Neural Network Deconvolution for Atomic Structure Determination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods in X-ray crystallography face challenges in determining the molecular structure of larger molecules and lower resolution data due to the inability to recover phase information, known as the 'phase problem', which limits the scalability and accuracy of structure determination.
Innovation Solution
Training neural networks, specifically convolutional neural networks, on known Patterson maps and atomic structures to deconvolve Patterson maps and obtain atomic structures, addressing issues like centrosymmetry and vector origin ambiguity, enabling the reconstruction of atomic coordinates from incomplete diffraction data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If direct methods are used to compute molecular structure from X-ray data, then structure determination is possible for small molecules, but it cannot scale to larger molecules or lower resolution data
Solution Approach 1:
The patent replaces traditional mechanical/mathematical iterative procedures (direct methods) with a neural network-based system. The neural network learns to directly map Patterson maps to atomic structures, substituting the complex iterative calculation process with a trained model that can handle larger molecules and lower resolution data effectively.
Solution Approach 2:
The invention changes the fundamental parameters of the problem-solving approach by transforming the input (Patterson map) and output (atomic structure) representations, and by using a neural network to learn the transformation rather than applying fixed mathematical rules. This allows the system to operate successfully at lower resolutions where traditional methods fail.
2Loss of information
If phase information is lost during X-ray crystallography, then diffraction magnitudes can be measured, but molecular structure cannot be directly reconstructed
Solution Approach 1:
The patent introduces Patterson maps as an intermediary representation that can be computed from diffraction magnitudes alone (without phases). The neural network then acts as a mediator to transform this intermediate representation back into atomic structure information, effectively recovering structural details despite the loss of phase information.
Solution Approach 2:
Instead of trying to recover phases directly from diffraction data (the traditional approach), the patent inverts the problem by computing Patterson maps from magnitudes and then using a neural network to infer atomic structures from Patterson maps. This reverse approach bypasses the phase problem entirely.
3Extent of automation
If neural networks are trained on known structures, then they can predict unknown structures, but the network must generalize beyond training data
Solution Approach 1:
The patent applies preliminary action by extensively training the neural network on a diverse set of known atomic structures and their corresponding Patterson maps before deployment. This pre-training ensures the network learns robust patterns and relationships that enable it to generalize reliably to unknown structures it has never seen before.
Data Source
AI summary
Atomic position data may be obtained from x-ray diffraction data. The x-ray diffraction data for a sample may be squared and/or otherwise operated on to obtain input data for a neural network. The input data may be input to a trained convolutional neural network. The convolutional neural network may have been trained based on pairs of known atomic structures and corresponding neural network inputs. For the neural network input corresponding to the sample and input to the trained convolutional neural network, the convolutional neural network may obtain an atomic structure corresponding to the sample.