Denatured

First EC number level accuracy (blue) and loss (orange) across training epochs.

Inspiration

Our team was drawn to this challenge due to the bioengineering backgrounds of some of our team members. In many of our chemistry and anatomy classes, we have learned about the application and importance of enzymes. Enzymes affect chemical reactions that are happening in our bodies every day. This project gave us the opportunity to further extend our knowledge and use our computer science skills to solve the challenge. This project is just one of the many pieces of the puzzle that come together to help us understand the complexity of living things.

What it does

Our project takes the reactants and products of a chemical reaction in the form of SMILES and translates it into a machine-readable format. This is done by tokenizing individual molecules for relevant chemical features. The tokenization is inspired by byte-pair-encoding that is utilized in NLP. It is able to recognize recurring subunits across databases that may be of importance. These tokens are then assigned unique IDs using a dictionary so that they can be input into our model. Each reaction is input as a series of tokens. We used a Long short-term memory (LSTM) artificial recurrent neural network (RNN) to train our classifier.

How we built it

We used Google Colab to work in a collaborative coding setting. To translate the given data to meaningful input to our machine, we used SMILES Pair Encoding. SMILES PE learns the vocabulary of SMILES substrings from the ChEMBL dataset and uses the vocabulary to tokenize the values. We then created our own library of the tokens and assigned unique IDs to each token. The IDs/tokens present in each reaction were then input into our model. The model was trained using a LSTM RNN. A model was used to predict the class of the enzyme (the first EC number). The training set was then split by the main class of each reaction and 7 subclass models were trained to predict the subclasses. When evaluating the models, the main model first predicted the class of the reaction; the reaction was then fed into the corresponding submodel to determine the second EC number.

Challenges we ran into

Determining a way to represent and quantize the SMILES input

Our team looked at several different methods to extract key features from SMILES such as using DeepChem featurizers. Ultimately we chose to use SMILES PE as it creates chemically explainable and human-readable SMILES substrings to be used as tokens.

Determining good hyperparameter values for our RNN

Our team experimented with using different batch sizes, epochs and other hyperparameters to see what gave the best results. A particular challenge was balancing the accuracy of the model with both the risk of overfitting, as well as the runtime of fitting the model.

Accomplishments that we're proud of

This was everyone on the team's first time making a deep learning (DL) model. It was exciting to see everything come together!
This was also our team's first time working with NLP and tokenization! Finding a way to preserve key information and features while quantizing the data was something new and of great importance to us.
Our model was able to classify reactions for the first EC number with an accuracy of 79%! We were surprised and excited about this result!
For our team, this was one of the first projects that we worked on to bridge the gap between chemistry and computer science. It was interesting to see the intersection between the two fields and the breadth of applications for DL.

What we learned

We learned about processing language and string inputs to extract useful and quantifiable features.
We learned how to build and test deep learning models using TensorFlow Keras.
We learned about different tools and APIs for working with ML such as Google Colab.

Using Google Colab, we were able to run our model using a NVIDIA Tesla K80 GPU. This was incredibly useful, as without the GPU our model would have taken several hours to fit. With the use of Google Colab, we were able to do this in a matter of minutes.