Chemical Structure Image Translation via Encoder-Decoder Tokenization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for translating images of chemical molecular structures into textual identifiers are inefficient, with low accuracy and high manual effort required, especially when dealing with historical data sets that are small or of poor quality, and are not suitable for automatic search due to limitations in machine learning model performance.
Innovation Solution
A system using an encoder-decoder architecture with an attention mechanism processes images of chemical molecular structures to generate standardized images, which are then translated into textual identifiers by predicting unique tokens and calculating correctness probabilities to select the most accurate sequence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If existing machine learning methods are used for translating chemical structure images to textual identifiers, then the process can be automated, but the accuracy is limited to maximum 90% and requires optimal conditions that are not met by historical data
Solution Approach 1:
The patent segments the translation process into distinct modules: image pre-processing stage, feature extraction stage, and textual identifier generation stage. This segmentation allows each module to be optimized independently, with pre-processing specifically addressing historical data quality issues through techniques like noise reduction and normalization, thereby improving overall accuracy while maintaining automation.
Solution Approach 2:
The patent applies preliminary action by implementing comprehensive pre-processing steps before the main translation task. Historical images undergo normalization, contrast enhancement, and artifact removal prior to feature extraction. This preliminary preparation ensures that suboptimal historical data is transformed into a format suitable for accurate machine learning processing, enabling high accuracy without requiring optimal original conditions.
2Measurement precision
If manual methods are used for converting scanned chemical structures to machine-readable format, then accuracy can be maintained, but the process becomes highly time-consuming and requires significant manual effort
Solution Approach 1:
The patent introduces an intermediary automated system that bridges manual accuracy requirements and productivity needs. The machine learning model acts as an intermediary that processes images through multiple stages: automated pre-processing, feature extraction with attention mechanisms, and identifier generation. This intermediary system achieves accuracy comparable to manual methods while providing hundreds of times faster processing throughput.
Solution Approach 2:
The patent replaces the mechanical manual process of chemists visually inspecting and transcribing chemical structures with an automated computer vision system. The system uses convolutional neural networks and attention mechanisms to automatically extract structural features and generate textual identifiers, substituting human visual-cognitive-motor processes with automated image processing and pattern recognition algorithms, thereby achieving both accuracy and high productivity.
3Extent of automation
If existing solutions are applied to historical data sets with image corruption, then the translation can proceed, but performance is significantly reduced and requires extensive manual intervention
Solution Approach 1:
The patent implements beforehand cushioning by incorporating robust pre-processing steps that anticipate and compensate for image corruption in historical data. Techniques include noise filtering, contrast normalization, and artifact removal applied before the main translation process. These cushioning measures ensure that corrupted historical images are repaired and standardized, providing a reliable input foundation that maintains high translation accuracy and enables fully automated processing without manual intervention.
Data Source
AI summary
Disclosed is a system and a method for translating an image of a structural formula of a chemical molecule into a textual identifier therefor utilizing unique tokens for each of known entities. The method comprises pre-processing the image of the structural formula to generate a standardized image; processing the standardized image using an encoder-decoder architecture, wherein an encoder generates embeddings for features in the standardized image and a decoder is implemented to associate each of the features to one of the unique tokens; recurrently processing each of the features in the standardized image for predicting corresponding unique token to generate multiple possible sequences, and dynamically calculating a correctness probability for each of the generated sequences; selecting one of the sequences with highest calculated correctness probability; and generating the textual identifier, as an output, based on the selected one of the sequences.