I've been working on some ML models for quite some time and recently had my curiosities about CNN's, I was wondering if there were ways to feed strings as inputs for a CNN model?
I worked in image classification with CNN, however, I couldn't find a straightforward way for utilizing strings directly or at least converting them to an image/graph intermediate, obviously taking a SS of the string and using it as input would be a silly approach.
The datasets I'm using are of chemical compounds (represented as strings of SMILES).
I would appreciate it if anyone could give some practical advice on how I could use these SMILES strings in a CNN or at least convert them into some form of an image/graph.
Thanks in advance :)
PS: I could also work with something like using X= [a1, a2, a3 ... an] as the input variable (where a's are different descriptors), which is what I did/do with ML models.
Dear Naeem A.
These papers might be helpful, have a look:
1. https://utorontomist.medium.com/cnn-based-on-smiles-representation-of-compounds-for-detecting-chemical-motif-f672bfd16fc3
2. https://pubs.acs.org/doi/10.1021/acs.jcim.0c01127
3. https://www.sciencedirect.com/science/article/pii/S1046202319303433
Kind Regards
Qamar Ul Islam
You could use 1D CNN or convert the smile to 2d images and use 2D CNN too.
Qamar Ul Islam I'm aware of some these papers, however, they're limited to the specific approaches they chose whereas I'm looking for a more generalized approach, moreover, most of these articles focus mainly on the results hence I find it hard to understand there code and/or reproduce it, Thanks though.
Vikas Ramachandra Interesting approach, could you elaborate a little more on the 1d CNN part?
For example, if X is my np.array with the smiles and y a np.array of the features.
Are you suggesting something like this;
- model = Sequential()
- model.add(Conv1D(32, (3, 3), input_shape=X.shape[1:]))
- model.add(Activation('relu'))
- model.add(MaxPooling1D(pool_size=(2, 2)))
- # some extra layers #
- model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
Did I get it right?
As for converting the smiles to 2D images, did you mean using the 2D chemical depictions as input? (If this is what you meant, I've tried, its accuracy is terrible :)
@Naeem Abdul Ghafoor i also have few experience with NLP. But it can process the string data like sentence or word. As you have string data you can find some NLP methods to solve your problem.
- Badges
- Science topic