Abstract:
Integrating the chemistry concepts about NMR and the deep learning technology
advance concepts, in here predict the SMILES structure of molecules for the given 1H
NMR spectrum. The NMR is the principal method to identify the molecular structure
and identify the content and purity of a sample.
1H NMR which is also called proton
NMR is more popular for identifying organic molecule structures among chemistry
experts and analyzing it manually, so it is time-consuming. However, it is really
challenging to implement a system to predict molecular structure for a given NMR
because it is required to identify the 1H NMR image correctly and required to generate
a correct sequence of SMILES with a meaningful molecule. That means this research
is a combination of computer vision and natural language processing (NLP)
capabilities.
Manually downloaded 1H NMR public data set, and open-source SMILES data set was
used to carry out this research. The resized NMR image was converted into its feature
vector using the InceptionV3 model in CNN referring to the transfer learning
techniques. Before training the model, SMILES was tokenized into the characters and
combined with the 1H NMR image feature vector. Then SMILES was converted into a
vector. The LSTM technique in deep learning was used to identify the hidden pattern
of SMILES characters. The Seq2seq approach is utilized to merge this CNN and RNN
model and return a single model following the encoder-decoder concepts to predict
molecule structure based on the given 1H NMR spectrum. The hyper-parameter
techniques were carried out to maximize the predictivity of this deep learning model.
This implemented model was evaluated using ROUGH techniques and with the
benchmark study. The implemented system predicts molecular structure in SMILES
notation for provided 1H NMR spectrum less than 2.5 seconds and the model accuracy
around 90%.
