:
The patent replaces hand-designed fold representation methods with a learned representation system. A transformer encoder automatically learns the optimal fold representation from structural data, substituting manual feature engineering with a data-driven approach that captures the complete fold space and improves generalization capacity
3Adaptability or versatility
If energy minimization-based approaches are used, then the method can sample both sequence and conformational space, but the accuracy of energy functions and efficiency in sequence and conformational search are low
Solution Approach 1:
The patent replaces energy minimization-based approaches with a transformer-based generative model. Instead of relying on inaccurate energy functions and slow conformational sampling, the model directly generates sequences conditioned on fold representations, achieving both versatility in space sampling and high accuracy in protein design
Data Source
AI summary
A system and method for designing protein sequences conditioned on a specific target fold. The system is a transformer-based generative framework for modeling a complex sequence-structure relationship. To mitigate the heterogeneity between the sequence domain and the fold domain, a Fold-to-Sequence model jointly learns a sequence embedding using a transformer and a fold embedding from the density of secondary structural elements in 3D voxels. The joint sequence-fold representation through novel intra-domain and cross-domain losses with an intra-domain loss forces two semantically similar (where the proteins should have the same fold(s)) samples from the same domain to be close to each other in a latent space, while a cross-domain loss forces two semantically similar samples in different domains to be closer. In an embodiment, the Fold-to-Sequence model performs design tasks that include low resolution structures, structures with a region of missing residues, and NMR structural ensembles.