Protein Structure Prediction Using MSA and Pair Embedding Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting protein structures are computationally intensive and time-consuming, often requiring extensive search processes that consume significant computational resources and are not efficient.
Innovation Solution
A neural network system that utilizes both multiple sequence alignments (MSA) and pair embeddings to predict protein structures by alternating updates between MSA representations and pair embeddings, enabling a single forward pass through trained neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional search-based methods are used for protein structure prediction, then prediction accuracy can be achieved, but computational resource consumption and time are excessive
Solution Approach 1:
The patent replaces conventional search-based mechanical methods with a neural network-based computational system. The neural network learns protein structure prediction from training data and directly outputs predictions without performing exhaustive searches, substituting the mechanical search process with an intelligent computational model that achieves both accuracy and speed.
Solution Approach 2:
The neural network is trained in advance on a large dataset of known protein structures and sequences. This preliminary training action enables the network to encode learned patterns and relationships, allowing it to make accurate predictions rapidly during inference without needing to perform computationally intensive searches at prediction time.
2Reliability
If conventional search-based methods are used for protein structure prediction, then comprehensive structure analysis can be performed, but computational resource consumption increases significantly
Solution Approach 1:
The patent substitutes energy-intensive search-based computational methods with a neural network system that has been pre-trained to perform structure prediction. The neural network processes input sequences and generates structure predictions with significantly lower computational resource consumption while maintaining reliability through the patterns learned during training.
Solution Approach 2:
The neural network creates a computational model that copies and encodes the relationships between protein sequences and structures learned from training data. This copied knowledge allows the system to make reliable predictions without repeatedly performing the full computational search process, reducing energy consumption while preserving prediction reliability.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for predicting protein structure. An initial single embedding and initial structure parameters are obtained for each amino acid. A folding neural network, comprising multiple update blocks, processes these inputs to generate final structure parameters that characterize the predicted structure. Each update block is configured to receive an input comprising final pair embeddings, a current single embedding, and current structure parameters for each amino acid. The block processes this input to update the current single embedding and structure parameters. This iterative refinement yields the final structure parameters. The resulting predicted protein structure is then used to evaluate interactions with one or more candidate ligands, and one or more ligands are selected based on the evaluation.