ChemPix: automated recognition of hand-drawn hydrocarbon structures using deep learning

Area of Science:

• Computational chemistry
• Machine learning applications in chemistry
• cheminformatics

Background:

• Current methods for inputting molecular structures into chemistry software demand specialized knowledge and resources.
• There is a need for user-friendly tools to bridge the gap between chemical drawings and digital representations.

Purpose of the Study:

• To develop ChemPix, an offline tool for recognizing hand-drawn hydrocarbon structures using machine learning.
• To simplify the process of converting chemical drawings into machine-readable formats like SMILES (Simplified Molecular Input Line Entry System).

Main Methods:

• A neural image captioning model, combining a convolutional neural network (CNN) encoder and a long short-term memory (LSTM) decoder, was employed.
• Training datasets were created using RDKit-generated molecular images with augmentations and a crowd-sourced set of hand-drawn structures.
• An ensemble approach, using a committee of neural networks, was utilized to improve prediction accuracy and confidence.

Main Results:

• The ensemble model achieved 76% accuracy in recognizing hand-drawn hydrocarbon structures.
• Considering the top 3 predictions increased the accuracy to 86%.
• The system provides a confidence value for each prediction based on the agreement within the network committee.

Conclusions:

• ChemPix effectively recognizes hand-drawn hydrocarbon structures, significantly lowering the barrier to entry for computational chemistry.
• The developed tool demonstrates the potential of machine learning in automating chemical structure representation.
• This approach enhances accessibility to chemistry software by removing the need for domain expertise or expensive tools for structure input.